CAE-LSTM-based unsupervised structural damage identification method

Abstract

The invention relates to the technical field of structural damage identification, in particular to an unsupervised structural damage identification method based on CAE-LSTM. The method comprises the following steps: training a CAE-LSTM model by using a training set to obtain a trained model, and reconstructing unknown data including health data and damage data by using the trained model to obtain a reconstruction error of the health data and a reconstruction error of the damage data; determining a damage sensitivity factor of the acceleration response signal of each batch by combining a probability density function of health data, a probability density function of damage data and a reconstruction error in the acceleration response signal of each batch of the undamaged structure so as to determine a damage threshold and screen a damage position; acquiring a damage factor according to the health state data and the damage state data of the damage position sensor; and judging the damage degree based on the size of the damage factor. According to the invention, the accuracy and reliability of the structural damage identification result are improved.

Description

Technical Field

[0001] The present invention relates to the technical field of structural damage identification, and particularly relates to an unsupervised structural damage identification method based on CAE-LSTM. Background Art

[0002] At present, unsupervised learning is widely used in the field of structural health monitoring. Different unsupervised learning methods, such as autoencoders, variational autoencoders, K-means clustering, etc., are used to detect damage in structures and conduct preliminary analysis. However, these methods have problems of insufficient positioning accuracy and inability to quantify the degree of damage. Summary of the Invention

[0003] In order to solve the problems of insufficient positioning accuracy and inability to quantify the degree of damage when the existing methods identify structural damage, the purpose of the present invention is to provide an unsupervised structural damage identification method based on CAE-LSTM, and the specific technical solutions adopted are as follows:

[0004] The present invention provides an unsupervised structural damage identification method based on CAE-LSTM, and the method includes the following steps:

[0005] Training the CAE-LSTM model with a training set to obtain a trained model, where the training set is composed of acceleration response signals of the structure in a healthy state;

[0006] Reconstructing unknown data including healthy data and damaged data with the trained model to obtain the reconstruction error of the healthy data and the reconstruction error of the damaged data;

[0007] Combining the probability density function of the healthy data, the probability density function of the damaged data, and the reconstruction error in the acceleration response signal of each batch of the undamaged structure to determine the damage sensitivity factor of the acceleration response signal of each batch; determining the damage threshold based on the size distribution of all the damage sensitivity factors, and screening the damage location based on the damage threshold;

[0008] Obtaining a damage factor according to the healthy state data and the damaged state data of the sensors at the damage location; judging the degree of damage based on the size of the damage factor.

[0009] Preferably, the CAE part of the CAE-LSTM model includes multiple convolutional layers, and each convolutional layer extracts multi-scale spatial features of the signal through convolutional kernels of different sizes.

[0010] Preferably, the reconstruction error of the healthy data and the reconstruction error of the damaged data include:

[0011] For any piece of health data or any piece of damage data: Denote the absolute value of the difference between the original sample data and the reconstructed sample data of this data as the first difference of this data.

[0012] Take the average value of the first differences of all health data as the reconstruction error of the health data.

[0013] Take the average value of the first differences of all damage data as the reconstruction error of the damage data.

[0014] Preferably, determining the damage sensitivity factor of the acceleration response signal of each batch by combining the probability density function of the health data, the probability density function of the damage data, and the reconstruction error in the acceleration response signal of each batch of the undamaged structure includes:

[0015] For any batch of acceleration response signals of the undamaged structure:

[0016] Denote the ratio between the standard deviation of the reconstruction errors of all health data in the acceleration response signal of this batch and the maximum value of the probability density function of all health data in the acceleration response signal of this batch as the first ratio.

[0017] Denote the ratio between the standard deviation of the reconstruction errors of all damage data in the acceleration response signal of this batch and the maximum value of the probability density function of all damage data in the acceleration response signal of this batch as the second ratio.

[0018] Calculate the ratio of the second ratio to the first ratio; Determine the difference between this ratio and the constant 1 as the damage sensitivity factor of the acceleration response signal of this batch.

