Structural damage detection system based on asymmetric multi-task learning

Abstract

The invention discloses 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 structure damage identification model, is provided, in a multi-task learning framework, identification of damage positions and sizes is regarded as two related tasks, the advantages of multi-task learning can be exerted, and meanwhile the performance of the two tasks is improved. By adopting the asymmetric method, the influence of a negative migration phenomenon existing in multi-task learning can be effectively reduced, and the identification performance of the damage position and size is improved. Besides, in order to improve the effective utilization of the damage position features, a position attention mechanism is also provided, the mechanism can adaptively allocate weights for the damage position features and the damage size features, and after the two features are weighted and summed, the damage size features containing rich position information can be extracted from the two features. Therefore, the effect of the damage size identification task is improved.

Description

Technical Field

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

[0002] Structural health monitoring (SHM) refers to the use of sensors and data acquisition systems installed on the structure to monitor the working status and health of the structure in real time, identify the damage and its development trend of the structure, and thus improve the safety, reliability and service life of the structure. SHM technology is widely used in civil engineering, aerospace, mechanical engineering, marine 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 has become particularly important. Traditional structural 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. Through continuous online monitoring, the SHM system can detect early damage to the structure in time and avoid catastrophic consequences caused by accumulated damage. In addition, SHM can also provide detailed structural response data, providing a scientific basis for the evaluation and optimization design of the structure. The implementation of structural health detection depends on sensor technology, data acquisition and transmission technology, signal processing and data analysis technology and other aspects. Common sensor types include strain sensors, accelerometers, vibration sensors, fiber optic 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 anomalies 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 of the model and the efficiency of data utilization by learning multiple related tasks at the same time. In the field of SHM, there are already some applications that use multi-task learning for structural damage identification. These methods aim to utilize the function of multi-task learning, share the feature information of different tasks, and improve 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, there is still no good solution for how to combine multiple tasks and what information to share to maximize the advantages of multi-task learning. In the field of structural health monitoring, how to ensure the effectiveness and stability of multi-task learning models and improve the reliability of SHM systems in complex environments still requires in-depth research and the development of new models and systems.

[0004] In structural damage detection, in addition to detecting the damage location, sometimes in order to more accurately determine the damage situation, it is also necessary to detect the degree / size of the damage. The detection of multiple damage locations and damage degrees 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 a low correlation or even negative correlation between different tasks, which leads to a decrease in the performance of the detection model. This phenomenon is called multi-task negative transfer. Solving the multi-task negative transfer phenomenon is of great significance to ensure the effectiveness and stability of the multi-task learning model. Improve the reliability of the SHM system in complex environments, ensure the accuracy and timeliness of structural health monitoring, and thus avoid 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 those caused by different damage locations, and it is found that the vibration response collected by sensors close to the damage location has a larger amplitude and richer details. This requires finding more efficient and accurate detection methods to meet these challenges, so as to better ensure the safety and reliability of various structures. Summary of the invention

[0006] The purpose of the present invention is to address the above-mentioned technical problems existing in the prior art and provide a structural damage detection system based on asymmetric multi-task learning that can effectively extract the location and size information of damage from vibration response data, thereby improving the accuracy of damage detection. That is, an asymmetric multi-task structural damage identification model, in the multi-task learning framework, regards the identification of damage location and size as two related tasks, giving full play to the advantages of multi-task learning and improving the performance of the two tasks at the same time. It effectively reduces the impact of the negative transfer phenomenon existing in multi-task learning and improves the recognition performance of damage location and size.

[0007] In order to achieve the above-mentioned invention object, the present invention provides the following technical solutions:

[0008] The present invention provides a structural damage detection system based on asymmetric multi-task learning, which is based on an asymmetric multi-task learning framework and is used to achieve accurate positioning and size assessment of structural damage by integrating multiple sub-networks; the overall structure of the system includes two core parts, namely a backbone network and an attention sub-network;

[0009] The backbone network is responsible for extracting and identifying features of damage locations from input vibration response data; the backbone network is composed of four cascaded convolution modules, each of which includes a convolution layer and a pooling layer; the four cascaded convolution modules are cascaded in sequence, 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 passed to the pooling layer for dimensionality reduction processing; the four cascaded convolution modules are used to extract valuable features from the input data, and finally generate a classification result of the damage location through a soft maximum layer;

