Pressure stress / damage identification method and device based on transfer learning admittance characteristics

Abstract

The invention belongs to the technical field related to structure health monitoring, and discloses a pressure stress / damage identification method and equipment based on transfer learning admittance characteristics, and the method comprises the steps: (1) constructing a structure health monitoring system of a piezoelectric sensing system based on a piezoelectric admittance technology, acquiring original admittance data of the structure in different pressure stress states in a preset frequency band range; (2) enhancing the original admittance data based on a Python language; (3) dividing the partially enhanced admittance data into a plurality of sub-frequency bands, calculating an RMSD value of each sub-frequency band, and performing matrix multiplication on the RMSD values to obtain a similar matrix so as to remodel the similar matrix into a two-dimensional data form; (4) inputting the similar matrix into an RB-DANN network model in the form of two-dimensional data, so that the RB-DANN network model realizes data migration between two domains and realizes compressive stress / damage identification at the same time; wherein the RB-DANN network model is constructed based on a residual error principle and domain adaptation. According to the invention, the cost is reduced.

Description

Technical Field

[0001] The present invention belongs to the technical field related to structural health monitoring, and more specifically, relates to a method and device for compressive stress / damage identification based on transfer learning admittance features. Background Art

[0002] In recent years, with the continuous development of seismic isolation technology, base isolation has been widely applied in building and bridge structures; it dissipates seismic energy to reduce the seismic response transmitted to the superstructure to protect the corresponding structure. Rubber bearings have become the most commonly used and efficient key components in base isolation due to their excellent elasticity, energy dissipation capacity, durability, and economy. However, when subjected to a large axial pressure from the superstructure, the performance of rubber isolation bearings may be affected. First of all, the rubber material itself has a certain deformability. Under axial pressure, the rubber isolation bearing may undergo compressive deformation, resulting in an increase in the stiffness of the bearing, which in turn affects its original vibration isolation effect. With the continuous action of pressure, the material properties of the rubber may undergo fatigue damage, and even indentations, cracks and other phenomena may occur. In addition, if the axial pressure is too large, exceeding the design load-bearing capacity of the rubber isolation bearing, it may cause the rubber bearing to degenerate due to permanent deformation, stiffness reduction, material fatigue, and potential damage at the interface between rubber and steel. In severe cases, it may lead to the failure of the bearing, thus affecting the stability of the overall structure and even causing structural damage. In practical engineering, generally speaking, most structural damages are a gradual process, and structural damages generally cannot be identified by the naked eye. However, as the damage intensifies, the structural safety, serviceability, and durability will all be affected, thus affecting life and property safety. And rubber isolation bearings belong to the hidden works in construction engineering, and it is more difficult to monitor their damages. Monitoring the magnitude of the compressive stress borne by the rubber bearing is an important measure to reflect its structural performance. By monitoring the change of its compressive stress to judge whether the rubber bearing has reached its ultimate bearing capacity and is damaged, and providing effective early warning information, it provides important reference information for the later maintenance and damage early warning of the structure.

[0003] The identification of the compressive stress / damage of rubber bearings based on the piezoelectric admittance (the reciprocal of impedance) technology verifies the feasibility of the method by analyzing the characteristics and differences of the conductance spectrum curves of rubber bearings under different compressive stresses. The entire piezoelectric sensing and collection system consists of a main structure, piezoelectric materials (PZT), an impedance analyzer, and a signal processing computer. The piezoelectric patches that act as both sensors and actuators are pasted on the main structure, and a corresponding excitation voltage is applied through the impedance analyzer to make the piezoelectric patches vibrate, thereby exciting the main structure to vibrate, and the mechanical impedance of the two is coupled. Therefore, the change of the mechanical impedance of the main structure can be reflected by collecting the impedance spectrum curve of the piezoelectric patch within a certain frequency band range, and further the change of the internal performance of the structure, such as stiffness, mass, damping, etc. When analyzing the change characteristics of the conductance spectrum under different working conditions, the conductance spectrum in the undamaged state is generally selected as the reference standard for compressive stress identification, and compared with the conductance spectra under other working conditions to judge the change of the internal performance of the main structure, so as to achieve the purpose of monitoring the structure.