[0019] Preferably, determining the damage threshold by using the magnitude distribution of all the damage sensitivity factors includes:

[0020] Sort the damage sensitivity factors of the acceleration response signals of all batches of the undamaged structure, and select the maximum damage sensitivity factor at a preset confidence level as the damage threshold.

[0021] Preferably, screening the damage location based on the damage threshold includes:

[0022] Determine the sensor location corresponding to when the damage sensitivity factor is greater than the damage threshold as the damage location.

[0023] Preferably, obtaining the damage factor according to the health state data and damage state data of the sensors at the damage location includes:

[0024] The absolute value of the difference between the average value of the health state data of the damage location sensor and the average value of the damage state data is denoted as the second difference;

[0025] Obtain the maximum value among the average value of the health state data of the damage location sensor and the average value of the damage state data;

[0026] Take the ratio between the second difference and the maximum value as the damage factor.

[0027] Preferably, the training batch size of the CAE-LSTM model is 64, and the key parameters are selected according to Bayesian hyperparameter optimization.

[0028] The present invention has at least the following beneficial effects:

[0029] 1. The CAE-LSTM network in the present invention can effectively extract spatio-temporal features and can accurately locate the damage area when damage occurs. Compared with traditional unsupervised learning methods, the present invention can reduce the positioning error caused by uneven sensor layout or insufficient data collection;

[0030] 2. Through the combination of reconstruction error calculation and damage-sensitive feature extraction technology, the method of the present invention has strong adaptability under different structures and load conditions, can identify damage without relying on specific load conditions, and improves the generalization ability of the model;

[0031] 3. By quantifying the damage factor, the present invention can divide the damage degree into multiple levels, provide more detailed damage assessment information, and help subsequent maintenance and repair decisions;

[0032] 4. Traditional damage identification methods often rely heavily on the layout of sensors, and unreasonable layout will affect the detection effect. However, the spatial features and time series features extracted by the CAE-LSTM network of the present invention can better handle different sensor layout situations, reduce the dependence on sensor layout, and improve the robustness of damage identification. Since the method provided by the present invention does not depend on specific load conditions and sensor configurations, it can be widely applied to the health monitoring of various structures, efficiently and accurately perform damage identification and quantitative evaluation, and improve the accuracy and reliability of the damage identification results under various sensor layout conditions. BRIEF DESCRIPTION OF THE DRAWINGS

[0033] In order to more clearly illustrate the technical solutions and advantages in the embodiments of the present invention or the prior art, the following will briefly introduce the drawings required for the description of the embodiments or the prior art. Obviously, the following drawings are only some embodiments of the present invention. For those of ordinary skill in the art, other drawings can be obtained based on these drawings without creative efforts.

[0034] Figure 1 Flow chart of an unsupervised structural damage identification method based on CAE-LSTM provided by an embodiment of the present invention;

[0035] Figure 2 Comparison chart of reconstruction errors of sensor 1 provided by an embodiment of the present invention in healthy and damaged states;

[0036] Figure 3 Comparison chart of reconstruction errors of sensor 5 provided by an embodiment of the present invention in healthy and damaged states;

[0037] Figure 4 Comparison chart of reconstruction errors of sensor 11 provided by an embodiment of the present invention in healthy and damaged states;

[0038] Figure 5 Comparison chart of reconstruction errors of sensor 15 provided by an embodiment of the present invention in healthy and damaged states;

[0039] Figure 6 Comparison chart of reconstruction errors of sensor 21 provided by an embodiment of the present invention in healthy and damaged states;

[0040] Figure 7 Comparison chart of reconstruction errors of sensor 25 provided by an embodiment of the present invention in healthy and damaged states;

[0041] Figure 8 Result chart of damage sensitivity factors of sensors obtained by using a CAE model;

[0042] Figure 9 Result chart of damage sensitivity factors of sensors obtained by using a CAE-LSTM model;

[0043] Figure 10 Result chart of damage identification obtained by using a CAE model;

[0044] Figure 11 Result chart of damage identification obtained by using a CAE-LSTM model. Detailed implementation manners

[0045] In order to further elaborate on the technical means and effects adopted by the present invention to achieve the predetermined invention purpose, the following provides a detailed description of an unsupervised structural damage identification method based on CAE-LSTM proposed according to the present invention in conjunction with the accompanying drawings and preferred embodiments.

