Earthquake damage assessment method based on wavelet transform and deep convolutional neural network

Abstract

The present invention discloses a method for earthquake damage assessment based on wavelet transform and deep convolutional neural network, which relates to the field of earthquake engineering technology and includes the following steps: Step 1: Obtain building information in the target area, simulate the target area, generate several earthquake motion records, and add damage labels to each building in the earthquake records; Step 2: Perform wavelet transform on the earthquake motion time history in the earthquake motion records to generate a multi-scale two-dimensional time-frequency energy map; Step 3: Construct a data set by one-to-one correspondence between building information, multi-scale two-dimensional time-frequency energy map, and damage labels; Step 4: Construct and train an earthquake damage assessment model; Step 5: Perform earthquake damage assessment based on the trained earthquake damage assessment model. The method of the present invention can realize intelligent identification of earthquake damage levels and is suitable for application scenarios such as rapid post-earthquake assessment, key area screening, and auxiliary emergency decision-making.

Description

Technical Field

[0001] The present invention relates to the field of earthquake engineering technology, and in particular to an earthquake damage assessment method based on wavelet transform and deep convolutional neural network. Background Art

[0002] In urban areas prone to frequent strong earthquakes, characterized by diverse building types and high population density, assessing damage to buildings after an earthquake is crucial for emergency response, resource allocation, and ensuring life safety. However, post-earthquake environments often present obstacles such as communication disruptions and restricted roads, as well as a large number of buildings, posing significant challenges to damage assessment.

[0003] Currently, there are three main types of post-earthquake building damage assessment methods: 1) manual inspection and identification methods. Although this method can intuitively reflect the damage situation, it relies on manpower and is time-consuming, making it difficult to quickly promote it on a large scale; 2) vulnerability model analysis methods based on earthquake intensity and building structure type. Although this method has established a theoretical model, it is usually based on the assumption of a linear system, making it difficult to accurately reflect the nonlinear behavior of the structure under complex vibrations; 3) nonlinear time-history analysis methods. This method can simulate structural responses with high precision, but requires the establishment of a detailed finite element model, consumes huge computing resources, and is not suitable for emergency assessment.

[0004] With the development of artificial intelligence, data-driven models have gradually been used in earthquake damage identification tasks. For example, machine learning methods such as SVM and RF attempt to integrate building attributes and seismic motion characteristics for classification, but their feature extraction capabilities are limited. Although convolutional neural networks (CNNs) have shown good performance in image recognition, they directly use raw acceleration time history inputs, fail to effectively capture the time-varying characteristics of the signal, and have poor generalization capabilities.

[0005] Based on the above, existing methods still face the following key challenges in practical applications: 1) Current models fail to effectively capture the nonlinear evolution and energy accumulation effects of earthquake signals during the learning process of time-frequency features, resulting in an incomplete understanding of the building damage process; 2) Multi-scale modeling capabilities are limited, especially in extracting short-term impacts and long-term pattern changes, which affects the grasp of the full picture of earthquake effects; 3) Training data samples mainly come from single scenarios or historical earthquakes, failing to fully simulate and cover the building response characteristics under diverse scenarios, limiting the model's migration and generalization capabilities.

[0006] In view of this, this invention is proposed. Summary of the Invention

[0007] The purpose of this invention is to address the shortcomings of the existing technology and propose an earthquake damage assessment method based on wavelet transform and deep convolutional neural network, which can realize the intelligent identification of earthquake damage level and is suitable for application scenarios such as rapid post-earthquake assessment, key area screening and auxiliary emergency decision-making.

[0008] In order to achieve the above object, the present invention adopts the following technical solutions:

[0009] A method for earthquake damage assessment based on wavelet transform and deep convolutional neural network, characterized by comprising the following steps:

[0010] Step 1: Obtain building information in the target area, simulate the target area, generate several earthquake motion records, and add damage labels to each building in the earthquake records.

[0011] Step 2: Perform wavelet transform on the seismic time history in the seismic records to generate a multi-scale two-dimensional time-frequency energy map;

[0012] Step 3: Construct a dataset by corresponding building information, multi-scale two-dimensional time-frequency energy maps, and damage labels one by one;

[0013] Step 4: Build and train an earthquake damage assessment model. The earthquake damage assessment model includes a deep convolution module and a fully connected module. The multi-scale two-dimensional time-frequency energy map is input into the deep convolution module, which outputs the extracted features. The extracted features are combined with building information and then input into the fully connected module, which outputs the earthquake damage assessment results.

[0014] Step 5: Perform earthquake damage assessment based on the trained earthquake damage assessment model.

[0015] Furthermore, the step 1 includes the following steps:

[0016] Step 1.1: Obtain building information in the target area;

[0017] Step 1.2: Use the response spectrum-based artificial seismic wave synthesis method to simulate and generate several ground motion records in the target area under multiple earthquake intensities;

[0018] Step 1.3: Based on the generated local earthquake records, calculate the exceedance probability of each damage level for each building using the building vulnerability function. Calculate the actual occurrence probability of each damage level based on the difference in exceedance probabilities between adjacent damage levels. Determine the cumulative probability interval for each damage level, starting from no damage, based on the actual occurrence probability of each damage level. The exceedance probability calculation formula is as follows:

[0019] ;

[0020] In the formula, P[D≥dk|IM] is the probability that a building reaches or exceeds the kth damage level d under a given earthquake intensity index IM. k The probability of θ k is the median intensity of the kth damage level; β k is the logarithmic standard deviation of the kth damage level; Φ(·) is the standard normal distribution function;

[0021] Step 1.4: Use the first-order Monte Carlo sampling method to generate a random number γ that follows a uniform distribution in the interval [0, 1] and map it to the cumulative probability interval of each damage level to determine the unique damage level.

[0022] Furthermore, the step 1.1 includes the following steps:

[0023] Step 1.1: Based on the target acceleration response spectrum Sa(ω), construct the power spectrum density function S(ω) of the ground motion. The calculation formula is as follows:

[0024] ;

[0025] In the formula, S(ω) is the power spectrum density function of the ground motion, and its unit is rad. 2 / s 3 ; ξ is the damping ratio; ω is the angular frequency, in rad / s; S a (ω) is the target acceleration response spectrum value, in g or m / s 2 ;T d is the earthquake duration, in seconds; p is the confidence probability coefficient that controls the matching degree of the response spectrum;

[0026] Step 1.1.2: Based on the power spectrum density S(ω) of the ground motion, construct the frequency domain amplitude spectrum A(ω) and determine the frequency resolution Δω. The calculation formula is as follows:

[0027] ;

[0028] ;

[0029] In the formula, A(ω) is the frequency domain amplitude spectrum; Δω is the frequency resolution; fs is the signal sampling frequency, in Hz; nfft is the number of fast Fourier transform points;

[0030] Step 1.1.3: Superimpose the random phase Φ(ω) on the frequency-domain amplitude spectrum A(ω) and perform Fourier inversion to obtain the initial artificial ground motion acceleration time history α(t). The calculation formula is as follows:

[0031] ;

[0032] In the formula, α(t) is the initial artificial ground motion acceleration time history, in units of g or m / s 2 Φ(ω) is the random phase at the angular frequency ω, which is uniformly distributed in [0, 2π]. t is the time variable in seconds, which is used to represent the time axis of the earthquake history after Fourier inversion. i is the imaginary unit.