[0004] However, this method for identifying the compressive stress / damage of rubber bearings has certain limitations. It only analyzes the compressive stress and damage borne by its main structure from a qualitative perspective, which often relies on empirical knowledge and is affected by subjective factors. Since the admittance data volume is large, a large amount of time and labor costs are required for its collection. In recent years, with the continuous development of deep learning, relevant means can be provided to quantitatively analyze the admittance data and ensure a high accuracy rate. However, for the admittance data collected by different piezoelectric patches of the same structure, multiple network models need to be trained multiple times to achieve the purpose of stress / damage identification, which reduces the utilization rate of network resources and increases the economic and time costs. Summary of the Invention

[0005] In view of the above defects or improvement requirements of the prior art, the present invention provides a method and device for identifying compressive stress / damage based on transfer learning admittance features, aiming to solve the problem of high costs caused by the need for multiple network models to be trained multiple times for the admittance data collected by different piezoelectric patches of the same structure in the existing compressive stress / damage identification method.

[0006] To achieve the above object, according to one aspect of the present invention, a method for identifying compressive stress / damage based on transfer learning admittance features is provided. The method includes the following steps:

[0007] (1) Based on the piezoelectric admittance technology, construct a structural health monitoring system for the piezoelectric sensing system, and then collect the original admittance data of the structure under different compressive stress states within a predetermined frequency band range;

[0008] (2) Enhance the original admittance data based on the Python language;

[0009] (3) Divide the partially enhanced admittance data into multiple sub - frequency bands, calculate the RMSD value of each sub - frequency band, perform matrix multiplication on the RMSD values to obtain a similarity matrix, and thus reshape it into the form of two - dimensional data;

[0010] (4) Input the similarity matrix in the form of two - dimensional data into the RB - DANN network model, and then the RB - DANN network model realizes data migration between the two domains and simultaneously realizes compressive stress / damage identification; among them, the RB - DANN network model is constructed based on the residual principle and domain adaptation.

[0011] Furthermore, the original admittance data includes source - domain data and target - domain data; the predetermined frequency band is 40 Hz - 500 kHz.

[0012] Furthermore, step (2) includes the following sub - steps:

[0013] 2.1, Calculate the mean value of the admittance data under each working condition and construct the standard admittance data. The corresponding calculation formula is:

[0014]

[0015] In the formula, represents the j - th dimensional value of the i - th working - condition data of the original admittance data; n represents the number of admittance data under the i - th working condition, is the constructed standard admittance data;

[0016] 2.2, Calculate the relative difference and RMSD value between the existing admittance data and the standard admittance data under each working condition;

[0017] 2.3, Generate an array conforming to Gaussian random noise based on the original admittance data, multiply it by the relative difference and RMSD value, and finally add it to the original admittance data to obtain the enhanced admittance data.

[0018] Furthermore, the calculation formulas for the relative difference and RMSD value are:

[0019]

[0020]

[0021] In the formula, is the relative difference of the j - th dimension of the admittance data under the i - th working condition; RMSD i represents the mean square deviation under the j - th working condition.

[0022] Furthermore, the calculation formula for the enhanced admittance data is:

[0023] ε i ~N(0,σ); ε' i= ε i × RMSD i

[0024]

[0025] Wherein, ε i represents randomly generated noise; ε i ' is the noise after multiplying with RMSD i ; X aug represents the generated enhanced admittance data.

[0026] Furthermore, add labels from 0 to 11 to the processed source domain data, and do not add labels to the target domain data. Divide the data into a training set and a test set according to an 8:2 ratio.

