Structural damage identification method and system based on improved channel attention mechanism and one-dimensional convolutional neural network

Abstract

The invention belongs to the field of structural damage identification, and discloses a structural damage identification method and system based on an improved channel attention mechanism and a one-dimensional convolutional neural network, and the method specifically comprises the steps: building an experimental model of a lossless structure and different damage position structures; applying excitation to the to-be-measured structure, measuring acceleration time history data of each data acquisition point, and performing data preprocessing; constructing a data set, and dividing the data set into a training set and a test set; constructing a neural network model based on an improved channel attention mechanism and a one-dimensional convolutional neural network, and training by using the training set; and inputting the data of the test set into the trained convolutional neural network model for calculation to obtain a structural damage identification result. According to the invention, the structure damage position can be accurately identified, and the importance degree of different sensor channels can be quantified.

Description

Technical Field

[0001] The present invention belongs to the field of structural damage identification, and particularly relates to a structural damage identification method and system based on an improved channel attention mechanism and a one-dimensional convolutional neural network. Background Art

[0002] In recent years, people have gradually paid more attention to the health detection of structures. Compared with traditional static non-destructive testing methods (such as acoustic or ultrasonic waves, thermal imaging, ray method, pulse echo method, Lamb wave method, etc.), vibration-based structural health detection has significant advantages. This method does not require prior knowledge of the approximate location of structural damage, and at the same time has the characteristics of low detection cost and does not affect the normal use of the structure, making it a research hotspot in the field of structural damage identification.

[0003] The damage of a structure is mainly manifested in the changes of its physical parameters (such as stiffness, mass, damping), and the physical parameters of the structure are directly related to its measured dynamic response. After damage occurs, the change of the physical parameters of the structure will cause a certain change in its dynamic response. The vibration-based identification method precisely utilizes this characteristic to judge the existence, location and severity of structural damage by analyzing the dynamic response of the structure.

[0004] Deep learning algorithms play an increasingly important role in the field of structural health monitoring with their efficient computing power and strong generalization performance. Compared with traditional machine learning methods, deep learning methods can automatically learn abstract features from input signals without manually selecting features through iterative training. The vibration signals measured in actual engineering contain a large amount of noise, and traditional damage identification methods may not be able to detect damage information due to this. At the same time, deep learning algorithms have high robustness and have significant advantages in anti-noise ability, making them more suitable for actual engineering applications.

[0005] For the response data of structural health monitoring, such as displacement, acceleration, stress and strain, etc., are mostly one-dimensional time series. One-dimensional convolutional neural network (1D-CNN) can directly process these raw signals without additional dimensional conversion, and is particularly suitable for analyzing one-dimensional time series data collected by sensors. Compared with two-dimensional convolutional neural network (2D-CNN), 1D-CNN significantly reduces the complexity of the model and improves the computing efficiency.

[0006] Since the importance of damage information contained in different positions of a structure is different, during the training process, it is necessary to focus on the extraction of features of key parts. By introducing the channel attention mechanism (SENet), the convolutional neural network (CNN) can focus on the sensor channels that are most important for damage identification, thereby enhancing the network's ability to extract damage information and further improving the recognition efficiency and accuracy.

[0007] At present, certain progress has been made in the research on structural damage identification based on CNN, but there are still limitations, which restrict its wide application in the field of structural damage identification. First, in multi-sensor signals, there are significant differences in the contributions of different sensors to damage information. However, existing methods uniformly process all channels, resulting in the model failing to effectively focus on the features of key sensors. Second, the embedding position of the traditional SENet module is usually relatively backward, so it is difficult to fully capture the importance differences between channels. Its feature extraction method uses global average pooling, which is insensitive to positive and negative characteristics, and the Sigmoid activation function easily leads to over-concentration of weights, weakening the discrimination ability of channel weights. In addition, the computational efficiency of 2D-CNN in processing time series data is low, which also limits the application performance. Summary of the Invention

[0008] To solve the problems existing in the prior art, the present invention provides a structural damage identification method and system based on an improved channel attention mechanism and a one-dimensional convolutional neural network, which can accurately identify the structural damage position and quantify the importance of different sensor channels.