[0033] Step 1.1.4: Modulate the initial artificial ground motion acceleration time history α(t) with the envelope function to obtain the final ground motion time history α s (t), the calculation formula is as follows:

[0034] ;

[0035] In the formula, α s (t) is the final ground motion time history, in g or m / s 2 ; g(t) is the envelope function;

[0036] Step 1.1.5: Calculate the response spectrum S of the current synthetic ground motion a,k (ω), the calculation formula is as follows:

[0037] ;

[0038] In the formula, k is the number of iterations, S a,k (ω) is the response spectrum of the current synthetic ground motion, α s,k (t) is the current final earthquake time history;

[0039] Step 1.1.6: If S a,k (ω) and S a If the error of (ω) is greater than the preset threshold, the spectrum is corrected according to the frequency ratio formula and the process returns to step 1.1.3 until the error is less than or equal to the preset threshold:

[0040] .

[0041] Furthermore, step 2 includes the following steps:

[0042] Step 2.1: Get the Morlet wavelet as the mother wavelet function. The formula of the Morlet wavelet is as follows:

[0043] ;

[0044] In the formula, Ψ(t) is the Morlet wavelet function; ω0 is the central angular frequency of the wavelet, in rad / s; t is the time variable, in s; i is the imaginary unit;

[0045] Step 2.2: Perform wavelet transform on the input signal and calculate its wavelet transform coefficient W(a,b) at different scales a and time points b. The calculation formula is as follows:

[0046] ;

[0047] In the formula, W(a,b) is the response coefficient; a is the scale parameter; b is the translation parameter; x(t) is the input signal; Ψ*(·) is the complex conjugate function of the Morlet wavelet;

[0048] Step 2.3: Square the modulus of the wavelet transform coefficient W(a,b) to obtain the multi-scale two-dimensional time-frequency energy map E(a,b). The calculation formula is as follows:

[0049] ;

[0050] In the formula, E(a,b) is the multi-scale two-dimensional time-frequency energy map.

[0051] Furthermore, in step 2, the actual frequency f is calculated using the following scale-frequency relationship:

[0052] ;

[0053] In the formula, f is the actual frequency; fc is the center frequency of the Morlet wavelet, fc=ω0 / 2π=0.955Hz;

[0054] The range of scale a is a min =0.06,a max =3.0, and is geometrically divided into N=32 scales, with the scale value a of the i-th layer. i The calculation formula is as follows:

[0055] ;

[0056] In the formula, a i is the scale value of the i-th layer;

[0057] In the time domain, with a sampling interval of ∆t=0.02s, 2000 time shift points are generated for the 40s seismic signal. The jth time shift point b j The calculation formula is as follows:

[0058] ;

[0059] In the formula, b j is the jth time shift point, and △t is the sampling time interval.

[0060] Furthermore, in step 3, the building information includes construction year x1, building purpose x2, number of floors x3, building height x4, building area x5, building length x6, building width x7, wall thickness x8, floor spacing x9, number of columns x1, and building height x2. 10 , and plane regularity x 11 ;

[0061] The continuous variables were standardized using the Z-score standardization method, and the calculation formula is as follows:

[0062] ;

[0063] In the formula, x i is the original variable value, μ is the mean value of the variable in all samples, and σ is the standard deviation;

[0064] The discrete variables are converted into sparse binary vectors through one-hot encoding and then standardized.

[0065] Furthermore, in step 4, the fully connected module includes a first fully connected layer, a first Dropout layer, a second fully connected layer, and a second Dropout layer that are sequentially connected.

[0066] Furthermore, in step 4, when training the earthquake damage assessment model, the multi-scale two-dimensional time-frequency energy map in the sample is enhanced by a lightweight geometric perturbation method, including the following steps: performing a small-range time axis translation on the image with a probability of 50% to simulate the slight uncertainty of the starting time of the shaking; introducing an affine transformation within a limited range, including a horizontal displacement within ±10 pixels and a proportional scaling within the range of 0.95 to 1.05.

[0067] Furthermore, in step 4, an early stopping mechanism is introduced when training the earthquake damage assessment model. After each round of training, the performance of the earthquake damage assessment model on the validation set is checked in real time. If the effect of the earthquake damage assessment model on the validation set does not improve in five consecutive rounds of training, the training is automatically terminated; and the earthquake damage assessment model is restored to the round with the best verification performance.

[0068] Furthermore, step 5 includes the following steps:

[0069] Step 5.1: Receive the earthquake acceleration time history;

[0070] Step 5.2: Receive building structural attribute data;

[0071] Step 5.3: The standardized building structural attributes and energy images are combined into input sample pairs, which are then fed into the trained earthquake damage assessment model to output the corresponding building damage level prediction results.

[0072] Step 5.4: Map the prediction results to the spatial locations of buildings to form a spatial layer that can represent different damage levels. Output the post-earthquake building damage distribution map, administrative area earthquake damage summary results, and key facility screening information to assist in earthquake damage analysis and emergency response assessment.

[0073] Compared with the prior art, the present invention has the following beneficial effects:

[0074] 1. In step 1, the scenario database is constructed by collecting the building structure information and synthetic seismic acceleration time history in the target area, and generating structural damage level labels based on the building vulnerability matrix and Monte Carlo sampling; in step 2, the pre-processed seismic acceleration time history data is used to obtain its time-frequency response spectrum using Morlet continuous wavelet transform, and the unified format is used as the input of the earthquake damage assessment model, which solves the problem that the current model fails to effectively capture the nonlinear evolution and energy accumulation effects of seismic signals in the learning process of time-frequency features; in step 3, the building structure data is standardized and encoded, paired with the time-frequency diagram, and used as part of the joint input of the earthquake damage assessment model, which solves the problem that the training data samples mainly come from a single scenario or historical earthquakes, and failed to fully simulate and cover the building response characteristics under diverse scenarios; in step 4, based on the above data, a GoogLeNet-FC neural network containing a multi-layer Inception module was constructed to automatically extract multi-scale features from the time-frequency diagram, and supplement the individual difference information of buildings through the structural attribute vector, thereby improving the model's ability to discriminate damage levels and solving the problem of limited multi-scale modeling capabilities, especially in extracting short-term impacts and long-term pattern changes; in step 4, image enhancement technology was introduced during the training process to improve the robustness of the model. After the training was completed, the classification accuracy and ROC curve classification performance indicators were used to comprehensively evaluate the model performance, effectively improving the model's adaptability to complex post-earthquake scenarios.