[0027] Furthermore, the RB-DANN network model includes a feature extraction section, a label prediction section, and a domain adaptation section. The feature extraction section consists of four residual blocks, two pooling layers, and a flattening layer. The residual block consists of two paths. One path consists of two convolutional layers. When passing from convolutional layer 1 to convolutional layer 2, it passes through a batch normalization layer and an activation layer. When passing from convolutional layer 2 to the pooling layer / output layer, there is also a batch normalization layer and an activation layer. The other path consists of one convolutional layer and one activation layer.

[0028] Furthermore, the label prediction section includes two fully connected layers. The first fully connected layer performs a linear transformation on the one-dimensional data output by the feature extraction section, and then activates it with the Relu function. The second fully connected layer only performs a linear transformation and then outputs the compressive stress / damage recognition result. The domain adaptation section consists of a domain classifier and a gradient reversal layer. The domain classifier includes six fully connected layers. The first to fifth fully connected layers perform a linear transformation on the one-dimensional data output by the feature extraction section and then activate it with the Relu function. The sixth fully connected layer only performs a linear transformation and then outputs the domain adaptation result. Among them, the residual blocks are connected by skip connections.

[0029] The present invention also provides a compressive stress / damage recognition system based on transfer learning admittance features. The system includes a memory and a processor. The memory stores a computer program. When the processor executes the computer program, it executes the compressive stress / damage recognition method based on transfer learning admittance features as described above.

[0030] The present invention also provides a computer-readable storage medium. The computer-readable storage medium stores machine-executable instructions. When the machine-executable instructions are called and executed by the processor, the machine-executable instructions cause the processor to implement the compressive stress / damage recognition method based on transfer learning admittance features as described above.

[0031] Generally speaking, compared with the prior art through the above technical solutions conceived by the present invention, the method and device for compressive stress / damage identification based on transfer learning admittance features provided by the present invention mainly have the following beneficial effects:

[0032] 1. The present invention combines the residual idea and the domain adaptation idea to process the admittance data collected by piezoelectric patches at different parts of the same structure, and then realizes the compressive stress / damage identification of multiple parts through one training, improving the utilization rate of network resources and reducing the economic cost and time cost.

[0033] 2. The labeled source domain data and the unlabeled target domain data are used as inputs, while the compressive stress and the domain adaptation result between the two are used as outputs; through continuous training and testing among the three modules of the RB-DANN network, the compressive stress / damage prediction of the source domain and target domain data and the judgment of the domain adaptation result are carried out; the structural health monitoring method provided by the present invention has high economy and practicability, can accurately identify the structural performance of different parts of the same structure, and makes the structural health monitoring intelligent and efficient.

[0034] 3. The enhancement of the original admittance data based on the Python language can enhance multiple groups on the basis of collecting a set of original admittance data under different working conditions, which is of great significance for saving the labor cost generated during the collection of admittance data.

[0035] 4. On the basis of ensuring that the compressive stress borne by the structure is accurately identified, the method provided by the present invention realizes the compressive stress / damage prediction of multiple parts of the structure through one training, which can greatly improve the utilization rate of network resources and reduce the time and economic costs in the long-term structural health monitoring.

[0036] 5. The method proposed by the present invention has good effects on the compressive stress / damage identification of rubber bearings, can identify the magnitude of the compressive stress borne by the structure under multiple working conditions, and the residual block avoids the problems of gradient explosion and gradient disappearance in the network through skip connections. Description of the Drawings

[0037] Figure 1 is a flowchart of a method for compressive stress / damage identification based on transfer learning admittance features provided by the present invention;

[0038] Figure 2 is a schematic diagram of the principle of the residual block network structure;

[0039] Figure 3 is a schematic diagram of the overall network architecture of the residual-domain adaptation neural network;

[0040] Figure 4 is a schematic diagram of the training and learning of the residual-domain adaptation neural network;

[0041] Figure 5 It is a schematic diagram of the admittance collection system and result monitoring;