[0046] Unless otherwise defined, all technical and scientific terms used herein have the same meaning as commonly understood by those skilled in the technical field to which the present invention belongs.

[0047] The following specifically describes the specific solution of an unsupervised structural damage identification method based on CAE-LSTM provided by the present invention in conjunction with the accompanying drawings.

[0048] An embodiment of an unsupervised structural damage identification method based on CAE-LSTM:

[0049] This embodiment proposes an unsupervised structural damage identification method based on CAE-LSTM. As Figure 1 shown, an unsupervised structural damage identification method based on CAE-LSTM in this embodiment includes the following steps:

[0050] Step S1, training the CAE-LSTM model using a training set to obtain a trained model, where the training set is composed of acceleration response signals of the structure in a healthy state.

[0051] First, collect the acceleration response signals of the structure in a healthy state and a damaged state, and perform preprocessing such as normalization on the collected acceleration response signals to reduce the influence of data deviation on the damage identification result. Use the normalization result of the acceleration response signal of the structure in a healthy state to train the CAE-LSTM model, so that it learns the optimal feature representation of the data in a healthy state, and thus provides a basis for subsequent damage identification. Before the original acceleration signal is input into the CAE-LSTM model, empirical mode decomposition is first used for denoising processing to improve the signal quality and damage identification accuracy. The convolutional autoencoder (CAE) part of the CAE contains multiple convolutional layers. Each convolutional layer extracts multi-scale spatial features of the signal through convolutional kernels of different sizes, and improves the feature expression ability through normalization and activation functions to improve the identification accuracy of damage information; during the training process of the CAE-LSTM model, the CAE part is used to extract the spatial features of the signal, reduce data redundancy, and improve the feature representation ability; the long short-term memory network (LSTM) part uses bidirectional LSTM units, and by simultaneously processing the forward and backward time series information, enhances the model's ability to model the time dependence of the signal and improves the perception ability of the structural state change; during the training process, the LSTM part is used to learn the time series features of the signal and enhance the perception ability of time-varying patterns. During the training process of the CAE-LSTM model, the acceleration response signals in the healthy state are divided into several small batches, and the batch size is 64. The training batch size of the CAE-LSTM model is 64, and the key parameters are automatically adjusted by the Bayesian optimization method. The optimization includes parameters such as the CAE convolutional kernel size, the number of convolutional layers, the number of LSTM units, the learning rate, and the batch size, etc., to improve the training effect and damage identification ability of the model.

[0052] So far, this embodiment has obtained a trained CAE-LSTM model.

[0053] Step S2: Use the trained model to reconstruct the unknown data containing healthy data and damaged data, and obtain the reconstruction error of the healthy data and the reconstruction error of the damaged data.

[0054] Next, use the trained CAE-LSTM model to reconstruct the unknown data containing healthy data and damaged data. By calculating the reconstruction error, damage-sensitive features are extracted, and the damage-sensitive features are used to reflect the change of the signal reconstruction deviation under the damaged state. It should be noted that: the healthy data is the data under the healthy state, and the damaged data is the data under the damaged state.

[0055] Specifically, use the characteristics of CAE to reconstruct the input data to obtain the reconstruction results of the structure in healthy data and the structure in damaged data.

[0056] For any healthy data or any damaged data: The absolute value of the difference between the original sample data of the data and the reconstructed sample data is denoted as the first difference of the data. By using this method, the first difference of each healthy data and the first difference of each damaged data can be obtained. The average value of the first differences of all healthy data is used as the reconstruction error of the healthy data; the average value of the first differences of all damaged data is used as the reconstruction error of the damaged data, and the reconstruction error is used to reflect the change of the signal reconstruction deviation under the damaged state. In this embodiment, the calculation formula of the reconstruction error is given, and the reconstruction error can be expressed as:

[0057]

[0058] where MAE represents the reconstruction error, n represents the number of data, represents the reconstructed sample data of the i-th data, y i represents the original sample data of the i-th data, and || represents the absolute value symbol. represents the first difference of the i-th data.