[0009] To achieve the above object, the present invention provides the following solutions:

[0010] A structural damage identification method based on an improved channel attention mechanism and a one-dimensional convolutional neural network, the method includes:

[0011] Establish experimental models of undamaged structures and structures with different damage positions;

[0012] Apply excitation to the structure to be measured, measure the acceleration time history data of each data acquisition point and transmit it to a multi-channel sensor;

[0013] Slice the multi-channel sensor data into acceleration time series segments of a fixed length;

[0014] Construct a data set based on the acceleration time series segments and divide the data set into a training set and a test set;

[0015] Construct a convolutional neural network model based on the improved channel attention mechanism and a one-dimensional convolutional neural network, and use the training set for training;

[0016] Input the data of the test set into the trained convolutional neural network model for calculation to obtain the structural damage identification result.

[0017] Preferably, the acceleration time history data received by the sensor is , where represents the acceleration time history data received by the c-th sensor, and the total number of acceleration sensors is C.

[0018] Preferably, the improved channel attention mechanism includes:

[0019] S41: Improve the insertion position of the channel attention mechanism module from after the convolution operation to before the convolution operation, use the channel attention mechanism to weight the C sensor data, and then judge the importance of the data at different measurement points;

[0020] S42: Improve the global average pooling to global max pooling;

[0021] S43: Improve the compression of the number of neurons in the middle layer to appropriate expansion;

[0022] S44: Improve the activation function to the PReLU function.

[0023] Preferably, the convolutional neural network model includes: an improved channel attention mechanism module, a first convolutional layer, a first max pooling layer, a second convolutional layer, a second max pooling layer, a global average pooling layer, and an output layer connected in sequence.

[0024] Preferably, the loss function adopted in the convolutional neural network model is the cross-entropy loss function, and the formula is:

[0025]

[0026] In the formula, represents the loss function, represents the number of categories, represents the sign function. If the true category of the sample is equal to take 1, otherwise take 0; represents the observed sample belongs to the category the predicted probability of, is the number of samples.

[0027] The present invention also provides a structural damage identification system based on an improved channel attention mechanism and a one-dimensional convolutional neural network. The system is used to implement any one of the above methods. The system includes: an experimental model construction module, a data measurement module, a data segmentation module, a data set construction module, a network model construction module, and a damage calculation module;

[0028] The experimental model construction module is used to establish experimental models of non-damaged structures and structures with different damage positions;

[0029] The data measurement module is used to apply an excitation to the structure to be measured, measure the acceleration time history data of each data acquisition point, and transmit it to the multi-channel sensor;

[0030] The data segmentation module is used to segment the multi-channel sensor data into acceleration time series segments with a fixed length;

[0031] The dataset construction module is used to construct a dataset based on the acceleration time series segments and divide the dataset into a training set and a test set;

[0032] The network model construction module is used to construct a convolutional neural network model based on an improved channel attention mechanism and a one-dimensional convolutional neural network, and train it using the training set;

[0033] The damage calculation module is used to input the data of the test set into the trained convolutional neural network model for calculation to obtain the structural damage identification result.

[0034] Preferably, the acceleration time history data received by the sensor is where represents the acceleration time history data received by the c-th sensor, and the total number of acceleration sensors is C.

[0035] Preferably, the improved channel attention mechanism includes:

[0036] S41: Change the insertion position of the channel attention mechanism module from after the convolution operation to before the convolution operation, use the channel attention mechanism to weight the data of C sensors, and then judge the importance of the data at different measuring points;

[0037] S42: Change the global average pooling to global max pooling;

[0038] S43: Change the compression of the number of neurons in the middle layer to appropriate expansion;

[0039] S44: Change the activation function to the PReLU function.

[0040] Preferably, the convolutional neural network model includes: an improved channel attention mechanism module, a first convolutional layer, a first max pooling layer, a second convolutional layer, a second max pooling layer, a global average pooling layer, and an output layer connected in sequence.

[0041] Preferably, the loss function adopted in the convolutional neural network model is the cross-entropy loss function, and the formula is:

[0042]

[0043] In the formula, represents the loss function, represents the number of categories, represents the sign function. If the true category of the sample is equal to take 1, otherwise take 0; represents the observed sample belongs to the category of the predicted probability, is the number of samples.

