Structural damage detection system based on asymmetric multi-task learning

Abstract

The application relates to a structural damage detection system based on asymmetric multi-task learning, and relates to structural health monitoring. A new detection system, namely an asymmetric multi-task structural damage identification model, is provided. In a multi-task learning framework, the identification of damage location and size is regarded as two related tasks, the advantages of multi-task learning can be brought into play, and the performance of the two tasks can be improved. An asymmetric method can effectively reduce the influence of the negative transfer phenomenon existing in multi-task learning, and improve the identification performance of the damage location and size. In addition, in order to improve the effective use of the damage location feature, a position attention mechanism is further provided. The mechanism can adaptively assign weights to the damage location feature and the damage size feature. After the two features are weighted and summed, the damage size feature containing rich position information can be extracted from the damage size feature, so that the effect of the damage size identification task is improved.

Description

TECHNICAL FIELD

[0001] The present application relates to structural health monitoring, and in particular to a structural damage detection system based on asymmetric multi-task learning. BACKGROUND

[0002] Structural Health Monitoring (SHM) refers to the real-time monitoring of the working state and health condition of a structure through sensors installed on the structure and a data acquisition system, identifying the damage and its development trend of the structure, so as to improve the safety, reliability and service life of the structure. SHM technology is widely used in civil engineering, aerospace, mechanical engineering, ocean engineering and other fields, and is an important means of modern engineering structure management and maintenance. With the increasing scale and complexity of modern engineering structures, structural health detection becomes particularly important. Traditional structure maintenance and inspection methods, such as visual inspection, regular maintenance and local non-destructive testing, have the problems of low detection frequency, insufficient coverage and inability to monitor in real time. The SHM system can discover early damage of the structure in time through continuous online monitoring, avoiding catastrophic consequences caused by damage accumulation. In addition, SHM can also provide detailed structural response data to provide scientific basis for the evaluation and optimal design of the structure. The realization of structural health detection relies on sensor technology, data acquisition and transmission technology, signal processing and data analysis technology and other aspects. Commonly used sensor types include strain sensors, acceleration sensors, vibration sensors, optical fiber sensors, ultrasonic sensors, etc. These sensors can accurately capture the dynamic response and environmental changes of the structure. Data acquisition and transmission technology ensures the real-time collection and transmission of sensor data, and transmits the data to the central processing unit through wired or wireless networks. Signal processing and data analysis technology processes and analyzes the collected data to identify abnormalities and damage in the structure and predict the health status of the structure.

[0003] Multi-Task Learning (MTL) is a widely used machine learning method that can improve the generalization ability and data utilization efficiency of the model by simultaneously learning multiple related tasks. In the field of SHM, there are some applications of multi-task learning for structural damage identification. These methods aim to utilize the capabilities of multi-task learning to share feature information across different tasks, improving the detection accuracy and efficiency of each task. Although multi-task learning has significant advantages in improving model performance and data utilization, in practical applications, how to combine multiple tasks and share information to maximize the advantages of multi-task learning is still not well solved. In the field of structural health monitoring, how to ensure the effectiveness and stability of the multi-task learning model and improve the reliability of the SHM system in complex environments still needs further research and development of new models and systems.

[0004] In structural damage detection, in addition to detecting the damage location, sometimes in order to determine the damage condition more accurately, it is also necessary to detect the extent / size of the damage. The detection of multiple damage locations and damage extents constitutes a multi-task learning problem. Although multi-task learning has significant advantages in improving model performance and data utilization, in practical applications, there is low correlation between different tasks, or even negative correlation, which leads to a decline in the performance of the detection model, which is called multi-task negative transfer phenomenon. Solving the multi-task negative transfer phenomenon is of great significance to ensure the effectiveness and stability of the multi-task learning model. Improving the reliability of the SHM system in complex environments ensures the accuracy and timeliness of structural health monitoring, thereby avoiding potential structural failures and safety hazards.

[0005] In some actual damage detection problems, the changes in vibration response caused by different damage sizes are much weaker than the changes in vibration response caused by different damage locations, and it is found that the vibration response collected by the sensors close to the damage location has larger amplitude and more details. It is necessary to find a more efficient and accurate detection method to cope with these challenges, so as to better guarantee the safety and reliability of various structures. SUMMARY

[0006] The purpose of the present application is to provide a structural damage detection system based on asymmetric multi-task learning, which can effectively extract the position and size information of the damage from the vibration response data, thereby improving the accuracy of damage detection, in order to solve the above technical problems existing in the prior art. That is, the asymmetric multi-task structural damage identification model regards the identification of damage location and size as two related tasks in the multi-task learning framework, takes advantage of multi-task learning, and improves the performance of the two tasks. Effectively reduce the influence of negative transfer phenomenon existing in multi-task learning, improve the identification performance of damage location and size.