[0059] By using the above formula, the reconstruction error of the healthy data and the reconstruction error of the damaged data can be obtained. The reconstruction error calculation uses the mean absolute error for evaluation, which is used to quantify the deviation between the healthy state and the damaged state data. The size of the reconstruction error directly affects the extraction effect of the damage-sensitive features. It should be noted that: when calculating the reconstruction error of the healthy data, substitute the healthy data into the above formula to obtain the reconstruction error of the healthy data; when calculating the reconstruction error of the damaged data, substitute the damaged data into the above formula to obtain the reconstruction error of the damaged data.

[0060] Step S3: Determine the damage sensitivity factor of the acceleration response signal for each batch by combining the probability density functions of the healthy data, the probability density functions of the damaged data, and the reconstruction error in the acceleration response signal of each batch of undamaged structures; determine the damage threshold using the magnitude distribution of all the damage sensitivity factors, and screen the damage location based on the damage threshold.

[0061] Since the characteristics of the damaged data deviate from the feature space of the healthy state, the reconstruction error is relatively large, and the reconstruction error follows a normal distribution. That is, by calculating the probability density functions of the healthy data and the damaged data and comparing the differences in their means and standard deviations, the ratio of the standard deviation to the probability density function is used to extract the damage-sensitive features.

[0062] Specifically, first, calculate the probability density function of the healthy data and the probability density function of the damaged data according to the normal distribution model of the reconstruction error. The specific calculation formulas are as follows:

[0063]

[0064] where f(MAE) represents the value of the probability density function, σ represents the mean of the reconstruction error, μ represents the standard deviation of the reconstruction error, π represents the pi, and E represents the reconstruction error.

[0065] Using the above formulas, the value of the probability density function of the healthy data and the value of the probability density function of the damaged data can be obtained. It should be noted that: when calculating the value of the probability density function of the healthy data, substitute the reconstruction error of the healthy data into the above formula to calculate the value of the probability density function of the healthy data; when calculating the value of the probability density function of the damaged data, substitute the reconstruction error of the damaged data into the above formula to calculate the value of the probability density function of the damaged data.

[0066] Divide the acceleration response signal of the undamaged structure into multiple batches, with each batch size of 64. The initial batch is regarded as the healthy state, and the subsequent batches are regarded as unknown states.

[0067] For any batch of acceleration response signals of the undamaged structure: Denote the ratio between the standard deviation of the reconstruction errors of all the healthy data in this batch of acceleration response signals and the maximum value of the probability density function of all the healthy data in this batch of acceleration response signals as the first ratio. Denote the ratio between the standard deviation of the reconstruction errors of all the damaged data in this batch of acceleration response signals and the maximum value of the probability density function of all the damaged data in this batch of acceleration response signals as the second ratio. Calculate the ratio of the second ratio to the first ratio; Determine the difference between the ratio and the constant 1 as the damage sensitivity factor of this batch of acceleration response signals.

[0068] In this embodiment, a specific calculation formula for the damage sensitivity factor is given. The damage sensitivity factor of the acceleration response signals of this batch can be expressed as:

[0069]

[0070] Where DSF represents the damage sensitivity factor of the acceleration response signals of this batch, and SDPP D represents the second ratio, and SDPP U represents the first ratio.

[0071] Sort the damage sensitivity factors of the acceleration response signals of all batches of the undamaged structure, and select the maximum damage sensitivity factor at the preset confidence level as the damage threshold. In this embodiment, the change range is statistically analyzed based on the damage-sensitive characteristics in the healthy state, and the damage threshold is set as the benchmark for damage identification. In this embodiment, the preset confidence level is 99%. In specific applications, the implementer can set it according to specific circumstances.

