Building earthquake damage assessment method, earthquake damage assessment model training method and electronic system

Abstract

The invention relates to the technical field of buildings, in particular to a building earthquake damage assessment method, an earthquake damage assessment model training method and an electronic system. According to the building earthquake damage assessment method, the real-time building vibration data of the target building are acquired; extracting characteristic parameters of the real-time building vibration data based on a time domain analysis and frequency domain analysis method; and inputting the characteristic parameters of the real-time building vibration data into the earthquake damage evaluation model to obtain an earthquake damage evaluation result of the target building. According to the earthquake damage evaluation model training method, the building vibration database is established; extracting characteristic parameters of the building vibration data based on a time domain analysis method and a frequency domain analysis method; acquiring parameter information of the artificial neural network; and based on the parameter information of the artificial neural network, training to obtain an earthquake damage evaluation model. According to the method, the model input data can fully reflect the key information of the real-time building vibration data in a mode of extracting the characteristic parameters, so that the reliability of the earthquake damage assessment result is ensured.

Description

Technical Field

[0001] The present invention relates to the field of building technology, and in particular to a building earthquake damage assessment method, an earthquake damage assessment model training method and an electronic system. Background Art

[0002] Earthquakes, one of the most severe natural disasters facing humanity, often cause structural damage and significant casualties. Therefore, rapid and accurate assessment of post-earthquake building damage is crucial for rapid emergency response and the rapid restoration of building functions. This includes providing crucial information for emergency response decisions and ensuring the rational allocation of rescue resources. It also identifies buildings that are already in a dangerous state and unusable, helping to prevent further casualties from aftershocks. Furthermore, it provides timely assessment recommendations for less severely damaged buildings, facilitating the rapid restoration of urban building functionality.

[0003] Among the existing building earthquake damage assessment methods: manual assessment methods are relatively accurate and reliable, but inefficient and require trained professionals or structural engineers to conduct assessments, which are highly professional and subjective; vulnerability assessment methods are based on historical earthquake damage data and are highly reliable and stable, but it is difficult to consider the time-frequency characteristics of earthquake motion, which will affect the final damage result of the structure and have low accuracy; elastic-plastic time-history analysis methods can consider the characteristics of the structure and earthquake motion, and have high accuracy, but the calculation volume is large and the assessment efficiency is relatively low.

[0004] Therefore, there is an urgent need in the prior art for a building earthquake damage assessment method that can quickly and accurately assess the damage to a building after an earthquake and solve at least one of the problems existing in the prior art. Summary of the Invention

[0005] In order to solve at least one of the problems existing in the prior art, the present invention provides a building earthquake damage assessment method, an earthquake damage assessment model training method and an electronic system.

[0006] An embodiment of the present invention provides a method for assessing earthquake damage to a building, the method comprising the following steps: acquiring real-time building vibration data of a target building; extracting characteristic parameters of the real-time building vibration data based on time domain analysis and frequency domain analysis methods, the characteristic parameters comprising the root mean square amplitude of a time domain graph, the frequency and amplitude corresponding to the peak of a frequency domain amplitude spectrum, and the power bucket of a frequency domain power spectrum; wherein the power bucket is the average value of the power spectrum density within each pre-divided frequency band; inputting the characteristic parameters of the real-time building vibration data into an earthquake damage assessment model to obtain an earthquake damage assessment result for the target building; wherein the earthquake damage assessment model is an artificial neural network.

[0007] Optionally, the feature extraction calculation formula of the root mean square amplitude is:

[0008]

[0009] Among them, F RMS is the RMS amplitude, x h is the amplitude corresponding to the discrete sampling sequence number h in the time domain graph, H is the total number of discrete sampling sequences in the time domain graph, and the value range of h is [1, H];

[0010] The feature extraction calculation formula for the frequency and amplitude corresponding to the peak is:

[0011] Fp eaj (i)f req =k i

[0012] FP eak (i) amp =X i (k i )

[0013] F Peak ={F Peak (1) freq ,···,F Peak (I) freq ; F Peak (1) amp ,···,F Peak (I) amp}

[0014] Among them, F Peak (i) freq is the frequency characteristic parameter corresponding to the i-th peak in the frequency domain amplitude spectrum, F Peak (i) amp is the amplitude characteristic parameter corresponding to the i-th peak in the frequency domain amplitude spectrum, k i is the frequency corresponding to the i-th peak, X i (k i ) is the amplitude corresponding to the i-th peak, F Peak is a set of the amplitude characteristic parameter and the frequency characteristic parameter, I is the total number of peaks in the high amplitude frequency band in the frequency domain amplitude spectrum, and the value range of i is [1, I];

[0015] The feature extraction calculation formula of the power bucket is:

[0016]

[0017] Among them, k n is the frequency corresponding to the discrete sampling sequence number n in the frequency domain power spectrum, N is the total number of discrete sampling sequences in the frequency domain power spectrum, P(k n ) is the frequency kn The corresponding power density, X(k n ) is the frequency k n The corresponding amplitude, is the characteristic parameter of the mth power bucket, k min is the lower frequency limit of the power bucket, k max is the upper frequency limit of the power bucket, mean is the calculation frequency k min With frequency k max The average power density between PSD is the set of the power buckets, M is the total number of power buckets, and the value range of m is [1, M].

[0018] Optionally, the earthquake damage assessment result of the target building includes one of multiple preset vibration labels; the total number of peaks in the frequency domain amplitude spectrum is the maximum number of peaks in the high amplitude frequency band in the frequency domain amplitude spectrum corresponding to the vibration sample data in the building vibration database, and the vibration sample data in the building vibration database is the basic data for training the earthquake damage assessment model; the total number of power buckets is the number of frequency band divisions of the frequency domain power spectrum corresponding to the vibration sample data in the building vibration database, and the number of frequency band divisions is used to divide out the high power level frequency bands corresponding to the preset vibration labels.

[0019] Optionally, the step of inputting the characteristic parameters of the real-time building vibration data into the earthquake damage assessment model to obtain the earthquake damage assessment result of the target building includes: inputting the characteristic parameters of the real-time building vibration data into the earthquake damage assessment model to obtain the confidence of the preset vibration label; wherein the earthquake damage assessment model includes a preset confidence threshold; if the confidence of the vibration label is greater than or equal to the confidence threshold, outputting the corresponding preset vibration label.

