Unsupervised damage identification method, device and product based on macro-strain wavelet packet energy

Abstract

The present disclosure provides an unsupervised damage identification method, device and product of macro strain wavelet packet energy, and relates to the technical field of computer. The method comprises: acquiring a measured macro strain response signal of a structure under unknown multiple load actions; extracting a target macro strain response signal of a target measuring point from the measured macro strain response signal; respectively performing wavelet packet decomposition on the target macro strain response signal and the measured macro strain response signal to determine wavelet packet energy feature vectors of the target macro strain response signal and the measured macro strain response signal; inputting the wavelet packet energy feature vector of the target macro strain response signal into a deep learning network to eliminate the influence of unknown multiple loads and simultaneously obtain a predicted wavelet packet energy feature vector; calculating a change rate between the predicted wavelet packet energy feature vector and the wavelet packet energy feature vector of the measured macro strain response signal to determine damage change rates of each measuring point; and determining a damage position of the structure object based on the amplitude distribution of the damage change rates of each measuring point.

Description

Technical Field

[0001] This disclosure relates to the field of computer technology, and in particular to an unsupervised damage identification method, device and product based on macro-strain wavelet packet energy. Background Technology

[0002] Existing structural damage identification technologies mostly rely on supervised learning, which requires a large amount of labeled damage data for training. They also generally assume that the load type is the same before and after damage, making it difficult to adapt to the complex working conditions in actual engineering where the load is unknown, variable, and multiple loads act simultaneously. At the same time, the acceleration or displacement response of traditional frequency domain transfer rate methods is not sensitive to local minor damage, and effective frequency bands need to be selected manually. Summary of the Invention

[0003] This disclosure provides an unsupervised damage identification method, device, and product based on macro-strain wavelet packet energy.

[0004] According to one aspect of this disclosure, an unsupervised damage identification method based on macro-strain wavelet packet energy is provided, comprising: acquiring measured macro-strain response signals collected from all measuring points on a structural object to be monitored under unknown multi-load conditions; extracting a target macro-strain response signal corresponding to at least one target measuring point on the structural object from the measured macro-strain response signals, wherein the number of target measuring points is greater than or equal to the number of actual loads on the structural object; performing wavelet packet decomposition on the target macro-strain response signal and the measured macro-strain response signal respectively to determine the wavelet packet energy feature vector of the target macro-strain response signal and the wavelet packet energy feature vector of the measured macro-strain response signal; and performing wavelet packet decomposition based on macro-strain wavelet packet energy. The energy transfer ratio matrix is ​​obtained by inputting the wavelet packet energy feature vector of the target macro-strain response signal into a deep learning network to eliminate the influence of unknown multi-loads and predict the wavelet packet energy feature vectors of all measuring points. The deep learning network is a deep learning network based on a convolutional autoencoder, and the macro-strain wavelet packet energy transfer ratio matrix represents the feature mapping relationship between macro-strain response signals acquired by different sensors. For each measuring point, the rate of change between the predicted wavelet packet energy feature vector and the wavelet packet energy feature vector of the measured macro-strain response signal is calculated to determine the damage change rate of each measuring point. Based on the amplitude distribution of the damage change rate of each measuring point, the damage location of the structural object is determined.

[0005] Based on an unsupervised damage identification method using macro-strain wavelet packet energy, this paper addresses a structural object under unknown multi-load conditions. By comparing the damage change rate between the predicted and measured values ​​of the wavelet packet eigenvectors of the macro-strain response signal at each measuring point, the method effectively amplifies weak anomalous signals caused by local physical damage. Analysis of the spatial amplitude distribution of the damage change rate enables the localization of damage locations within the structural object. This method overcomes the dependence of traditional techniques on precise dynamic models and dense sensor arrangements, exhibits strong robustness to unknown load conditions, and improves the accuracy and reliability of damage identification for structural objects under limited measuring point conditions.

[0006] According to at least one embodiment of the unsupervised damage identification method for macro-strain wavelet packet energy of this disclosure, the training process of the deep learning network includes: applying at least one load excitation to a structural object in a lossless state, and dividing all measurement points into a first measurement point set containing the target measurement point and a second measurement point set containing all measurement points; acquiring a first macro-strain response signal of the first measurement point set and a second macro-strain response signal of the second measurement point set; performing wavelet packet decomposition on the first macro-strain response signal and the second macro-strain response signal to determine the wavelet packet energy feature vector of the first macro-strain response signal and the wavelet packet energy feature vector of the second macro-strain response signal; constructing a convolutional autoencoder framework based on the macro-strain wavelet packet energy transfer ratio matrix, and training the deep learning network with the wavelet packet energy feature vector of the first macro-strain response signal as the input sample and the wavelet packet energy feature vector of the second macro-strain response signal as the output sample.

[0007] According to at least one embodiment of the unsupervised damage identification method based on wavelet packet energy of macro-strain according to this disclosure, for each measurement point, the rate of change between the predicted wavelet packet energy feature vector and the wavelet packet energy feature vector of the measured macro-strain response signal is calculated to determine the damage change rate of each measurement point. This includes: for each measurement point on the structural object, based on the frequency band order formed after wavelet packet decomposition, matching the predicted value in the predicted wavelet packet energy feature vector with the corresponding measured value in the wavelet packet energy feature vector of the measured macro-strain response signal; calculating the difference between each pair of predicted and measured values ​​at the same frequency band position, and summing the squares of the differences across all frequency bands to obtain an error value; dividing the error value by the sum of all measured values ​​in the wavelet packet energy feature vector of the measured macro-strain response signal to determine the damage change rate of each measurement point.

[0008] According to at least one embodiment of the unsupervised damage identification method for macro-strain wavelet packet energy of this disclosure, the wavelet packet decomposition employs orthogonal wavelet basis functions to perform three-level wavelet packet decomposition on the target macro-strain response signal and the measured macro-strain response signal, respectively.

[0009] According to at least one embodiment of the unsupervised damage identification method for macro-strain wavelet packet energy of the present disclosure, the convolutional autoencoder includes an encoder and a decoder, a skip connection is provided between the encoder and the decoder, the encoder includes a convolutional layer and a pooling layer, and the decoder includes a deconvolutional layer or an upsampling layer plus a convolutional layer.

[0010] According to at least one embodiment of the unsupervised damage identification method of macro-strain wavelet packet energy of this disclosure, the damage location of the structural object is determined based on the amplitude distribution of the damage change rate of each measuring point, including: based on the amplitude distribution of the damage change rate of each measuring point, determining the region corresponding to the measuring point whose damage change rate value is greater than or equal to a target threshold and whose damage change rate is the largest as the damage location of the structural object.

[0011] According to at least one embodiment of the unsupervised damage identification method for macro-strain wavelet packet energy of the present disclosure, before inputting the wavelet packet energy feature vector of the target macro-strain response signal into a deep learning network, the method includes: normalizing the wavelet packet energy feature vector of the target macro-strain response signal; and inputting the normalized wavelet packet energy feature vector of the target macro-strain response signal into the deep learning network.

[0012] According to at least one embodiment of the unsupervised damage identification method for macro-strain wavelet packet energy of this disclosure, before determining the damage change rate, the method includes: denormalizing the predicted wavelet packet energy feature vector; and calculating the damage change rate by using the denormalized predicted wavelet packet energy feature vector and the wavelet packet energy feature vector of the measured macro-strain response signal.

[0013] According to another aspect of this disclosure, an electronic device is provided, comprising: a memory storing execution instructions; and a processor executing the execution instructions stored in the memory, such that the processor performs an unsupervised damage identification method for macro-strain wavelet packet energy according to any embodiment of this disclosure.

[0014] According to another aspect of this disclosure, a readable storage medium is provided, wherein execution instructions are stored therein, which, when executed by a processor, are used to implement the unsupervised damage identification method for macro-strain wavelet packet energy according to any embodiment of this disclosure.