[0007] In order to achieve the above-mentioned purpose of the application, the present application provides the following technical solutions:

[0008] The present application provides a structural damage detection system based on asymmetric multi-task learning, the overall structure of the system includes two core parts, namely the backbone network and the attention subnetwork;

[0009] The backbone network is used to extract and identify the features of the damage location from the input vibration response data; the backbone network is composed of four cascaded convolution modules, each of which contains a convolution layer and a pooling layer; the four cascaded convolution modules are sequentially cascaded, and the output of each module is used as the input of the next module; within each convolution module, the output of the convolution layer is transmitted to the pooling layer for dimension reduction processing; the four cascaded convolution modules are used to extract valuable features from the input data, and finally a soft-max layer is used to generate the classification result of the damage location;

[0010] The attention sub-network is used to further learn and extract the size features of the damage size based on the damage location features obtained from the backbone network; the attention sub-network is composed of multiple cascaded attention modules, each of which uses adaptive attention masks to extract the features of the damage size, and each attention module is connected with the corresponding convolution module in the backbone network and also cascaded with the output of the previous attention module; this design enables the attention sub-network to pass and share information between the attention modules, thereby performing feature learning and optimization in a global range;

[0011] The system input is a vibration response matrix composed of vibration response data recorded by multiple sensors and aligned and spliced along the time axis; the backbone network is connected with the attention sub-network, and the internal part of the attention module accepts two inputs: one is the output from the previous attention module (or the initial input), and the other is the output from the corresponding convolution module in the backbone network; this connection ensures that the attention sub-network can make full use of the damage location features extracted by the backbone network;

[0012] The damage location classification result is generated by the soft-max layer in the backbone network, indicating the damage location classification corresponding to the input sample;

[0013] The damage size evaluation result is output by the last attention module in the attention sub-network, indicating the damage size evaluation value corresponding to the input sample.

[0014] Further, the attention module mainly consists of two parts: a mask generator and a feature extractor; the core of the mask generator is two cascaded convolution layers, which are used to learn and generate adaptive attention masks from two inputs; the attention mask is used to weight the input features so that the model can focus on the most important information; the feature extractor extracts features related to the damage size from the weighted features.

[0015] Further, within each attention module, an adaptive attention mask is generated by the mask generator, and then the mask is multiplied element by element with the input features to obtain weighted features, which are transmitted to the feature extractor for further processing.

[0016] The application provides a structural damage detection method based on asymmetric multi-task learning, comprising the following steps:

[0017] 1) Data preprocessing: aligning and splicing vibration response data recorded by multiple sensors along the time axis into a vibration response matrix;

[0018] 2) Main network processing: inputting the vibration response matrix into the main network, sequentially passing through four convolution modules, each convolution module extracting different levels of features, and the output of the last convolution module generating a damage location classification result through a soft maximum layer;

[0019] 3) Attention subnetwork processing: using the output of each convolution module in the main network as input to construct an attention subnetwork, each attention module generating an adaptive attention mask to weight the input features; the weighted features are further processed by a feature extractor to generate a damage size evaluation result;

[0020] 4) Multi-task loss optimization: using a cross-entropy loss function to calculate the loss of the main network and the attention subnetwork, combining the two through a multi-task loss function, and optimizing the model parameters through minimizing the loss function.

[0021] In step 1), the data preprocessing, k sensors are used to record the vibration response of the structure, and during the collection of vibration response data, the sampling frequency is f s , and the number of sampling points is N; the response recorded by each sensor can be represented as where s i，t is the data recorded by the i-th sensor at time step t; to integrate these data, align the vibration responses recorded by all sensors along the time axis, and splice them to form a vibration response matrix In this matrix, s i ∈R N×1 represents the vibration response recorded by the i-th sensor.