[0020] Compared with the existing technology, the building earthquake damage assessment method provided by the present invention has the following beneficial effects:

[0021] The building earthquake damage assessment method provided by the embodiment of the present invention obtains real-time building vibration data of the target building; based on the time domain analysis and frequency domain analysis methods, the characteristic parameters of the real-time building vibration data are extracted, and the above-mentioned characteristic parameters include the amplitude root mean square of the time domain graph, the frequency and amplitude corresponding to the peak of the frequency domain amplitude spectrum, and the power bucket of the frequency domain power spectrum; wherein the power bucket is the average value of the power spectrum density within each pre-divided frequency band; the characteristic parameters of the real-time building vibration data are input into the earthquake damage assessment model to obtain the earthquake damage assessment result of the target building; wherein the earthquake damage assessment model is an artificial neural network. The present invention reduces the complexity of calculation, improves the data quality, and enhances the interpretability of the data by extracting various characteristic data such as the amplitude root mean square of the time domain graph, the frequency and amplitude corresponding to the peak of the frequency domain amplitude spectrum, and the power bucket of the frequency domain power spectrum, thereby ensuring that the model input data can fully reflect the key information of the real-time building vibration data, thereby ensuring the reliability of the earthquake damage assessment result.

[0022] An embodiment of the present invention also provides a method for training an earthquake damage assessment model, the method comprising: establishing a building vibration database, the building vibration database comprising building vibration data corresponding to various vibration labels; extracting characteristic parameters of the building vibration data based on time domain analysis and frequency domain analysis methods; obtaining parameter information of an artificial neural network, the parameter information comprising the number of hidden layers, the number of neurons, a confidence threshold, an optimization algorithm, and a loss function; and based on the parameter information of the artificial neural network, training a earthquake damage assessment model with the characteristic parameters of the building vibration data as input and the vibration label as output.

[0023] Optionally, the method further includes: lightweight compiling the earthquake damage assessment model based on edge machine learning technology to obtain model firmware that can be oriented to embedded devices; and configuring the model firmware at the edge of the sensor.

[0024] Optionally, the building vibration data includes building static pulsation data under non-seismic excitation, elevator operation vibration data and building construction vibration data, as well as building non-damage data, building slightly damaged data and building seriously damaged data under seismic excitation; the vibration labels include static vibration, elevator vibration, construction vibration, green label, yellow label and red label.

[0025] Optionally, the establishment of a building vibration database includes: obtaining vibration sample data under non-seismic excitation, vibration sample data under initial seismic excitation and design parameters of a target building; classifying the vibration sample data under non-seismic excitation according to whether the elevator and construction equipment in the target building are in operation, and obtaining the static pulsation data of the building under non-seismic excitation, the elevator operation vibration data and the building construction vibration data; establishing a nonlinear multi-degree-of-freedom shear model according to the design parameters of the target building; amplitude modulating the vibration sample data under the initial seismic excitation to obtain vibration sample data under seismic excitation; performing nonlinear time history analysis on the vibration sample data under seismic excitation based on the nonlinear multi-degree-of-freedom shear model to obtain building response data under seismic excitation; classifying the vibration data under seismic excitation based on a trilinear skeleton line model and the building response under seismic excitation into non-damaged building data, slightly damaged building data and severely damaged building data under seismic excitation.

[0026] Optionally, the amplitude modulating the vibration sample data under the initial earthquake excitation includes: multiplying the amplitudes of small vibration samples and / or medium vibration samples in the vibration sample data under the initial earthquake excitation by a preset multiple.

[0027] Compared with the existing technology, the earthquake damage assessment model training method provided by the present invention has the following beneficial effects:

[0028] The earthquake damage assessment model training method provided by an embodiment of the present invention establishes a building vibration database, which includes building vibration data corresponding to various vibration labels; extracts characteristic parameters of the building vibration data based on time domain analysis and frequency domain analysis methods; obtains parameter information of an artificial neural network, which includes the number of hidden layers, the number of neurons, the confidence threshold, the optimization algorithm and the loss function; based on the parameter information of the artificial neural network, uses the characteristic parameters of the building vibration data as input and the vibration labels as output to train an earthquake damage assessment model, so that the trained building earthquake damage model can accurately evaluate various types of vibration data.

[0029] An embodiment of the present invention provides an electronic system, comprising: an acquisition module for acquiring real-time building vibration data of a target building; a feature extraction module for extracting characteristic parameters of the real-time building vibration data based on time domain analysis and frequency domain analysis methods, the characteristic parameters including the root mean square amplitude of a time domain graph, the frequency and amplitude corresponding to the peak of a frequency domain amplitude spectrum, and the power bucket of a frequency domain power spectrum; wherein the power bucket is the average value of the power spectrum density within each pre-divided frequency band; and an evaluation module for inputting the characteristic parameters of the real-time building vibration data into a seismic damage assessment model to obtain a seismic damage assessment result for the target building; wherein the seismic damage assessment model is an artificial neural network.

[0030] The electronic system provided by the present invention has the beneficial effect of being able to achieve the same technical effect as the above-mentioned building earthquake damage assessment method. To avoid repetition, it will not be described here. BRIEF DESCRIPTION OF THE DRAWINGS

[0031] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the following briefly introduces the drawings required for use in the embodiments or the description of the prior art. Obviously, the drawings described below are merely embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on the provided drawings without paying any creative work.

[0032] Figure 1 A schematic flow chart of a building earthquake damage assessment method provided by an embodiment of the present invention;

[0033] Figure 2 A schematic flow chart of a method for training an earthquake damage assessment model provided in an embodiment of the present invention;

[0034] Figure 3 Schematic diagram of a nonlinear multi-degree-of-freedom shear model in an embodiment of the present invention;

[0035] Figure 4 Schematic diagram of a trilinear skeleton model according to an embodiment of the present invention;

[0036] Figure 5 A schematic flowchart of a near-real-time building earthquake damage assessment method based on edge machine learning provided by an embodiment of the present invention;

[0037] Figure 6 Schematic diagram of the hardware structure of a wireless intelligent sensor node in an embodiment of the present invention;

[0038] Figure 7 This is a schematic diagram of the internal logical architecture of the sensor in an embodiment of the present invention;

[0039] Figure 8Schematic diagram of the data structure of the building vibration database in an embodiment of the present invention;

[0040] Figure 9 Schematic diagram of a time-frequency diagram in an embodiment of the present invention;

[0041] Figure 10 Schematic diagram of the frequency domain amplitude spectrum in an embodiment of the present invention;

[0042] Figure 11 Schematic diagram of frequency domain power spectrum in an embodiment of the present invention;

[0043] Figure 12 Schematic diagram of an artificial neural network architecture model in an embodiment of the present invention;

[0044] Figure 13 This is a schematic diagram of the model performance preliminarily estimated on the platform in an embodiment of the present invention;

[0045] Figure 14 A schematic structural diagram of an electronic system provided by an embodiment of the present invention. DETAILED DESCRIPTION

[0046] In order to make the above-mentioned objects, features and advantages of the present invention more obvious and easy to understand, specific embodiments of the present invention are described in detail below with reference to the accompanying drawings.