[0075] 2. Without relying on post-earthquake images or on-site inspections, the system can combine building structural information with seismic motion input to realize automatic identification of earthquake damage levels for a large range of building groups; the model prediction results can be mapped to the spatial location of the building to form a layer representation reflecting different damage levels, and output post-earthquake building damage distribution map, administrative area earthquake damage summary results and key facility screening information to support earthquake damage analysis, key area identification and emergency response assessment; fully exploit the non-stationary characteristics of earthquake signals, integrate structural properties for personalized assessment, and rely on multi-sample training data generated by simulation to enhance the model's generalization and practical application capabilities, with good engineering applicability and promotion value. BRIEF DESCRIPTION OF THE DRAWINGS

[0076] Figure 1This is a flowchart of an earthquake damage assessment method based on wavelet transform and deep convolutional neural network;

[0077] Figure 2 Schematic diagram of wavelet transform of earthquake acceleration signal, where A is earthquake acceleration time history diagram, B is Fourier amplitude spectrum diagram, and C is time-frequency distribution diagram of wavelet change;

[0078] Figure 3 Structural diagram of the model for earthquake damage assessment;

[0079] Figure 4 Performance evaluation diagram of the earthquake damage assessment model, where A is the confusion matrix diagram of the training set, B is the ROC diagram of the training set, C is the confusion matrix diagram of the validation set, and D is the ROC diagram of the validation set;

[0080] Figure 5 is a processing flow chart of the method of this embodiment in an example application;

[0081] Figure 6 Schematic diagram of the spatial visualization of building damage levels and the vulnerability matrix, where A is the spatial visualization and B is the vulnerability matrix. DETAILED DESCRIPTION

[0082] The technical solutions in the embodiments of the present invention will be clearly and completely described below in conjunction with the drawings in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, rather than all the embodiments.

[0083] Example 1:

[0084] A seismic damage assessment method based on wavelet transform and deep convolutional neural network, such as Figure 1 As shown, the following steps are included:

[0085] Step 1: Obtain building information in the target area, simulate the target area, generate several earthquake motion records, and add damage labels to each building in the earthquake records.

[0086] In this embodiment, the damage label represents the damage level, which may include five categories: no damage, slight damage, moderate damage, severe damage, and collapse.

[0087] Step 2: Perform wavelet transform on the seismic time history in the seismic records to generate a multi-scale two-dimensional time-frequency energy map;

[0088] In this embodiment, the earthquake time history is transformed into a high-resolution time-frequency diagram by wavelet transformation, which captures its instantaneous frequency variation and energy distribution, and reflects the non-stationary and frequency-dependent characteristics of the earthquake acceleration signal.

[0089] Step 3: Construct a dataset with one-to-one correspondence between building information, multi-scale two-dimensional time-frequency energy maps, and damage labels.

[0090] Step 4: Build and train an earthquake damage assessment model. The earthquake damage assessment model includes a deep convolution module and a fully connected module. The multi-scale two-dimensional time-frequency energy map is input into the deep convolution module, which outputs the extracted features. The extracted features are combined with building information and then input into the fully connected module, which outputs the earthquake damage assessment results.

[0091] In this embodiment, the multi-scale two-dimensional time-frequency energy map is first subjected to a deep convolution module to extract features. The extracted features are then combined with building information and input into a fully connected module to output earthquake damage assessment results. This constructs a unified multimodal feature space and realizes the effective fusion of image information and prior structural knowledge.

[0092] Step 5: Perform earthquake damage assessment based on the trained earthquake damage assessment model.

[0093] In this embodiment, the trained earthquake damage assessment model is applied to the task of identifying building damage. It can support the joint input of seismic motion and building structure information, realize the rapid prediction and spatial visual expression of the post-earthquake damage level of buildings in the regional scope, and assist in post-earthquake emergency response and risk investigation.

[0094] In an optional embodiment, step 1 includes the following steps:

[0095] Step 1.1: Obtain building information in the target area;

[0096] In this optional embodiment, building information of the target area may be collected from municipal archives. The building information may include one or more of the number of floors, structure type, height, construction year, and building purpose.

[0097] Step 1.2: Use the response spectrum-based artificial seismic wave synthesis method to simulate and generate several ground motion records in the target area under multiple earthquake intensities;

[0098] Since buildings respond differently to different peak ground accelerations (PGAs), dominant frequencies, and specific seismic wave characteristics, this optional embodiment uses an artificial seismic wave synthesis method based on a response spectrum for simulation. Specifically, 0.05g, 0.1g, 0.2g, 0.4g, and 0.8g can be selected as PGA values, corresponding to earthquake intensities VI to X, respectively, to comprehensively evaluate the impact of different earthquake magnitudes on building structures.

[0099] Step 1.3: Based on the generated local earthquake records, calculate the exceedance probability of each damage level for each building using the building vulnerability function. Calculate the actual occurrence probability of each damage level based on the difference in exceedance probabilities between adjacent damage levels. Determine the cumulative probability interval for each damage level, starting from no damage, based on the actual occurrence probability of each damage level. The exceedance probability calculation formula is as follows:

[0100] ;

[0101] In the formula, P[D≥dk|IM] is the probability that a building reaches or exceeds the kth damage level d under a given earthquake intensity index IM (such as PGA). k The probability of θ is obtained from the peak acceleration (PGA) of the generated ground motion record. k is the median intensity of the kth damage level; β k is the logarithmic standard deviation of the kth damage level, reflecting the uncertainty of the fragility curve; Φ(·) is the standard normal distribution function.

[0102] Step 1.4: Use the first-order Monte Carlo sampling method to generate a random number γ that follows a uniform distribution in the interval [0, 1] and map it to the cumulative probability interval of each damage level to determine the unique damage level.

[0103] For example, the PGA value of a certain earthquake motion record is 0.35. First, the exceedance probability of each damage level is calculated, then the actual occurrence probability of each damage level is calculated, and finally, the cumulative probability interval of each damage level is determined in sequence starting from no damage, as shown in Table 1.