[0042] Figure 6 It is a schematic diagram for comparing the admittance data of the source domain data under the non-damaged condition and other different working conditions. Among them, (a), (b), (c), (d), (e), (f), (g), (h), (i), (j), (k) correspond to working conditions 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, and 12 respectively;

[0043] Figure 7 It is a schematic diagram for comparing the admittance data of the target domain data under the non-damaged condition and other different working conditions. Among them, (a), (b), (c), (d), (e), (f), (g), (h), (i), (j), (k) correspond to working conditions 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, and 12 respectively;

[0044] Figure 8 In (a) and (b) of it are schematic diagrams of the damage indicators of the source domain data and the target domain data under all working conditions respectively;

[0045] Figure 9 It is a schematic diagram of the training and test losses and accuracies of the label predictor. Among them, (a), (b), (c), (d) correspond to the training set label loss, the test set label loss, the training set label accuracy, and the test set label accuracy respectively;

[0046] Figure 10 It is a schematic diagram of the training and test losses and accuracies of the domain adaptation section. Among them, (a), (b), (c), (d) correspond to the training set domain loss, the test set domain loss, the training set accuracy, and the test set accuracy respectively. Specific implementation manners

[0047] In order to make the objectives, technical solutions, and advantages of the present invention clearer, the present invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain the present invention and are not used to limit the present invention. In addition, the technical features involved in the various embodiments of the present invention described below can be combined with each other as long as they do not conflict with each other.

[0048] Please refer to Figure 1 , the present invention provides a method for identifying compressive stress / damage based on transfer learning admittance features. The identification method can accurately and quantitatively identify the magnitudes of the compressive stress and damage in two domains with only one network training, greatly improving the utilization rate of network resources.

[0049] The recognition method mainly includes the following steps:

[0050] Step 1: Based on the piezoelectric admittance technology, construct a structural health monitoring system for the piezoelectric sensing system, and then collect the original admittance data of the structure under different compressive stress states within a predetermined frequency band range.

[0051] The original admittance data includes source domain data and target domain data. The predetermined frequency band is 40 Hz - 500 kHz. In one embodiment, based on the piezoelectric admittance technology, a structural health monitoring system for the piezoelectric sensing system is constructed. Under the action of a 1 V excitation voltage of an impedance meter, 1 set of original admittance signals of the structure within the frequency band range (40 Hz - 500 kHz) under different compressive stress states is collected (801 dimensions for each working condition, 1 set for source domain data and 1 set for target domain data).

[0052] Step 2: Enhance the original admittance data based on the Python language.

[0053] Step 2 specifically includes the following sub-steps:

[0054] 2.1 Calculate the mean of the admittance data for each working condition and construct the standard admittance data. The corresponding calculation formula is:

[0055]

[0056] In the formula, represents the j-th dimensional value of the i-th working condition data of the original admittance data; n represents the number of admittance data under the i-th working condition, is the constructed standard admittance data; in the present invention, i = 1, 2,..., 12; j = 1, 2,..., 801; n = 801.

[0057] 2.2 Calculate the relative difference and RMSD value between the existing admittance data and the standard admittance data for each working condition. The corresponding calculation formula is:

[0058]

[0059] In the formula, is the relative difference of the j-th dimension of the admittance data under the i-th working condition; RMSD i represents the mean square deviation under the j-th working condition.

[0060] 2.3 Generate an array conforming to Gaussian random noise based on the original admittance data, multiply it by the relative difference and RMSD value, and finally add it to the original admittance data to obtain the enhanced admittance data. The corresponding calculation formula is:

[0061] ε i ~N(0,σ); ε' i = ε i×RMSD i

[0062]

[0063] where ε i represents randomly generated noise, and in this embodiment, it follows Gaussian noise with a mean of 0 and a standard deviation of 0.00003; ε i ′ is the noise after multiplication with RMSD i ; X aug represents the generated enhanced admittance data.