[0044] Compared with the prior art, the beneficial effects of the present invention are as follows:

[0045] By optimizing the embedding position of the channel attention mechanism module, the present invention significantly improves the model's attention ability to key sensor channels, enhances the CNN's ability to extract structural damage information, and has higher recognition efficiency and accuracy;

[0046] The present invention can better extract the significant features of acceleration signals through global max pooling, and uses the PReLU activation function to enhance the discrimination ability of channel weight distribution;

[0047] The present invention weights multi-channel sensors by improving the channel attention mechanism module, and visualizes the weights of sensor channels, providing a reliable basis for optimizing sensor layout. BRIEF DESCRIPTION OF THE DRAWINGS

[0048] In order to more clearly illustrate the technical solutions of the present invention, the accompanying drawings required for use in the embodiments will be briefly introduced below. Obviously, the accompanying drawings in the following description are only some embodiments of the present invention. For those of ordinary skill in the art, other drawings can be obtained based on these drawings without creative efforts.

[0049] Figure 1 is an exemplary method flow chart of a structural damage identification method based on an improved channel attention mechanism and a one-dimensional convolutional neural network in an embodiment of the present invention;

[0050] Figure 2 is a layout diagram of measuring points of a steel frame structure in an embodiment of the present invention;

[0051] Figure 3 is a curve graph of the accuracy before the improvement of the channel attention mechanism in an embodiment of the present invention;

[0052] Figure 4 is a curve graph of the loss before the improvement of the channel attention mechanism in an embodiment of the present invention;

[0053] Figure 5 is a curve graph of the accuracy after the improvement of the channel attention mechanism in an embodiment of the present invention;

[0054] Figure 6 is a curve graph of the loss after the improvement of the channel attention mechanism in an embodiment of the present invention;

[0055] Figure 7 is a visualization graph of the weights of 13-channel sensors in an embodiment of the present invention;

[0056] Figure 8The result graph of selecting the CH.1 sensor data to train the model in the embodiment of the present invention;

[0057] Figure 9 The result graph of selecting the CH.1 and 3 sensor data to train the model in the embodiment of the present invention;

[0058] Figure 10 The result graph of selecting the CH.5, 8, 11, and 13 sensor data to train the model in the embodiment of the present invention;

[0059] Figure 11 The result graph of selecting the CH.1, 3, 6, and 9 sensor data to train the model in the embodiment of the present invention. Detailed implementation manners

[0060] Next, the technical solutions in the embodiments of the present invention will be clearly and completely described in conjunction with the accompanying drawings in the embodiments of the present invention. Obviously, the described embodiments are only a part of the embodiments of the present invention, rather than all of the embodiments. All other embodiments obtained by those of ordinary skill in the art based on the embodiments of the present invention without creative efforts shall fall within the protection scope of the present invention.

[0061] To make the above objects, features, and advantages of the present invention more obvious and understandable, the present invention will be further described in detail below in conjunction with the accompanying drawings and specific implementation manners.

[0062] Embodiment 1

[0063] The present invention discloses a structural damage identification method based on an improved channel attention mechanism and a one-dimensional convolutional neural network, including the following contents:

[0064] S1 Model establishment: Establish experimental models of undamaged structures and structures with different damage positions;

[0065] S2 Data sampling: Apply excitation to the structure to be measured, and measure the acceleration time history data of each data acquisition point;

[0066] S3 Data preprocessing: Cut the multi-channel sensor data into acceleration time series segments with a fixed length;

[0067] S4 Create a data set: Construct a data set based on the acceleration data samples obtained in S3, and divide the data set into a training set and a test set;

[0068] S5 Improve the channel attention mechanism: To efficiently extract the damage information of the structure, improve the channel attention mechanism module;

[0069] S6 Construct a neural network model: Embed the improved channel attention mechanism module into a one-dimensional CNN model, and use the training set in S4 to train the neural network model;

[0070] S7 recognition result verification: substitute the data of the test set in S4 into the trained neural network model in S6 to obtain the prediction result; if the damage recognition accuracy of the test set is higher than the set threshold, save the training parameters to obtain a neural network model that can be applied to structural damage recognition; otherwise, adjust the parameters of the neural network model and continue training;

[0071] S8 Structural damage identification: Structural damage identification results are obtained based on the neural network model in S7.

[0072] Furthermore, in step S1, the lossy structure is a lossless structure and its damage is simplified by using an added mass method.