[0047] The present invention provides a method for assessing earthquake damage to buildings. Figure 1 The schematic flow chart of a building earthquake damage assessment method shown in FIG. 1 includes the following steps:

[0048] S110, acquiring real-time building vibration data of the target building.

[0049] S120 , extracting characteristic parameters of the real-time building vibration data based on time domain analysis and frequency domain analysis methods, wherein the characteristic parameters include the root mean square amplitude of the time domain graph, the frequency and amplitude corresponding to the peak of the frequency domain amplitude spectrum, and the power bucket of the frequency domain power spectrum.

[0050] The power bucket is the average value of the power spectrum density within each pre-divided frequency band.

[0051] It should be noted that the real-time building vibration data may include multiple sample data, each of which has a corresponding time domain graph, frequency domain amplitude spectrum, and frequency domain power spectrum.

[0052] Optionally, the feature extraction calculation formula of the above amplitude root mean square is:

[0053]

[0054] Among them, F RMS is the root mean square amplitude, xh is the amplitude corresponding to the discrete sampling sequence number h in the time domain graph, H is the total number of discrete sampling sequences in the time domain graph, and the value range of h is [1, H];

[0055] The feature extraction calculation formula for the frequency and amplitude corresponding to the above peak is:

[0056] F Peak (i) freq =k i (2)

[0057] F Peak (i) amp =X i (k i ) (3)

[0058] F Peak =PF Peak (1) freq ,···,F Peak (I) freq ,F Peak (1) amp ,···,F Peak (I) amp}(4)

[0059] Among them, F Peak (i) freq is the frequency characteristic parameter corresponding to the i-th peak in the frequency domain amplitude spectrum, F Peak (i) amp is the amplitude characteristic parameter corresponding to the i-th peak in the frequency domain amplitude spectrum, k i is the frequency corresponding to the i-th peak, X i (k i ) is the amplitude corresponding to the i-th peak, F Peak is a set of amplitude characteristic parameters and frequency characteristic parameters, I is the total number of peaks in the high amplitude frequency band in the frequency domain amplitude spectrum, and the value range of i is [1, I];

[0060] The feature extraction calculation formula of the above power bucket is:

[0061]

[0062] Among them, k n is the frequency corresponding to the discrete sampling sequence number n in the above frequency domain power spectrum, N is the total number of discrete sampling sequences in the above frequency domain power spectrum, P(k n ) is the frequency k n The corresponding power density, X(k n ) is the frequency k n The corresponding amplitude, is the characteristic parameter of the mth power bucket, kmin is the lower frequency limit of the power bucket, k max is the frequency upper limit of the above power bucket, mean is the calculation frequency k min With frequency k max The average power density between PSD is the set of power buckets, M is the total number of power buckets, and the value range of m is [1, M].

[0063] It should be noted that the above characteristic parameters can be extracted either in advance or in real time. These characteristic parameters can effectively distinguish different types of vibrations, providing an effective guarantee for saving computing power resources of edge computing devices.

[0064] Optionally, the earthquake damage assessment result of the target building includes one of multiple preset vibration labels; the total number of peaks in the frequency domain amplitude spectrum is the maximum number of peaks in the high amplitude frequency band in the frequency domain amplitude spectrum corresponding to the vibration sample data in the building vibration database, and the vibration sample data in the building vibration database is the basic data for training the earthquake damage assessment model; the total number of power buckets is the number of frequency band divisions of the frequency domain power spectrum corresponding to the vibration sample data in the building vibration database, and the number of frequency band divisions is used to divide the high power level frequency bands corresponding to the preset vibration labels.

[0065] It should be noted that the high-power level frequency band corresponding to the above-mentioned preset vibration label is: the high-power level frequency band corresponding to the preset vibration label is obtained by classifying the vibration label and analyzing the high-power corresponding frequency based on the frequency domain power spectrum of the training sample data of the above-mentioned earthquake damage assessment model.

[0066] S130, inputting characteristic parameters of the real-time building vibration data into an earthquake damage assessment model to obtain an earthquake damage assessment result of the target building.

[0067] Optionally, the earthquake damage assessment model is an artificial neural network (ANN).

[0068] Optionally, the above-mentioned step S130 includes: inputting the characteristic parameters of the real-time building vibration data into the earthquake damage assessment model to obtain the confidence of the above-mentioned preset vibration label; wherein the above-mentioned earthquake damage assessment model includes a pre-set confidence threshold; if the confidence of the above-mentioned preset vibration label is greater than or equal to the above-mentioned confidence threshold, then outputting the corresponding above-mentioned preset vibration label.

[0069] It should be noted that if the confidence level of the preset vibration tag is less than the above confidence level threshold, the above preset vibration tag will not be output or the vibration abnormality information will not be output.

[0070] The present invention also provides a method for training a earthquake damage assessment model. Figure 2 A schematic flow chart of a method for training an earthquake damage assessment model is shown, the method comprising:

[0071] S210: Establish a building vibration database, where the building vibration database includes building vibration data corresponding to various vibration tags.

[0072] Optionally, the above-mentioned building vibration data includes building static pulsation data under non-seismic excitation, elevator operation vibration data and building construction vibration data, as well as building non-damage data, building slightly damaged data and building seriously damaged data under seismic excitation; the above-mentioned vibration labels include static vibration, elevator vibration, construction vibration, green label, yellow label and red label.