[0022] In step 2), the specific steps of the main network processing can be:

[0023] A one-dimensional convolution network is used as the main network, in which the input data is first processed by the convolution layer, this process involves the convolution filter sliding along the input sequence, while performing element-wise multiplication on the elements covered by the filter and accumulating the results, thereby generating a new sequence; this process can be represented by the mathematical formula of the convolution layer, as shown in the following formula:

[0024]

[0025] where h represents the subsequence, u(i) represents the kernel (filter), y represents the output signal, k represents the index of the subsequence, and N represents the length of the kernel; after being processed by the convolution layer, the data is transmitted to the pooling layer, which reduces the dimensionality of the features;

[0026] The backbone network is built by four closely connected convolution modules, each of which contains a convolution layer and a pooling layer, and the output features generated by each convolution layer are shared, which contain rich details of the vibration signal and important information of the damage location.

[0027] In step 3), the specific steps of the attention subnetwork processing can be:

[0028] The attention subnetwork is composed of a series of attention modules connected in cascade; each module is connected with the corresponding convolution module in the backbone network, and also cascaded with the output of the previous attention module; the attention subnetwork can pass and share information between the attention modules at each level, thereby learning and optimizing the features globally;

[0029] The internal of the attention module accepts two inputs: one from the output of the previous attention module, and the other from the damage location feature of the corresponding convolution module in the backbone network; the attention module is mainly composed of two parts: a mask generator and a feature extractor;

[0030] The core of the mask generator is two cascaded convolution layers, and the main task of these two convolution layers is to learn and generate an adaptive attention mask A n from the two inputs; the mask is used to weight the input features so that the model can focus on the most important information; the calculation process of A n is as follows:

[0031]

[0032] where F n represents the mask generator of the nth attention module, Y n-1 is the output of the previous attention module, and X n is the output of the corresponding convolution module; the symbol represents the concatenation of the two inputs in the channel dimension direction; after obtaining the adaptive attention mask A n , it can be multiplied element by element with the input value ; in this way, a new feature is obtained, in which each feature dimension is weighted by the corresponding attention weight; next, the feature extractor G n extracts the feature about the damage size from this weighted feature; in order to achieve this goal, the feature extractor G na combination of a one-dimensional convolution layer with a size of 3 and a one-dimensional pooling layer with a size of 2;

[0033]

[0034] wherein, The symbol represents an element-level multiplication operation; the first attention module accepts two inputs; the first input is derived from the output of the last convolution layer in the backbone network, which contains high-level features after multiple layers of convolution and nonlinear transformation; the second input is from the first convolution block of the backbone network, which retains some basic features of the original data; the combination of the two inputs retains both the intuitive features of the original data and the high-level features processed by the network, thereby providing rich information and helping to build a more powerful attention subnetwork.

[0035] In step 4), the specific steps of the multi-task loss optimization can be:

[0036] A multi-task loss function is used to optimize the parameters of the model by minimizing the loss function, which can consider the performance of multiple tasks at the same time, so that the model can optimize the performance of multiple tasks during training, thereby making the model have better generalization ability and higher performance;

[0037] L = L location + λL size

[0038] The variables L location and L size represent the losses for optimizing the damage location recognition ability of the backbone network and the damage size recognition ability of the attention subnetwork, respectively, for measuring the performance of the model in locating the damage location and estimating the damage size;

[0039] The parameter λ is a weight factor for adjusting the relative importance of L location and L size in the total loss, which can be set to 1; L location and L size both use cross-entropy loss as the loss function, and the specific calculation formula is as follows:

[0040]

[0041] where y represents the actual label of sample i, represents the predicted label of sample i, y i,c represents the value of the actual label of sample i on class C (usually 0 or 1, indicating whether it belongs to the class), represents the prediction result of the model on sample i in class C, represents the probability that the sample belongs to class C; N is the number of samples, and C is the total number of classes;

[0042] For each class, when the model predicts that a certain instance belongs to the class with a higher probability, the corresponding loss will decrease accordingly; this means that if the actual label y i,c is 1, that is, the instance does indeed belong to the class, then the loss function will encourage the model to increase the probability of predicting the class as the correct label; on the contrary, if the actual label y i,c is 0, that is, the instance does not belong to the class, the loss function will encourage the model to reduce the probability of predicting the class as the correct label.

[0043] Compared with the prior art, the technical effects and advantages of the present application are that:

[0044] 1. The backbone network of the present application is constructed by four closely connected convolution modules, each of which includes a convolution layer and a pooling layer. The backbone network is used to output the classification result of the damage position and share information with the attention subnetwork. The output features generated by each convolution layer are shared, and these features contain rich details of the vibration signal and important information of the damage position. The attention subnetwork uses these shared features obtained from the backbone network to accurately evaluate the damage size.