[0064] Step 3: Divide the partially enhanced admittance data into multiple sub-bands, calculate the RMSD value of each sub-band, perform matrix multiplication on the RMSD values to obtain a similarity matrix, and thus reshape it into the form of two-dimensional data.

[0065] Specifically, perform data preprocessing on 50 groups of enhanced admittance data, remove the unstable frequency bands at the front and back, select 784 data from them, and divide the selected data into 56 sub-bands, with 14 data in each sub-band. Calculate the RMSD values of 56 sub-bands of 50 groups of data. Obtain a similarity matrix by performing matrix multiplication on the RMSD values, and thus reshape the matrix into the form of 56×56, and use it as the input data of the RB-DANN network in the form of a two-dimensional array. Add labels from 0 to 11 to the processed source domain data, do not add labels to the target domain data, and divide the data into a training set and a test set according to the ratio of 8:2.

[0066] Step 4: Input the similarity matrix in the form of two-dimensional data into the RB-DANN network model, and then the RB-DANN network model realizes data migration between the two domains and simultaneously realizes compressive stress / damage identification; among them, the RB-DANN network model is constructed based on the residual principle and domain adaptation.

[0067] The RB-DANN network model includes a feature extraction section, a label prediction section, and a domain adaptation section. The feature extraction section consists of four residual blocks, two pooling layers, and one flattening layer. The residual block consists of two paths. One path consists of two convolutional layers. When passing from convolutional layer 1 to convolutional layer 2, it passes through a batch normalization layer and an activation layer. When passing from convolutional layer 2 to the pooling layer / output layer, there is also a batch normalization layer and an activation layer. The other path consists of one convolutional layer and one activation layer.

[0068] The label prediction section includes two fully connected layers. The first fully connected layer performs a linear transformation on the one-dimensional data output by the feature extraction section, and then activates it with the Relu function. The second fully connected layer only performs a linear transformation and then outputs the compressive stress / damage identification result.

[0069] The domain adaptation block consists of a domain classifier and a gradient reversal layer. The domain classifier includes six fully connected layers. The first to fifth fully connected layers perform linear transformation on the one-dimensional data output by the feature extraction block and then activate it with the Relu function. The sixth fully connected layer only performs linear transformation and then outputs the domain adaptation result.

[0070] Among them, the residual block avoids the problems of gradient explosion and gradient disappearance in the network through skip connections. Domain adaptation is achieved among the feature extraction block, the label prediction block, and the domain adaptation block, thereby achieving the purpose of transfer learning. The labeled source domain data and the unlabeled target domain data are input into the feature extraction block to extract domain-invariant features and compressive stress / damage features. The labeled source domain data is predicted in the label prediction block to achieve compressive stress / damage recognition. Based on the gradient reversal layer, the source domain data and the target domain data achieve transfer learning in the domain adaptation block. The network learning performance and generalization ability are judged by label loss, domain loss, label accuracy, and domain accuracy, and a RB-DANN network model that meets the requirements is obtained by adjusting hyperparameters.

[0071] During the training and learning process of the RB-DANN network model, for the network architecture and the data set partitioning method, the partitioned training set and test set are input into the RB-DANN network model. The labeled source domain data and the unlabeled target domain data first enter the feature extractor for training and testing to extract the domain-invariant features between the two domains. The labeled source domain data realizes label classification through the label predictor and outputs the label loss and accuracy. In the forward propagation process of the domain adaptation block, the domain classifier is used to distinguish the features of the two domains, and in the backward propagation process, the gradient reversal layer is used to confuse the features between the two domains, realizing domain adaptation and outputting the domain loss and accuracy.

[0072] According to the output results of the RB-DANN network training and learning, hyperparameters such as the learning rate, number of training epochs, batch size, and loss function in the network are modified, and the data is input into the RB-DANN network for training and testing again. When both the label accuracy and the domain adaptation accuracy of the RB-DANN network reach over 95%, the model ends the training and testing, and a suitable RB-DANN network model is obtained.