[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 an adaptive attention mask to extract the features of the damage size, and each attention module is connected to the corresponding convolution module in the backbone network, and is also cascaded with the output of the previous attention module; this design enables the attention sub-network to transfer and share information between attention modules at all levels, thereby performing feature learning and optimization on a global scale;

[0011] The input of the system is a vibration response matrix, which is formed by aligning and splicing the vibration response data recorded by multiple sensors along the time axis; the backbone network is connected to the attention sub-network, and 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 method ensures that the attention sub-network can fully utilize the damage location features extracted by the backbone network;

[0012] The damage location classification result is generated by the softmax layer in the backbone network, which represents the damage location classification corresponding to the input sample;

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

[0014] Furthermore, the attention module is mainly composed of two parts: 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 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 about the size of the damage from the weighted features.

[0015] Furthermore, inside each attention module, an adaptive attention mask is generated by a mask generator, and then the mask is element-wise multiplied with the input feature to obtain the weighted feature, which is passed to the feature extractor for further processing.

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

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

[0018] 2) Backbone network processing: The vibration response matrix is ​​input into the backbone network and passes through four convolution modules in sequence. Each convolution module extracts features at different levels. The output of the last convolution module generates the damage location classification result through the soft max layer.

[0019] 3) Attention sub-network processing: Using the output of each convolutional module in the backbone network as input, an attention sub-network is constructed. Each attention module generates an adaptive attention mask and performs weighted processing on the input features. The weighted features are further processed by the feature extractor to generate damage size assessment results.

[0020] 4) Multi-task loss optimization: The cross entropy loss function is used to calculate the losses 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.

[0021] In step 1), the data is preprocessed by using k sensors to record the vibration response of the structure. During the collection of the vibration response data, the sampling frequency is f s , the number of sampling points is N; the response recorded by each sensor can be expressed as Among them, s i，t is the data recorded by the i-th sensor at time step t; to integrate these data, the vibration responses recorded by all sensors are aligned along the time axis and spliced ​​together 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 backbone network processing may be:

[0023] 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 convolution layer. This process involves sliding the convolution filter along the input sequence, performing element-by-element multiplication on the elements covered by the filter and accumulating the results to generate a new sequence. This process can be expressed by the mathematical formula of the convolutional layer, as shown below:

[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 transferred to the pooling layer, which reduces the dimensionality of the features.

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

[0027] In step 3), the specific steps of the attention sub-network processing may be:

[0028] 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 can transfer and share information between attention modules at all levels, thereby performing feature learning and optimization on a global scale;

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

[0030] The core of the mask generator is two cascaded convolutional layers, whose main task is to learn and generate an adaptive attention mask A from two inputs. n ; The role of this mask 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:

[0031]

[0032] Among them, F n represents the mask generator of the nth attention module, Y n-1 is the output of the previous attention module, X n is the output of the corresponding convolution module; the symbol Represents the concatenation of two inputs in the channel dimension direction; obtains the adaptive attention mask A n Then, you can compare it with the input value Perform element-by-element multiplication; this will result in a new feature in which each feature dimension is weighted by the corresponding attention weight; next, the feature extractor G n Features about the size of the damage will be extracted from this weighted feature; to achieve this goal, the feature extractor G nA combination of a one-dimensional convolutional layer of size 3 and a one-dimensional pooling layer of size 2 is used;

[0033]

[0034] in, The symbol represents 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 layers of convolution and nonlinear transformation; the second input comes from the first convolutional block of the backbone network, which retains some basic features of the original data; the combination of these two inputs not only retains the intuitive features of the original data, but also contains high-level features after network processing, thereby providing rich information and helping to build a more powerful attention sub-network.