[0073] Further, in step S2, the acceleration time history data received by the sensor is ,in Represents the acceleration time history data received by the cth sensor, and the total number of acceleration sensors is C.

[0074] Furthermore, in step S3, the acceleration data collected by each sensor is processed, invalid data is removed, and valid acceleration data is cut into data sequences of equal length.

[0075] Furthermore, in step S4, the channel attention mechanism module is improved as follows:

[0076] S41 Embedding position: The conventional embedding position of the channel attention mechanism module is after the convolution operation. This approach combines all sensor channels and fails to consider the importance of sensor channels at different measurement points. However, the importance of damage information contained in different positions of the structure is often different. Therefore, the present invention improves the insertion position of the channel attention mechanism module from after the convolution operation to before the convolution operation, and uses the channel attention mechanism to weight the C sensor data, thereby judging the importance of data at different measurement points;

[0077] S42 Feature extraction method: In the original channel attention mechanism, global average pooling is used to process the data. However, the acceleration data has both positive and negative values. Global average pooling cannot effectively extract the features of the data. Therefore, global average pooling is improved to global maximum pooling.

[0078] S43 Channel Expansion: Instead of compressing the number of intermediate-layer neurons, it is improved to appropriate expansion. SENet is located after the convolutional operation in computer vision applications. At this time, it often faces several hundred or even more than 1000 feature channels. Therefore, the number of neurons in the first fully connected layer is relatively small, which plays a role in controlling the model complexity. However, this may also lead to the loss of some effective information. In the application of structural damage identification, the number of sensor channels is often small. Therefore, it is proposed to replace the operation of compressing the number of neurons in SENet with expanding to the nearest power of 2, which can not only ensure the integrity of information but also control the complexity of the model;

[0079] S44 Activation Function: The activation function used in the original channel attention mechanism is Sigmoid. However, when using the Sigmoid function in this problem, the neural network model will converge the weights of the vast majority of channels to 1 after training, which is not conducive to comparing the weights of different sensor channels. Therefore, the activation function is improved to the PReLU function, which is more conducive to linear expansion.

[0080] The formula derivation of the improved attention mechanism is as follows:

[0081] Input Tensor: (Batch size is B, the number of sensor channels is C, and the sample time length is L)

[0082] 1. Global Max Pooling

[0083] Take the maximum value of the time dimension for each sensor channel:

[0084]

[0085] Output Tensor:

[0086] 2. Neuron Number Expansion Strategy

[0087] Neuron Dimension Conversion Formula:

[0088]

[0089] Where E is the nearest power of 2 greater than C.

[0090] 3. MLP Weight Generation

[0091] Generate channel weights through a fully connected layer and a non-linear activation:

[0092]

[0093]

[0094]

[0095] Among them, is the first-layer weight matrix (channel expansion), is the first-layer weight matrix (channel restoration).

[0096] Output weight tensor:

[0097] 4. Feature Recalibration

[0098] Broadcast the weights to the time dimension and multiply them with the input channel by channel:

[0099]

[0100] Output tensor:

[0101] Furthermore, in step S5, the neural network model includes an improved channel attention mechanism module, a first convolutional layer, a first max pooling layer, a second convolutional layer, a second max pooling layer, a global average pooling layer, and an output layer connected in sequence.

[0102] The model structure formula is as follows:

[0103] Attention enhancement:

[0104] First-layer convolution and pooling:

[0105] Second-layer convolution and pooling:

[0106] Global average pooling layer:

[0107] Output layer:

[0108] The overall forward propagation formula of the model is:

[0109]

[0110] Furthermore, the loss function adopted in the neural network model is the cross-entropy loss function, and the formula is as follows:

[0111]

[0112] In the formula, represents the loss function, represents the number of categories, represents the sign function. If the sample 's true category is equal to Take 1, otherwise take 0; Indicates the observed sample Belongs to the category Of the predicted probability, Is the number of samples.