[0073] It should be noted that "static vibration" corresponds to the static pulsation data of the building, "elevator vibration" corresponds to the elevator operation vibration data, "construction vibration" corresponds to the building construction vibration data, "green label" corresponds to the building undamaged data, "yellow label" corresponds to the building slightly damaged data, and "red label" corresponds to the building seriously damaged data.

[0074] Optionally, the above step S210 includes: acquiring vibration sample data under non-seismic excitation, vibration sample data under initial seismic excitation, and design parameters of the target building.

[0075] Based on whether the elevators and construction equipment in the target building are in operation, the vibration sample data under non-seismic excitation is classified to obtain the building static pulsation data, elevator operation vibration data and building construction vibration data under non-seismic excitation; thus, the vibration sample data under non-seismic excitation is classified.

[0076] A nonlinear multi-degree-of-freedom (MDOF) shear model is established based on the design parameters of the target building; the vibration sample data under the initial seismic excitation is amplitude modulated to obtain the vibration sample data under the seismic excitation; a nonlinear time history analysis is performed on the vibration sample data under the seismic excitation based on the above nonlinear multi-degree-of-freedom shear model to obtain the building response data under the seismic excitation; the vibration data under the seismic excitation is classified into non-damaged building data, slightly damaged building data and severely damaged building data under the seismic excitation based on the trilinear skeleton line model and the building response under the seismic excitation; the vibration sample data under the seismic excitation is classified so that the classified vibration sample data corresponds to the corresponding label.

[0077] Optionally, the step of amplitude modulating the vibration sample data under the initial earthquake excitation includes: multiplying the amplitude of the small vibration samples or medium vibration samples in the vibration sample data under the initial earthquake excitation by a preset multiple, so that the vibration sample data under the earthquake excitation are evenly distributed in the Peak Ground Acceleration (PGA) level table. Among them, some small vibration samples or medium vibration samples do not need to be adjusted to large vibration samples and need to be multiplied by a multiple of 1; another part of the small vibration samples or medium vibration samples needs to be adjusted to large vibration samples and needs to be multiplied by the corresponding required multiples; the set of multiples by which the small vibration samples or medium vibration samples are multiplied constitutes the preset multiples.

[0078] It should be noted that the number of large vibration samples in the actual vibration sample data collected under earthquake excitation is far less than the number of small and medium vibration samples. Amplitude modulation of the vibration sample data under initial earthquake excitation can make the number of the three vibration samples similar, ensuring that the model does not cause differences in the accuracy of the three vibration type labels due to uneven sample distribution. The amplitude of the vibration sample is the peak ground acceleration.

[0079] S220 , extracting characteristic parameters of the building vibration data based on time domain analysis and frequency domain analysis methods.

[0080] S230, obtaining parameter information of the artificial neural network.

[0081] Among them, the above parameter information includes the number of hidden layers, the number of neurons, the confidence threshold, the optimization algorithm and the loss function.

[0082] S240 , based on the parameter information of the artificial neural network, a earthquake damage assessment model is obtained by training with the characteristic parameters of the building vibration data as input and the vibration labels as output.

[0083] Optionally, the method further includes lightweight compilation of the earthquake damage assessment model using edge machine learning (EML) technology to generate model firmware suitable for embedded devices; and deploying the model firmware on the sensor edge. This avoids the transmission of large amounts of post-earthquake time series data, providing a useful reference for near-real-time building damage assessment and post-earthquake emergency decision-making.

[0084] The earthquake damage assessment model training method provided in this embodiment, by establishing a building vibration database under earthquake and non-earthquake conditions, extracting representative characteristic parameters of vibration data based on time domain and frequency domain analysis methods, can better divide data of different vibration types. The above sample data is further trained and the model architecture is optimized through machine learning algorithms to obtain a neural network model that performs well in near-real-time assessment results of building earthquake damage. Finally, after evaluating and compiling the neural network model, it is uploaded to the edge of the wireless intelligent sensor node. This method directly obtains building response data based on sensors, and the wireless intelligent sensor node only needs to upload the evaluation results to the cloud server, avoiding the transmission of large amounts of time series data in post-earthquake emergency scenarios, while ensuring accuracy while also greatly improving timeliness.

[0085] In one embodiment, to illustrate the contents of the building vibration database in detail, this embodiment provides a specific implementation method for establishing the building vibration database under earthquake and non-earthquake conditions, which can be performed by referring to the following method:

[0086] On the one hand, wireless intelligent sensor nodes are deployed near the elevator shaft on the top floor of the target building to collect daily vibration data. The sensors can collect three main types of non-seismic vibration: static vibration, elevator vibration, and construction vibration. It should be noted that in actual applications, the non-seismic vibration classification can be determined based on the actual equipment conditions and sensor location within the building.

[0087] On the other hand: referring to the local fault type and site conditions of the target building, a seismic motion set is selected. According to the design parameters of the target building, a corresponding nonlinear multi-degree-of-freedom shear model is established. The above seismic motion set is further input into the nonlinear multi-degree-of-freedom shear model for nonlinear time history analysis to obtain the building response under different seismic motion sequences. The building damage level is divided into three levels: green label (label for building data with no damage), yellow label (label for building data with slight damage), and red label (label for building data with severe damage) through the trilinear skeleton line model. Finally, the acceleration time history response of the top floor is output to the database together with the damage status label.

[0088] Specifically, see Figure 3The schematic diagram of the nonlinear multi-degree-of-freedom shear model shown in the figure is that the nonlinear multi-degree-of-freedom shear model concentrates the mass of each floor of the building on a mass point, and the adjacent floors are simplified to shear springs for connection. The nonlinear multi-degree-of-freedom shear model can better reflect the nonlinear seismic performance of multi-story buildings, predict the engineering demand parameters of each floor (maximum inter-story displacement angle and floor peak acceleration, etc.), and consider the damage concentration of different floors. The building damage levels under the above seismic excitation include three levels: green label, yellow label and red label. After using the simplified nonlinear multi-degree-of-freedom shear model for nonlinear time history analysis, not only the top floor acceleration response will be obtained, but also the top floor maximum inter-story displacement ratio will be obtained. The above-mentioned top floor maximum inter-story displacement ratio is crucial for determining the damage state of the building. See Figure 4 The schematic diagram of the trilinear skeleton model is shown in Figure 1. The vibration classification method under earthquake excitation is based on the trilinear skeleton line. If the maximum inter-story displacement of the top floor output by the nonlinear multi-degree-of-freedom shear model analysis does not reach the corresponding building stress at its yield strength V y , the structure is considered to be in a non-damaged state and the vibration is classified as green. If the maximum inter-story displacement of the top floor output by the nonlinear multi-degree-of-freedom shear model is greater than the corresponding building stress at the yield strength V y and yield peak strength V p If the maximum inter-story displacement output by the nonlinear multi-degree-of-freedom shear model analysis is greater than the corresponding building stress reaching the yield peak strength V p After that, the structure is considered to be in a state of severe damage and the vibration is classified as red label. It should be noted that in order to ensure the conservatism of the evaluation results, 4 (V p +V y ) / 5 is the boundary between the yellow and red labels, where 4V y / 5 is less than V p / 5.