[0045] 2. The present application adopts an asymmetric multi-task learning framework and a position attention mechanism, which can fully utilize the advantages of multi-task learning methods in improving the performance of damage position and damage size recognition tasks, and can also minimize the adverse effects of negative transfer, so as to make the recognition of damage position and size more accurate. The attention subnetwork of the present application includes a series of cascaded attention modules, which can extract important information related to the damage size from the damage position features obtained from the backbone network, thereby helping to accurately identify the size of the damage.

[0046] 3. The design of the cross-entropy loss function of the present application not only focuses on the prediction ability of the model for positive samples of each class, but also focuses on the prediction ability of the model for negative samples. This method ensures that the model can exhibit excellent prediction performance on all possible classes. By minimizing this loss function, the model can more effectively learn and distinguish the features of different classes, thereby improving its overall classification performance.

[0047] 4. The present application can adaptively assign weights to damage position features and damage size features, and after weighted summation of the two features, damage size features containing rich position information can be extracted, thereby improving the effect of the damage size recognition task. Experimental results show that the present application has obvious improvement in accurately identifying the size of the damage, and is superior to other methods. BRIEF DESCRIPTION OF DRAWINGS

[0048] Figure 1 The system overall architecture diagram of the embodiment of the present application.

[0049] Figure 2 The structure diagram of the backbone network.

[0050] Figure 3 The structure diagram of the attention module.

[0051] Figure 4 The task data description. DETAILED DESCRIPTION

[0052] In order to make the objects, technical solutions and advantages of the present application clearer, the following embodiments will further illustrate the present application with the accompanying drawings. It should be understood that the specific embodiments described herein are only used to explain the present application and not to limit the present application.

[0053] In the embodiment, k sensors are used to record the vibration response of the structure. During the collection of vibration response data, the sampling frequency is f s and the number of sampling points is N. The response recorded by each sensor can be represented as where s i，t is the data recorded by the i-th sensor at time step t. In order to integrate these data, the vibration responses recorded by all sensors are aligned along the time axis and spliced to form a matrix In this matrix, s i ∈R N×1 represents the vibration response recorded by the i-th sensor. By this method, the data of multiple sensors can be effectively integrated for further analysis and processing. Considering that the damage has different positions and sizes, N l damage positions and N s damage sizes are defined. A model F is designed and trained. The purpose of the model is to determine and output the specific damage position and damage size by analyzing the input vibration response matrix S. Therefore, the function of the model can be represented as:

[0054] The present application proposes an asymmetric multi-task structure damage identification system, the backbone network of which is composed of four cascaded convolution modules (CMs) for extracting valuable features from input data. Then, a soft-max layer is used to generate the classification results of damage location. In addition to the backbone network, an attention sub-network is also constructed, which is tasked with further learning the size features of damage from the extracted location features. In this attention sub-network, cascaded attention modules (AMs) are connected to the corresponding convolution modules in the backbone network. These attention modules use adaptive attention masks to extract damage size features. For damage location and damage size recognition, the model uses a cross-entropy loss function. This loss function can optimize the model parameters to accurately estimate the location and size of damage. Experimental results show that this method effectively retains the advantages of multi-task learning while reducing the adverse effects of negative transfer. It performs very well in damage location and damage size recognition tasks. The framework diagram of the entire system is shown in Figure 1 This design structure enables the model to effectively extract the location and size information of damage from the input data, thereby improving the accuracy of damage detection.

[0055] 1. Backbone network for identifying damage location

[0056] In the structural damage detection method proposed in the present application, the damage location is first accurately identified, because accurately identifying the damage location helps to improve the performance of damage size recognition. Since one-dimensional convolutional networks are very suitable for analyzing and processing sequential data such as audio signals and time series, they are used as the backbone network in the present application to identify the damage location corresponding to the input sample. In a one-dimensional convolutional network, the input data is first processed by a convolutional layer. This process involves sliding the convolutional filter along the input sequence while performing element-wise multiplication on the elements covered by the filter and accumulating the results to produce a new sequence. This process can be represented by the mathematical formula of the convolutional layer as follows.

[0057]

[0058] where h represents a sub-sequence, u(i) represents a kernel (filter), y represents an output signal, k represents the index of the sub-sequence, and N represents the length of the kernel. After processing by the convolutional layer, the data is transmitted to the pooling layer, which is tasked with reducing the dimensionality of the features. This step is crucial because it not only reduces the dimensionality of the data but also highlights the most critical features, laying the foundation for subsequent processing.