[0015] According to another aspect of this disclosure, a computer program product is provided, including a computer program that, when executed by a processor, implements an unsupervised damage identification method for macro-strain wavelet packet energy according to any embodiment of this disclosure. Attached Figure Description

[0016] The accompanying drawings illustrate exemplary embodiments of the present disclosure and, together with the description thereof, serve to explain the principles of the present disclosure. These drawings are included to provide a further understanding of the present disclosure and are incorporated in and constitute a part of this specification.

[0017] Figure 1 This is a schematic diagram of the overall process of an unsupervised damage identification method based on macro-strain wavelet packet energy according to one embodiment of the present disclosure.

[0018] Figure 2 This is a flowchart illustrating the training process of a deep learning network in an unsupervised damage identification method based on macro-strain wavelet packet energy according to one embodiment of the present disclosure.

[0019] Figure 3 This is a flowchart illustrating the normalization of the wavelet packet energy feature vector of the target macro-strain response signal in an unsupervised damage identification method based on the wavelet packet energy of macro-strain according to one embodiment of the present disclosure.

[0020] Figure 4 This is a flowchart illustrating the process of determining the rate of change of damage in an unsupervised damage identification method based on macro-strain wavelet packet energy according to one embodiment of the present disclosure.

[0021] Figure 5 This is a flowchart illustrating the process of inverse normalization of the predicted wavelet packet energy feature vector in an unsupervised damage identification method for macro-strain wavelet packet energy according to one embodiment of the present disclosure.

[0022] Figure 6 This is a schematic diagram of the CAE network architecture in an unsupervised damage identification method based on macro-strain wavelet packet energy according to one embodiment of the present disclosure.

[0023] Figure 7 This is a schematic diagram of the framework of an unsupervised damage identification method based on macro-strain wavelet packet energy according to one embodiment of the present disclosure.

[0024] Figure 8 This is a schematic diagram of a simply supported beam model in an unsupervised damage identification method based on macro-strain wavelet packet energy according to one embodiment of the present disclosure.

[0025] Figure 9 This is a schematic diagram of the damage identification results under test condition 1 in an unsupervised damage identification method based on macro-strain wavelet packet energy according to an embodiment of the present disclosure.

[0026] Figure 10 This is a schematic diagram of the damage identification results under test condition 2 in an unsupervised damage identification method based on macro-strain wavelet packet energy according to an embodiment of this disclosure.

[0027] Figure 11This is a schematic diagram of the damage identification results under harmonic and stationary loads in an unsupervised damage identification method based on macro-strain wavelet packet energy according to one embodiment of the present disclosure.

[0028] Figure 12 This is a schematic diagram of damage identification results under harmonic and non-stationary loads in an unsupervised damage identification method based on macro-strain wavelet packet energy according to one embodiment of the present disclosure.

[0029] Figure 13 This is a schematic diagram of a two-dimensional finite element model of a cable-stayed bridge structure in an unsupervised damage identification method based on macro-strain wavelet packet energy according to one embodiment of this disclosure.

[0030] Figure 14 This is a schematic diagram of the damage monitoring results of 15% damage to 15 elements of the main beam under 10% noise in an unsupervised damage identification method based on macro-strain wavelet packet energy according to one embodiment of the present disclosure.

[0031] Figure 15 This is a schematic diagram of the damage monitoring results of main beam elements 15 and 41 with 15% damage and element 82 with 15% damage under 10% noise in an unsupervised damage identification method based on macro-strain wavelet packet energy according to an embodiment of this disclosure.

[0032] Figure 16 This is a schematic structural block diagram of a structural damage identification device according to one embodiment of the present disclosure.

[0033] Figure 17 This is a schematic structural block diagram of an electronic device according to one embodiment of the present disclosure. Detailed Implementation

[0034] The present disclosure will now be described in further detail with reference to the accompanying drawings and examples. It should be understood that the specific examples described herein are for illustrative purposes only and are not intended to limit the scope of the disclosure. Furthermore, it should be noted that, for ease of description, only the parts relevant to the present disclosure are shown in the accompanying drawings.

[0035] It should be noted that, where there is no conflict, the embodiments and features described in this disclosure can be combined with each other. The technical solutions of this disclosure will now be described in detail with reference to the accompanying drawings and embodiments.

[0036] In the actual service life of large bridges or high-rise buildings, the structures are subjected to multiple complex and time-varying unknown loads such as vehicle movement, wind loads, and temperature cycles. Their dynamic response data exhibits significant non-stationarity and diverse operating conditions. Under these unknown multiple loads, the uncertainty of load type, location, and magnitude often leads to the masking or misjudgment of damage characteristics. When attempting to monitor damage using a limited number of sensor points, it is difficult to reliably separate weak anomalies caused by localized physical damage (such as cracks or stiffness degradation) and accurately locate them on specific structural components. Furthermore, due to the inability to establish a mapping relationship between damage characteristics and specific sensor point locations, it is difficult to determine whether an anomaly at a particular sensor point originates from damage in its own region or is an effect transmitted from distant damage through structural coupling, especially when sensor coverage is incomplete.

[0037] To address this, this disclosure proposes an unsupervised damage identification method based on macro-strain wavelet packet energy. The macro-strain wavelet packet energy transfer ratio matrix is ​​derived, and a deep learning network based on a convolutional autoencoder is constructed accordingly. This network effectively learns the strain response transfer law between local and global measurement points on a structural object under unknown multi-load conditions. Using measured signals from some target measurement points as input, the response feature vectors of all measurement points under the lossless assumption are predicted. By calculating the spatial distribution of the rate of change between the predicted wavelet packet energy feature vector and the actual feature vector, the mode mismatch signal caused by local physical damage is accurately separated.

[0038] The unsupervised damage identification method based on macro-strain wavelet packet energy disclosed herein can be deployed in health monitoring systems for large-scale infrastructure. Sensors at key locations on the structure transmit raw macro-strain response signals in real-time to a high-performance server cluster in a remote data center via wired or wireless networks. The servers utilize powerful computing capabilities to run trained convolutional autoencoder models, performing batch processing and damage index calculations on massive amounts of data. For moving or critical components requiring rapid response, such as wind turbine blades or high-speed rail bogies, lightweight prediction models can be deployed on embedded edge computing terminals locally on the device. This unsupervised damage identification method based on macro-strain wavelet packet energy can be applied to scenarios requiring the perception of the overall structural state through a limited number of measurement points, such as post-flight health assessment of aerospace vehicles, long-term micro-damage monitoring of historical buildings, and indirect leak location in industrial pipeline networks.

[0039] Figure 1 A schematic diagram illustrating the overall flow of an unsupervised damage identification method based on macro-strain wavelet packet energy according to one embodiment of this disclosure is shown. Figure 1 The method M100 shown includes steps S110 to S140. This method can be executed by a server computing node.

[0040] In step S110, for the structural object to be monitored under unknown multi-load conditions, the measured macro-strain response signals collected at all measuring points on the structural object are obtained.

[0041] The aforementioned macro-strain response signals reflect the dynamic changes in macroscopic-scale strain over time in structural components under multiple unknown external loads, either as a whole or locally. Macro-strain response signals directly characterize the elongation or shortening of material fibers and are important physical quantities for assessing the structural stress state, stiffness changes, and damage development. Unlike micro-strain, macro-strain focuses on the measurable overall deformation behavior of engineering structures under actual service conditions.

[0042] By utilizing multiple sensors deployed at key locations on a structure, dynamic macro-strain response signals generated by the structure under various unknown excitations are simultaneously acquired, with each sensor corresponding to a measurement point. The sensors form a spatially distributed monitoring array, covering the main load-bearing components and potential weak points of the structure. The acquired set of measured macro-strain response signals includes energy redistribution caused by potential local damage, providing a complete data view for global analysis.