[0089] In addition, the problem of insufficient number of samples in the above-mentioned seismic motion collection can be solved by the following method: the selected initial seismic motion records are amplitude modulated according to different levels of ground peak acceleration, so as to obtain a sufficient number of seismic motion sequences to be input.

[0090] Specifically, 22 earthquake motion records were selected with reference to the local fault type and site conditions of the target building. The 22 earthquake motion records were amplitude modulated according to 20 different levels of PGA, that is, between 0.1 and 1 m / s. 2 0.1m / s 2 Set 10 PGA levels for the gradient, between 2 and 11 m / s 2 1m / s 2Ten PGA levels are set for the gradient. By multiplying the amplitudes of small and / or medium vibration samples in the earthquake motion concentration by a preset multiple, there are 1 to 2 samples distributed at each PGA level, ultimately obtaining a sufficient number of earthquake motion records to be input.

[0091] In one embodiment, to conserve computing resources on edge computing devices and improve computational efficiency, this embodiment provides a specific implementation for extracting effective eigenvalues ​​of sample data from a building vibration database. The following method can be used:

[0092] First, the root mean square (RMS) of the amplitude is selected as the time-domain characteristic parameter of vibration to characterize the time-domain vibration characteristics of building vibration data. The RMS calculation formula for each sample data is shown in Equation (1). Compared with other instantaneous characteristic parameters such as maximum, minimum, and peak values, RMS is not significantly affected by instantaneous interference and is more sensitive to distinguishing different types of vibration signals.

[0093] The second aspect is to select the peak point amplitude and the corresponding frequency as the characteristic parameters of the frequency domain amplitude spectrum to characterize the vibration characteristics of the high amplitude frequency band of the building vibration data. The calculation formula of the corresponding characteristic of the frequency domain peak of each sample data is shown in formulas (2)-(4).

[0094] Third aspect: Compared with the spectrum diagram (i.e., the frequency domain amplitude spectrum), the analysis of the power spectrum diagram can identify the power level of each frequency band of the building vibration data. Therefore, the power bucket is used as the frequency domain dimension feature parameter to characterize the power characteristics of the building vibration data in a specific frequency band. According to the upper and lower boundaries of the pre-divided frequency band range, the average value of the power spectrum density within this range is calculated as the extracted feature parameter. The power bucket calculation formula for each sample data is shown in (5)-(7).

[0095] Among them, the division of frequency bands is mainly based on the frequency domain power spectrum density characteristics of the building vibration data in the database, so as to divide the frequency band range of different types of vibration and the specific power bucket boundaries, and extract the frequency domain power spectrum characteristic parameters.

[0096] In one embodiment, to improve the accuracy and efficiency of building damage assessment results, this embodiment provides a specific implementation method for using a machine learning algorithm to train a high-performance neural network model. The following method can be used as a reference:

[0097] Step (1): Set the input of the ANN model architecture to the time domain features and frequency domain features extracted from the building vibration data signal; set the output to the vibration classification label; set different parameters, including the number of hidden layers, the number of neurons, the confidence threshold, the optimization algorithm and the loss function.

[0098] Step (2): Train the model architecture under different parameter settings, compare the accuracy of the near-real-time assessment results of building earthquake damage by different training models, and obtain the neural network model with excellent performance.

[0099] In one embodiment, to avoid the transmission of large amounts of time series data in post-earthquake emergency scenarios, this embodiment provides a specific implementation method for uploading a neural network model to the edge of a wireless intelligent sensor node. The following method can be used:

[0100] Step (1): Upload the obtained neural network model to the development platform in the EML field, and the platform will preliminarily estimate the model performance.

[0101] It should be noted that EML technology, by combining edge computing and machine learning, provides a solution for building earthquake damage assessment that is both timely and accurate.

[0102] Step (2): Optimize the model memory using the compiler provided by the platform, then compile the firmware using programming software and upload it to the edge node. Deploy the model firmware, which can be used for embedded devices, at the edge of the sensor, ultimately enabling near-real-time assessment of building earthquake damage at the edge of the sensor.

[0103] The embodiment of the present invention also provides a specific method for near real-time assessment of building earthquake damage based on edge machine learning, see Figure 5 The schematic flow chart of the near-real-time building earthquake damage assessment method based on edge machine learning is shown in the figure. The specific steps of the method are as follows:

[0104] Step S1: deploy wireless intelligent sensor nodes to collect building vibration data under non-seismic conditions.

[0105] Specifically, wireless intelligent sensor nodes are deployed next to the elevator shaft on the top floor of the target building to collect daily vibration data. The types of non-seismic vibrations that the sensor can collect mainly include static vibration, elevator vibration, and construction vibration. The sensor collects building vibration data under non-seismic conditions, that is, the vibration sample data under the above-mentioned non-seismic excitation. The window length of each sample data is set to 20s, with an overlap of 18s, and some samples with poor data quality are deleted. In order to ensure the reliability and generalization ability of the subsequent training model, each type of vibration data is collected twice independently. Figure 6 The hardware structure of the wireless intelligent sensor node shown in the figure includes power supply components, memory, accelerometer, microcontroller and NB-IoT module. Figure 7The sensor's internal logic architecture is shown in the figure. The sensor is powered by a battery module and a solar module. An accelerometer acquires digital signals of building vibration acceleration, which are then uploaded to a single-chip microcontroller for near-real-time earthquake damage assessment. The NB-IoT module connects the sensor to a cloud server.

[0106] Step S2: Acquire a sufficient number of earthquake motion records.