[0059] The backbone network is composed of four closely connected convolution modules, as shown in Figure 2Each convolutional module contains one convolutional layer and one pooling layer, and the output channel number and kernel size of each module are labeled in the figure.

[0060] The function of the backbone network is not only limited to the classification result of the damage location, but also plays another key role: sharing information with the attention subnetwork. The output features generated by each convolutional layer are shared, which contain rich details of the vibration signal and important information of the damage location. The attention subnetwork uses these shared features obtained from the backbone network to accurately assess the damage size.

[0061] 2. Attention subnetwork for identifying damage size

[0062] Inspired by the attention mechanism, the present application introduces an attention mechanism to construct an attention subnetwork. The main goal of this subnetwork is to adaptively learn and extract damage size features from the damage location features obtained by the backbone network. As shown in Figure 1 , the attention subnetwork is composed of a series of attention modules connected in cascade. Each module is connected to the corresponding convolutional module in the backbone network and also cascaded with the output of the previous attention module. This design allows the attention subnetwork to pass and share information between attention modules, enabling global feature learning and optimization.

[0063] As shown in Figure 3 , the attention module accepts two inputs: one from the output of the previous attention module and the other from the damage location features of the corresponding convolutional module in the backbone network. The attention module is mainly composed of two parts: a mask generator and a feature extractor.

[0064] The core of the weight generator (also called mask generator) is two cascaded convolutional layers, whose main task is to learn and generate an adaptive attention mask A n from the two inputs. The mask serves to weight the input features so that the model can focus on the most important information. n The calculation process of the attention mask A

[0065]

[0066] where F n represents the mask generator of the nth attention module, Y n-1 is the output of the previous attention module, and X n is the output of the corresponding convolutional module. The symbol represents the concatenation of the two inputs in the channel dimension. After obtaining the adaptive attention mask A n , it can be combined with the input value An element-wise multiplication is performed. This way, a new feature is obtained, where each feature dimension is weighted by the corresponding attention weight. Next, the feature extractor G n A feature about the damage size will be extracted from this weighted feature. To achieve this goal, the feature extractor G n A combination of a one-dimensional convolution layer with a size of 3 and a one-dimensional pooling layer with a size of 2 is adopted.

[0067]

[0068] where, The symbol represents an element-wise multiplication operation. The first attention module accepts two inputs. The first input Y0 is derived from the output of the last convolution layer in the backbone network, which contains high-level features after multiple layers of convolution and nonlinear transformation. The second input X1 is from the first convolution block of the backbone network, which preserves some basic features of the original data. The combination of these two inputs not only preserves the intuitive features of the original data, but also contains high-level features processed by the network, thus providing rich information and helping to build a more powerful attention subnetwork.

[0069] 3. Loss function

[0070] During the training phase of the model, a multi-task loss function is used to optimize the parameters of the model by minimizing the loss function. This multi-task loss function can consider the performance of multiple tasks at the same time, so that the model can optimize the performance of multiple tasks during the training process, thereby making the model have better generalization ability and higher performance.

[0071] L = L location + λL size

[0072] In the model, the variables L location and L size represent the losses used to optimize the damage location recognition ability of the backbone network and the damage size recognition ability of the attention subnetwork, respectively. These two loss functions are key indicators in the training process, used to measure the performance of the model in locating damage locations and estimating damage sizes.

[0073] The parameter λ is a weight factor used to adjust the relative importance of L location and L size in the total loss. In this embodiment, λ is set to 1.

[0074] L location and L sizeBoth use cross-entropy loss as the loss function, which is an effective method to evaluate the accuracy of multi-classification problem model, and also provides gradient information for model parameter optimization. The specific calculation formula is as follows:

[0075]

[0076] Where y represents the actual label of sample i, represents the predicted label of sample i, y i,c represents the actual label of sample i in class C (usually 0 or 1, indicating whether it belongs to this class), represents the model's prediction of sample i in class C, indicating the probability that the sample belongs to class C. N is the number of samples, and C is the total number of classes.

[0077] In the model of the present application, for each class, when the model predicts that a certain instance belongs to this class with increasing probability, the corresponding loss will decrease accordingly. This means that if the actual label y i,c is 1, i.e. the instance indeed belongs to this class, then the loss function will encourage the model to increase the probability of predicting this class as the correct label. Conversely, if the actual label y i,c is 0, i.e. the instance does not belong to this class, the loss function will encourage the model to reduce the probability of predicting this class as the correct label.