[0043] In step S120, for at least one target measuring point on the structural object, the target macro-strain response signal corresponding to the target measuring point is extracted from the measured macro-strain response signal, and the number of target measuring points is greater than or equal to the number of actual loads on the structural object.

[0044] From the globally perceived measured macro-strain response signal, the target macro-strain response signal corresponding to the representative target measurement point is extracted, the information boundary of the prediction network (i.e., deep learning network) is defined, and the adaptability to different practical engineering constraints is also reflected.

[0045] Preferably, a target measuring point is selected from all measuring points according to a predetermined measuring point selection scheme. The target macro-strain response signal corresponding to the target measuring point is located and extracted from the measured macro-strain response signal.

[0046] Optionally, the selection of measurement points can be based on the importance of structural mechanics, or on modal participation or historical damage sensitivity.

[0047] In step S130, wavelet packet decomposition is performed on the target macro-strain response signal and the measured macro-strain response signal respectively to determine the wavelet packet energy feature vector of the target macro-strain response signal and the wavelet packet energy feature vector of the measured macro-strain response signal.

[0048] By employing multi-resolution analysis techniques, non-stationary dynamic signals containing complex interference are decoupled into multiple independent frequency band components, and energy distribution patterns characterizing the intrinsic dynamic properties of the structure are extracted from them. This transforms lengthy time-domain macro-response signals into compact, robust, and diagnostically informative low-dimensional feature vectors.

[0049] Preferably, wavelet packet decomposition employs orthogonal wavelet basis functions to perform three-level wavelet packet decomposition on both the target macro-strain response signal and the measured macro-strain response signal. Specifically, by selecting orthogonal wavelet basis functions and performing three-level wavelet packet decomposition on both signals, the signal is sequentially decomposed into multiple non-overlapping sub-signals. The energy of each sub-signal's corresponding frequency band is calculated, and the energy of all frequency bands of the same target or measured macro-strain response signal is arranged in ascending order to form the corresponding wavelet packet energy feature vector of the target or measured macro-strain response signal. This transforms the non-stationary time-domain macro-response signal into a low-dimensional feature vector characterizing its frequency-domain energy distribution, effectively extracting key information reflecting the dynamic characteristics of the structural object.

[0050] In step S140, based on the macro-strain wavelet packet energy transfer ratio matrix, the wavelet packet energy feature vector of the target macro-strain response signal is input into the deep learning network to eliminate the influence of unknown multi-loads and predict the wavelet packet energy feature vectors of all measuring points, thereby obtaining the predicted wavelet packet energy feature vectors of each measuring point. The deep learning network is a deep learning network based on a convolutional autoencoder.

[0051] Preferably, the convolutional autoencoder includes an encoder and a decoder, with a skip connection between the encoder and the decoder. The encoder includes a convolutional layer and a pooling layer, and the decoder includes a deconvolutional layer or an upsampling layer plus a convolutional layer.

[0052] The wavelet packet energy feature vector of the target macro-strain response signal is input into a pre-trained deep learning network. The encoder of the deep learning network abstracts and compresses the wavelet packet energy feature vector of the target macro-strain response signal layer by layer, extracting a high-order representation, which is then passed to the decoder through a bottleneck layer. The decoder, combined with the shallow feature information introduced by skip connections, progressively upsamples and reconstructs the output to generate a feature vector set equal in number to all measurement points of the structural object under unknown multi-load conditions, i.e., the predicted wavelet packet energy feature vector.

[0053] In step S150, for each measuring point, the rate of change between the predicted wavelet packet energy eigenvector and the measured macro-strain response signal wavelet packet energy eigenvector is calculated to determine the damage change rate of each measuring point.

[0054] By comparing the difference between the predicted wavelet packet energy feature vector generated by the deep learning network under healthy conditions and the actual wavelet packet energy feature vector observed by the sensor, potential physical damage is transformed into a measurable numerical indicator, amplifying and locating weak signals that deviate from the normal behavior pattern of the structure caused by local structural deterioration.

[0055] In step S160, the damage location of the structural object is determined based on the amplitude distribution of the damage change rate at each measuring point.

[0056] After obtaining the damage change rate at all measuring points, it is treated as a scalar field defined on the structural space grid. The amplitude distribution characteristics of this scalar field are analyzed to identify extreme points or high-value clusters, from which measuring points are determined to locate the damage location of the structural object.

[0057] Preferably, based on the amplitude distribution of the damage change rate at all measuring points, the region corresponding to the measuring point with a damage change rate greater than or equal to the target threshold and the largest damage change rate is determined as the damage location of the structural object. By filtering out significantly abnormal measuring points using the target threshold, and further locating the region with the strongest response, the influence of weak interference and noise is effectively eliminated. Utilizing the physical characteristic that damage typically causes the greatest mode mismatch at the source point, accurate and reliable location of the damage location on the structural object is achieved.

[0058] The aforementioned target threshold is based on statistical analysis of long-term monitoring data of the structured object under known health conditions (such as the mean plus several times the standard deviation) or obtained through simulation and experimental calibration, and represents the upper limit of acceptable reconstruction error.

[0059] Therefore, the unsupervised damage identification method based on macro-strain wavelet packet energy disclosed in this paper utilizes wavelet packet decomposition to extract the time-frequency energy features of the macro-strain response signal, and combines this with the stable transmission law learned by a deep learning network under healthy conditions to effectively filter out interference from unknown multi-loads. By calculating the mode mismatch between the predicted wavelet packet energy feature vector and the measured macro-strain response signal's wavelet packet energy feature vector and performing spatial distribution analysis, the damage source caused by local stiffness degradation can be accurately located. This solves the problem of traditional methods' dependence on densely deployed sensors and precise physical models, and improves the accuracy, robustness, and engineering practicality of damage identification for large and complex structures under limited measurement point conditions.

[0060] Regarding step S140, based on the macro-strain wavelet packet energy transfer ratio matrix, the wavelet packet energy feature vector of the target macro-strain response signal is input into the deep learning network to predict the wavelet packet energy feature vectors of all measurement points, thereby obtaining the predicted wavelet packet energy feature vectors for each measurement point. In some embodiments of this disclosure, the training process of the deep learning network may include, for example... Figure 2Steps S1401 to S1404 are shown.

[0061] In step S1401, at least one load excitation is applied to the structural object in an undamaged state, and all measuring points are divided into a first measuring point set containing the target measuring point and a second measuring point set containing all measuring points.

[0062] The aforementioned load excitations include moving vehicles, pulse hammering, and / or wind vibration simulation.

[0063] After confirming that the structure is intact, one or more external load excitations that can induce the typical dynamic behavior of the structure are applied, and the measuring points are divided into a first measuring point set containing the target measuring point and a second measuring point set containing all measuring points.

[0064] In step S1402, the first macro-strain response signal of the first measurement point set and the second macro-strain response signal of the second measurement point set are obtained respectively.

[0065] For each measuring point in the first set of measuring points, the corresponding first macro-strain response signal is acquired. Simultaneously, for each measuring point in the second set of measuring points, the corresponding second macro-strain response signal is acquired.

[0066] In step S1403, wavelet packet decomposition is performed on the first macro-strain response signal and the second macro-strain response signal respectively to determine the wavelet packet energy feature vectors of the first macro-strain response signal and the second macro-strain response signal.

[0067] Step S1403 is the same as step S130, and will not be described in detail here.

[0068] In step S1404, a convolutional autoencoder framework is constructed based on the macro-strain wavelet packet energy transfer ratio matrix. The wavelet packet energy feature vector of the first macro-strain response signal is used as the input sample, and the wavelet packet energy feature vector of the second macro-strain response signal is used as the output sample to train the deep learning network to eliminate the influence of unknown multi-load.