[0073] In one embodiment, first, the original admittance data (divided into conductance data and susceptance data) under different compressive stress conditions of different parts of the rubber bearing (source domain data and target domain data) is collected by using a piezoelectric sensing collection system. In this embodiment, in a specific example, the conductance data is processed and analyzed, and the collected original conductance data is enhanced. The method of generating Gaussian random noise is used to enhance the admittance data under 1 group of different working conditions to 50 groups. Then, the enhanced conductance data is preprocessed, that is, the conductance data under each working condition is divided into sub-bands, and the corresponding RMSD value is calculated. The admittance data is converted into a two-dimensional matrix through matrix multiplication operation as the input data of the residual-domain adaptation neural network; a residual-domain adaptation neural network model is constructed to achieve feature learning of multiple parts of the structure in one training and automatically judge the compressive stress magnitude of the structure, and then damage identification is carried out. In this embodiment, the feasibility and accuracy are verified by conducting a limit compression test on the rubber bearing. The results show that this embodiment can accurately and quantitatively identify the compressive stress magnitude of the rubber bearing and has good practical applicability.

[0074] Among them, Table 1 shows the overall architecture parameter settings of the constructed residual-domain adaptation neural network (RB-DANN); Table 2 shows the specific parameter settings of the residual blocks in the feature extractor. See Table 1 and Table 2 for details.

[0075] Table 1 Residual-domain adaptation neural network architecture parameter settings

[0076]

[0077]

[0078] Table 2 Specific network parameter settings of the residual block

[0079]

[0080] Figure 2 It is a schematic diagram of the principle of the residual block network, which introduces the feature extraction process of the data set based on the residual block skip connection in the present invention. Figure 3 It represents a schematic diagram of the principle of the overall network architecture of the residual-domain adaptation neural network, which introduces the positions of the three blocks in the network architecture of the present invention. Figure 4 It is the overall flowchart of the training and learning of the residual-domain adaptation network to extract features and output results. The present invention adjusts the model network architecture according to this process to obtain a suitable network model. Figure 5It is a admittance data collection system for rubber bearings under compressive stress. First, two piezoelectric wafers are bonded to the rubber bearing, and it is placed on a microcomputer-controlled electro-hydraulic servo universal testing machine to gradually apply axial pressure to the rubber bearing. At the same time, an Agilent 4294A commercial impedance analyzer is used to measure the admittance signal on the piezoelectric wafer. During the test, the rubber bearing shows obvious external morphological changes under the action of axial pressure. As the axial pressure increases, the height of the bearing gradually decreases, showing obvious compressive deformation. At the same time, the lateral dimension of the bearing expands. Especially under the action of large pressure, the peripheral part of the bearing becomes wider. As the pressure further increases, slight stress concentration areas appear on the surface of the bearing, and the local deformation is more significant. Tiny depressions or protrusions may appear in individual areas, indicating that the bearing material is subjected to large stress concentration. Due to the sensitivity of the piezoelectric admittance technology to local changes in the structure, its conductance spectrum curve can reflect the changes in the structural performance.

[0081] According to the specification of "JT / T 4 - 2019 - Highway Bridge Plate Rubber Bearings", the ultimate compressive bearing capacity of rubber is 70 MPa. Therefore, for the ultimate compressive test of rubber bearings, 12 working conditions are set, which are: 0 MPa, 15.9 MPa, 31.8 MPa, 47.7 MPa, 63.6 MPa, 79.5 MPa, 95.4 MP, 111.3 MPa, 127.2 MPa, 143.1 MPa, 159 MP, 174.9 MPa; to analyze the changes in the structural performance of rubber bearings under different compressive stresses and before and after damage. The piezoelectric wafers are respectively pasted on both sides of the rubber bearing and placed on a microcomputer-controlled electro-hydraulic servo universal testing machine to be gradually loaded according to the above-designed working conditions to obtain the original admittance data under different working conditions, and the admittance data in the frequency band of 40 Hz - 500 kHz is selected for the admittance spectrum. There are 801 admittance signal points in this frequency band.