[0104] Table 1 Exceedance probability, actual occurrence probability and cumulative probability interval of each damage level of a building under a certain earthquake record

[0105]

[0106] The random number γ = 0.73 is sampled, which falls into the interval [0.6, 0.8), corresponding to the moderate damage label. The final result is that the building is judged to have "moderate damage" under the current earthquake conditions, and the damage label is "moderate damage."

[0107] In this optional embodiment, steps 1.3 and 1.4 combine the building vulnerability function with the Monte Carlo sampling probability judgment method to determine the damage label of each building under different earthquake motion inputs. This method can determine a unique damage label for each building under different earthquake motion inputs while reflecting the uncertainty of building damage. In other words, although the input is uncertain, the labeling result itself is unique, which plays a crucial role in the subsequent classification model training.

[0108] In an optional embodiment, step 1.1 includes the following steps:

[0109] Step 1.1: Based on the target acceleration response spectrum Sa(ω), construct the power spectrum density function S(ω) of the ground motion. The calculation formula is as follows:

[0110] ;

[0111] In the formula, S(ω) is the power spectrum density function of the ground motion, and its unit is rad. 2 / s 3 ; ξ is the damping ratio; ω is the angular frequency, in rad / s; S a (ω) is the target acceleration response spectrum value, in g or m / s 2 ;T d is the earthquake duration in seconds; p is the confidence probability coefficient that controls the matching degree of the response spectrum, indicating that the generated earthquake motion response spectrum does not exceed the target design spectrum under the specified probability p, and its value is usually 0.85-0.9.

[0112] Step 1.1.2: Based on the power spectrum density S(ω) of the ground motion, construct the frequency domain amplitude spectrum A(ω) and determine the frequency resolution Δω. The calculation formula is as follows:

[0113] ;

[0114] ;

[0115] In the formula, A(ω) is the frequency domain amplitude spectrum, that is, the amplitude value of each frequency point in the frequency domain; Δω is the frequency resolution, that is, the frequency domain sampling interval; fs is the signal sampling frequency, in Hz; nfft is the number of fast Fourier transform points.

[0116] Step 1.1.3: Superimpose the random phase Φ(ω) on the frequency-domain amplitude spectrum A(ω) and perform Fourier inversion to obtain the initial artificial ground motion acceleration time history α(t). The calculation formula is as follows:

[0117] ;

[0118] In the formula, α(t) is the initial artificial ground motion acceleration time history, in units of g or m / s 2 ; Φ(ω) is the random phase at the angular frequency ω, which is uniformly distributed in [0,2π]; t is the time variable in seconds, which is used to represent the time axis of the earthquake history after Fourier inversion; i is the imaginary unit, which satisfies i 2 =-1.

[0119] Step 1.1.4: Modulate the initial artificial ground motion acceleration time history α(t) with the envelope function to obtain the final ground motion time history α s (t), the calculation formula is as follows:

[0120] ;

[0121] In the formula, α s (t) is the final ground motion time history, in g or m / s 2 ; g(t) is the envelope function, which is used to control the time function of the seismic wave energy release process. Usually, a function with rise-peak-attenuation characteristics is selected.

[0122] In this optional embodiment, in order to simulate the characteristics of earthquake energy release over time, the initial artificial earthquake acceleration time history α(t) is modulated by an envelope function to obtain a more physically consistent earthquake waveform, that is, the final earthquake time history α s (t).

[0123] Step 1.1.5: Calculate the response spectrum S of the current synthetic ground motion a,k (ω), the calculation formula is as follows:

[0124] ;

[0125] In the formula, k is the number of iterations, S a,k (ω) is the response spectrum of the current synthetic seismic motion, that is, the response spectrum of the kth iterative synthetic seismic motion, α s,k (t) is the current final earthquake motion time history, that is, the final earthquake motion time history obtained in the kth iteration;

[0126] In this optional embodiment, the “acceleration response spectrum calculation module” is preset, and α s,k (t) After input, get S a,k (ω); This module simulates a "unit mass, fixed damping" single degree of freedom oscillator system (SDOF), α s,k (t) is used as the seismic input, and the response analysis is performed on multiple vibration periods T. The Newmark-β numerical method is used to calculate the response of the system to the seismic motion. The maximum acceleration response value under the American and Russian periods is obtained, which is the corresponding response spectrum corresponding to this frequency point.

[0127] Step 1.1.6: If S a,k (ω) and S a If the error of (ω) is greater than the preset threshold, the spectrum is corrected according to the frequency ratio formula and the process returns to step 1.1.3 until the error is less than or equal to the preset threshold:

[0128] .

[0129] In a specific implementation process, in order to fully express the earthquake response characteristics under different intensity scenarios, according to the specifications of the "Seismic Parameter Zoning Map (GB 18306-2015)", 3590 synthetic seismic acceleration records covering five levels of design intensity from VI to X were generated. The simulation process took into account the site category (Class I, Class II), seismic zoning and grouping scheme to ensure the engineering applicability of seismic motion. Each seismic record was fitted with the corresponding design acceleration response spectrum S a (ω), the damping ratio ξ is set to 3.5%, the sampling time interval is 0.02 seconds, and the frequency resolution is controlled by the Fourier interval Δω = 0.005·2π for frequency domain discretization. In the construction process of the power spectrum density function S(ω), the peak exceedance probability control coefficient p = 0.9 is introduced to control the synthetic ground motion to not exceed the target response spectrum S at a 90% confidence level. a (ω) to ensure the controllability and physical rationality of the statistical response characteristics. Subsequently, the frequency domain amplitude spectrum A(ω) is combined with the random phase Φ(ω) to generate the preliminary seismic time history α(t) through Fourier inversion. The seismic signal is modulated using the tri-phased envelope function g(t) to obtain a more physically realistic time history record α s (t) to simulate the time-dependent release of actual earthquake energy. Finally, to unify the data format and accommodate subsequent modeling requirements, all ground motion records were standardized to a fixed length of 40 seconds. The processing rules are as follows: 1) If the original signal length is less than 40 seconds, the entire signal is retained and zero-padded at the end; 2) If the signal length is greater than 40 seconds and the PGA occurs within the first or last 20 seconds, the corresponding 40 seconds before and after are retained; 3) In all other cases, the 20 seconds before and after the PGA are truncated, for a total of 40 seconds.

[0130] In an optional embodiment, step 2 includes the following steps:

[0131] Step 2.1: Get the Morlet wavelet as the mother wavelet function. The formula of the Morlet wavelet is as follows:

[0132] ;

[0133] In the formula, Ψ(t) is the Morlet wavelet function; ω0 is the central angular frequency of the wavelet, in rad / s, which can be set to 6; t is the time variable, in seconds; and i is the imaginary unit.