[0069] The wavelet packet energy eigenvectors of the first and second macro-strain response signals are normalized, and the deep learning network is trained to learn the feature mapping relationship between macro-strain response signals acquired by different sensors, namely the macro-strain wavelet packet energy transfer ratio matrix.

[0070] Therefore, it not only effectively captures the complex dynamic coupling characteristics between various measuring points, but also enhances the robustness to unknown loads, laying a solid and reliable intelligent diagnostic foundation for the subsequent accurate location of full-structure damage using only a few key measuring point data.

[0071] Regarding step S140, the wavelet packet energy feature vector of the target macro-strain response signal is input into the deep learning network to predict the wavelet packet energy feature vectors of all measurement points, thereby obtaining the predicted wavelet packet energy feature vectors for each measurement point. In some embodiments of this disclosure, before inputting the wavelet packet energy feature vector of the target macro-strain response signal into the deep learning network, the following steps may be included: Figure 3 Steps S310 to S320 are shown.

[0072] In step S310, the wavelet packet energy eigenvector of the target macro-strain response signal is normalized.

[0073] To eliminate the deviation caused by the difference in amplitude scale between different physical quantities or different measurement points, the wavelet packet energy feature vector of the target macro-strain response signal is uniformly adjusted in numerical range. This ensures that the deep learning network can converge more stably and efficiently, and improves its robustness to changes in input data.

[0074] Preferably, normalization is performed using the following formula: in, This represents the normalized energy value of the i-th measurement point in the j-th frequency band. This represents the unnormalized energy value of the i-th measurement point in the j-th frequency band. This represents the maximum original energy across all frequency bands at the i-th measurement point. This leads to the wavelet packet energy eigenvector of the normalized target macro-strain response signal. .

[0075] In step S320, the wavelet packet energy feature vector of the normalized target macro-strain response signal is input into the deep learning network.

[0076] Therefore, by normalizing the wavelet packet energy feature vector of the target macro-strain response signal, the signal amplitude scale deviation caused by load intensity differences or sensor location is effectively eliminated. The distribution of the wavelet packet energy feature vector of the target macro-strain response signal is consistent with that during the model training phase, improving the stability and robustness of the deep learning network during online inference and avoiding prediction distortion caused by abnormally high amplitude inputs. By preserving the relative distribution pattern of energy in each frequency band and suppressing the influence of absolute values, it ensures that the model can focus on learning and identifying the healthy baseline patterns determined by the intrinsic dynamic characteristics of the structural object.

[0077] Regarding step S150, for each measuring point, the rate of change between the predicted wavelet packet energy eigenvector and the wavelet packet energy eigenvector of the measured macro-strain response signal is calculated to determine the damage change rate at each measuring point. In some embodiments of this disclosure, this may include, for example... Figure 4Steps S1501 to S1503 are shown.

[0078] In step S1501, for each measurement point on the structural object, based on the frequency band order formed after wavelet packet decomposition, the predicted value in the predicted wavelet packet energy feature vector is matched with the corresponding measured value in the wavelet packet energy feature vector of the measured macro-strain response signal.

[0079] For each measurement point, based on the predicted wavelet packet energy feature vector generated using the same wavelet packet decomposition parameters and the wavelet packet energy feature vector of the measured macro-strain response signal, the predicted values ​​in the predicted wavelet packet energy feature vector are paired one by one with the corresponding measured values ​​in the wavelet packet energy feature vector of the measured macro-strain response signal in a fixed order from low to high or from high to low frequency bands.

[0080] In step S1502, the difference between the predicted value and the measured value for each pair of positions in the same frequency band is calculated, and the sum of the squares of the differences in all frequency bands is obtained to obtain the error value.

[0081] For each pair of matched predicted and measured values ​​at each measurement point, calculate the difference between the two values ​​at the same frequency band location. Square the differences in all frequency bands and sum them up to obtain the total error value for that measurement point.

[0082] In step S1503, the error value is divided by the sum of all measured values ​​in the wavelet packet energy feature vector of the measured macro-strain response signal to determine the damage change rate at each measuring point.

[0083] Specifically, the rate of change of damage is: in, Indicates the first Damage change rate at each measuring point The eigenvector representing the predicted wavelet packet energy is the th The first measuring point The energy value (i.e., the predicted value) of each frequency band. The wavelet packet energy eigenvector representing the measured macro-strain response signal is the first... The first measuring point Energy values ​​(i.e., measured values) of each frequency band. Indicates the number of frequency bands. This indicates the number of layers in the wavelet packet decomposition.

[0084] Therefore, by precisely matching the predicted wavelet packet energy eigenvectors with the measured macro-strain response signal's wavelet packet energy eigenvectors in both the measurement point and frequency band dimensions, and calculating the error value, the deviation of the structural response mode from the healthy baseline is effectively quantified, and the energy distribution anomalies caused by local damage are sensitively captured. Using the square of the sum of the total energies of the measured macro-strain response signal's wavelet packet energy eigenvectors as a normalization factor eliminates the influence of unknown multi-load intensity variations on error calculation.

[0085] Regarding step S150, for each measurement point, the rate of change between the predicted wavelet packet energy eigenvector and the wavelet packet energy eigenvector of the measured macro-strain response signal is calculated to determine the damage change rate for each measurement point. In some embodiments of this disclosure, before determining the damage change rate, the following may be included: Figure 5 Steps S510 to S520 are shown.

[0086] In step S510, the predicted wavelet packet energy eigenvector is inversely normalized.

[0087] The maximum energy value of each target measurement point is recorded when the wavelet packet energy feature vector of the target macro-strain response signal is normalized. The predicted value in the predicted wavelet packet energy feature vector output by the deep learning network is multiplied by the maximum energy value of the corresponding measurement point. The normalized predicted energy values ​​in all frequency bands are restored to the same order of magnitude and amplitude scale as the wavelet packet energy feature vector of the measured macro-strain response signal, and the inverse normalized predicted wavelet packet energy feature vector is obtained.

[0088] In step S520, the damage change rate is calculated by comparing the predicted wavelet packet energy eigenvector after inverse normalization with the wavelet packet energy eigenvector of the measured macro-strain response signal.

[0089] Therefore, by inverse-normalizing the predicted wavelet packet energy eigenvector, it is restored to the original energy scale consistent with the wavelet packet energy eigenvector of the measured macro-strain response signal. This ensures that both have the same physical dimensions and amplitude benchmark when calculating the damage change rate through point-by-point comparison. This eliminates the computational bias that may be introduced by the normalization / inverse-normalization process, allowing the difference between the predicted and measured values ​​to truly and accurately reflect the degree of mode mismatch in the structural response. It not only retains the high sensitivity to energy redistribution caused by local damage but also guarantees the physical consistency and numerical reliability of the error measurement, thereby improving the accuracy and reliability of the damage identification results.

[0090] The technical solution of this disclosure will be further explained below with a specific implementation example.

[0091] In the time domain, if the impulse response function generated by the impulse load is Then from the measuring point Load at the location The system caused at the measurement point Response at the location It can be expressed as the integral form of a series of unit impulse response functions: (1) Equation (1) can be written in matrix form as shown in equation (2): (2) in The total number of sampling points. For time sampling points, Given the sampling interval, equation (2) can be written in the following form: (3) Formula (3) expresses the relationship between a single load action and a single response, i.e., the measurement point Load time history at the location and measuring points Response time at the location The relationship between structural objects. When multiple loads act simultaneously, and measurements are taken. The structural response at each location, and the relationship between load and structural response, can be further expressed as: (4) It can be further written as follows: (5) in This is the transfer matrix between load and response, which is related to the location of the load application and the location of the measuring point, but is independent of the load time history.