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

[0036] A multi-task loss function is used to optimize the model parameters by minimizing the loss function. This multi-task loss function can take into account the performance of multiple tasks at the same time, so that the model can optimize the performance of multiple tasks at the same time during the training process, thus making the model have better generalization ability and higher performance;

[0037] L=L location +λL size

[0038] Variable L location and L size They represent the losses used to optimize the damage location recognition ability of the backbone network and the damage size recognition ability of the attention sub-network, respectively, and are used to measure the performance of the model in locating the damage location and estimating the damage size;

[0039] The parameter λ is a weight factor used to adjust L location and L size The relative importance in the total loss, λ can be set to 1;

[0040] L location and L size Cross entropy loss is used as the loss function. The specific calculation formula is shown as follows:

[0041]

[0042] Among them, 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 in category C (usually 0 or 1, indicating whether it belongs to the category), It represents the prediction result of the model for sample i in category C, which means the probability that the sample belongs to category C; N is the number of samples, and C is the total number of categories;

[0043] For each category, as the model predicts that the probability of an instance belonging to that category increases, the corresponding loss decreases accordingly; this means that if the actual label y i,c is 1, that is, the instance does belong to this category, then the loss function will encourage the model to increase the probability of predicting this category as the correct label; conversely, if the actual label y i,c If it is 0, that is, the instance does not belong to this category, the loss function will encourage the model to reduce the probability of predicting this category as the correct label.

[0044] Compared with the prior art, the technical effects and advantages of the present invention are:

[0045] 1. The backbone network of the present invention 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 results of the damage location and share information with the attention sub-network. The output features generated by each convolution layer are shared, and these features contain rich details of the vibration signal and important information about the damage location. The attention sub-network uses these shared features obtained from the backbone network to accurately evaluate the damage size.

[0046] 2. The present invention adopts an asymmetric multi-task learning framework and a position attention mechanism, which can give full play to the advantages of multi-task learning methods in improving the performance of damage location and damage size identification tasks, while also minimizing the adverse effects of negative transfer phenomena, thereby making the identification of damage location and size more accurate. The attention subnetwork of the present invention includes a series of cascaded attention modules, which can extract important information related to the damage size from the damage location features obtained from the backbone network, thereby facilitating the accurate identification of the damage size.

[0047] 3. The design of the cross entropy loss function adopted by the present invention not only focuses on the model's prediction ability for positive samples of each category, but also focuses on its prediction ability for negative samples. This method ensures that the model can show excellent prediction performance on all possible categories. By minimizing this loss function, the model can more effectively learn and distinguish the features of different categories, thereby improving its overall classification performance.

[0048] 4. The present invention can adaptively assign weights to the damage location feature and the damage size feature. After weighted summing the two features, the damage size feature containing rich location information can be extracted, thereby improving the effect of the damage size identification task. Experimental results show that the present invention has a significant improvement in accurately identifying the damage size, which is superior to other methods. BRIEF DESCRIPTION OF THE DRAWINGS

[0049] Figure 1 Schematic diagram of the overall system architecture of an embodiment of the present invention.

[0050] Figure 2 This is a schematic diagram of the backbone network structure.

[0051] Figure 3 Schematic diagram of the structure of the attention module.

[0052] Figure 4 Description of the task data. DETAILED DESCRIPTION

[0053] In order to make the purpose, technical solution and advantages of the present invention more clearly understood, the following embodiments will further illustrate the present invention in conjunction with the accompanying drawings. It should be understood that the specific embodiments described herein are only used to explain the present invention and are not used to limit the present invention.

[0054] In this 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 , the number of sampling points is N. The response recorded by each sensor can be expressed 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 ​​together to form a matrix In this matrix, s i ∈R N×1 represents the vibration response recorded by the i-th sensor. In this way, the data of multiple sensors can be effectively integrated for further analysis and processing. Considering that the damage has different locations and sizes, define N l damage locations and N s A model F is designed and trained. The purpose of this model is to determine and output the specific damage location by analyzing the input vibration response matrix S. and damage size Therefore, the functionality of the model can be expressed as:

[0055] The present invention proposes an asymmetric multi-task structural damage identification system, whose backbone network consists of four cascaded convolutional modules (CM), which are used to extract valuable features from the input data. Then, through a soft max layer, the classification result of the damage location is generated. In addition to the backbone network, an attention sub-network is constructed. The task of this network is to further learn the size features of the damage from the extracted position features. In this attention sub-network, the cascaded attention modules (AM) are connected to the corresponding convolution modules in the backbone network. These attention modules use adaptive attention masks to extract the features of the damage size. For the identification of damage location and damage size, the model adopts a cross entropy loss function. This loss function can optimize the model parameters to accurately estimate the location and size of the damage. Experimental results show that this method effectively retains the advantages of multi-task learning while reducing the adverse effects of the negative transfer phenomenon. It performs very well in the tasks of identifying damage location and damage size. 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 the damage from the input data, thereby improving the accuracy of damage detection.

