A tensile membrane structure damage identification method based on transfer learning and residual network

Abstract

The application discloses a kind of tensioned membrane structure damage identification method based on transfer learning and residual network, compared with traditional structure dynamic detection modal parameter identification method, by utilizing the powerful feature learning ability of deep learning model, it can realize from mass data automatically extract deeper damage feature, and it is well applicable to tensioned membrane structure dynamic damage detection with dense self-vibration frequency, high-order modal shape complex.It largely gets rid of the dependence of traditional structure dynamic detection signal processing technology, artificial feature extraction and expert experience, and is one of ideal methods for realizing tensioned membrane structure intelligent dynamic detection and health monitoring.In addition, targeted improvement strategy is proposed, so that the method has the characteristics of high efficiency, convenience and low cost, and has broad application prospects in the field of comprehensive dynamic detection and health monitoring of cable membrane structure.

Description

Technical Field

[0001] This invention relates to the fields of damage identification and artificial intelligence technology for tensioned cable-membrane structures, specifically to a method for damage identification of tensioned membrane structures based on transfer learning and residual networks. Background Technology

[0002] Tensile membrane structures (also known as tensioned cable-stayed membrane structures) are spatial structural systems composed of a stable spatial hyperbolic tensile membrane surface, a supporting mast system, supporting cables, and edge cables. They are primarily formed by applying an initial pretension to the membrane material and supporting components (steel frames, steel columns, or steel cables) to create a stable spatial shape capable of withstanding external loads. Due to their strong plasticity and high flexibility, tensile membrane structures represent a perfect combination of "force" and "form" in architectural structures, making them extremely widely used in public buildings. However, in practical engineering applications, the main loads borne by tensile membrane structures are highly random, leading to a high sensitivity to defects. Initial defects and local component damage can significantly affect the structure's load-bearing capacity and dynamic characteristics. Therefore, the mechanical properties of tensile membrane structures are very sensitive to membrane surface damage. Damage to the membrane surface can cause the tensile membrane structure to lose its stress equilibrium state, and if further subjected to external loads, the tensile membrane structure is highly likely to experience instantaneous failure. Tensile membrane structures are often local landmark buildings; their destruction not only causes significant casualties and huge economic losses but also brings extremely negative social impacts. In order to ensure the quality of tensile membrane structure projects, extend the service life of the structure, avoid premature failure of tensile membrane structures due to membrane surface damage, and further reduce the risk of disasters and accidents, it is necessary to conduct regular structural health monitoring and maintenance on the membrane surface of tensile membrane structures.

[0003] Structural damage identification is currently a crucial research topic in the field of structural health monitoring. Generally, when an engineering structure suffers damage, it exhibits abnormalities in certain characteristics compared to a normal structure. These abnormalities manifest as changes in structural properties, such as dynamic and static characteristics, surface condition, shape, size, and other characteristic parameters. Structural damage identification primarily assesses the health of an engineering structure by measuring and identifying abnormal changes in these characteristic parameters. This helps determine the presence and type of damage, as well as its location and extent, the current condition of the structure, its functional use, and the trend of damage changes. To date, rapidly developing structural dynamic response measurement technology has provided robust and effective technical support for traditional dynamic testing methods in civil engineering structures. Most overall testing techniques are based on the analysis of structural vibration signals, and dynamic testing primarily relies on the structure's dynamic response to identify its current state. Therefore, considering only the static characteristics of a structure is insufficient to fully reflect its essential features. Comprehensive dynamic testing and health monitoring of cable-membrane structures holds significant research importance and broad application prospects.

[0004] Structural dynamic response measurement and detection essentially involves detecting changes in the physical parameters of a structure. These changes primarily reflect variations in structural physical properties (stiffness, mass, and damping) through modal parameters (natural frequencies, mode shapes, etc.). However, unlike traditional civil engineering structures, tensile membrane structures are spatial structural systems composed of a stable spatial hyperbolic tensile membrane surface, a supporting mast system, supporting cables, and edge cables. These structures are highly nonlinear systems integrating material and geometric nonlinearity, characterized by dense natural frequencies and complex high-order mode shapes, making traditional modal parameter identification methods difficult to apply. While using vibration signals for structural health monitoring of tensile membrane structures is theoretically feasible, challenges remain, such as difficulties in identifying vibration signal modal parameters and the inability to accurately characterize the damage features of this type of structure. With the development of artificial intelligence and machine vision technologies, data processing and image recognition methods are widely applied to damage detection in civil structures. Deep learning algorithms, as a typical method of artificial intelligence, possess exceptionally powerful feature learning capabilities, automatically extracting deeper features from massive amounts of data. Therefore, they largely eliminate reliance on traditional signal processing techniques, manual feature extraction, and expert experience. Therefore, it is imperative to use deep learning methods to identify damage and monitor the health of tensile membrane structures based on vibration signals. Utilizing the powerful multi-level feature learning and high-dimensional feature extraction capabilities of deep learning can overcome the current difficulties in identifying modal parameters of vibration signals from tensile membrane structures and accurately characterizing the damage features of the membrane structure. This makes it one of the ideal methods for realizing intelligent dynamic detection and health monitoring of tensile membrane structures.