[0092] Because strain is more sensitive to structural damage, and the measurement using a long gauge length strain sensor (macro-strain) can compensate for the limitations of point strain sensors in identifying localized damage far from the measurement point, this disclosure uses the macro-strain response of the observation unit on the upper surface of the structural object for damage identification, thereby improving the accuracy of damage identification. Macro-strain is the average strain of the structural object over a specified length. The relationship between the macro-strain response of a single element and the angular displacements at its two left and right nodes is as follows: (6) in, For the first Each unit in The macro-strain response measured at any given time; For the first The length of each unit (i.e., the gauge length of the sensor); For the first The average distance from each unit sensor to the neutral axis; and They represent the first The angular displacement of the left and right nodes of the unit. Similar to formula (1), based on the impulse response function, the first... angular displacement at the left and right nodes of each unit and This can be expressed using the Duhamel integral as follows: , (7) in, and They represent the measurement points respectively. Apply a unit pulse at the location, Time of the first The pulse rotation response function at the left and right nodes of the unit. Substituting into formula (6), the first... The macro-strain response of a single element can be rewritten as: (8) Similar to formula (5), we can obtain: (12) Furthermore, observations under multiple loads will be conducted. The structural responses at each measuring point are denoted as the first set of responses. Then, the structural responses at some measurement points are extracted from the first set of responses to form the second set of response vectors. Then the macro-strain responses of the first and second groups can be expressed as: ; (9) in and These are the transfer matrices corresponding to the first group of responses and the second group of macro-strain responses, respectively. Then, according to the left side of (9), the following equation can be obtained: (10) in represent The pseudo-inverse, the right-hand side of equation (9), can be further expressed as: (11) in Therefore, when observing the macro-strain response of the structure, The transfer ratio matrix between the two sets of macro-strain responses is only related to the structure. Equation (11) represents the relationship between the two sets of long gauge length strain responses of the undamaged structure. Furthermore, when the linear structure is damaged, the structural parameters will change. At this time, the relationship between the two sets of long gauge length strain responses of the damaged structure can be expressed as: (12) in Let be the transfer ratio matrix between the two sets of macrostrain responses of the structure under damaged conditions. Based on formulas (11) and (12), the transfer ratio matrix between the first set of macrostrain responses of the damaged structure can be obtained. and the macro-strain response transfer ratio matrix based on lossless structures Predict the second set of macro-strain responses Therefore, this disclosure defines the following relation: (13) at this time, and Differences exist, and these differences can be used to identify structural damage. When a structural object is damaged, its dynamic response changes accordingly, but minor localized damage is difficult to identify directly from the response time history. Therefore, it is necessary to extract more damage-sensitive characteristic indicators from the structural dynamic response for damage identification.

[0093] Furthermore, wavelet packet analysis possesses strong time-frequency domain analysis capabilities and is highly sensitive to abrupt changes in vibration response signals. Wavelet packet energy is an effective tool for analyzing the local characteristics of signals; through multi-scale decomposition, it can reflect the energy distribution of a signal at different frequencies and time scales. Wavelet packet decomposition transforms the original signal into multiple wavelet packet coefficients, thereby revealing its energy characteristics within a specific frequency band and time window. Therefore, the... Macro-strain response at individual measuring points go through Layer wavelet packet decomposition can be expressed as: (14) wavelet packet frequency band signal It can be represented as a linear combination of wavelet packet basis functions: (15) Indicates the first The expansion coefficients of the k-th basis function in the j-th frequency band at each measurement point. Let k represent the basis function of the j-th frequency band.

[0094] Furthermore, the energy of each frequency band of the wavelet packet can be expressed as: (16) This represents the total number of sampling points contained in the wavelet packet frequency band signal within the j-th frequency band.

[0095] Wavelet packet frequency band energy This reveals the energy distribution of the signal across different frequency ranges. In network training, normalization can unify feature dimensions, avoid gradient instability, and improve model convergence speed and training stability. Simultaneously, it helps enhance sensitivity to minute damage features, improving recognition accuracy and robustness. Therefore, the wavelet packet energy features of the macro-strain response at each measurement point are normalized using the following formula: (17) in, This represents the normalized energy value of the i-th measurement point in the j-th frequency band. This represents the unnormalized energy value of the i-th measurement point in the j-th frequency band. This represents the maximum value of the original energy across all frequency bands at the i-th measurement point. This leads to the normalized wavelet packet energy vector (e.g., the wavelet packet energy eigenvector of the target macro-strain response signal). Therefore, by The normalized wavelet packet energy characteristics corresponding to the second set of macro-strain responses formed by the measurement points can be expressed as follows: ,from Extracting the normalized wavelet packet energy features corresponding to a portion of the sensors constitutes the first set of normalized wavelet packet energy features for the macro-strain response. .

[0096] Furthermore, referring to formula (11), a similar relationship is established between the normalized wavelet packet energies of the two sets of macro-strain responses for the undamaged and damaged structures: , (18) in, and These represent the macro-strain response wavelet packet energy transfer ratio matrices corresponding to the normalized wavelet packet energy of the macro-strain response of the undamaged and damaged structures, respectively. and The distribution represents the normalized wavelet packet energy corresponding to the second and first groups of macro-strain responses of the lossless structure; and Let represent the normalized wavelet packet energies corresponding to the second and first groups of macro-strain responses of the damaged structure, respectively. Based on equation (13), the following relationship is constructed: (19) in The predicted normalized wavelet packet energy of the second set of macro-strain responses is generated using an unsupervised network (i.e., a deep learning network) based on the response transfer ratio matrix of the undamaged structure in the wavelet packet energy domain, utilizing the wavelet packet energy of the first set of macro-strain responses of the damaged structure. After the network prediction is completed, the predicted normalized wavelet packet energy is... Inverse normalization is performed to obtain the wavelet packet energy. .

[0097] Therefore, based on the lossless structure, any number of loads can be applied at the load location, and the wavelet packet energy can be normalized using the two sets of macro-strain responses of the corresponding lossless structure. and Construct an unsupervised network to enable the network to autonomously learn two sets of macro-strain response wavelet packet energy transfer ratio matrices of lossless structure. The mapping relationship is shown in Equation (18), which eliminates the need for complex inverse matrices and regularization techniques found in theoretical methods. After network training is complete, the model is saved, and the first set of normalized wavelet packet energy with a lossy structure is used during testing. The trained network is input, and the network predicts the energy of the second set of normalized wavelet packets. As shown in equation (19). Furthermore, inverse normalization yields... , Normalized wavelet packet energy corresponding to the second set of macro-strain responses measured in the lossy structure There will inevitably be differences, and these differences can be used to determine the changes in the structural state and to locate the damage.

[0098] Furthermore, structural damage causes changes in the wavelet packet energy of its response signal in a specific frequency band. Based on this feature, a wavelet packet energy change rate sum-of-squares index can be effectively used for damage identification and localization. Therefore, wavelet packet energy predicted by unsupervised networks... and measured wavelet packet energy Construct the first The wavelet packet energy change rate square index (DI) corresponding to the macro-strain response at each measuring point, i.e., the damage change rate: (20) Indicates the first The square index of wavelet packet energy change rate at each measurement point (i.e., damage change rate). This represents the second set of responses predicted by the network. The corresponding measurement point is the first The macro-strain wavelet packet energy of each frequency band, for example, the first wavelet packet energy eigenvector in the prediction of the wavelet packet energy eigenvector. The first measuring point Predicted values ​​for each frequency band. Represented as the second group of measured responses of the damaged structure The first sensor corresponding to the The wavelet packet energy of the macro-strain in each frequency band, for example, the wavelet packet energy eigenvector of the measured macro-strain response signal. The first measuring point Measured values ​​for each frequency band Indicates the number of frequency bands. This indicates the number of wavelet packet decomposition levels. It should be noted that the selection of the wavelet mother function and the number of decomposition levels has a significant impact on the feature extraction effect. However, given that this disclosure focuses on unsupervised damage identification under unknown multi-load conditions, the selection of relevant parameters is based on existing literature and will not be discussed further. The Daubechies (DB) family of wavelet packet functions are widely used in structural damage identification due to their good regularity and high-order vanishing moments. Therefore, the Db20 wavelet function is selected to perform three-level wavelet packet decomposition on the macro-strain response, meaning that each sensor measurement point can obtain a wavelet packet energy feature vector with a dimension of 1×8.