[0082] Figure 6 and Figure 7Schematic diagram of the comparison of conductance spectra of admittance data under the non-destructive condition (compressive stress is 0 MPa) of the source domain data and the target domain data, respectively, and under other different working conditions. For the source domain data: (1) The change in the conductance spectrum is small under working conditions 2 - 3, and the change in the structural performance of the rubber bearing cannot be seen from it; (2) Under working conditions 4 - 12, the conductance spectrum curve shows an obvious upward right shift compared with the conductance spectrum in the non-destructive state, which indicates that the structural performance of the rubber bearing shows a change of increased stiffness and decreased damping under the action of axial pressure, which indicates that the energy dissipation capacity of the rubber bearing will be affected under the action of compressive stress. For the target domain data: (1) The change in its conductance spectrum characteristics is still relatively small under working conditions 2 - 3, and the change in the structural performance of the rubber bearing cannot be seen from it either; (2) Under working conditions 4 - 12, its conductance spectrum curve also shows an obvious upward right shift, and the reason is the same as the analysis result of the source domain conductance spectrum curve. It can be seen from this that the conductance spectrum curve can qualitatively reflect the change in structural performance.

[0083] Figure 8 It shows the curve relationship constructed by using the RMSD and MAPD damage indicators to analyze the conductance data measured under two piezoelectric wafers under different working conditions. It can be found that the damage indicator and the corresponding working condition are linearly related before 79.5 Mpa in the curve in the figure, while after 79.5 Mpa, the relationship between the damage indicator and the corresponding working condition tends to be stable or undergoes a mutation, and it can be seen that the structural performance of the rubber bearing changes before and after 79.5 Mpa, but it is still an analysis of the structural performance at the qualitative level.

[0084] Figure 9 and Figure 10 Then, a group of original admittance data of the rubber bearing collected by two piezoelectric wafers under different working conditions is enhanced based on Gaussian random noise to obtain 50 groups of enhanced data, and data preprocessing is performed on the sub-bands under each working condition to obtain 600 groups of 56×56 two-dimensional data as the input data of the RB-DANN network. The input data set is trained and learned through a feature extractor, a label predictor, and a domain adaptor to obtain the corresponding damage and accuracy rate. The label prediction accuracy rates of its training and testing are both 100%; while the domain adaptation accuracy rates of training and testing can reach 97.29% and 100% respectively, indicating that the RB-DANN network model can better achieve label prediction and domain adaptation so as to achieve the purpose of transfer learning, and realize the identification of compressive stress / damage at different parts of the same structure through one training.

[0085] The present invention also provides a compressive stress / damage identification system based on transfer learning admittance features. The system includes a memory and a processor. The memory stores a computer program, and when the processor executes the computer program, it executes the compressive stress / damage identification method based on transfer learning admittance features as described above.

[0086] The present invention also provides a computer-readable storage medium storing machine-executable instructions, which, when called and executed by a processor, cause the processor to implement the above-described method for identifying compressive stress / damage based on transfer learning admittance features.

[0087] Those skilled in the art can easily understand that the above are only the preferred embodiments of the present invention, and are not intended to limit the present invention. Any modifications, equivalent replacements, and improvements made within the spirit and principles of the present invention should all be included within the protection scope of the present invention.

Claims

1. A compressive stress / damage identification method based on transfer learning admittance features, characterized in that: The method comprises the following steps: (1) A structural health monitoring system for a piezoelectric sensing system is constructed based on piezoelectric admittance technology, thereby collecting the original admittance data of the structure under different compressive stress states within a predetermined frequency band; (2) Enhance the original admittance data based on Python language; (3) Divide the partially enhanced admittance data into multiple sub-bands, calculate the RMSD value of each sub-band, perform matrix multiplication on the RMSD value to obtain a similarity matrix, and thus reshape it into a two-dimensional data form; (4) The similarity matrix is ​​input into the RB-DANN network model in the form of two-dimensional data, and then the RB-DANN network model realizes data migration between the two domains and simultaneously realizes compressive stress / damage identification; wherein, the RB-DANN network model is constructed based on the residual principle and domain adaptation.