[0005] Traditional deep learning neural network models suffer from gradient vanishing, gradient exploding, and network degradation as network depth increases, severely impacting feature extraction and learning capabilities and leading to a sharp decline in network performance. Residual networks, however, break this traditional model's construction pattern. Compared to the single-connection model of ordinary neural networks, residual neural networks utilize identity mappings to achieve skip connections. The residual modules of residual networks alleviate the performance drop problem caused by increasing depth in traditional neural networks, making it possible to build deeper network models. This allows for the full extraction of time-frequency information contained in vibration signals, effectively improving the accuracy of tensile membrane structure damage identification. In recent years, transfer learning, a learning method that transfers knowledge learned in a source domain to a target domain, has significantly improved the accuracy of target damage identification in small sample situations by using large datasets to pre-train model parameters. Furthermore, since the characteristics exhibited by different samples of tensile membrane structure vibration signals vary considerably at different times, attention mechanisms can be used to focus the model's attention on different locations of target information, improving feature acquisition capabilities and reducing the model's sensitivity to input data. Summary of the Invention

[0006] To address the aforementioned technical shortcomings, the purpose of this invention is to provide a damage identification method for tensile membrane structures based on transfer learning and residual networks. This method can solve the problems that traditional structural dynamics detection methods are difficult to apply to tensile membrane structures, which are highly nonlinear systems that integrate material nonlinearity and geometric nonlinearity, and cannot accurately characterize the damage features of tensile membrane structures.

[0007] To solve the above-mentioned technical problems, the present invention adopts the following technical solution:

[0008] This invention provides a method for damage identification of tensile membrane structures based on transfer learning and residual networks, comprising the following steps:

[0009] Step 1: Apply an external excitation signal to the membrane surface and collect the original vibration signals of the membrane surface under different damage conditions;

[0010] Step 2: Preprocess the membrane vibration signal data under different damage conditions, and use wavelet transform to convert the one-dimensional membrane vibration sequence data into two-dimensional image data to construct a membrane vibration signal image dataset under different damage conditions.

[0011] Step 3: Introduce the attention mechanism into the ResNet50 residual network architecture. Embed the attention mechanism in both the shallow and deep layers of the ResNet50 network to complete the construction of the improved ResNet50 residual network model.

[0012] Step 4: Train the improved ResNet50 residual network model on the constructed dataset of membrane vibration signal images under different damage conditions, and use the transfer learning method to transfer the parameter weights of the ImageNet pre-trained model to the improved ResNet50 residual network model for training, while retaining the optimal model parameter weights during model training.

[0013] Step 5: Based on the pre-trained optimal improved ResNet50 residual network model, classify and detect membrane vibration signals under different damage conditions, and output the membrane damage classification and detection results and attention feature map.

[0014] Preferably, in step 1, the original vibration signals of the membrane surface under different damage conditions are collected using an accelerometer. The specific steps are as follows:

[0015] Step 21: Divide the membrane surface to be tested into a grid to determine the measurement points, ensuring that the measurement points are relatively densely and evenly distributed on the surface of the membrane surface to be tested;

[0016] Step 22: In the dynamic characteristic test of different damaged membrane surfaces, an external excitation signal is applied to the tested damaged membrane surface to make it undergo forced vibration. The external excitation signal includes steady-state sinusoidal excitation, sinusoidal frequency sweep excitation, random excitation and hammer impact excitation.

[0017] Step 23: Place the accelerometer at the measurement point to measure the forced vibration signal of the membrane surface under different damage conditions, and export the .xlsx files sequentially after the measurement is completed.

[0018] Preferably, noise points in the .xlsx files of membrane vibration signals under different damage conditions are manually deleted sequentially, including noise signals collected by the accelerometer before applying an external excitation signal to the tested membrane surface, and noise signals collected by the accelerometer after the membrane surface has undergone forced vibration attenuation under the applied external excitation signal; the data length of the .xlsx files of one-dimensional vibration signals of different damaged membrane surfaces after manual preprocessing is the same.

[0019] Preferably, in step 2, the method of converting one-dimensional membrane vibration sequence data into two-dimensional image data using wavelet transform to construct a membrane vibration signal image dataset under different damage conditions is as follows: Based on the preprocessed one-dimensional vibration signal .xlsx file data of different damaged membranes, the collected damaged membrane vibration signals are processed using Morlet continuous wavelet transform to convert the one-dimensional vibration time-domain signal of the damaged membrane into two-dimensional damaged membrane image data including time-domain and frequency-domain information, and divided into training set, validation set and test set.

[0020] Preferably, by utilizing the adjustable time-frequency window of continuous wavelet transform, the width of the time window changes accordingly with the frequency when performing time-frequency analysis on the vibration signal of the damaged membrane surface.

[0021] When the vibration frequency of the damaged membrane is high, the time window width in the wavelet transform process will decrease accordingly, and the frequency resolution will increase accordingly.

[0022] Meanwhile, when the vibration frequency of the damaged membrane is low, the time window width in the wavelet transform process will increase accordingly, and the corresponding frequency resolution will decrease.