[0099] Furthermore, given the difficulty in obtaining structural damage label data in actual engineering projects, and based on the characteristics of unsupervised learning, the network is trained using only the response dataset under the undamaged state of the structure, without the need for damage labeling. Combining the derivation process of the above physical formulas, an unsupervised learning network based on the convolutional autoencoder (CAE) architecture was constructed according to formula (23), and... and Using these datasets as input and output data for a CAE network, respectively, feature mapping relationships are learned. This method imbues the network training process with explicit physical meaning, thereby effectively eliminating the interference of different load types before damage under unknown multi-load conditions on structural damage identification, while improving the interpretability and applicability of the model.

[0100] Traditional autoencoders consist of an encoder and a decoder. The encoder extracts high-level features through dimensionality reduction, while the decoder reconstructs the input data based on these features. Convolutional autoencoders introduce convolution and deconvolution operations into the autoencoder structure, effectively extracting local structural features and improving data reconstruction and feature mapping capabilities. The encoder progressively compresses spatial dimensions and extracts deep features through multiple convolutions, while the decoder progressively restores the output dimension through multiple deconvolutions, thus learning the mapping relationship between input and output.

[0101] To enhance the nonlinear expressive power of the network and avoid underfitting, this disclosure introduces multi-layer convolutional and deconvolutional structures in the encoder and decoder, controlling the model depth and output dimension through the kernel size and stride. Simultaneously, skip connections are employed to transfer shallow features from the encoder to deeper layers of the decoder, preventing feature loss, improving gradient propagation, and accelerating training convergence, thereby enhancing network stability and recognition performance.

[0102] like Figure 6 As shown, the constructed network has The normalized wavelet packet energy feature input corresponding to each sensor data and The network outputs normalized wavelet packet energy features corresponding to each sensor data point, with each sensor data point having energy across 8 frequency bands. To preserve the spatial characteristics of the sensors, the network's input and output dimensions are as follows: and When compressing input data using convolutional layers, Conv_1 to Conv_2 compresses the data height to [value missing]. The data length remains at 8. Useful low-level information is preserved in Conv_11 to Conv_33 and directly transmitted to the decoder layer via Add_1 to Add_3. Conv_3 is the bottleneck layer with an output dimension of... This allows the most effective feature information to be passed to the decoder. After the bottleneck layer, the deconvolutional layers recover the data features layer by layer and finally generate the target dimension. .

[0103] Therefore, during the network training phase, multiple random white noises are applied to the undamaged structure to divide the observed macro-strain responses of the structure in the undamaged state into the first group of macro-strain responses. Second set of macro-strain responses Then, the wavelet packet energy vector corresponding to each sensor in the group is used to construct features by estimating the wavelet packet energy. and And normalize to obtain the corresponding and The datasets are used as input and output datasets, respectively, to train the constructed CAE network and learn the mapping relationship. After the network training is completed, untrained load types are applied to the damaged structure and the corresponding macro-strain responses are observed. The network is also divided into two groups, and the wavelet packet energy feature matrix of each sensor is estimated. and , will the corresponding Predicting from the input of a pre-trained CAE network And perform inverse normalization to obtain the wavelet packet energy. .

[0104] like Figure 7 As shown, during the network training phase, multiple sets of random white noise excitations are applied under a lossless structural state, and multiple loads are considered to act simultaneously to obtain the corresponding macro-strain responses. The observed macro-strain responses are divided into two groups, denoted as the first group of macro-strain responses. (i.e., macro-strain response signal) and the second set of macro-strain responses (Second macro-strain response signal). Wavelet packet decomposition was performed on both sets of macro-strain responses, and the wavelet packet energy at each measurement point was estimated to construct the first wavelet packet energy feature matrix. (i.e., the wavelet packet energy eigenvector of the first macro-strain response signal) and the second wavelet packet energy eigenma matrix (i.e., the wavelet packet energy eigenvector of the second macro-strain response signal). Normalization is performed using equation (17), and then... and These are used as the input and output of the constructed CAE network, respectively, to train the network and learn the feature mapping relationship between the responses of different sensors. .

[0105] During the testing phase, arbitrary multi-load excitations are applied under possible damage states of the structure, and macro-strain responses are collected. The test responses are grouped in the same manner as in the training phase, and wavelet packet decomposition is performed on each group to obtain the first wavelet packet energy feature matrix for the testing phase. (i.e., the wavelet packet energy eigenvector of the target macro-strain response signal) and the second wavelet packet energy eigenma matrix during the testing phase. (i.e., the wavelet packet energy eigenvector of the measured macro-strain response signal). Normalization yields The trained CAE network is then input to predict the normalized wavelet packet energy features of the corresponding second set of macro-strain responses. (i.e., predicting the wavelet packet energy eigenvector), and then inversely normalizing it to obtain... .

[0106] Wavelet packet energy predicted by CAE network wavelet packet energy of the measured signal A wavelet packet energy change rate sum of squares index is constructed. When the wavelet packet energy change rate sum of squares at a certain location is significantly greater than the wavelet packet energy change rate sum of squares at other locations, it is determined that structural damage has occurred in the region corresponding to that location.

[0107] To illustrate the technical effectiveness of the unsupervised damage identification method based on macro-strain wavelet packet energy disclosed in this paper, the following verification case is provided.

[0108] like Figure 8 The simply supported beam shown was used to verify the effectiveness of the unsupervised damage identification method based on macro-strain wavelet packet energy disclosed herein. The simply supported beam consists of 10 elements (measuring locations) with a total of 11 nodes and a total length of 10m. The beam's cross-section has a height of 0.15m and a width of 0.06m. The material's mass density is 7800 kg / m³. 3 The elastic modulus is 20 GPa and the damping ratio is 2%. Two independent loads, F1 and F2, are applied simultaneously at nodes 4 and 10. The macro-strain of the upper surface of each element of the structure is observed using a long gauge length sensor with a sampling frequency of 100 Hz and a sampling duration of 10 s.

[0109] To construct the training dataset, two independent white noise excitations were applied under a structurally intact state, generating 2000 sets of macro-strain response data. Based on the macro-strain responses obtained from long-gauge sensors, the observation units were divided into two groups: the first group consisted of macro-strain data numbered [2, 4, 6, 9], and the second group consisted of macro-strain data from all observation units. To simulate measurement error, Gaussian white noise with an amplitude of 5% of the signal standard deviation was added to the strain data from each sensor. Subsequently, a three-level wavelet packet decomposition was performed on the two sets of strain responses using a dB20 wavelet as the mother function to extract the frequency band energy and normalize it, resulting in wavelet packet energy features with a dimension of 1×8.

[0110] During training, the first set of features was used as CAE input and the second set of features was used as output to perform unsupervised learning of the structural feature mapping relationship. A total of 2000 sets of samples were divided into training set, validation set and test set in a ratio of 7:2:1. The network was trained with Adam optimizer for 500 epochs, with an initial learning rate of 0.01, which decayed to 0.1 times the original value every 10 epochs, and a batch size of 32.

[0111] After network training, damage is simulated by reducing the stiffness of specific units. Two independent harmonic loads are then applied to the damaged structure to obtain the strain response, with 5% noise added. Finally, the extracted wavelet packet energy is input into the training network, and the predicted and measured features are compared to verify the effectiveness of the proposed technique under different load conditions than the training excitation.