[0107] Specifically, 22 earthquake motion records were selected with reference to the local fault type and site conditions of the target building. The 22 earthquake motion records were amplitude modulated according to 20 different levels of PGA, that is, between 0.1 and 1 m / s. 2 0.1m / s 2 Set 10 PGA levels for the gradient, between 2 and 11 m / s 2 1m / s 2 Ten PGA levels are set for the gradient, and finally, a sufficient number of ground motion records to be input are obtained at each PGA level, which are the vibration sample data under the above earthquake excitation.

[0108] Step 3: Perform nonlinear time-history analysis to obtain building structure response data.

[0109] Specifically, the earthquake motion records are input into a nonlinear multi-degree-of-freedom shear model established according to the design parameters of the target building, and nonlinear time history analysis is performed to obtain the building response data under different earthquake sequences. It should be noted that the above-mentioned building response data is the building force.

[0110] Step 4: Differentiate the building damage levels using the trilinear skeleton model.

[0111] Specifically, a trilinear skeleton model was used to categorize building damage levels into green, yellow, and red. A 20-second window was taken with the time corresponding to the building's peak response as the midpoint, and the ground motion records were entered into the database along with the damage status labels.

[0112] Step 5: Summarize the building vibration data under earthquake and non-earthquake conditions to form a building vibration database.

[0113] Specifically, see Figure 8The data structure diagram of the building vibration database is shown in the figure. For the three types of vibration data under non-seismic conditions, the data collected for the first time are divided into training vibration data and verification vibration data in a ratio of 8:2, and the data collected for the second time are specifically used to construct test vibration data. For the vibration data under seismic conditions, the first 17 of the 22 seismic motions are selected as training seismic motions, and the last 5 are selected as test seismic motions. After the training seismic motions are amplitude modulated according to the above 20 PGA levels, the training and verification vibration data are established in a ratio of 8:2, and the test seismic motions are amplitude modulated in the same way to obtain the test vibration data. According to the above steps, a building vibration database containing 6 types of vibration types under seismic and non-seismic conditions is established, laying the foundation for the subsequent machine learning work to divide the training set, verification set and test set.

[0114] Step 6: Calculate the root mean square of the sample data.

[0115] Specifically, the RMS of each sample data is calculated as the time domain characteristic parameter of the building vibration data to characterize the vibration characteristics of the building vibration data.

[0116] A time-frequency plot visualizes the changes in a vibration signal over time and frequency. It displays the changes in the signal's frequency components at different time points on the same graph. The plotting principle involves setting the window length and overlap ratio, segmenting each sample data point in time, and performing a Fast Fourier Transform (FFT) on each segment to extract the frequency components and corresponding amplitude information within that time window. By traversing all time windows and calculating the spectrum of each window, the spectral characteristics of the signal at different time points can be obtained.

[0117] For example, see Figure 9 The diagram shows a schematic diagram of a time-frequency graph. When plotting a time-frequency graph, the horizontal axis represents time (Time) and the vertical axis represents frequency (Frequency). Color or brightness represents the intensity, or amplitude, of a frequency component at a given moment. A time-frequency analysis of the data in the database reveals that the amplitude bands under non-seismic excitation are relatively stable. Idel and elevator vibration exhibit stable high-amplitude bands at one or two locations, while construction vibration caused by equipment such as hammer drills is mostly located in high-frequency bands with larger amplitudes. Vibration types under seismic excitation include green, yellow, and red tags. The high-amplitude bands under seismic excitation exhibit significant temporal variation, with core high-amplitude components located at locations 1 to 3. As the damage level increases, the high-amplitude bands broaden and the building's natural frequency decreases significantly.

[0118] Step 7: Calculate the peak value and corresponding frequency of the amplitude spectrum.

[0119] Specifically, FFT analysis is performed on each sample data, and the amplitude spectrum part selects the peak point amplitude and corresponding frequency of the preset number of groups of amplitude spectra as characteristic parameters, which are used to characterize the vibration characteristics of the high-amplitude frequency band in the frequency domain of the building vibration data; wherein the preset number of groups is the above-mentioned total number of peaks, and the above-mentioned total number of peaks is the maximum number of peaks in the high-amplitude frequency band in the frequency domain amplitude spectrum corresponding to the vibration sample data in the building vibration database.

[0120] For example, see Figure 10 The diagram shows a frequency domain amplitude spectrum, where the horizontal axis is frequency (Frequencey) and the vertical axis is amplitude (Amplitude), and it includes three peaks (Peak). The data samples in the building vibration database are subjected to amplitude spectrum analysis to determine the location of the peak point, and extract its frequency and corresponding amplitude, that is, the horizontal and vertical coordinates are used as amplitude features and frequency features. In addition, the time-frequency graph can be analyzed to determine the number of groups of peak points in the frequency domain amplitude spectrum to be extracted. There are at most three high-amplitude frequency bands of vibration corresponding to each data label. Therefore, according to the analysis of the time-frequency graph, the number of groups of peak points in the frequency domain amplitude spectrum to be extracted is determined to be 3, that is, three peak points and their corresponding frequencies. That is, by extracting the three highest peak features in the frequency domain amplitude spectrum, the frequency domain peak features of each sample data can be extracted.

[0121] Step 8: Calculate the average value of the power spectrum density of the frequency domain power spectrum within the frequency band.

[0122] Specifically, the power bucket is used as a frequency domain feature parameter. Based on the upper and lower boundaries of the pre-defined frequency band, the average value of the power spectrum density within this range is calculated as the extracted feature parameter. This is used to characterize the power characteristics of the building vibration data within a specific frequency band. A total of 14 frequency domain power spectrum feature parameters were extracted.

[0123] Table 1

[0124]

[0125]

[0126] Furthermore, it can be seen from Table 1 that the high-power level frequency bands of different vibrations are different, and the segments corresponding to the green, yellow, and red labels are 1.5-2.5, 1.0-2.5, and 0-2.0, respectively, which are relatively close. In order to enable the subsequent training model to effectively distinguish these three vibrations, a segment is divided every 0.5Hz in the 0-3.0Hz frequency band; after 3Hz, there are static vibration and elevator vibration, and their high-power level frequency bands are [5,7] and [4,7] respectively. The span is large, but there are also overlapping parts, and it is necessary to divide the difference parts. In the segmentation, the subsequent segments are divided into [3.0, 5.0], [5.0, 7.5], and [7.5, 10]. The final construction vibration segment is [14, 20], so the segment is divided into [10, 15] and [15, 20]. The subsequent 20-50 Hz frequency band does not have a high power level frequency band, but it cannot be divided into only one frequency band. This is because the feature extracted from each power bucket is the power mean. If the frequency band span is too large, the power mean will become larger and may be close to the feature values ​​extracted from some previous frequency bands, which is not conducive to model differentiation.