[0134] Step 2.2: Perform wavelet transform on the input signal and calculate its wavelet transform coefficient W(a,b) at different scales a and time points b. The calculation formula is as follows:

[0135] ;

[0136] In the formula, W(a,b) is the response coefficient, which represents the response value at a certain scale and time; a is the scale parameter that controls the frequency, with a range of [0.06, 3.0] and a total of 32 layers; b is the translation parameter that controls the time position, with a total of 2000 sampling points and a step size of 0.02 seconds; x(t) is the input signal, i.e., the earthquake time history, with a length of 40 seconds and 2000 sampling points; and Ψ*(·) is the complex conjugate function of the Morlet wavelet.

[0137] Step 2.3: Square the modulus of the wavelet transform coefficient W(a,b) to obtain the multi-scale two-dimensional time-frequency energy map E(a,b). The calculation formula is as follows:

[0138] ;

[0139] In the formula, E(a,b) is the multi-scale two-dimensional time-frequency energy map.

[0140] In an optional embodiment, in step 2, the actual frequency f is calculated using the following scale-frequency relationship:

[0141] ;

[0142] In the formula, f is the actual frequency; fc is the center frequency of the Morlet wavelet, fc=ω0 / 2π=0.955Hz.

[0143] The range of scale a is a min =0.06,a max =3.0, and is geometrically divided into N=32 scales, with the scale value a of the i-th layer. i The calculation formula is as follows:

[0144] ;

[0145] In the formula, a i is the scale value of the i-th layer.

[0146] In this optional embodiment, the range of the scale a is set to cover the main seismic frequency band of 0.5 Hz to 25 Hz.

[0147] In the time domain, with a sampling interval of ∆t=0.02s, 2000 time shift points are generated for the 40s seismic signal. The jth time shift point b j The calculation formula is as follows:

[0148] ;

[0149] In the formula, b jis the jth time shift point, and Δt is the sampling time interval.

[0150] In a specific implementation process (linked to the previous specific implementation process), the seismic records can be unified into 40s length, and the sampling interval is set to ∆t=0.02s, that is, each record contains 2000 sampling points, and the frequency range is controlled by the Fourier frequency increment ∆ω=0.005·2π=0.0314, corresponding to the Nyquist frequency of 25Hz. Therefore, the wavelet scale setting covers the main seismic frequency band in the range of 0.5Hz to 25Hz; the time-frequency response matrix obtained by multi-scale expansion of the seismic time history is 32×2000 in size, which is normalized and interpolated to convert it into a two-dimensional time-frequency energy map with a unified image input format (such as 224×224). The specific wavelet transform time-frequency map results are as follows: Figure 2 shown.

[0151] In an optional embodiment, in step 3, the building information includes construction year x1, building purpose x2, number of floors x3, building height x4, building area x5, building length x6, building width x7, wall thickness x8, floor spacing x9, number of columns x1, and building height x2. 10 , and plane regularity x 11 ;

[0152] The continuous variables were standardized using the Z-score standardization method, and the calculation formula is as follows:

[0153] ;

[0154] In the formula, x i is the original variable value, μ is the mean of the variable in all samples, and σ is the standard deviation.

[0155] In this optional embodiment, continuous variables such as building height and number of floors are standardized to eliminate the impact of different dimensions. After standardization, each continuous feature is mapped to the same dimensional range to avoid weight imbalance between features, which helps to improve the convergence speed and stability of the deep network during training.

[0156] The discrete variables are converted into sparse binary vectors through one-hot encoding and then standardized.

[0157] In this optional embodiment, discrete variables such as the construction age and building purpose are converted into sparse binary vectors through one-hot encoding and standardized; for example, the construction age is divided into four categories: before 1979, 1980s, 1990s, and after 2001, and their encoding is in the form of [1,0,0,0], [0,1,0,0], etc., indicating the mutual exclusion relationship between different categories, thereby improving the model's ability to recognize classification boundaries.

[0158] In this optional embodiment, the continuous and discrete building information are normalized and spliced ​​to form a unified structural feature vector Si, and its corresponding multi-scale two-dimensional time-frequency energy map Ti together constitute the input sample X i ={T i ,S i}, label Yi is the corresponding damage label.

[0159] In an optional embodiment, if Figure 3 As shown, in step 4, the deep convolution module is the GoogLeNet network, and the features of the Softmax2 output of the GoogLeNet network are combined with the building information and input into the fully connected module.

[0160] In this optional embodiment, the average pooling layer before the softmax output layer of the traditional GoogLeNet network is removed, allowing high-dimensional representations to be retained after feature extraction, making it suitable for multimodal fusion input. The GoogLeNet network includes multiple Inception modules, each of which contains four parallel branches: 1×1 convolution, 3×3 convolution, 5×5 convolution, and max pooling + 1×1 convolution. 1) All convolution branches use ReLU activation functions and batch normalization (BN). 2) Multiple Inception modules are stacked to form the backbone network, facilitating the extraction of multi-scale time-frequency features from different receptive fields.

[0161] In an optional embodiment, if Figure 3 As shown, in step 4, the fully connected module includes a first fully connected layer, a first Dropout layer, a second fully connected layer and a second Dropout layer connected in sequence.

[0162] In this optional embodiment, a Dropout layer operation is introduced to suppress overfitting, keep the network parameters moderate, and ensure that the overall structure has both expressive power and deployment flexibility.

[0163] In an optional embodiment, in step 4, when training the earthquake damage assessment model, a unified training sample pair X is constructed. i ={T i ,S i}, Yi=Damage Label; T i After size normalization, the image is uniformly scaled to 224×224 pixels in grayscale format, and the pixel values ​​are normalized to the range [0,1]. The calculation formula is as follows:

[0164] ;

[0165] In the formula, x is the original pixel value, xmin and x max are the minimum and maximum grayscale values ​​in the image, respectively.

[0166] In an optional embodiment, in step 4, when training the earthquake damage assessment model, the multi-scale two-dimensional time-frequency energy map in the sample is enhanced by a lightweight geometric perturbation method, including the following steps: performing a small-range time axis translation (equivalent to a lateral translation) on the image with a probability of 50% to simulate the slight uncertainty of the starting time of the vibration; introducing an affine transformation within a limited range, including a horizontal displacement within ±10 pixels and a proportional scaling within the range of 0.95 to 1.05, to enhance the spatial invariance modeling capability of the model.

[0167] In this optional embodiment, to improve the generalization and robustness of the earthquake damage assessment model, an online enhancement strategy is implemented on the input multi-scale 2D time-frequency energy image during each training cycle. Given that the ground motion time-frequency image has a clear correspondence between time and frequency axes, the enhancement operation must be performed without destroying its physical structure. Therefore, the enhancement is performed using a lightweight geometric perturbation approach.