[0056] 1. Backbone network for identifying damage locations

[0057] In the structural damage detection method proposed in the present invention, the damage location must be accurately identified first, because accurately identifying the damage location helps to improve the performance of damage size identification. Since the one-dimensional convolutional network is very suitable for analyzing and processing sequence data, such as audio signals and time series, it is used as a backbone network in the present invention 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 convolution filter along the input sequence while performing element-by-element multiplication on the elements covered by the filter and accumulating the results to generate a new sequence. This process can be represented by the mathematical formula of the convolutional layer, as shown below.

[0058]

[0059] Among them, 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 transferred to the pooling layer, the main task of which is to reduce the dimension of the features. This step is crucial because it not only reduces the dimension of the data, but also highlights the most critical features, laying the foundation for subsequent processing.

[0060] The backbone network is constructed from four closely connected convolutional modules, such as Figure 2As shown in the figure, each convolution module contains one convolution layer and one pooling layer. The number of output channels and convolution kernel size of each module are marked in the figure.

[0061] The function of the backbone network is not limited to outputting the classification results of the damage location. It also plays another key role: sharing information with the attention sub-network. The output features generated by each convolutional layer are shared, and these features contain rich details of the vibration signal and important information about the damage location. The attention sub-network uses these shared features obtained from the backbone network to accurately assess the damage size.

[0062] 2. Attention subnetwork for identifying lesion size

[0063] Inspired by the attention mechanism, this paper introduces the attention mechanism and constructs an attention sub-network. The main goal of this sub-network is to adaptively learn and extract the features of the damage size from the damage location features obtained from the backbone network. Figure 1 As shown in Figure 1, the attention subnetwork is composed of a series of cascaded attention modules. Each module is connected to the corresponding convolution module in the backbone network and is also cascaded with the output of the previous attention module. This design allows the attention subnetwork to transfer and share information between attention modules at all levels, thereby performing feature learning and optimization on a global scale.

[0064] like Figure 3 As shown in Figure 1, the attention module accepts two inputs: one is the output of the previous attention module, and the other is the damage location feature from the corresponding convolution module in the backbone network. The attention module mainly consists of two parts: mask generator and feature extractor.

[0065] The core of the weight generator (also called mask generator) is two cascaded convolutional layers. The main task of these two convolutional layers is to learn and generate an adaptive attention mask A from two inputs. n The purpose of this mask is to weight the input features so that the model can focus on the most important information. n The calculation process is as follows:

[0066]

[0067] Among them, F n represents the mask generator of the nth attention module, Y n-1 is the output of the previous attention module, X n is the output of the corresponding convolution module. Symbol Represents the concatenation of two inputs in the channel dimension. Obtain the adaptive attention mask A n Then, you can compare it with the input value Perform element-by-element multiplication. This results in a new feature in which each feature dimension is weighted by the corresponding attention weight. Next, the feature extractor G n The feature extractor G is used to extract the damage size from the weighted feature. n A combination of a one-dimensional convolutional layer of size 3 and a one-dimensional pooling layer of size 2 is used.

[0068]

[0069] in, The symbol represents an element-wise multiplication operation. The first attention module accepts two inputs. The first input, Y0, comes from the output of the last convolutional layer in the backbone network. This output contains high-level features after multiple layers of convolution and nonlinear transformation. The second input, X1, comes from the first convolutional block of the backbone network. This output retains some basic features of the original data. The combination of these two inputs not only retains the intuitive features of the original data, but also contains high-level features after network processing, which can provide rich information and help build a more powerful attention sub-network.

[0070] 3. Loss Function

[0071] In the training phase of the model, a multi-task loss function is used to optimize the model parameters by minimizing the loss function. This multi-task loss function can take into account the performance of multiple tasks at the same time, so that the model can optimize the performance of multiple tasks at the same time during the training process, thus making the model have better generalization ability and higher performance.

[0072] L=L location +λL size

[0073] In the model, the variable L location and L size They represent the losses used to optimize the damage location recognition ability of the backbone network and the damage size recognition ability of the attention sub-network. These two loss functions are key indicators in the training process and are used to measure the performance of the model in locating the damage location and estimating the damage size.