[0127] For example, see Figure 11 The schematic diagram of the frequency domain power spectrum shown in the figure divides the 0-50 Hz frequency domain segment into four power buckets, and calculates the power spectrum density mean of each power bucket.

[0128] It should be noted that according to the principle of FFT, the bandwidth is 50 Hz after the 100 Hz sampling rate is converted into the frequency domain amplitude spectrum and the frequency domain power spectrum. That is to say, the 100 Hz sampling rate can only collect vibrations within 50 Hz. Therefore, the 50-100 Hz graph and the 0-50 Hz graph are symmetrical about x = 50 Hz.

[0129] In step 9, the representative characteristic parameters of the vibration data are divided into a training set, a validation set, and a test set according to different proportions.

[0130] Specifically, according to the vibration data types and proportions divided in step 5, the representative characteristic parameters corresponding to the vibration data are divided into a training set, a validation set, and a test set for model training.

[0131] Step 10: Set the artificial neural network model architecture under different parameters to complete model training; compare the prediction accuracy under different architectures and optimize the model.

[0132] Specifically, see Figure 12The following is a schematic diagram of the artificial neural network architecture model. The input is set to 21 time-domain and frequency-domain features extracted from building vibration signals. The architecture is optimized with 1-2 hidden layers and 1-20 neurons per hidden layer. A dropout layer with a probability of 0.1 is added after each hidden layer to prevent overfitting. The output is the vibration classification label, with a confidence threshold of 60%. If no category exceeds the confidence threshold, the vibration is considered abnormal. The optimization algorithm uses the Adam algorithm with a learning rate of 0.001 and a cross-entropy loss function for multi-class classification. Training is performed on each architecture to select the neural network model that performs best for building damage assessment.

[0133] Step 11: Upload the neural network model to the edge of the wireless intelligent sensor node to achieve near real-time assessment of earthquake damage at the sensor edge.

[0134] Specifically, we use a development platform that has excellent performance in the EML field. Figure 13 The following diagram shows the preliminary model performance estimates from the platform. The model inference time is 10ms, meeting the requirements for near-real-time earthquake damage assessment. Peak memory usage is 1.4KB, and flash memory usage is 15.2KB, making it suitable for smart nodes with limited computing resources. The platform then uses an edge-optimized neural compiler to optimize the model memory and output a header file for customized programming. Finally, after the code is written, the programming software compiles the firmware and uploads it to the edge node. This ultimately enables near-real-time earthquake damage assessment at the sensor edge.

[0135] The above-mentioned near-real-time assessment method for building earthquake damage based on edge machine learning provided in this embodiment establishes a building vibration database under earthquake and non-earthquake conditions, extracts the time domain characteristics and frequency domain characteristics of the sample data, and applies the machine learning algorithm to obtain a neural network model with excellent performance in building earthquake damage assessment results. After evaluating and compiling the neural network model, it is uploaded to the edge of the wireless intelligent sensor node, avoiding the transmission of large amounts of time series data in post-earthquake emergency scenarios. While ensuring accuracy, it also greatly improves timeliness, providing a useful reference for near-real-time assessment of building earthquake damage and rapid post-earthquake emergency response.

[0136] An embodiment of the present invention provides an electronic system, see Figure 14 The structure diagram of an electronic system shown in FIG. 1 includes:

[0137] The acquisition module 1402 is used to acquire real-time building vibration data of the target building.

[0138] Feature extraction module 1404 is used to extract characteristic parameters of real-time building vibration data based on time domain analysis and frequency domain analysis methods. The above characteristic parameters include the root mean square amplitude of the time domain graph, the frequency and amplitude corresponding to the peak of the frequency domain amplitude spectrum, and the power bucket of the frequency domain power spectrum.

[0139] The power bucket is the average value of the power spectrum density within each pre-divided frequency band.

[0140] The evaluation module 1406 is used to input the characteristic parameters of the real-time building vibration data into the earthquake damage assessment model to obtain the earthquake damage assessment result of the target building.

[0141] Among them, the above-mentioned earthquake damage assessment model is an artificial neural network.

[0142] The electronic system provided by the present invention has the beneficial effect of being able to achieve the same technical effect as the above-mentioned building earthquake damage assessment method. To avoid repetition, it will not be described here.

[0143] Of course, those skilled in the art will understand that all or part of the processes in the above-mentioned embodiment methods can be implemented by instructing the control device through a computer program, and the program can be stored in a computer-readable storage medium. When the program is executed, it may include the processes of the above-mentioned method embodiments, wherein the storage medium may be a memory, a disk, an optical disk, etc.

[0144] Finally, it should be noted that, in this document, relational terms such as first and second, etc., are used only to distinguish one entity or operation from another entity or operation, and do not necessarily require or imply any actual relationship or order between these entities or operations. Moreover, the terms "comprises," "comprising," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or device comprising a series of elements includes not only those elements, but also other elements not explicitly listed, or elements inherent to such process, method, article, or device. In the absence of further limitations, an element defined by the phrase "comprising a ..." does not exclude the presence of additional identical elements in the process, method, article, or device comprising the element.

[0145] The various embodiments in this specification are described in a progressive manner, and each embodiment focuses on the differences from other embodiments. The same or similar parts between the various embodiments can be referenced to each other.

[0146] Although the present invention is disclosed as above, the present invention is not limited thereto. Any person skilled in the art can make various changes and modifications without departing from the spirit and scope of the present invention. Therefore, the scope of protection of the present invention should be based on the scope defined by the claims.

Claims

1. A building earthquake damage assessment method, characterized in that: The method comprises the following steps: Obtain real-time building vibration data of the target building; Based on time domain analysis and frequency domain analysis methods, characteristic parameters of the real-time building vibration data are extracted. The characteristic parameters include the root mean square amplitude of the time domain graph, the frequency and amplitude corresponding to the peak of the frequency domain amplitude spectrum, and the power bucket of the frequency domain power spectrum. The power bucket is the average value of the power spectrum density within each pre-divided frequency band. The characteristic parameters of the real-time building vibration data are input into a seismic damage assessment model to obtain a seismic damage assessment result of the target building; wherein the seismic damage assessment model is an artificial neural network.