[0078] In short, the design of cross-entropy loss function not only focuses on the model's prediction ability for positive samples of each class, but also focuses on its prediction ability for negative samples. This method ensures that the model can exhibit excellent prediction performance on all possible classes. By minimizing this loss function, the model can more effectively learn and distinguish the features of different classes, thereby improving its overall classification performance.

[0079] In order to more clearly illustrate the process and advantages of the method of the present application, a specific embodiment is given below.

[0080] Lamb waves are ultrasonic guided waves propagating in plate-like structures, and are often used in structural health monitoring and damage detection because they can provide important information about potential structural damage. The present application collects Lamb wave data as training and testing data for the model to verify the effectiveness of the proposed method.

[0081] The experimental device for data collection is as follows:

[0082] Firstly, the hot spot ultrasonic guided wave damage diagnosis system developed by Junsheng Technology Co. Ltd. was used to excite and receive signals. The diagnosis system was connected to four PZT sensors. During the collection of Lamb wave data, PZT-1 served as the excitation source to emit excitation signals, while the remaining sensors (PZT-2, PZT-3, PZT-4) were responsible for recording the corresponding Lamb wave signals. On the other side, a notebook computer was connected to display and save the data.

[0083] Figure 4 A task data description is shown.

[0084] A square of size 160 mm x 160 mm was set as the damaged area, which contained 16 sub-areas of size 40 mm x 40 mm. Different sub-areas showed different damage locations, with location numbers L1-L16. Damage in real structures is usually diverse and complex, and obtaining a real damage dataset for a structure is difficult and expensive. Therefore, a solid adhesive tape was attached to the surface of the sample to block the propagation of the wave, which had a similar effect to actual damage. In this embodiment, the solid tape was cut into squares to simulate surface damage such as delamination or debonding of composite laminates. Three square tapes with sizes of 2 cm, 3 cm, and 4 cm were used to simulate different sizes of damage in composite laminates. Instead of manually pressing, a 500-gram weight was placed on the tape to ensure uniform force application. In this embodiment, four piezoelectric sensors (PZT-1, PZT-2, PZT-3, PZT-4) with diameters of 6 mm and thicknesses of 0.25 mm were used to generate and collect Lamb wave data. To ensure that the damage area was fully covered, the four PZT sensors were arranged in a square array with a distance of 180 mm between each sensor.

[0085] The network training process is as follows:

[0086] Firstly, the training data came from the one-dimensional time series data collected by PZT-2, PZT-3, and PZT-4. The excitation signal was a five-cycle sinusoidal pulse with a center frequency of 125 kHz, and the sensor signal sampling frequency was 24 MHz with 8000 sampling points. Therefore, the signals from the three sensors used as receiving signals were used as input training data, each containing 8000 sampling points. The three time series data were aligned by time and concatenated to form an 8000 x 3 matrix sequence.

[0087] After obtaining the training data, the training data first enters the backbone network and undergoes three layers of convolution to extract the feature information of the time sequence. Meanwhile, the features after each layer of convolution are retained as inputs of the attention subnetwork. The output of the backbone network is used as the prediction of the damage location and as part of the first layer input of the attention subnetwork. With the output of the backbone network and the output of the first layer of convolution as inputs, the attention subnetwork first generates a mask, then weights the input features, and then uses a convolution layer to extract features from the weighted results. The final feature extraction result is an output of the attention layer and an input of the next attention layer. There are three such attention layers, and after three layers, the final damage size feature is output. After processing by the backbone network and the attention subnetwork, the model obtains the final damage location feature and the damage size feature, and both features are processed by the tiling layer and the linear layer to make a prediction. The cross-entropy loss is used as the loss function in the method, the total loss is the sum of the loss of the damage location and the loss of the damage size, and the backpropagation gradient descent is based on the total loss.

[0088] To comprehensively evaluate the effect of the present application, experiments were conducted on the Lamb wave dataset. Single-task learning, multi-task learning and the method of the present application were used to conduct experiments on the test set, and the experimental results are shown in Tables 1-3. In the case of single damage, all methods can accurately detect the location and size of the damage. In the case of double damage, all methods can accurately identify the location of the damage. In identifying the damage size, the accuracy of the single-task learning method is 87.9%, the accuracy of the multi-task learning method is 90.7%, and the accuracy of the method proposed in the present application is 93.5%. In the case of three damages, all methods can accurately identify the location of the damage. In identifying the damage size, the accuracy of the single-task learning method is 89.3%, the accuracy of the multi-task learning method is 92.7%, and the accuracy of the method proposed in the present application is 97.4%.