[0074] The parameter λ is a weight factor used to adjust L location and L size The relative importance of λ in the total loss. In this embodiment, λ is set to 1.

[0075] 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 models and also provides gradient information for the optimization of model parameters. The specific calculation formula is shown below:

[0076]

[0077] Among them, 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 in category C (usually 0 or 1, indicating whether it belongs to the category), It represents the prediction result of the model for sample i in category C, which means the probability that the sample belongs to category C. N is the number of samples, and C is the total number of categories.

[0078] In the model of the present invention, for each category, when the probability that the model predicts that an instance belongs to that category increases, the corresponding loss will decrease accordingly. This means that if the actual label y i,c is 1, that is, the instance does belong to this category, then the loss function will encourage the model to increase the probability of predicting this category as the correct label. On the contrary, if the actual label y i,c If it is 0, that is, the instance does not belong to this category, the loss function will encourage the model to reduce the probability of predicting this category as the correct label.

[0079] In short, the design of the cross entropy loss function focuses not only on the model's ability to predict positive samples of each category, but also on its ability to predict negative samples. This approach ensures that the model can show excellent prediction performance on all possible categories. By minimizing this loss function, the model can more effectively learn and distinguish the features of different categories, thereby improving its overall classification performance.

[0080] In order to more clearly illustrate the process and advantages of the method of the present invention, a specific example is given below.

[0081] Lamb waves are ultrasonic guided waves that propagate in plate-like structures. They are often used in structural health detection and damage detection because they can provide important information about potential structural damage. The present invention collects Lamb wave data as training and test data for the model to verify the effectiveness of the proposed method.

[0082] The experimental setup for data collection is as follows:

[0083] First, the hotspot ultrasonic guided wave damage diagnosis system developed by Joyson Technology was used to excite and receive signals. One side of the diagnosis system was connected to four PZT sensors. During the process of collecting Lamb wave data, PZT-1 was used as an excitation source to send out excitation signals, while the remaining sensors (PZT-2, PZT-3, and PZT-4) were responsible for recording the corresponding Lamb wave signals. The other side was connected to a laptop computer to display and save data.

[0084] Figure 4 Displays the task data description.

[0085] A square of 160 mm × 160 mm in size is set as the damaged area, which contains 16 sub-areas of 40 mm × 40 mm in size. Different sub-areas show different damage locations, and the locations are numbered L1-L16. The damage in real structures is usually diverse and complex, and it is difficult and expensive to obtain a real damage data set for the structure. Therefore, a solid tape is attached to the surface of the sample to block the propagation of the wave, which has a similar effect to the actual damage. In this embodiment, the solid tape is 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 are used to simulate damage of different sizes in composite laminates. A 500-gram weight is placed on the tape instead of manually pressing to ensure that the force is evenly applied. In this embodiment, four piezoelectric sensors (PZT-1, PZT-2, PZT-3, PZT-4) with a diameter of 6 mm and a thickness of 0.25 mm are used to generate and collect Lamb wave data. To ensure that the damaged area is adequately covered, four PZT sensors are arranged in a square array with a distance of 180 mm between each sensor.

[0086] The network training process is as follows:

[0087] First, the training data comes from the one-dimensional time series data collected from the three sensors PZT-2, PZT-3, and PZT-4. The excitation signal is a five-cycle sine pulse with a center frequency of 125kHz, the sensor signal sampling frequency is 24MHz, and the number of sampling points is 8000. Therefore, the three sensor signals used as receiving signals are used as input training data, each containing 8000 sampling points. These three time series data are spliced ​​in a time-aligned form to form an 8000×3 matrix sequence.

[0088] 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 series. At the same time, the features after each layer of convolution will be retained as the input of the attention sub-network. 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 sub-network. The output of the backbone network and the output of the first layer of convolution are taken as input. The attention sub-network first generates a mask, then weights the input features, and then uses the convolution layer to extract features after the weighting. The final feature extraction result is used as an output of the attention layer and as the input of the next attention layer. There are three layers of such attention layers. After three layers, the final damage size feature is output. After being processed by the backbone network and the attention sub-network, the model obtains the final damage location feature and damage size feature respectively. Both features must be passed through the flattening layer and the linear layer to make a final prediction. This method uses cross entropy loss as the loss function. The total loss is equal to the sum of the loss of the damage location and the loss of the damage size. The reverse gradient descent is based on the total loss.