[0023] Based on the adaptive nature of wavelet transform for signal analysis, and considering that wavelet transform essentially matches the target signal with a scaled and translated copy of the mother wavelet, by comparing the variation characteristics of the vibration signal of the damaged membrane and the similarity between different mother wavelets and their variation characteristics, the Morlet mother wavelet is selected to perform continuous wavelet transform on the vibration signal of the damaged membrane. This achieves the best time-frequency analysis effect on the vibration of the damaged membrane, thereby reflecting the rapid change characteristics of information in various frequency bands under forced vibration conditions of different damaged membranes in a short period of time.

[0024] Preferably, in step 3, attention mechanisms are embedded into both the shallow and deep layers of the ResNet50 network to complete the construction of the improved ResNet50 residual network model. The specific method is as follows:

[0025] The ResNet50 residual network model is divided into 5 stages. The input data passes through the 5 stages of ResNet50, namely Stage 1, Stage 2, Stage 3, Stage 4, and Stage 5, and then the corresponding output results are obtained.

[0026] Stage 1 has a simple structure and can be regarded as a preprocessing of the input. The other stages are all composed of Bottleneck and have similar structures.

[0027] Stage 2 contains 3 Bottlenecks, while the remaining Stages 3, 4, and 5 contain 4, 6, and 3 Bottlenecks respectively.

[0028] The improved ResNet50 residual network model is constructed by mixing channel attention and spatial attention mechanisms in Stages 2-5. Specifically, the attention mechanism is mixed between the shallow network Stage 2 and Stage 3, between the shallow network Stage 3 and the deep network Stage 4, and between the deep network Stage 4 and Stage 5.

[0029] Preferably, channel attention and spatial attention mechanisms are hybridized in Stages 2-5 to construct an improved ResNet50 residual network model, specifically as follows:

[0030] The hybrid embedding attention mechanism consists of two attention modules: the channel attention module (CAM) and the spatial attention module (SAM). The channel attention module (CAM) is responsible for allocating attention resources to each convolutional channel, while the spatial attention module (SAM) transfers spatial information from the original image to a specific space and extracts key information, enabling the convolutional neural network to more accurately focus on the regions that play a decisive role in the classification of two-dimensional images of damaged membrane surfaces. By hybrid embedding the attention mechanism into both shallow and deep networks, the ResNet50 residual network model is improved to enhance its ability to extract time-frequency features of different damaged membrane surfaces.

[0031] Preferably, the hybrid embedding attention mechanism consists of two attention modules: a channel attention module and a spatial attention module. The specific method is as follows:

[0032] Step 81: The Channel Attention Mechanism (CAM) first uses average pooling (AvgPool) and max pooling (MaxPool) to aggregate spatial information, and aggregates the initial feature dimensions H×W×C into two 1×1×C vectors. These two vectors represent the channel average pooling feature and the channel max pooling feature, respectively.

[0033] Step 82: Pass the two channel feature vectors above through a fully connected layer network with shared weights, and merge the output vectors using the element-wise summation method to generate channel-dimensional attention weights;

[0034] Step 83: Compared to the channel attention mechanism, the spatial attention mechanism SAM focuses on positional information and complements the channel attention mechanism in the H×W dimension; similarly, max pooling and average pooling are used on the channel to generate two H×W×1 feature vectors, which represent the spatial average pooling feature and the spatial max pooling feature, respectively.

[0035] Step 84: Pass the two spatial feature vectors above through a fully connected layer network with shared weights, and merge the output vectors using the element-wise summation method to generate spatial dimension attention weights;

[0036] Step 85: Generate a single-channel spatial attention map by passing the stacked channel dimension and spatial dimension feature maps through a convolutional layer, and assign attention weights to the locations that need to be suppressed or enhanced in the two-dimensional images of different damaged membrane surfaces.

[0037] Preferably, in step 4, the specific method for transferring the parameter weights of the ImageNet pre-trained model to the improved ResNet50 residual network model for training using transfer learning is as follows:

[0038] An improved ResNet50 residual network model was pre-trained using the ImageNet public dataset. The ImageNet pre-trained model was then used as a feature extraction network to train the improved ResNet50 residual network model on a two-dimensional image dataset with different damaged membrane surfaces, in order to extract the time-frequency features of the two-dimensional images with different damaged membrane surfaces.

[0039] By applying transfer learning, the amount of training data on two-dimensional images of vibration signals of the damaged membrane surface and the computational resources of the experimental platform required to improve the ResNet50 residual network model are reduced, so that the improved ResNet50 residual network model does not need to be retrained and learned from scratch when training on two-dimensional images of the damaged membrane surface.

[0040] Preferably, in step 5, a pre-trained model is obtained by training the two-dimensional image of the damaged membrane surface based on the improved ResNet50 residual network model to classify and detect the membrane surface vibration signals under different damage conditions.

[0041] Based on the classification results, six evaluation parameters are calculated: accuracy on the training set and test set, loss on the training set and test set, number of correct classifications (true positives TP), number of false normal classifications (false negatives FN), number of correct normal classifications (true negatives TN), and number of false classifications (false positives FP).

[0042] By combining the attention mechanism of the improved ResNet50 residual network model with hybrid embedding, the final model's attention weight allocation on the test set is visualized, and attention feature maps of two-dimensional images of different damaged membrane surfaces are output.