[0168] The enhanced multi-scale two-dimensional time-frequency energy map maintains the original size and number of channels, which neither affects the input specifications of the downstream network structure nor improves the robustness and adaptability of the model in the face of different vibration scenarios and signal perturbations without sacrificing the accuracy of the vibration energy distribution.

[0169] In an optional embodiment, in step 4, when training the earthquake damage assessment model, the Adam optimizer is used to update the weight parameters. Adam combines the advantages of momentum gradient descent and adaptive learning rate. The weight update rule is:

[0170] ;

[0171] In the formula, θ t is the network parameter for the tth iteration; α is the learning rate, which is initially set to 0.0125 and adopts a step decay mechanism, that is, the learning rate is reduced to half of its original value every 10 epochs; is the first-order moment estimate (sliding average of the bias-corrected gradient); is the second-order moment estimate (the sliding average of the squared gradient after bias correction); Ɛ is a small constant to prevent division by zero, usually 10 -8 The batch size of the batch training is set to 32, and the He normal distribution is used to initialize the initial weights of the network to ensure that the output distribution of each layer of the network is balanced under the ReLU activation function.

[0172] In an optional embodiment, in step 4, when training the earthquake damage assessment model, the loss function uses categorical cross-entropy to optimize the classification performance of the model, which is defined as follows:

[0173] ;

[0174] In the formula, N represents the total number of samples, C=5 represents the number of damage level categories, is the true label value of sample i under category j (represented by One-hot encoding), The probability output of the model prediction belonging to category j.

[0175] In an optional embodiment, in step 4, an early stopping mechanism is introduced when training the earthquake damage assessment model. After each round of training, the performance of the earthquake damage assessment model on the validation set (i.e., data that does not participate in training and is used to test the model's capabilities) is checked in real time. If the effect of the earthquake damage assessment model on the validation set does not improve (i.e., the loss value does not decrease) in five consecutive rounds of training, the training is automatically terminated to prevent the model from continuing to "hard-learn" useless or even harmful details; and the earthquake damage assessment model is restored to the round with the best verification performance to ensure that the model version with the best performance is ultimately retained.

[0176] In this optional embodiment, the maximum number of training rounds can be set to 45, but this is only an upper limit. The actual number of training rounds depends on the performance of the earthquake damage assessment model on the validation set. If the earthquake damage assessment model reaches the optimal state in the 20th round and does not continue to improve in the next 5 rounds, the training will automatically terminate at the 25th round.

[0177] In this optional embodiment, an early stopping mechanism is introduced to further suppress overfitting and improve training efficiency.

[0178] In an optional embodiment, in step 4, when training the earthquake damage assessment model, a comprehensive model evaluation system is constructed based on accuracy, precision, recall, F1 score, and AUROC classification performance indicators.

[0179] In this optional embodiment, in order to comprehensively and systematically evaluate the performance of the proposed wavelet transform and earthquake damage assessment model, a scientific and reasonable evaluation process is established from multiple levels, covering classification accuracy evaluation, category discrimination ability verification, spatial distribution consistency analysis and numerical prediction deviation measurement, etc., to ensure that the earthquake damage assessment model has engineering usability and generalization capabilities in post-earthquake damage identification tasks.

[0180] In terms of classification accuracy, based on the predicted output of the earthquake damage assessment model on the validation set and the corresponding true labels, a confusion matrix of five types of damage levels was constructed, and multiple performance indicators were calculated based on this, including accuracy, precision, recall, and F1-score. Among them, accuracy is used to measure the proportion of correct judgments among all predictions; precision represents the proportion of samples predicted to be in a certain category that are actually in that category, reflecting the credibility of the model's output; recall measures the proportion of all samples that actually belong to that category that are successfully identified, reflecting the model's coverage capability; and the F1-score comprehensively balances precision and recall, unifying the evaluation indicators through harmonic averaging, which is particularly important in the context of multi-category imbalance. The relevant calculation formula is as follows:

[0181] ;

[0182] ;

[0183] ;

[0184] ;

[0185] In the formula, TP is the number of true positive cases, TN is the number of true negative cases, FP is the number of false positive cases, and FN is the number of false negative cases.

[0186] To evaluate the earthquake damage assessment model's ability to discriminate between damage level categories, this optional embodiment introduces the ROC curve (Receiver Operating Characteristic Curve) and AUC (Area Under Curve) for auxiliary analysis. Specifically, each damage level category is considered a positive class, while the remaining categories are considered negative. An ROC curve is constructed for each category, and the AUC is calculated. The closer the AUC is to 1, the stronger the model's ability to distinguish between categories with fuzzy damage level boundaries and similar characteristics, providing a basis for subsequent model structure optimization and label boundary processing.

[0187] In a specific implementation process (here is related to the previous specific implementation process), some results are as follows Figure 4 shown. Figure 4 A in Figure 4Figure C shows that the earthquake damage assessment model proposed in this example achieved overall classification accuracy of 93.04% on the training set and 96.19% on the test set, respectively. Furthermore, the model achieved precision, recall, and F1 scores exceeding 93% for each category, with the maximum deviation between precision and recall being only 0.0037, demonstrating the model's stable classification performance with well-calibrated boundaries.

[0188] Figure 4 B in Figure 4 D in further shows the ROC curves and corresponding AUC values ​​of the five damage levels in the case of multi-category classification. The results show that the AUC values ​​of all categories exceed 0.95, indicating that the model has good recognition ability in all categories, especially in minority categories such as "mild damage" and "severe damage", where the performance is significantly improved, and no obvious recognition bias is generated due to category imbalance. The earthquake damage assessment method proposed in this embodiment retains the joint time-frequency features of the input signal by introducing continuous wavelet transform, and combines the multi-scale Inception structure in the GoogLeNet backbone network to model the multi-level earthquake damage morphology. It not only enhances the model's expression of non-stationary features, but also improves its ability to distinguish categories with blurred boundaries and overlapping features. The above experimental results fully demonstrate that the method of this embodiment has strong discrimination stability, generalization ability and practical effectiveness in the multi-category imbalanced earthquake damage identification task, verifying the adaptability and engineering application value of this model in complex earthquake damage scenarios.

[0189] In an optional embodiment, step 5 includes the following steps:

[0190] Step 5.1: Receive the earthquake acceleration time history;

[0191] In this optional embodiment, the earthquake acceleration time history may be a measured or simulated synthetic signal.

[0192] Step 5.2: Receive building structural attribute data;

[0193] In this optional embodiment, the building structure attribute data includes location, height, number of floors, structure type, construction year and usage function.