[0089] In order to fully evaluate the effect of the present invention, experiments were conducted on the Lamb wave data set. Single-task learning, multi-task learning and the method of the present invention were used to conduct experiments on the test set, and the experimental results are shown in Tables 1 to 3 respectively. 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 terms of identifying the size of the damage, the accuracy of the single-task learning method reached 87.9%, the accuracy of the multi-task learning method reached 90.7%, and the accuracy of the method proposed in the present invention reached 93.5%. In the presence of three damages, all methods can accurately identify the location of the damage. In terms of identifying the size of the damage, the accuracy of the single-task learning method reached 89.3%, the accuracy of the multi-task learning method reached 92.7%, and the accuracy of the method proposed in the present invention reached 97.4%.

[0090] Table 1

[0091]

[0092] Table 2

[0093]

[0094] Table 3

[0095]

[0096] As can be seen from Tables 1 to 3, the embodiments of the present invention provide an innovative damage detection method. The method adopts an asymmetric multi-task learning framework and a position attention mechanism, which can give full play to the advantages of the multi-task learning method in improving the performance of damage location and damage size recognition tasks, and can also minimize the adverse effects of the negative transfer phenomenon, thereby making the identification of damage location and size more accurate. The method mainly includes two parts: first, the location of the damage is determined by using a backbone network based on a one-dimensional convolutional neural network; second, the size of the damage is identified by combining the damage location features with a special position attention sub-network. The attention sub-network contains a series of cascaded attention modules, which can extract important information related to the damage size from the damage location 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 were conducted on the Lamb wave dataset. The experimental results show that the method has a significant improvement in accurately identifying the size of the damage, which is better than other methods.

[0097] The above embodiments are only preferred embodiments of the present invention and cannot be considered to limit the scope of the present invention. All equivalent changes and improvements made within the scope of the present invention should still fall within the scope of the present invention.

Claims

1. Structural damage detection system based on asymmetric multi-task learning, characterized by Based on an asymmetric multi-task learning framework, it is used to accurately locate and size structural damage by integrating multiple sub-networks; the overall structure of the system includes two core parts, namely the backbone network and the attention sub-network; The backbone network is responsible for extracting and identifying features of damage locations from input vibration response data; The backbone network consists of four cascaded convolutional modules, each of which contains a convolutional layer and a pooling layer. The four cascaded convolutional modules are cascaded in sequence, and the output of each module is used as the input of the next module. Inside 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 pass through a soft max layer to generate the classification results of the damage location. 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 consists of multiple cascaded attention modules. Each attention module uses an adaptive attention mask to extract the features of the lesion size. Each attention module is connected to the corresponding convolution module in the backbone network and is also cascaded with the output of the previous attention module. This design enables the attention sub-network to transfer and share information between attention modules at all levels, thereby performing feature learning and optimization on a global scale. The system input is a vibration response matrix, which is formed by aligning and splicing the vibration response data recorded by multiple sensors along the time axis; the backbone network is connected to the attention sub-network, and 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 method ensures that the attention sub-network can fully utilize the damage location features extracted by the backbone network; The damage location classification result is generated by the softmax layer in the backbone network, which represents the damage location classification corresponding to the input sample; The damage size assessment result is output by the last attention module in the attention sub-network, which represents the damage size assessment value corresponding to the input sample.

2. The structural damage detection system based on asymmetric multi-task learning as claimed in claim 1, characterized in that 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 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 about the size of the damage from the weighted features.

3. The structural damage detection system based on asymmetric multi-task learning as claimed in claim 1, characterized in that Inside each attention module, an adaptive attention mask is generated by a mask generator, and then the mask is multiplied element-by-element with the input feature to obtain the weighted feature, which is passed to the feature extractor for further processing.