[0043] The beneficial effects of this invention are as follows:

[0044] (1) This invention combines the characteristics of the vibration signal of the membrane surface of the tensile membrane structure and selects Morlet continuous wavelet transform to convert the collected one-dimensional vibration sequence signal of the damaged membrane surface into two-dimensional image data of the damaged membrane surface. For the first time, it innovatively applies the deep learning convolutional neural network image processing model to the field of vibration damage identification and health detection of tensile membrane structures.

[0045] (2) The attention mechanism is applied to the ResNet50 residual network architecture. The attention mechanism is embedded in both the shallow and deep ResNet50 networks respectively to complete the construction of an improved ResNet50 residual network model and enhance the residual network model's ability to identify membrane vibration signals under different damage conditions.

[0046] (3) Applying the transfer learning method to the training of the improved ResNet50 residual network model can not only reduce the amount of training data of two-dimensional images of vibration signals of damaged membrane surface and the computing resources of the experimental platform required to build the improved ResNet50 residual network model, thus reducing the training cost of damage identification of tensile membrane structures, but also enable the improved ResNet50 residual network model to be trained on two-dimensional images of damaged membrane surface without having to retrain and learn from scratch, thereby speeding up the network training speed. It has the characteristics of simplified and fast training process, good scalability and high damage identification accuracy.

[0047] (4) The method proposed in this invention combines time-frequency analysis, image recognition, attention mechanism, residual network and transfer learning method, which has the characteristics of high efficiency, convenience and low cost. It largely gets rid of the dependence on traditional structural dynamic detection signal processing technology, manual feature extraction and expert experience. It is an ideal method to realize intelligent dynamic detection and health monitoring of tensile membrane structures. It has broad application prospects in the field of comprehensive dynamic detection and health monitoring of cable membrane structures and can be widely used in engineering practice. Attached Figure Description

[0048] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0049] Figure 1 A flowchart illustrating a method for damage identification of tensile membrane structures based on transfer learning and residual networks, provided in an embodiment of the present invention;

[0050] Figure 2 This is a schematic diagram of a planar unidirectional tension membrane structure under different damage conditions provided in the embodiments of the present invention;

[0051] Figure 3 This is a schematic diagram of wavelet transform of a one-dimensional vibration sequence signal of a tension membrane structure provided in an embodiment of the present invention;

[0052] Figure 4 A schematic diagram of the ResNet residual network model architecture with a hybrid embedded attention mechanism provided in an embodiment of the present invention;

[0053] Figure 5 A schematic diagram of the channel feature attention mechanism (CAM) provided in an embodiment of the present invention;

[0054] Figure 6 This is a schematic diagram of the Spatial Feature Attention (SAM) mechanism provided in an embodiment of the present invention;

[0055] Figure 7 Schematic diagrams of Hybrid Channel Feature Attention (CAM) and Spatial Feature Attention (SAM) provided in embodiments of the present invention;

[0056] Figure 8 This is a schematic diagram illustrating the damage identification of tensile membrane structures using an improved ResNet residual network model employing transfer learning, as provided in an embodiment of the present invention. Detailed Implementation

[0057] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0058] like Figure 1As shown in the figure, this embodiment discloses a damage identification method for tensile membrane structures based on transfer learning and residual networks. It can solve the problems that traditional structural dynamic detection methods are difficult to apply to tensile membrane structures, which are highly nonlinear systems that integrate material nonlinearity and geometric nonlinearity, and cannot accurately characterize the damage characteristics of tensile membrane structures. It belongs to the field of damage monitoring of cable membrane structure building materials.

[0059] Specifically, it includes:

[0060] Step 1: Apply an external excitation signal to the membrane surface and use an accelerometer to collect the original vibration signals of the membrane surface under different damage conditions;

[0061] Step 2: Preprocess the membrane vibration signal data under different damage conditions, and use wavelet transform to convert the one-dimensional membrane vibration sequence data into two-dimensional image data to construct a membrane vibration signal image dataset under different damage conditions.

[0062] Step 3: Introduce the attention mechanism into the ResNet50 residual network architecture. Embed the attention mechanism in both the shallow and deep layers of the ResNet50 network to complete the construction of the improved ResNet50 residual network model.

[0063] Step 4: Train the improved ResNet50 residual network model on the constructed dataset of membrane vibration signal images under different damage conditions, and use the transfer learning method to transfer the parameter weights of the ImageNet pre-trained model to the improved ResNet50 residual network model for training, while retaining the optimal model parameter weights during model training.

[0064] Step 5: Based on the pre-trained optimal improved ResNet50 residual network model, classify and detect membrane vibration signals under different damage conditions, and output the membrane damage classification and detection results and attention feature maps. Compared with traditional structural dynamic detection modal parameter identification methods, the method of this invention utilizes the exceptionally powerful feature learning capability of deep learning models to automatically extract deeper damage features from massive amounts of data, and is well-suited for dynamic damage detection of tensile membrane structures with dense natural frequencies and complex high-order mode shapes.