[0194] Step 5.3: The standardized building structural attributes and energy images are combined into input sample pairs, which are then input into the trained earthquake damage assessment model to output the corresponding building damage level prediction results.

[0195] Step 5.4: Map the prediction results to the spatial locations of buildings to form a spatial layer that can represent different damage levels. Output the post-earthquake building damage distribution map, administrative area earthquake damage summary results, and key facility screening information to assist in earthquake damage analysis and emergency response assessment.

[0196] In this optional embodiment, the earthquake damage assessment model is applied to actual scenarios, and a building information and earthquake data fusion processing framework is constructed under the GIS environment using the earthquake resistance performance survey and risk investigation database, realizing an integrated processing flow of earthquake motion input, structural information loading, model reasoning and prediction result visualization. Figure 5 The figure shows a schematic diagram of the processing flow in an example application, which displays the multi-source input of building structure properties and seismic motion records. After data preprocessing, wavelet transform and sample construction, it is input into the earthquake damage assessment model to complete the damage level prediction, and finally outputs a variety of results such as layer rendering, report statistics and key building screening.

[0197] The interface receives structural attribute data of buildings in the target area, including location coordinates, height, number of floors, structure type, construction year, and building purpose. After standardization and one-hot encoding, this data forms a structural feature vector and is paired one-to-one with seismic motion records. The seismic motion input uses acceleration time-history signals measured or simulated by the seismic network. The wavelet transform module is automatically called to perform multi-scale time-frequency analysis, generating a multi-scale two-dimensional time-frequency energy map. This map is then converted into a standard-sized image as input according to the requirements of the earthquake damage assessment model. The structural features and image samples are then input into the deployed earthquake damage assessment model to complete batch predictions of building damage levels.

[0198] The prediction results are rendered in the platform map interface in the form of layers, and different damage levels are marked with different color codes to form a distribution map of building damage levels after the earthquake and calculate the vulnerability matrix, such as Figure 6 The framework supports spatially categorized display, attribute association, and damage level statistics for buildings within a spatial scope. It automatically generates a variety of outputs, including building damage level maps, administrative region damage summary reports, and critical facility screening results, for post-earthquake assessment and decision support. The output data can be exported as image files, statistical tables, or structured documents, making it suitable for information aggregation and communication during the earthquake response process.

[0199] In summary, the earthquake damage assessment method based on wavelet transform and deep convolutional neural network of this embodiment, in step 1, collects the building structure information and synthetic earthquake acceleration time history of the target area, and generates structural damage level labels based on the building vulnerability matrix and Monte Carlo sampling to build a scenario database; in step 2, uses the preprocessed earthquake acceleration time history data, adopts Morlet continuous wavelet transform to obtain its time-frequency response spectrum, and uses it as the input of the earthquake damage assessment model in a unified format, which solves the problem that the current model fails to effectively capture the nonlinear evolution and energy accumulation effect of the earthquake signal in the process of learning the time-frequency characteristics; in step 3, the building structure data is standardized and encoded, and paired with the time-frequency diagram as part of the joint input of the earthquake damage assessment model, which solves the problem. The training data samples mainly come from single scenarios or historical earthquakes, and fail to fully simulate and cover the building response characteristics under diverse scenarios. In step 4, based on the above data, a GoogLeNet-FC neural network containing a multi-layer Inception module is constructed to automatically extract multi-scale features from the time-frequency diagram, and supplement the individual difference information of buildings through structural attribute vectors to improve the model's ability to discriminate damage levels. This solves the problem of limited multi-scale modeling capabilities, especially the inadequacy in extracting short-term impacts and long-term pattern changes. In step 4, image enhancement technology is introduced during the training process to improve the robustness of the model. After training, the classification accuracy and ROC curve classification performance indicators are used to comprehensively evaluate the model performance, effectively improving the model's adaptability to complex post-earthquake scenarios.

[0200] The earthquake damage assessment method based on wavelet transform and deep convolutional neural network in this embodiment can combine building structural information with seismic motion input to achieve automatic identification of earthquake damage levels for a large range of building groups without relying on post-earthquake images or on-site inspections. The model prediction results can be mapped to the spatial location of the buildings to form a layer representation reflecting different damage levels, and output post-earthquake building damage distribution maps, administrative region earthquake damage summary results, and key facility screening information to support earthquake damage analysis, key area identification, and emergency response assessment. The method fully exploits the non-stationary characteristics of earthquake signals, integrates structural properties for personalized assessment, and relies on multi-sample training data generated by simulation to enhance the model's generalization and practical application capabilities, showing good engineering applicability and promotion value.

[0201] The above description is only a preferred specific embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any technician familiar with the technical field, within the technical scope disclosed by the present invention, who makes equivalent replacements or changes based on the technical solution and inventive concept of the present invention, should be covered by the scope of protection of the present invention.