4. Structural damage detection method based on asymmetric multi-task learning, characterized by The following steps are involved: 1) Data preprocessing: align the vibration response data recorded by multiple sensors along the time axis and splice them into a vibration response matrix; 2) Backbone network processing: The vibration response matrix is ​​input into the backbone network and passes through four convolution modules in sequence. Each convolution module extracts features at different levels. The output of the last convolution module generates the damage location classification result through the soft max layer. 3) Attention sub-network processing: Use the output of each convolution module in the backbone network as input to build an attention sub-network. Each attention module generates an adaptive attention mask to perform weighted processing on the input features. The weighted features are further processed by a feature extractor to generate damage size assessment results; 4) Multi-task loss optimization: The cross entropy loss function is used to calculate the losses 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.

5. The structural damage detection method based on asymmetric multi-task learning as claimed in claim 4, characterized in that In step 1), the data is preprocessed by using k sensors to record the vibration response of the structure. During the collection of the vibration response data, the sampling frequency is f s , the number of sampling points is N; the response recorded by each sensor is expressed as Among them, s i，t is the data recorded by the i-th sensor at time step t; to integrate these data, the vibration responses recorded by all sensors are aligned along the time axis and spliced ​​together to form a vibration response matrix In this matrix, s i ∈R N×1 represents the vibration response recorded by the i-th sensor.

6. The structural damage detection method based on asymmetric multi-task learning as claimed in claim 1, characterized in that In step 2), the specific steps of the backbone network processing are: 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 convolution layer. This process involves sliding the convolution filter along the input sequence, performing element-by-element 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 convolution layer, as shown below: 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 transferred to the pooling layer, which reduces the dimensionality of the features. The backbone network is built from four closely connected convolutional modules. Each convolutional module contains a convolutional layer and a pooling layer. The output features generated by each convolutional layer are shared. These features contain rich details of the vibration signal and important information about the damage location.

7. The structural damage detection method based on asymmetric multi-task learning as claimed in claim 4, characterized in that In step 3), the specific steps of the attention sub-network processing are: 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 transmits and shares information between attention modules at all levels, thereby learning and optimizing features globally; The attention module accepts two inputs: one is the output from the previous attention module, and the other is the damage location feature from the corresponding convolution module in the backbone network. The attention module mainly consists of two parts: mask generator and feature extractor; The core of the mask generator is two cascaded convolutional layers, whose main task is to learn and generate an adaptive attention mask A from two inputs. n ; The role of this mask 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 represents the mask generator of the nth attention module, Y n-1 is the output of the previous attention module, X n is the output of the corresponding convolution module; the symbol Represents the concatenation of two inputs in the channel dimension direction; obtains the adaptive attention mask A n Then, compare it with the input value Perform element-by-element multiplication; obtain a new feature in which each feature dimension is weighted by the corresponding attention weight; next, the feature extractor G n Features about the size of the damage will be extracted from this weighted feature; to achieve this goal, 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 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 layers of convolution and nonlinear transformation; the second input comes from the first convolutional block of the backbone network, which retains some basic features of the original data; the combination of these two inputs not only retains the intuitive features of the original data, but also contains high-level features after network processing, thereby providing rich information and helping to build a more powerful attention sub-network.

8. The structural damage detection method based on asymmetric multi-task learning as claimed in claim 4, characterized in that In step 4), the specific steps of the multi-task loss optimization are: A multi-task loss function is used to optimize the model parameters by minimizing the loss function. This multi-task loss function can take into account the performance of multiple tasks at the same time, so that the model can optimize the performance of multiple tasks at the same time during the training process, thus making the model have better generalization ability and higher performance; L=L location +λL size Variable L location and L size They represent the losses used to optimize the damage location recognition ability of the backbone network and the damage size recognition ability of the attention sub-network, respectively, and are used to measure the performance of the model in locating the damage location and estimating the damage size; The parameter λ is a weight factor used to adjust L location and L size Relative importance in total losses; L location and L size Cross entropy loss is used as the loss function. The specific calculation formula is shown as follows: Among them, 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 in category C, It represents the prediction result of the model for sample i in category C, which means 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 that the probability of an instance belonging to that category increases, the corresponding loss decreases accordingly; this means that if the actual label y i,c is 1, that is, the instance does belong to this category, then the loss function will encourage the model to increase the probability of predicting this category as the correct label; conversely, if the actual label y i,c If it is 0, that is, the instance does not belong to this category, the loss function will encourage the model to reduce the probability of predicting this category as the correct label.

Images