[0065] This invention largely eliminates the reliance on traditional structural dynamic detection signal processing techniques, manual feature extraction, and expert experience, making it one of the ideal methods for intelligent dynamic detection and health monitoring of tensile membrane structures. Furthermore, targeted improvement strategies are proposed, making the tensile membrane structure damage identification method based on transfer learning and residual networks described in this invention highly efficient, convenient, and low-cost, with broad application prospects in the field of comprehensive dynamic detection and health monitoring of cable-membrane structures.

[0066] The specific implementation process of applying an external excitation signal to the membrane surface and collecting the original vibration signals of the membrane surface under different damage conditions using an accelerometer is as follows:

[0067] Step 1: Divide the membrane surface to be tested into a grid to determine the measurement points, ensuring that the measurement points are relatively densely and evenly distributed on the surface of the membrane surface to be tested;

[0068] Step 2: In the dynamic characteristic test of different damaged membrane surfaces, an external excitation signal is applied to the damaged membrane surface under test to make it undergo forced vibration. The external excitation signal includes, but is not limited to, steady-state sinusoidal excitation, sinusoidal frequency sweep excitation, random excitation and hammer impact excitation.

[0069] Step 3: Place the accelerometer at the measurement point to measure the forced vibration signal of the membrane surface under different damage conditions, and export the .xlsx files sequentially after the measurement is completed.

[0070] Specifically, the planar uniaxial tension membrane structure on the laboratory experimental platform was divided into a cross-shaped grid, with the cross intersections serving as sensor measurement points. Considering that the planar uniaxial tension membrane structure is a rectangular membrane surface with a length × width of 795mm × 765mm, the grid was divided along the length and width directions at intervals of 53mm and 51mm, respectively, so that the sensor measurement points are relatively densely and evenly distributed on the surface of the planar uniaxial tension membrane structure.

[0071] Specifically, such as Figure 2 As shown, the experiment was designed with four damage conditions for the membrane surface: damage condition 1 (intact membrane surface), damage condition 2 (1 crack), damage condition 3 (1 crack and 1 hole), and damage condition 4 (2 cracks and 2 holes). An external impact hammer excitation signal was applied to the planar uniaxial tensioned membrane structure under different damage conditions to induce forced vibration. Donghua sensors were placed at the measurement points on the membrane surface of the planar uniaxial tensioned membrane structure to measure the vibration signals. Data was measured three times at different measurement points on the membrane surface. After testing all measurement points on the membrane surface, .xlsx files were exported sequentially. 570 sets of vibration data were collected for each damage condition, for a total of 2280 sets of membrane surface vibration signal data collected for all conditions.

[0072] Specifically, noise points in a total of 2280 sets of membrane surface vibration signal data collected under all working conditions were manually removed. This included noise signals collected by the sensors before applying external excitation signals to the membrane surface under test, and noise signals collected by the sensors after the membrane surface underwent forced vibration attenuation following the application of external excitation signals. The .xlsx files containing one-dimensional vibration signals of different damaged membrane surfaces after manual preprocessing had the same data length. The processed vibration sequence signal length for each measurement point of the planar unidirectional tensioned membrane structure was uniformly set to 3000 values.

[0073] The method of converting one-dimensional membrane vibration sequence data into two-dimensional image data using wavelet transform to construct a membrane vibration signal image dataset under different damage conditions is as follows:

[0074] Based on the preprocessed one-dimensional vibration signal .xlsx file data of different damaged membrane surfaces, the collected vibration signals of the damaged membrane surfaces are processed using Morlet continuous wavelet transform to convert the one-dimensional vibration time-domain signal of the damaged membrane surface into two-dimensional damaged membrane surface image data including time-domain and frequency-domain information, and divided into training set, validation set and test set.

[0075] Specifically, such as Figure 3 As shown, based on the preprocessed one-dimensional vibration signal (.xlsx file) data of the damaged membrane surface of a planar uniaxial tensioned membrane structure, the collected vibration signals of the damaged membrane surface were processed using Morlet continuous wavelet transform, converting them into two-dimensional image data with a pixel value of 512×512. This achieves the conversion of the one-dimensional vibration time-domain signal of the damaged membrane surface into two-dimensional image data of the damaged membrane surface including time-domain and frequency-domain information. The 570 sets of membrane surface vibration data collected under each damage condition were divided into training, validation, and test sets according to a 6:2:2 ratio.

[0076] Wavelet transform was used to convert one-dimensional membrane vibration sequence data into two-dimensional image data, which was then used to construct a membrane vibration signal image dataset under different damage conditions. The specific method is as follows:

[0077] By utilizing the adjustable time-frequency window of continuous wavelet transform, the width of the time window changes accordingly with the frequency when performing time-frequency analysis on the vibration signal of the damaged membrane. When the vibration frequency of the damaged membrane is high, the width of the time window in the wavelet transform process will decrease accordingly, thereby improving the frequency resolution.

[0078] Meanwhile, when the vibration frequency of the damaged membrane is low, the time window width during wavelet transform increases accordingly, resulting in a decrease in frequency resolution. Based on the adaptive nature of wavelet transform for signal analysis, and considering that wavelet transform essentially matches the target signal with a scaled and translated copy of the mother wavelet, by comparing the variation characteristics of the damaged membrane vibration signal and the similarity between different mother wavelets and their signal variation characteristics, the Morlet mother wavelet is selected to perform continuous wavelet transform on the damaged membrane vibration signal. This achieves the best time-frequency analysis effect on the damaged membrane vibration, thus reflecting the rapid changes in information across various frequency bands under different forced vibration conditions of the damaged membrane within a short time.