[0112] As shown in Table 1 below, the normalized wavelet packet energy corresponding to the first group of macro-strain responses under each damage condition is input into the trained CAE model to obtain the predicted inverse normalized wavelet packet energy. Then, the wavelet packet energy obtained by observing the macro-strain response under actual damage conditions is combined with the wavelet packet energy. Estimate the corresponding damage location indicators to locate structural damage.

[0113] Table 1 Figure 9 and Figure 10 The damage localization results are shown for test conditions 1 and 2. Figure 9 As shown, in test condition 1, even a CAE network trained solely on wavelet packet energy under random white noise excitation on an undamaged structure can still accurately identify single damage locations under harmonic loading. Figure 10 As shown in test condition 2, this disclosure further demonstrates that it has a good ability to identify multiple damages of different degrees, and the DI amplitude can reflect the degree of damage, verifying the good adaptability and robustness of the method of this disclosure under load type changes and multiple damage identification conditions.

[0114] Furthermore, to verify the generalization ability of this disclosure under simultaneous application of different load types, harmonic loads were applied simultaneously to the damaged structure along with steady and non-steady loads. The aforementioned dual-damage condition was used as the test sample input to the trained CAE network, and the corresponding damage identification results are as follows: Figure 11 and Figure 12 As shown. It should be emphasized that neither harmonic loads nor uncertain loads appeared during the network training process, making it extremely challenging to identify structural damage under this combined load type.

[0115] Depend on Figure 11 and Figure 12 It is evident that, although the network was trained on data samples with no loss structure only under white noise excitation, it was still able to accurately identify damage at different locations of the structure under combined excitation during the testing phase, with small localization error and good sparsity in the identification results.

[0116] In addition, such as Figure 13 As shown, the method disclosed herein is verified using a two-dimensional finite element model of a cable-stayed bridge structure. The main girder of the cable-stayed bridge is 460m long (230m + 230m), and is discretized using 236 beam elements. The moment of inertia of the main girder is... Young's modulus and mass per unit length The values ​​are taken from the actual structural parameters. The bridge tower height is 150m, and 59 vertical elements are used for modeling. The Young's modulus is... .

[0117] One hundred element surfaces numbered [1–25, 35–84, 94–118] were selected as observation elements, and these elements were renumbered from 1 to 100 to obtain the corresponding macro-strain responses. The sampling frequency was set to 2000 Hz, and the sampling duration was 10 s. For the non-destructive structure, the long gauge length strain response was obtained by applying different random white noise excitations. To construct the input and output samples for the network, the observed responses were divided into two groups: the first group of macro-strain data came from observations of element numbers [10 20 30 40 50 60 70 80 90 100]; the second group of data came from the macro-strain of the 100 element surfaces of the observed elements. To simulate sensor noise and environmental interference, zero-mean Gaussian white noise with a standard deviation of 5% of the signal amplitude was considered in the response signal to verify the adaptability of the proposed method to actual working conditions.

[0118] To evaluate the noise robustness of the proposed method, damage identification analysis was conducted by superimposing Gaussian white noise with a standard deviation of 10% on the observed macro-strain response under three independent harmonic excitations. The identification results are as follows: Figure 14As shown. Furthermore, under the same load conditions and noise level, a set of three-damage conditions was also verified (i.e., components 15 and 41 have 15% stiffness damage, and component 82 has 15% stiffness damage), and the corresponding identification results are as follows. Figure 15 As shown.

[0119] like Figure 14 and Figure 15 As shown, even under 10% noise conditions, both single and multiple damage locations were accurately identified, and the DI value of the undamaged area remained at a low level. These results demonstrate that the proposed method possesses high damage identification accuracy and good noise robustness, and can effectively mitigate the impact of unknown multi-load changes before and after damage.

[0120] Based on any of the above embodiments, this disclosure also provides a structural damage identification device.

[0121] Figure 16 This is a schematic block diagram of a structural damage identification device according to one embodiment of the present disclosure.

[0122] like Figure 16 As shown, the structural damage identification device includes: Data acquisition module 1602 acquires the measured macro-strain response signals collected at all measuring points on the structural object to be monitored under unknown multi-load conditions. Data extraction module 1604 extracts the target macro-strain response signal corresponding to the target measurement point from the measured macro-strain response signal for at least one target measurement point on the structural object, wherein the number of target measurement points is greater than or equal to the number of actual loads on the structural object. The feature extraction module 1606 performs wavelet packet decomposition on the target macro-strain response signal and the measured macro-strain response signal respectively to determine the wavelet packet energy feature vector of the target macro-strain response signal and the wavelet packet energy feature vector of the measured macro-strain response signal. The data prediction module 1608, based on the macro-strain wavelet packet energy transfer ratio matrix, inputs the wavelet packet energy feature vector of the target macro-strain response signal into the deep learning network. While eliminating the influence of unknown multi-loads, it predicts the feature vectors of all measurement points to obtain the predicted wavelet packet energy feature vector. The deep learning network is built based on a convolutional autoencoder. The damage calculation module 1610 calculates the rate of change between the predicted wavelet packet energy eigenvector and the measured macro-strain response signal wavelet packet energy eigenvector for each measurement point, and determines the damage change rate for each measurement point. The damage location module 1612 determines the damage location of the structural object based on the amplitude distribution of the damage change rate at each measuring point.

[0123] The aforementioned structural damage identification device can be in the form of computer software, and each module of the aforementioned structural damage identification device can be implemented through computer software modules.

[0124] The specific implementation process of the functions and roles of each module in the above-mentioned structural damage identification device can be found in the implementation process of the corresponding steps in the above method, and will not be repeated here.

[0125] This disclosure also provides an electronic device 1000. Figure 17 A schematic diagram of the hardware implementation using the processing system is shown.

[0126] The hardware structure of electronic device 1000 can be implemented using a bus architecture. The bus architecture can include any number of interconnect buses and bridges, depending on the specific application and overall design constraints of the hardware. Bus 1100 connects various circuits including one or more processors 1200, memory 1300, and / or hardware modules. Bus 1100 can also connect various other circuits 1400 such as peripherals, voltage regulators, power management circuits, external antennas, etc. Bus 1100 can be an Industry Standard Architecture (ISA) bus, a Peripheral Component Interconnect (PCI) bus, or an Extended Industry Standard Component (EISA) bus, etc. Buses can be categorized as address buses, data buses, control buses, etc. For ease of representation, only one connection line is used in this figure, but this does not indicate that there is only one bus or one type of bus.

[0127] For ease of explanation, certain steps of the above method are described in relation to modules. It should be understood that the corresponding module performing one or more steps of the above method may be one or more hardware modules specifically configured to perform the corresponding step, or implemented by a processor configured to perform the corresponding step, or stored in a computer-readable medium for implementation by a processor, or implemented by some combination thereof.

[0128] This disclosure also provides a readable storage medium storing a computer program that, when executed by a processor, is used to implement the methods described above. A "readable storage medium" can be any means capable of containing, storing, communicating, propagating, or transmitting a program for use by or in conjunction with an instruction execution system, apparatus, or device. More specific examples of a readable storage medium include: an electrical connection with one or more wires (electronic device), a portable computer disk drive (magnetic device), random access memory (RAM), read-only memory (ROM), erasable and programmable read-only memory (EPROM or flash memory), fiber optic devices, and portable read-only memory (CDROM), etc.

[0129] This disclosure also provides a computer program product, the methods of which can be implemented wholly or partially through software, hardware, firmware, or any combination thereof. When implemented in software, it can be implemented wholly or partially as a computer program product. The computer program product includes one or more computer programs or instructions. When the computer program or instructions are loaded and executed, all or part of the processes or functions of this disclosure are performed.