Claims

1. A method for earthquake damage assessment based on wavelet transform and deep convolutional neural network, characterized in that: The steps include: Step 1: Obtain building information in the target area, simulate the target area, generate several earthquake motion records, and add damage labels to each building in the earthquake records. Step 2: Perform wavelet transform on the seismic time history in the seismic records to generate a multi-scale two-dimensional time-frequency energy map; Step 3: Construct a dataset by corresponding building information, multi-scale two-dimensional time-frequency energy maps, and damage labels one by one; Step 4: Build and train an earthquake damage assessment model. The earthquake damage assessment model includes a deep convolution module and a fully connected module. The multi-scale two-dimensional time-frequency energy map is input into the deep convolution module, which outputs the extracted features. The extracted features are combined with building information and then input into the fully connected module, which outputs the earthquake damage assessment results. Step 5: Perform earthquake damage assessment based on the trained earthquake damage assessment model; The step 1 comprises the following steps: Step 1.1: Obtain building information in the target area; Step 1.2: Use the response spectrum-based artificial seismic wave synthesis method to simulate and generate several ground motion records in the target area under multiple earthquake intensities; Step 1.3: Based on the generated local earthquake records, calculate the exceedance probability of each damage level for each building using the building vulnerability function. Calculate the actual occurrence probability of each damage level based on the difference in exceedance probabilities between adjacent damage levels. Determine the cumulative probability interval for each damage level, starting from no damage, based on the actual occurrence probability of each damage level. The exceedance probability calculation formula is as follows: ； In the formula, P[D≥dk|IM] is the probability that a building reaches or exceeds the kth damage level d under a given earthquake intensity index IM. k The probability of θ k is the median intensity of the kth damage level; β k is the logarithmic standard deviation of the kth damage level; Φ(·) is the standard normal distribution function; Step 1.4: Use the first-order Monte Carlo sampling method to generate a random number γ that follows a uniform distribution in the interval [0,1] and map it to the cumulative probability interval of each damage level to determine the unique damage level; The step 2 comprises the following steps: Step 2.1: Get the Morlet wavelet as the mother wavelet function. The formula of the Morlet wavelet is as follows: ； In the formula, Ψ(t) is the Morlet wavelet function; ω0 is the central angular frequency of the wavelet, in rad / s; t is the time variable, in s; i is the imaginary unit; Step 2.2: Perform wavelet transform on the input signal and calculate its wavelet transform coefficient W(a,b) at different scales a and time points b. The calculation formula is as follows: ； In the formula, W(a,b) is the response coefficient; a is the scale parameter; b is the translation parameter; x(t) is the input signal; Ψ*(·) is the complex conjugate function of the Morlet wavelet; Step 2.3: Square the modulus of the wavelet transform coefficient W(a,b) to obtain the multi-scale two-dimensional time-frequency energy map E(a,b). The calculation formula is as follows: ； In the formula, E(a,b) is the multi-scale two-dimensional time-frequency energy map; In step 3, the building information includes construction year x1, building purpose x2, number of floors x3, building height x4, building area x5, building length x6, building width x7, wall thickness x8, floor spacing x9, number of columns x1, and building height x2. 10 , and plane regularity x 11 ; The continuous variables were standardized using the Z-score standardization method, and the calculation formula is as follows: ； In the formula, x i is the original variable value, μ is the mean value of the variable in all samples, and σ is the standard deviation; The discrete variables are converted into sparse binary vectors through one-hot encoding and then standardized; In step 4, when training the earthquake damage assessment model, the multi-scale two-dimensional time-frequency energy map in the sample is enhanced by a lightweight geometric perturbation method, including the following steps: performing a small-range time axis translation on the image with a probability of 50% to simulate the slight uncertainty of the starting time of the shaking; introducing an affine transformation within a limited range, including a horizontal displacement within ±10 pixels and a proportional scaling within the range of 0.95 to 1.

05.

2. The earthquake damage assessment method based on wavelet transform and deep convolutional neural network according to claim 1 is characterized in that: The step 1.1 includes the following steps: Step 1.1: Based on the target acceleration response spectrum S a (ω), construct the power spectrum density function S(ω) of the earthquake motion, and the calculation formula is as follows: ； In the formula, S(ω) is the power spectrum density function of the ground motion, and its unit is rad. 2 / s 3 ; ξ is the damping ratio; ω is the angular frequency, in rad / s; S a (ω) is the target acceleration response spectrum value, in g or m / s 2 ; T d is the earthquake duration, in seconds; p is the confidence probability coefficient that controls the matching degree of the response spectrum; Step 1.1.2: Based on the power spectrum density S(ω) of the ground motion, construct the frequency domain amplitude spectrum A(ω) and determine the frequency resolution Δω. The calculation formula is as follows: ； ； In the formula, A(ω) is the frequency domain amplitude spectrum; Δω is the frequency resolution; fs is the signal sampling frequency, in Hz; nfft is the number of fast Fourier transform points; Step 1.1.3: Superimpose the random phase Φ(ω) on the frequency-domain amplitude spectrum A(ω) and perform Fourier inversion to obtain the initial artificial ground motion acceleration time history α(t). The calculation formula is as follows: ； In the formula, α(t) is the initial artificial ground motion acceleration time history, in units of g or m / s 2 Φ(ω) is the random phase at the angular frequency ω, which is uniformly distributed in [0, 2π]. t is the time variable in seconds, which is used to represent the time axis of the earthquake history after Fourier inversion. i is the imaginary unit. Step 1.1.4: Modulate the initial artificial ground motion acceleration time history α(t) with the envelope function to obtain the final ground motion time history α s (t), the calculation formula is as follows: ； In the formula, α s (t) is the final ground motion time history, in g or m / s 2 ; g(t) is the envelope function; Step 1.1.5: Calculate the response spectrum S of the current synthetic ground motion a,k (ω), the calculation formula is as follows: ； In the formula, k is the number of iterations, S a,k (ω) is the response spectrum of the current synthetic ground motion, α s,k (t) is the current final earthquake time history; Step 1.1.6: If S a,k (ω) and S a If the error of (ω) is greater than the preset threshold, the spectrum is corrected according to the frequency ratio formula and the process returns to step 1.1.3 until the error is less than or equal to the preset threshold: 。 3. The earthquake damage assessment method based on wavelet transform and deep convolutional neural network according to claim 1 is characterized in that: In step 2, the actual frequency f is calculated using the following scale-frequency relationship: ； In the formula, f is the actual frequency; fc is the center frequency of the Morlet wavelet, fc=ω0 / 2π=0.955Hz; The range of scale a is a min =0.06,a max =3.0, and is geometrically divided into N=32 scales, with the scale value a of the i-th layer. i The calculation formula is as follows: ； In the formula, a i is the scale value of the i-th layer; In the time domain, with a sampling interval of ∆t=0.02s, 2000 time shift points are generated for the 40s seismic signal. The jth time shift point b j The calculation formula is as follows: ； In the formula, b j is the jth time shift point, and ∆t is the sampling time interval.

4. The earthquake damage assessment method based on wavelet transform and deep convolutional neural network according to claim 1 is characterized in that: In step 4, the fully connected module includes a first fully connected layer, a first Dropout layer, a second fully connected layer, and a second Dropout layer that are sequentially connected.

5. The earthquake damage assessment method based on wavelet transform and deep convolutional neural network according to claim 1 is characterized in that: In step 4, an early stopping mechanism is introduced when training the earthquake damage assessment model. After each round of training, the performance of the earthquake damage assessment model on the validation set is checked in real time. If the performance of the earthquake damage assessment model on the validation set does not improve in five consecutive rounds of training, the training is automatically terminated, and the earthquake damage assessment model is restored to the round with the best verification performance.

6. The earthquake damage assessment method based on wavelet transform and deep convolutional neural network according to claim 1 is characterized in that: The step 5 includes the following steps: Step 5.1: Receive the earthquake acceleration time history; Step 5.2: Receive building structural attribute data; Step 5.3: The standardized building structural attributes and energy images are combined into input sample pairs, which are then fed into the trained earthquake damage assessment model to output the corresponding building damage level prediction results. Step 5.4: Map the prediction results to the spatial locations of buildings to form a spatial layer that can represent different damage levels. Output the post-earthquake building damage distribution map, administrative area earthquake damage summary results, and key facility screening information to assist in earthquake damage analysis and emergency response assessment.

Images