[0072] Using the above method, calculate the damage sensitivity factor for the acceleration response signals of all sensors. If the damage sensitivity factor calculated for the acceleration response signal of a certain sensor is greater than the damage threshold, it is determined that there is damage in the area where the sensor is located, and the position of the sensor is determined as the damage position.

[0073] Step S4: Obtain the damage factor based on the health state data and damage state data of the sensors at the damage position; judge the damage degree based on the magnitude of the damage factor.

[0074] In this embodiment, after determining the damage position, the damage factor is further calculated. The damage factor is used to characterize the severity of the damage, and the damage is divided into different levels based on the damage factor value to provide more refined damage assessment information.

[0075] Specifically, record the absolute value of the difference between the average value of the health state data and the average value of the damage state data of the sensors at the damage position as the second difference; obtain the maximum value of the average value of the health state data and the average value of the damage state data of the sensors at the damage position; use the ratio between the second difference and the maximum value as the damage factor.

[0076] In this embodiment, a specific calculation formula for the damage factor is given. The damage factor can be expressed as:

[0077]

[0078] Where α i represents the damage factor, represents the average value of the damage state data of the sensors at the damage position, represents the average value of the health status data of the damage location sensor, and max() represents the maximum value function, and are both greater than 0.

[0079] The calculation of the damage factor is not only based on the mean difference between the health status and the damage status data, but also corrected by combining the standard deviation change to improve the accuracy of damage quantitative assessment.

[0080] α i The value range of is between [0, 1], and the larger its value, the more serious the damage degree. Next, according to the value of the damage factor, the damage degree is divided into multiple levels, and then the severity of the damage can be judged.

[0081] In this embodiment, the first damage threshold, the second damage threshold, the third damage threshold, and the fourth damage threshold are respectively set, where the first damage threshold is less than the second damage threshold, the second damage threshold is less than the third damage threshold, and the third damage threshold is less than the fourth damage threshold.

[0082] If the damage factor is less than or equal to the first damage threshold, it is determined that the damage degree is level one. At this time, the structure is basically intact, the non-load-bearing components are slightly damaged, and the use function is not affected; if the damage factor is greater than the first damage threshold and less than or equal to the second damage threshold, it is determined that the damage degree is level two. At this time, the structure is slightly damaged, the non-load-bearing structure is damaged, and the load-bearing structure has allowable damage; if the damage factor is greater than the second damage threshold and less than or equal to the third damage threshold, it is determined that the damage degree is level three. At this time, the structure is moderately damaged, the main load-bearing structure is severely damaged, and the bearing capacity decreases; if the damage factor is greater than the third damage threshold and less than or equal to the fourth damage threshold, it is determined that the damage degree is level four. At this time, the structure is severely damaged, the main load-bearing structure is broken and fractured, the load-bearing capacity is significantly reduced, and the structure is in a dangerous state; if the damage factor is greater than the fourth damage threshold, it is determined that the damage degree is level five. At this time, the structure collapses, there is no possibility of repair, and the use function is lost. In this embodiment, the first damage threshold is 0.2, the second damage threshold is 0.4, the third damage threshold is 0.6, and the fourth damage threshold is 0.9. In specific applications, the implementer can set according to specific circumstances.

[0083] The method provided in this embodiment is applicable to the health monitoring of steel frame structures. By analyzing the acceleration response signals of the steel frame structure in different states, the location and quantitative assessment of damage are realized, and it can adapt to different sensor layout schemes and various external load conditions, improving the stability and accuracy of damage identification.

[0084] So far, the identification of structural damage has been completed by using the method provided in this embodiment.