2. The method according to claim 1, characterized in that The feature extraction calculation formula of the amplitude root mean square is: Among them, F RMS is the RMS amplitude, x h is the amplitude corresponding to the discrete sampling sequence number h in the time domain graph, H is the total number of discrete sampling sequences in the time domain graph, and the value range of h is [1, H]; The feature extraction calculation formula for the frequency and amplitude corresponding to the peak is: F Pea k (i) freq =k i F Pea k (i) amp =X i (k i ) F Peak ={F Peak (1) freq ,...,F Peak (I) freq ;F Peak (1) amp ,...,F Peak (I) amp } Among them, F Peak (i) freq is the frequency characteristic parameter corresponding to the i-th peak in the frequency domain amplitude spectrum, F Peak (i) amp is the amplitude characteristic parameter corresponding to the i-th peak in the frequency domain amplitude spectrum, k i is the frequency corresponding to the i-th peak, X i (k i ) is the amplitude corresponding to the i-th peak, F Peak is a set of the amplitude characteristic parameter and the frequency characteristic parameter, I is the total number of peaks in the high amplitude frequency band in the frequency domain amplitude spectrum, and the value range of i is [1, I]; The feature extraction calculation formula of the power bucket is: Among them, k n is the frequency corresponding to the discrete sampling sequence number n in the frequency domain power spectrum, N is the total number of discrete sampling sequences in the frequency domain power spectrum, P(k n ) is the frequency k n The corresponding power density, X(k n ) is the frequency k n The corresponding amplitude, is the characteristic parameter of the mth power bucket, k min is the lower frequency limit of the power bucket, k max is the upper frequency limit of the power bucket, mean is the calculation frequency k min With frequency k max The average power density between PSD is the set of the power buckets, M is the total number of power buckets, and the value range of m is [1, M].

3. The method according to claim 2, characterized in that The earthquake damage assessment result of the target building includes one of a plurality of preset vibration labels; The total number of peaks in the frequency domain amplitude spectrum is the maximum number of peaks in the high amplitude frequency band in the frequency domain amplitude spectrum corresponding to the vibration sample data in the building vibration database, and the vibration sample data in the building vibration database is the basic data for training the earthquake damage assessment model; The total number of power buckets is the number of frequency band divisions of the frequency domain power spectrum corresponding to the vibration sample data in the building vibration database, and the number of frequency band divisions is used to divide and distinguish the high-power level frequency bands corresponding to each of the preset vibration labels.

4. The method according to claim 3, characterized in that Inputting the characteristic parameters of the real-time building vibration data into a seismic damage assessment model to obtain a seismic damage assessment result for the target building includes: Inputting the characteristic parameters of the real-time building vibration data into the earthquake damage assessment model to obtain the confidence level of the preset vibration label; wherein the earthquake damage assessment model includes a preset confidence level threshold; If the confidence level of the preset vibration label is greater than or equal to the confidence threshold, the corresponding preset vibration label is output.

5. A method for training an earthquake damage assessment model, characterized in that: The method comprises: Establishing a building vibration database, wherein the building vibration database includes building vibration data corresponding to various vibration tags; Extracting characteristic parameters of the building vibration data based on time domain analysis and frequency domain analysis methods; Obtaining parameter information of an artificial neural network, the parameter information including the number of hidden layers, the number of neurons, a confidence threshold, an optimization algorithm, and a loss function; Based on the parameter information of the artificial neural network, an earthquake damage assessment model is obtained by training with the characteristic parameters of the building vibration data as input and the vibration label as output.

6. The method according to claim 5, characterized in that The method further comprises: Lightweight compilation of the earthquake damage assessment model is performed based on edge machine learning technology to obtain model firmware that can be oriented to embedded devices; The model firmware is configured at the sensor edge.

7. The method according to claim 5, characterized in that The building vibration data includes building static pulsation data under non-seismic excitation, elevator operation vibration data and building construction vibration data, as well as building non-damage data, building slightly damaged data and building seriously damaged data under seismic excitation; the vibration labels include static vibration, elevator vibration, construction vibration, green label, yellow label and red label.

8. The method according to claim 7, characterized in that The establishment of the building vibration database comprises: Obtain vibration sample data under non-seismic excitation, vibration sample data under initial seismic excitation, and design parameters of the target building; Classifying the vibration sample data under non-seismic excitation according to whether the elevator and construction equipment in the target building are in operation, to obtain the static pulsation data of the building under non-seismic excitation, the elevator operation vibration data, and the building construction vibration data; Establishing a nonlinear multi-degree-of-freedom shear model according to the design parameters of the target building; Amplitude modulating the vibration sample data under the initial earthquake excitation to obtain vibration sample data under the earthquake excitation; Performing nonlinear time history analysis on the vibration sample data under the earthquake excitation based on the nonlinear multi-degree-of-freedom shear model to obtain building response data under the earthquake excitation; The vibration data under the earthquake excitation is classified based on a trilinear skeleton line model and the building response under the earthquake excitation into the building non-damage data under the earthquake excitation, the building slightly damaged data and the building seriously damaged data.

9. The method according to claim 8, characterized in that The amplitude modulating the vibration sample data under the initial earthquake excitation includes: The amplitudes of the small vibration samples and / or the medium vibration samples in the vibration sample data under the initial earthquake excitation are multiplied by a preset multiple.

10. An electronic system, characterized in that: The electronic system comprises: An acquisition module is used to obtain real-time building vibration data of a target building; A feature extraction module is configured to extract characteristic parameters of the real-time building vibration data based on time domain analysis and frequency domain analysis methods. The characteristic parameters include the root mean square amplitude of the time domain graph, the frequency and amplitude corresponding to the peak of the frequency domain amplitude spectrum, and the power bucket of the frequency domain power spectrum. The power bucket is the average value of the power spectrum density within each pre-divided frequency band. An assessment module is used to input the characteristic parameters of the real-time building vibration data into a seismic damage assessment model to obtain a seismic damage assessment result of the target building; wherein the seismic damage assessment model is an artificial neural network.

Images