[0079] The method involves introducing an attention mechanism into the ResNet50 residual network architecture, embedding the attention mechanism into both the shallow and deep layers of the ResNet50 network, and constructing an improved ResNet50 residual network model.

[0080] Generally, the ResNet50 residual network model is divided into 5 stages. The input data passes through the 5 stages of ResNet50 (Stage 1, Stage 2, Stage 3, Stage 4, Stage 5) to obtain the corresponding output results. Among them, Stage 1 has a relatively simple structure and can be regarded as a preprocessing of the input. The latter 4 stages are all composed of Bottlenecks and have similar structures.

[0081] Stage 2 contains 3 bottlenecks, and the remaining 3 stages contain 4, 6, and 3 bottlenecks respectively. Figure 4 As shown, this invention integrates channel attention and spatial attention mechanisms in Stages 2-5. Specifically, it integrates the attention mechanism between shallow network Stage 2 and Stage 3, between shallow network Stage 3 and deep network Stage 4, and between deep network Stage 4 and Stage 5, thereby constructing an improved ResNet50 residual network model.

[0082] The improved ResNet50 residual network model is constructed by incorporating both channel attention and spatial attention mechanisms in Stages 2-5. The specific method is as follows:

[0083] The hybrid embedding attention mechanism of this invention consists of two attention modules: a channel attention module (CAM) and a spatial attention module (SAM). The CAM allocates attention resources to each convolutional channel, while the SAM transfers spatial information from the original image to a specific space and extracts key information, enabling the convolutional neural network to more accurately focus on regions that play a decisive role in the classification of two-dimensional images of damaged membrane surfaces. By hybrid embedding the attention mechanism into both shallow and deep networks, the ResNet50 residual network model is improved, enhancing its ability to extract time-frequency features from different damaged membrane surfaces.

[0084] The hybrid embedding attention mechanism consists of two attention modules: a channel attention module and a spatial attention module. The specific method is as follows:

[0085] Step 1: As Figure 5 As shown, the channel attention mechanism CAM first uses average pooling (AvgPool) and max pooling (MaxPool) to aggregate spatial information, and aggregates the initial feature dimension H×W×C into two 1×1×C vectors. These two vectors represent the channel average pooling feature and the channel max pooling feature, respectively.

[0086] Step 2: Pass the two channel feature vectors above through a fully connected layer network with shared weights, and combine the output vectors using the element-wise summation method to generate channel-dimensional attention weights;

[0087] Step 3: As Figure 6 As shown, compared to the channel attention mechanism, the spatial attention mechanism (SAM) focuses on positional information, complementing the channel attention mechanism in the H×W dimension. Similarly, max pooling and average pooling are used on the channels to generate two H×W×1 feature vectors, which represent the spatial average pooling feature and the spatial max pooling feature, respectively.

[0088] Step 4: Pass the two spatial feature vectors above through a fully connected layer network with shared weights, and merge the output vectors using the element-wise summation method to generate spatial dimension attention weights;

[0089] Step 5: As Figure 7 As shown, the stacked channel dimension and spatial dimension feature maps are used to generate a single-channel spatial attention map through a convolutional layer, and attention weights are assigned to the locations that need to be suppressed or enhanced in the two-dimensional images of different damaged membrane surfaces.

[0090] The method of transferring the parameter weights of the ImageNet pre-trained model to the improved ResNet50 residual network model for training using transfer learning is as follows:

[0091] like Figure 8 As shown, an improved ResNet50 residual network model is pre-trained using the publicly available ImageNet dataset. Transfer learning is then used to apply the ImageNet pre-trained model as a feature extraction network to train the improved ResNet50 residual network model on two-dimensional image datasets of different damaged membrane surfaces, thereby extracting the time-frequency features of the two-dimensional images of different damaged membrane surfaces. By employing transfer learning, the amount of training data on two-dimensional images of damaged membrane vibration signals and the computational resources of the experimental platform required to build the improved ResNet50 residual network model can be reduced, eliminating the need for the improved ResNet50 residual network model to be retrained and learned from scratch when training on two-dimensional images of damaged membrane surfaces.

[0092] The pre-trained model based on the optimal improved ResNet50 residual network classifies and detects membrane vibration signals under different damage conditions, and outputs membrane damage classification and detection results and attention feature maps. The specific method is as follows:

[0093] A pre-trained model, trained on two-dimensional images of damaged membrane surfaces using an improved ResNet50 residual network model, is used to classify and detect membrane vibration signals under different damage conditions. Based on the classification results, six evaluation parameters are calculated: accuracy on both the training and test sets, loss on both sets, number of correct classifications (true positives TP), number of false positives (false negatives FN), number of correct and normal classifications (true negatives TN), and number of false positives (false FP). The attention weight allocation of the final model on the test set is visualized using a hybrid embedding attention mechanism within the improved ResNet50 residual network model, and attention feature maps for two-dimensional images of damaged membrane surfaces are output.