[0130] Computer programs or instructions can be stored in a readable storage medium or transferred from one readable storage medium to another. For example, the computer program or instructions can be transferred from one website, computer, server, or data center to another website, computer, server, or data center via wired or wireless means. The readable storage medium can be any available medium capable of access, or a data storage device such as a server or data center that integrates one or more available media. The available medium can be a magnetic medium, such as a floppy disk, hard disk, or magnetic tape; an optical medium, such as a digital video optical disc; or a semiconductor medium, such as a solid-state drive. The computer-readable storage medium can be a volatile or non-volatile storage medium, or it can include both volatile and non-volatile types of storage media.

[0131] Those skilled in the art will understand that embodiments of this disclosure can be provided as methods, systems, or computer program products. Therefore, this disclosure can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, this disclosure can take the form of a computer program product embodied on one or more computer-usable storage media (including, but not limited to, disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0132] This disclosure is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to this disclosure. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, generate instructions for implementing the flowchart illustrations. Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0133] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes.

[0134] These computer program instructions may also be loaded onto a computer or other programmable data processing equipment to cause a series of operational steps to be performed on the computer or other programmable equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.

[0135] In the description of this specification, the references to terms such as "one embodiment / mode," "some embodiments / modes," "example," "specific example," or "some examples," etc., refer to specific features, structures, or characteristics described in connection with that embodiment / mode or example, which are included in at least one embodiment / mode or example of this disclosure. In this specification, the illustrative expressions of the above terms do not necessarily refer to the same embodiment / mode or example. Moreover, the specific features, structures, or characteristics described may be combined in any suitable manner in one or more embodiments / modes or examples. Furthermore, without contradiction, those skilled in the art can combine and integrate the different embodiments / modes or examples described in this specification, as well as the features of different embodiments / modes or examples.

[0136] Furthermore, the terms "first" and "second" are used for descriptive purposes only and should not be construed as indicating or implying relative importance or implicitly specifying the number of technical features indicated. Thus, a feature defined as "first" or "second" may explicitly or implicitly include at least one of that feature. In the description of this disclosure, "a plurality of" means at least two, such as two, three, etc., unless otherwise explicitly specified.

[0137] Those skilled in the art should understand that the above embodiments are merely for illustrating the present disclosure and are not intended to limit the scope of the disclosure. Those skilled in the art can make other changes or modifications based on the above disclosure, and these changes or modifications still fall within the scope of the present disclosure.

Claims

1. An unsupervised damage identification method based on macro-strain wavelet packet energy, characterized in that, include: For a structural object to be monitored under unknown multi-load conditions, the measured macro-strain response signals collected at all measuring points on the structural object are obtained. For at least one target measuring point on the structural object, the target macro-strain response signal corresponding to the target measuring point is extracted from the measured macro-strain response signal, wherein the number of target measuring points is greater than or equal to the number of actual loads on the structural object. Wavelet packet decomposition is performed on the target macro-strain response signal and the measured macro-strain response signal respectively to determine the wavelet packet energy feature vector of the target macro-strain response signal and the wavelet packet energy feature vector of the measured macro-strain response signal. Based on the macro-strain wavelet packet energy transfer ratio matrix, the wavelet packet energy feature vector of the target macro-strain response signal is input into a deep learning network to eliminate the influence of unknown multi-loads and predict the wavelet packet energy feature vectors of all measuring points, thereby obtaining the predicted wavelet packet energy feature vectors of each measuring point. The deep learning network is a deep learning network based on a convolutional autoencoder, and the macro-strain wavelet packet energy transfer ratio matrix is ​​the feature mapping relationship between macro-strain response signals obtained by different sensors. For each measurement point, the rate of change between the predicted wavelet packet energy eigenvector and the measured macro-strain response signal's wavelet packet energy eigenvector is calculated to determine the damage change rate at each measurement point; and The location of damage to the structural object is determined based on the amplitude distribution of the damage change rate at each measuring point.

2. The unsupervised damage identification method based on macro-strain wavelet packet energy as described in claim 1, characterized in that, The training process of the deep learning network includes: At least one load excitation is applied to a structural object in an undamaged state, and all measuring points are divided into a first measuring point set containing the target measuring point and a second measuring point set containing all measuring points. The first macro-strain response signal of the first measurement point set and the second macro-strain response signal of the second measurement point set are acquired respectively. Wavelet packet decomposition is performed on the first macro-strain response signal and the second macro-strain response signal respectively to determine the wavelet packet energy feature vector of the first macro-strain response signal and the wavelet packet energy feature vector of the second macro-strain response signal. Based on the macro-strain wavelet packet energy transfer ratio matrix, a convolutional autoencoder framework is constructed. The wavelet packet energy feature vector of the first macro-strain response signal is used as the input sample, and the wavelet packet energy feature vector of the second macro-strain response signal is used as the output sample to train the deep learning network.

3. The unsupervised damage identification method based on macro-strain wavelet packet energy as described in claim 1, characterized in that, For each measurement point, the rate of change between the predicted wavelet packet energy eigenvector and the measured macro-strain response signal's wavelet packet energy eigenvector is calculated to determine the damage change rate at each measurement point, including: For each measurement point on the structure, based on the frequency band order formed after wavelet packet decomposition, the predicted value in the predicted wavelet packet energy feature vector is matched with the corresponding measured value in the wavelet packet energy feature vector of the measured macro-strain response signal. Calculate the difference between the predicted and measured values ​​for each pair of values ​​in the same frequency band, and sum the squares of the differences across all frequency bands to obtain the error value; The damage change rate at each measuring point is determined by dividing the error value by the sum of all measured values ​​in the wavelet packet energy feature vector of the measured macro-strain response signal.

4. The unsupervised damage identification method based on macro-strain wavelet packet energy as described in claim 1, characterized in that, The wavelet packet decomposition employs orthogonal wavelet basis functions to perform three-level wavelet packet decomposition on both the target macro-strain response signal and the measured macro-strain response signal.

5. The unsupervised damage identification method based on macro-strain wavelet packet energy as described in claim 1, characterized in that, The convolutional autoencoder includes an encoder and a decoder, with a skip connection between the encoder and the decoder. The encoder includes a convolutional layer and a pooling layer, and the decoder includes a deconvolutional layer or an upsampling layer plus a convolutional layer.

6. The unsupervised damage identification method based on macro-strain wavelet packet energy as described in claim 1, characterized in that, Based on the amplitude distribution of the damage change rate at each measuring point, the damage location of the structural object is determined, including: Based on the amplitude distribution of the damage change rate at each measuring point, the region corresponding to the measuring point with the damage change rate value greater than or equal to the target threshold and the largest damage change rate is determined as the damage location of the structural object.

7. The unsupervised damage identification method based on macro-strain wavelet packet energy as described in claim 1, characterized in that, Before inputting the wavelet packet energy feature vector of the target macro-strain response signal into the deep learning network, the following steps are included: The wavelet packet energy eigenvector of the target macro-strain response signal is normalized; The wavelet packet energy feature vector of the normalized target macro-strain response signal is input into the deep learning network.

8. The unsupervised damage identification method for macro-strain wavelet packet energy as described in claim 7, characterized in that, Before determining the rate of change of damage, the following should be included: The predicted wavelet packet energy feature vector is inversely normalized; The damage change rate is calculated by comparing the predicted wavelet packet energy eigenvector after inverse normalization with the wavelet packet energy eigenvector of the measured macro-strain response signal.

9. An electronic device, characterized in that, include: The memory stores execution instructions; as well as A processor that executes the execution instructions stored in the memory, causing the processor to perform the unsupervised damage identification method of macro-strain wavelet packet energy according to any one of claims 1 to 8.

10. A computer program product, comprising a computer program, characterized in that, When the computer program is executed by the processor, it implements the unsupervised damage identification method for macro-strain wavelet packet energy as described in any one of claims 1 to 8.

Images