[0089] Table 1

[0090]

[0091] Table 2

[0092]

[0093] Table 3

[0094]

[0095] As can be seen from Tables 1 to 3, the embodiment of the present application provides an innovative damage detection method. The method adopts an asymmetric multi-task learning framework and a position attention mechanism, can fully exert the advantages of the multi-task learning method in improving the performance of the damage position and damage size identification tasks, and can also minimize the adverse effects of the negative transfer phenomenon, so as to make the identification of the damage position and size more accurate. The method mainly includes two parts: first, a backbone network based on a one-dimensional convolutional neural network is used to determine the position of the damage; and second, the size of the damage is identified by a special position attention subnetwork combined with the damage position features. The attention subnetwork includes a series of cascaded attention modules, which can extract important information related to the damage size from the damage position features obtained from the backbone network, thereby helping to accurately identify the size of the damage. In order to comprehensively evaluate the effect of the method, experiments are carried out on the Lamb wave data set. The experimental results show that the method has obvious improvement in accurately identifying the size of the damage, and is superior to other methods.

[0096] The above embodiments are only the preferred embodiments of the present application and should not be considered as limiting the scope of the present application. Any equivalent changes and improvements made according to the scope of the present application should still belong to the patent coverage of the present application.

Claims

1. A structural damage detection system based on asymmetric multi-task learning, characterized in that... The overall structure of the system comprises two core components: the backbone network and the attention subnetwork. The backbone network is responsible for extracting and identifying features of the damage location from the input vibration response matrix; The backbone network consists of four cascaded convolutional modules, each containing one convolutional layer and one pooling layer. The four cascaded convolutional modules are connected sequentially, with the output of each module serving as the input of the next module. Within each convolutional module, the output of the convolutional layer is passed to the pooling layer for dimensionality reduction. The four cascaded convolutional modules are used to extract valuable features from the input data, and finally, a soft max layer is used to generate the classification result of the damage location. The attention subnetwork is used to further learn and extract the size features of the damage based on the classification results of the damage location obtained from the backbone network; The attention subnetwork consists of multiple cascaded attention modules. Each attention module uses an adaptive attention mask to extract features of damage size. Each attention module is connected to the corresponding convolutional module in the backbone network and is also cascaded with the output of the previous attention module. This design enables the attention subnetwork to pass and share information between attention modules at each level, thereby enabling feature learning and optimization on a global scale. The system input is a vibration response matrix, which is formed by aligning and stitching together vibration response data recorded by multiple sensors along the time axis. The backbone network is connected to the attention sub-network. The attention module internally accepts two inputs: one is the output from the previous attention module or the initial input, and the other is the output from the corresponding convolution module in the backbone network. This connection method ensures that the attention sub-network can make full use of the damage location features extracted by the backbone network. The damage location classification result is generated by the soft maximum layer in the backbone network, representing the damage location classification corresponding to the input sample; The damage size assessment result is output by the last attention module in the attention subnetwork, representing the damage size assessment value corresponding to the input sample; The attention module includes a mask generator and a feature extractor; the core of the mask generator is two cascaded convolutional layers, which are used to learn and generate an adaptive attention mask from two inputs; the adaptive attention mask is used to weight the input features; the feature extractor extracts features about the damage size from the weighted features.