[0094] Specifically, a pre-trained model was trained on two-dimensional images of the membrane surface of a unidirectional tensile membrane structure with a damaged plane, based on an improved ResNet50 residual network model, to classify and detect membrane vibration signals under different damage conditions. Based on the classification results, the accuracy, loss, number of correct classifications (true positives TP), number of false positives (false negatives FN), number of correct and normal classifications (true negatives TN), and number of false positives (false positives FP) were calculated for both the training and test sets. A confusion matrix was then constructed by combining the calculated true positives TP, false negatives FN, true negatives TN, and false positives FP.

[0095] By combining the attention mechanism of hybrid embedding of the improved ResNet50 residual network model, the final model is visualized on the vibration test set of the planar uniaxial tensile membrane structure, and the attention feature map of the two-dimensional image of the planar uniaxial tensile membrane structure with different damage is output.

[0096] Obviously, those skilled in the art can make various modifications and variations to this invention without departing from its spirit and scope. Therefore, if these modifications and variations fall within the scope of the claims of this invention and their equivalents, this invention also intends to include these modifications and variations.

Claims

1. A method for damage identification of tensile membrane structure based on transfer learning and residual network, characterized in that, Includes the following steps: Step 1: Apply an external excitation signal to the membrane surface and collect the original vibration signals of the membrane surface under different damage conditions; Step 2: Preprocess the membrane vibration signal data under different damage conditions, and use wavelet transform to convert the one-dimensional membrane vibration sequence data into two-dimensional image data to construct a membrane vibration signal image dataset under different damage conditions. Step 3: Introduce the attention mechanism into the ResNet50 residual network architecture. Embed the attention mechanism in both the shallow and deep layers of the ResNet50 network to complete the construction of the improved ResNet50 residual network model. Step 4: Train the improved ResNet50 residual network model on the constructed dataset of membrane vibration signal images under different damage conditions, and use the transfer learning method to transfer the parameter weights of the ImageNet pre-trained model to the improved ResNet50 residual network model for training, while retaining the optimal model parameter weights during model training. Step 5: Based on the pre-trained optimal improved ResNet50 residual network model, classify and detect membrane vibration signals under different damage conditions, and output the membrane damage classification and detection results and attention feature map.

2. The method for damage identification of tensile membrane structures based on transfer learning and residual networks as described in claim 1, characterized in that, In step 1, the original vibration signals of the membrane surface under different damage conditions are collected using an accelerometer. The specific steps are as follows: Step 21: Divide the membrane surface to be tested into a grid to determine the measurement points, ensuring that the measurement points are relatively densely and evenly distributed on the surface of the membrane surface to be tested; Step 22: In the dynamic characteristic test of different damaged membrane surfaces, an external excitation signal is applied to the tested damaged membrane surface to make it undergo forced vibration. The external excitation signal includes steady-state sinusoidal excitation, sinusoidal frequency sweep excitation, random excitation and hammer impact excitation. Step 23: Place the accelerometer at the measurement point to measure the forced vibration signal of the membrane surface under different damage conditions, and export the .xlsx files sequentially after the measurement is completed.

3. The method for damage identification of tensile membrane structures based on transfer learning and residual networks as described in claim 2, characterized in that, The noise points in the .xlsx files of membrane vibration signals under different damage conditions were manually deleted sequentially, including the noise signals collected by the accelerometer before the external excitation signal was applied to the membrane surface under test, and the noise signals collected by the accelerometer after the membrane surface was subjected to forced vibration decay by the applied external excitation signal. The .xlsx files containing one-dimensional vibration signals from different damaged membrane surfaces after manual preprocessing have the same data length.

4. The method for damage identification of tensile membrane structures based on transfer learning and residual networks as described in claim 3, characterized in that, In step 2, the method of converting one-dimensional membrane vibration sequence data into two-dimensional image data using wavelet transform to construct a membrane vibration signal image dataset under different damage conditions is as follows: Based on the preprocessed one-dimensional vibration signal .xlsx file data of different damaged membranes, the collected damaged membrane vibration signals are processed using Morlet continuous wavelet transform to convert the one-dimensional vibration time-domain signal of the damaged membrane into two-dimensional damaged membrane image data including time-domain and frequency-domain information, and divided into training set, validation set and test set.

5. The method for damage identification of tensile membrane structures based on transfer learning and residual networks as described in claim 4, characterized in that, By utilizing the adjustable time-frequency window of continuous wavelet transform, the width of the time window changes accordingly with the frequency when performing time-frequency analysis on the vibration signal of the damaged membrane surface. When the vibration frequency of the damaged membrane is high, the time window width in the wavelet transform process will decrease accordingly, and the frequency resolution will increase accordingly. Meanwhile, when the vibration frequency of the damaged membrane is low, the time window width in the wavelet transform process will increase accordingly, and the corresponding frequency resolution will decrease. Based on the adaptive nature of wavelet transform for signal analysis, and considering that wavelet transform essentially matches the target signal with a scaled and translated copy of the mother wavelet, by comparing the variation characteristics of the vibration signal of the damaged membrane and the similarity between different mother wavelets and their variation characteristics, the Morlet mother wavelet is selected to perform continuous wavelet transform on the vibration signal of the damaged membrane. This achieves the best time-frequency analysis effect on the vibration of the damaged membrane, thereby reflecting the rapid change characteristics of information in various frequency bands under forced vibration conditions of different damaged membranes in a short period of time.