2. The compressive stress / damage identification method based on transfer learning admittance feature according to claim 1, characterized in that: The original admittance data includes source domain data and target domain data; the predetermined frequency band is 40 Hz-500 kHz.

3. The compressive stress / damage identification method based on transfer learning admittance feature according to claim 1, characterized in that: Step (2) includes the following sub-steps: 2.1, calculate the mean value of the admittance data under each working condition and construct the standard admittance data. The corresponding calculation formula is: In the formula, represents the jth dimension value of the original admittance data under the i-th working condition; n represents the number of admittance data under the i-th working condition, is the constructed standard admittance data; 2.2, calculate the relative difference and RMSD value between the existing admittance data under each working condition and the standard admittance data under that working condition; 2.3, based on the original admittance data, an array conforming to Gaussian random noise is generated and multiplied with the relative difference and RMSD value, and finally added to the original admittance data to obtain the enhanced admittance data.

4. The compressive stress / damage identification method based on transfer learning admittance feature according to claim 3, characterized in that: The calculation formula corresponding to the relative difference and RMSD value is: In the formula, is the relative difference of the jth dimension of the admittance data under the i-th condition; RMSD i represents the mean square deviation under the jth operating condition.

5. The compressive stress / damage identification method based on transfer learning admittance feature according to claim 4, characterized in that: The calculation formula corresponding to the enhanced admittance data is: e i ～ N(0,σ); e' i =e i ×RMSD i In the formula, ε i represents randomly generated noise; ε i ′ is the RMSD i Noise after multiplication; X aug represents the generated enhanced admittance data.

6. The compressive stress / damage identification method based on transfer learning admittance feature according to claim 2, characterized in that: The processed source domain data are labeled with 0 to 11, and the target domain data are not labeled. The data are divided into training set and test set in a ratio of 8:

2.

7. The compressive stress / damage identification method based on transfer learning admittance feature according to claim 2, characterized in that: The RB-DANN network model includes a feature extraction section, a label prediction section and a domain adaptation section. The feature extraction section consists of four residual blocks, two pooling layers and a flattening layer. The residual block consists of two paths. One path consists of two convolutional layers. When convolutional layer 1 is passed to convolutional layer 2, it passes through a batch normalization layer and an activation layer. When convolutional layer 2 is passed to the pooling layer / output layer, there is also a batch normalization layer and an activation layer. The other path consists of a convolutional layer and an activation layer.

8. The compressive stress / damage identification method based on transfer learning admittance feature according to claim 7, characterized in that: The label prediction section includes two fully connected layers. The first fully connected layer makes a linear change on the one-dimensional data output by the feature extraction section, and then activates the Relu function. The second fully connected layer only makes a linear transformation, and then outputs the compressive stress / damage recognition result. The domain adaptation section consists of a domain classifier and a gradient reversal layer. The domain classifier includes six fully connected layers. The first to fifth fully connected layers make a linear change on the one-dimensional data output by the feature extraction section, and then activate the Relu function. The sixth fully connected layer only makes a linear transformation, and then outputs the domain adaptation result. Among them, the residual block is connected through jumps.

9. A compressive stress / damage identification system based on transfer learning admittance features, characterized in that: The system includes a memory and a processor, wherein the memory stores a computer program, and when the processor executes the computer program, the compressive stress / damage identification method based on transfer learning admittance features according to any one of claims 1 to 8 is executed.

10. A computer-readable storage medium, characterized in that: The computer-readable storage medium stores machine-executable instructions. When the machine-executable instructions are called and executed by the processor, the machine-executable instructions prompt the processor to implement the compressive stress / damage identification method based on transfer learning admittance features as described in any one of claims 1-8.

Images