2. A structural damage detection method based on asymmetric multi-task learning, characterized in that... Includes the following steps: 1) Data preprocessing: Align the vibration response data recorded by multiple sensors along the time axis and stitch them together to form a vibration response matrix; 2) Backbone network processing: The vibration response matrix is ​​input into the backbone network and passed through four convolutional modules in sequence. Each convolutional module extracts features at different levels, and the output of the last convolutional module generates the damage location classification result through a soft maximum layer. The specific steps of the backbone network processing are as follows: A one-dimensional convolutional network is used as the backbone network. In a one-dimensional convolutional network, the input data is first processed by the convolutional layer. This process involves the convolutional filter sliding along the input sequence, while performing element-wise multiplication on the elements covered by the filter and accumulating the results to generate a new sequence. This process is represented by the mathematical formula of the convolutional layer, as shown in the following equation: Where h represents the subsequence, u(i) represents the filter, y represents the output signal, k represents the index of the subsequence, and N represents the length of the kernel; after being processed by the convolutional layer, the data is transmitted to the pooling layer, which reduces the dimensionality of the features. The backbone network is constructed from four closely connected convolutional modules, each of which contains one convolutional layer and one pooling layer. The output features generated by each convolutional layer are shared. 3) Attention sub-network processing: Using the outputs of each convolutional module in the backbone network as input, an attention sub-network is constructed. Each attention module generates an adaptive attention mask to weight the input features. The weighted features are further processed by a feature extractor to generate damage size assessment results. The specific steps of the attention sub-network processing are as follows: The attention subnetwork consists of a series of cascaded attention modules; each module is connected to the corresponding convolutional module in the backbone network, and is also cascaded with the output of the previous attention module; the attention subnetwork passes and shares information between attention modules at each level, thereby performing feature learning and optimization on a global scale; The attention module internally accepts two inputs: one from the output of the previous attention module, and the other from the damage location features of the corresponding convolutional module in the backbone network; the attention module consists of two parts: a mask generator and a feature extractor; The core of the mask generator consists of two cascaded convolutional layers. The task of these two convolutional layers is to learn from two inputs and generate an adaptive attention mask A. n The mask's function is to weight the input features so that the model can focus on the most important information; A n The calculation process is as follows: Among them, F n Y represents the mask generator for the nth attention module. n-1 It is the output of the previous attention module, X n This corresponds to the output of the convolution module; symbol This represents the concatenation of two inputs along the channel dimension; the adaptive attention mask A is obtained. n Then, compare it with the input value. Perform element-wise multiplication; this yields a new feature, where each feature dimension is weighted by its corresponding attention weight; next, the feature extractor G... n Features about the damage magnitude will be extracted from this weighted feature set; to achieve this, the feature extractor G... n A combination of a one-dimensional convolutional layer of size 3 and a one-dimensional pooling layer of size 2 is used; in, The symbol represents an element-level multiplication operation; the first attention module accepts two inputs; the first input comes from the output of the last convolutional layer in the backbone network, which contains high-level features after multiple convolutions and non-linear transformations; the second input comes from the first convolutional block of the backbone network. 4) Multi-task loss optimization: The cross-entropy loss function is used to calculate the loss of the backbone network and the attention sub-network. The two are combined through the multi-task loss function, and the model parameters are jointly optimized by minimizing the loss function.

3. The structural damage detection method based on asymmetric multi-task learning as described in claim 2, characterized in that... In step 1), the data preprocessing uses k sensors to record the vibration response of the structure. During the acquisition of vibration response data, the sampling frequency is f. s The number of sampling points is N; the response recorded by each sensor is represented as... Among them, s i，t This represents the data recorded by the i-th sensor at time step t. To integrate this data, the vibration responses recorded by all sensors are aligned along the time axis and stitched together to form a vibration response matrix. In this matrix, s i ∈R N×1 This represents the vibration response recorded by the i-th sensor.

4. The structural damage detection method based on asymmetric multi-task learning as described in claim 2, characterized in that... In step 4), the specific steps of the multi-task loss optimization are as follows: By employing a multi-task loss function, the model parameters are optimized by minimizing this loss function. This multi-task loss function can take into account the performance of multiple tasks at the same time, enabling the model to optimize the performance of multiple tasks simultaneously during training. L=L location +λL size variable L location and L size These represent the losses used to optimize the damage location recognition ability of the backbone network and the damage size recognition ability of the attention subnetwork, respectively, and are used to measure the model's performance in locating damage locations and estimating damage sizes. The parameter λ is a weighting factor used to adjust L location and L size Its relative importance in the total loss; L location and L size Both use cross-entropy loss as the loss function, and the specific calculation formula is shown in the following formula: Where y represents the actual label of sample i. Let y represent the predicted label of sample i. i,c This represents the actual label value of sample i in category C. This represents the model's prediction result for sample i in category C, and represents the probability that the sample belongs to category C; N is the number of samples, and C is the total number of categories; For each category, as the model predicts a higher probability that an instance belongs to that category, the corresponding loss decreases; this means that if the actual label y... i,c If the value is 1, meaning the instance actually belongs to that category, then the loss function will incentivize the model to increase the probability of predicting that category as the correct label; conversely, if the actual label y is 1, then the model will be incentivized to increase the probability of predicting that category as the correct label. i,c If the value is 0, it means the instance does not belong to that category, and the loss function will incentivize the model to reduce the probability of predicting that category as the correct label.

Images