[0085] Figure 2 、 Figure 3 、Figure 4 , Figure 5 , Figure 6 and Figure 7 shows the comparison of the reconstruction errors of six sensors in the healthy state and the damaged state, which are the comparison diagrams of the reconstruction errors of sensor 1, sensor 5, sensor 11, sensor 15, sensor 21, and sensor 25 in the healthy state and the damaged state respectively. In Figures 2 to 7 , the solid line represents the change curve of the reconstruction error in the healthy state, and the dashed line represents the change curve of the reconstruction error in the damaged state. It can be seen that the peak value of the reconstruction error of sensor 1 in the damaged state is significantly lower than that in the healthy state, and the standard deviation distribution range becomes wider. This phenomenon indicates that the signal characteristics in the damaged state have changed significantly. Therefore, based on this, it can be preliminarily speculated that the damage occurred near sensor 1. To further quantify the difference in the reconstruction error between the healthy state and the damaged state, the damage sensitivity factor of the damage-sensitive feature is calculated and used as a metric to obtain the damage threshold, as shown in Figure 9 . The results show that only the damage sensitivity factor value of sensor 1 is higher than the 99% confidence threshold, indicating that damage is very likely to have occurred at the position of this sensor.

[0086] To verify the superiority of the CAE-LSTM network in damage identification, its identification effect is compared with that of the traditional CAE network, as shown in Figure 8 . It is found that there are multiple phenomena in the CAE network where the damage sensitivity factor values of sensors are higher than the damage threshold, resulting in its inability to accurately identify the damage location.

[0087] Damage has occurred to the structure near sensor 1, and the reconstruction errors of the corresponding healthy and damaged state data conform to the normal distribution. The reconstruction errors of the two states are as shown in Figure 2 . The calculated damage factor is 0.2087, and its damage level is minor damage, indicating that the loosening of the bolt at node 1 has caused minor damage to the frame structure.

[0088] By calculating the reconstruction errors of the two structural states through the CAE-LSTM network, it is found that the peak values of the reconstruction errors of sensors 6, 9, 12, and 15 in the damaged state are all significantly reduced, and the standard deviation range increases, indicating that the signal characteristics have changed greatly in the damaged state. Calculate the damage sensitivity factors between the healthy state and the damaged state of all sensors, as shown in Figure 11 . The results verify that the damage sensitivity factors of sensors 6, 9, 12, and 15 exceed the damage threshold, that is, damage may have occurred near these sensors.

[0089] The damage identification results of the traditional CAE network are as shown in Figure 10As shown, it can be found that in addition to sensors 6, 9, 12, and 15, there are cases where the damage sensitivity factors of other sensors are higher than the damage threshold, resulting in the inability to accurately locate the damage position. This result indicates that there are certain deficiencies in the CAE network when dealing with multiple damage positions. It should be noted that: Figure 8 、 Figure 9 、 Figure 10 and Figure 11 The dashed lines in represent the damage threshold. Figure 8 、 Figure 9 、 Figure 10 and Figure 11 The ordinate DSF in represents the damage sensitivity factor.

[0090] The damage factors of sensors 6, 9, 12, and 15 are 0.6145, 0.6123, 0.6433, and 0.6469 respectively, and their damage levels belong to severe damage.

[0091] The CAE-LSTM network in this embodiment can effectively extract spatio-temporal features and can accurately locate the damage area when damage occurs. Compared with traditional unsupervised learning methods, this embodiment can reduce the positioning error caused by uneven sensor layout or insufficient data collection. In addition, by quantifying the damage factor in this embodiment, the damage degree can be divided into multiple levels, providing more detailed damage assessment information, which is helpful for subsequent maintenance and repair decisions. Through the combination of multiple technologies such as Bayesian hyperparameter optimization, reconstruction error calculation, and damage-sensitive feature extraction in this embodiment, the method has strong adaptability under different structural and load conditions, can identify damage without relying on specific load conditions, and improves the generalization ability of the model.

[0092] Traditional damage identification methods often rely heavily on the layout of sensors, and unreasonable layout will affect the detection effect. However, through the spatial features and time series features extracted by the CAE-LSTM network in this embodiment, it can better handle different sensor layout situations, reduce the dependence on sensor layout, and improve the robustness of damage identification. Since the method provided in this embodiment does not depend on specific load conditions and sensor configurations, it can be widely applied to the health monitoring of various structures, efficiently and accurately perform damage identification and quantitative assessment, and improve the accuracy and reliability of structural identification results under various sensor layout conditions.