6. The method for damage identification of tensile membrane structures based on transfer learning and residual networks as described in claim 1, characterized in that, In step 3, attention mechanisms are embedded into both the shallow and deep layers of the ResNet50 network to construct an improved ResNet50 residual network model. The specific method is as follows: The ResNet50 residual network model is divided into 5 stages. The input data goes through the 5 stages of ResNet50, namely Stage 1, Stage 2, Stage 3, Stage 4 and Stage 5, and then the corresponding output results are obtained. Stage 1 has a simple structure and can be regarded as a preprocessing of the input. The other stages are all composed of Bottleneck and have similar structures. Stage 2 contains 3 Bottlenecks, while the remaining Stage 3, Stage 4, and Stage 5 contain 4, 6, and 3 Bottlenecks respectively. The improved ResNet50 residual network model is constructed by mixing channel attention and spatial attention mechanisms in Stages 2-5. Specifically, the attention mechanism is mixed between the shallow network Stage 2 and Stage 3, between the shallow network Stage 3 and the deep network Stage 4, and between the deep network Stage 4 and Stage 5.

7. The method for damage identification of tensile membrane structures based on transfer learning and residual networks as described in claim 6, characterized in that, In Stages 2-5, a hybrid embedding channel attention mechanism and spatial attention mechanism are used to construct an improved ResNet50 residual network model, specifically as follows: The hybrid embedding attention mechanism consists of two attention modules: the channel attention module (CAM) and the spatial attention module (SAM). The channel attention module (CAM) is responsible for allocating attention resources to each convolutional channel, while the spatial attention module (SAM) transfers spatial information from the original image to a specific space and extracts key information, enabling the convolutional neural network to more accurately focus on the regions that play a decisive role in the classification of two-dimensional images of damaged membrane surfaces. By hybrid embedding the attention mechanism into both shallow and deep networks, the ResNet50 residual network model is improved to enhance its ability to extract time-frequency features of different damaged membrane surfaces.

8. The method for damage identification of tensile membrane structures based on transfer learning and residual networks as described in claim 7, characterized in that, The hybrid embedding attention mechanism consists of two attention modules: a channel attention module and a spatial attention module. The specific method is as follows: Step 81: The Channel Attention Mechanism (CAM) first uses average pooling and max pooling to aggregate spatial information, aggregating the initial feature dimensions H×W×C into two 1×1×C vectors, which represent the channel average pooling feature and the channel max pooling feature, respectively. Step 82: Pass the two channel feature vectors above through a fully connected layer network with shared weights, and merge the output vectors using the element-wise summation method to generate channel-dimensional attention weights; Step 83: Compared to the channel attention mechanism, the spatial attention mechanism SAM focuses on positional information and complements the channel attention mechanism in the H×W dimension; similarly, max pooling and average pooling are used on the channel to generate two H×W×1 feature vectors, which represent the spatial average pooling feature and the spatial max pooling feature, respectively. Step 84: Pass the two spatial feature vectors above through a fully connected layer network with shared weights, and merge the output vectors using the element-wise summation method to generate spatial dimension attention weights; Step 85: Generate a single-channel spatial attention map by passing the stacked channel dimension and spatial dimension feature maps through a convolutional layer, and assign attention weights to the locations that need to be suppressed or enhanced in the two-dimensional images of different damaged membrane surfaces.

9. The method for damage identification of tensile membrane structures based on transfer learning and residual networks as described in claim 8, characterized in that, In step 4, the specific method for transferring the parameter weights of the ImageNet pre-trained model to the improved ResNet50 residual network model for training using transfer learning is as follows: An improved ResNet50 residual network model was pre-trained using the ImageNet public dataset. The ImageNet pre-trained model was then used as a feature extraction network to train the improved ResNet50 residual network model on a two-dimensional image dataset with different damaged membrane surfaces, in order to extract the time-frequency features of the two-dimensional images with different damaged membrane surfaces. By applying transfer learning, the amount of training data on two-dimensional images of vibration signals of the damaged membrane surface and the computational resources of the experimental platform required to improve the ResNet50 residual network model are reduced, so that the improved ResNet50 residual network model does not need to be retrained and learned from scratch when training on two-dimensional images of the damaged membrane surface.

10. The method for damage identification of tensile membrane structures based on transfer learning and residual networks as described in claim 9, characterized in that, In step 5, a pre-trained model is obtained by training the two-dimensional image of the damaged membrane based on the improved ResNet50 residual network model to classify and detect the membrane vibration signal under different damage conditions. Based on the classification results, six evaluation parameters are calculated: the accuracy of the training set and the test set, the loss rate of the training set and the test set, the number of correct classifications, the number of false normal classifications, the number of correct normal classifications, and the number of false classifications. By combining the attention mechanism of the improved ResNet50 residual network model with hybrid embedding, the final model's attention weight allocation on the test set is visualized, and attention feature maps of two-dimensional images of different damaged membrane surfaces are output.

Images