[0093] It should be noted that: The above are only the preferred embodiments of the present invention and are not intended to limit the present invention. Any modifications, equivalent replacements, improvements, etc. made within the principles of the present invention shall be included in the protection scope of the present invention.

Claims

1. An unsupervised structural damage identification method based on CAE-LSTM, characterized in that, The method includes the following steps: Training the CAE-LSTM model using a training set to obtain a trained model, where the training set consists of acceleration response signals of the structure in a healthy state; Using the trained model to reconstruct unknown data including healthy data and damaged data, and obtaining the reconstruction error of the healthy data and the reconstruction error of the damaged data; Combining the probability density function of healthy data, the probability density function of damaged data, and the reconstruction error in each batch of acceleration response signals of the undamaged structure to determine the damage sensitivity factor of each batch of acceleration response signals; determining the damage threshold based on the magnitude distribution of all the damage sensitivity factors, and screening the damage location based on the damage threshold; Obtaining a damage factor according to the healthy state data and damaged state data of the sensors at the damage location; judging the damage degree based on the magnitude of the damage factor.

2. The unsupervised structural damage identification method based on CAE-LSTM according to claim 1, characterized in that The CAE part of the CAE-LSTM model includes multiple convolutional layers, and each convolutional layer extracts multi-scale spatial features of the signal through convolutional kernels of different sizes.

3. The unsupervised structural damage identification method based on CAE-LSTM according to claim 1, characterized in that The reconstruction error of the healthy data and the reconstruction error of the damaged data include: For any healthy data or any damaged data: Denote the absolute value of the difference between the original sample data and the reconstructed sample data of this data as the first difference of this data; Taking the average value of the first differences of all healthy data as the reconstruction error of the healthy data; Taking the average value of the first differences of all damaged data as the reconstruction error of the damaged data.

4. The unsupervised structural damage identification method based on CAE-LSTM according to claim 1, characterized in that The combining the probability density function of healthy data, the probability density function of damaged data, and the reconstruction error in each batch of acceleration response signals of the undamaged structure to determine the damage sensitivity factor of each batch of acceleration response signals includes: For any batch of acceleration response signals of the undamaged structure: Denote the ratio between the standard deviation of the reconstruction errors of all healthy data in the batch of acceleration response signals and the maximum value of the probability density function of all healthy data in the batch of acceleration response signals as the first ratio; Denote the ratio between the standard deviation of the reconstruction errors of all damaged data in the batch of acceleration response signals and the maximum value of the probability density function of all damaged data in the batch of acceleration response signals as the second ratio; Calculate the ratio of the second ratio to the first ratio; determine the difference between the ratio and the constant 1 as the damage sensitivity factor of the batch of acceleration response signals.

5. A method for unsupervised structural damage identification based on CAE-LSTM according to claim 1, characterized in that The determining the damage threshold using the magnitude distribution of all the damage sensitivity factors includes: Sorting the damage sensitivity factors of all batches of acceleration response signals of the undamaged structure, and selecting the maximum damage sensitivity factor at a preset confidence level as the damage threshold.

6. The unsupervised structural damage identification method based on CAE-LSTM according to claim 1, wherein, The screening the damage location based on the damage threshold includes: Determining the sensor location corresponding to when the damage sensitivity factor is greater than the damage threshold as the damage location.

7. A method for unsupervised structural damage identification based on CAE-LSTM according to claim 1, characterized in that The obtaining a damage factor according to the healthy state data and damaged state data of the sensors at the damage location includes: The absolute value of the difference between the average value of the health state data of the damage position sensor and the average value of the damage state data is denoted as the second difference; Obtain the maximum value among the average value of the health state data of the damage position sensor and the average value of the damage state data; Take the ratio between the second difference and the maximum value as the damage factor.

8. A method for unsupervised structural damage identification based on CAE-LSTM according to claim 1, characterized in that The training batch size of the CAE-LSTM model is 64, and the key parameters are selected according to Bayesian hyperparameter optimization.

Images