[0113] In the technical solution of the present invention, by designing an improved channel attention mechanism, the embedding position, feature extraction method and activation function of the channel attention mechanism are optimized. First, adjust the embedding position of the channel attention mechanism so that it can generate independent channel weights for each sensor, improving the ability to focus on important channels; second, improve the global average pooling to global maximum pooling to better extract the significant features of the acceleration signal; finally, use the PReLU activation function instead of the Sigmoid to avoid the weights converging to the same value, thereby enhancing the difference in weight distribution between sensors. Input the multi-channel acceleration data into the improved channel attention mechanism module, which can perform weighted processing on the multi-channel sensor data, enabling the convolutional neural network to focus on the sensor channels it most needs to pay attention to, enhancing the ability of the convolutional neural network to extract structural damage information, and thus improving the recognition efficiency and accuracy. The present invention explores and visualizes the weights of the sensor channels inside the model, which can provide a reliable reference for the optimal layout of sensors.

[0114] Embodiment 2

[0115] As Figure 1 Shown, the present invention proposes a structural damage identification method based on a channel attention mechanism and a one-dimensional convolutional neural network. The process of structural damage identification of an asymmetric steel frame structure based on this method is as follows:

[0116] Experimental condition introduction:

[0117] The experimental model of this embodiment is a 3-story 2-span asymmetric steel frame structure in the laboratory. This asymmetric steel frame structure is composed of 13 identical members, each member is 500 mm long, and the cross-section is Of the rectangular cross-section size. The initial elastic modulus of the steel used for the structural members is 200 Gpa, and the measured density is 7700 .

[0118] The working condition settings include a non-damaged working condition and 23 damaged working conditions. The non-damaged working condition corresponds to the complete state of the structure, while the damaged working conditions are achieved by attaching mass blocks in the middle of the structural members to simulate different damage positions. The specific details of the working condition settings are shown in Table 1:

[0119] Table 1 Details of working condition settings

[0120]

[0121] Data sampling:

[0122] An acceleration sensor is arranged in the middle of each member of the steel frame to collect the dynamic response signals of the structure under excitation. The data acquisition uses the Donghua DH5922D data acquisition instrument, and the sampling frequency is set to 1000 Hz. The layout of the structural measurement points and the positions of the impact hammer strikes are as Figure 2 shown. In the experiment, an impact hammer is used to apply excitation at specific positions, and the acceleration responses of each measurement point are measured in real time through the acceleration sensors. The acceleration signals collected by the sensors are transmitted to the data acquisition instrument through data transmission lines for recording.

[0123] Data preprocessing:

[0124] To ensure the data quality and the validity of the model input, the original acceleration data collected in the experiment needs to go through the following processing steps:

[0125] Rejection of abnormal data: Check the integrity and validity of the collected acceleration data, and reject the invalid data with noise interference, missing values, or abnormal fluctuations to ensure that the data used for analysis and training has high quality.

[0126] Data segmentation: For the acceleration time history signals after rejecting the abnormal data, segment them according to a fixed length. The 13-channel effective acceleration data is cut into data segments with a length of 1024. Each data segment contains the acceleration response signals of 13 channels. The size of each data sample is 13 , where: 13 represents the number of channels of the sensor (each member corresponds to one sensor), and 1024 represents the time step of the data of each channel, reflecting the dynamic response characteristics within a specific time window. Through the above segmentation method, multiple data samples can be generated from the original acceleration signals of each working condition.

[0127] Creation of a dataset:

[0128] The steel frame is divided into 13 units, each unit corresponding to an acceleration sensor, numbered from 1 to 13 in sequence. The damage location label is defined as a vector with a length of 13. Among them, when a certain unit number \(i\) is damaged, the \(i\)-th position of the label vector is set to 1, and the remaining positions are set to 0. For example, when unit 5 is damaged, the label is [0,0,0,0,1,0,0,0,0,0,0,0,0], and in the undamaged state, the label is [0,0,0,0,0,0,0,0,0,0,0,0,0].

[0129] Add the corresponding labels to all data samples of each working condition, and finally generate a complete dataset containing sample data and labels, where the format of each data sample is: (data sample, label vector) . After all the data samples are randomly shuffled, they are divided into a training set and a test set according to the ratio of 80% and 20%.

[0130] Improved Channel Attention Mechanism:

[0131] 1. Since the importance of damage information contained in different positions of the structure often varies, the insertion position of the channel attention mechanism module is improved from after the convolution operation to before the convolution operation. The channel attention mechanism is used to weight the data of 13 sensors, and then the importance of data at different measurement points is judged.

[0132] 2. Due to the positive and negative fluctuation characteristics of acceleration data, directly performing global average pooling may cause the positive and negative values to cancel each other out, thus failing to fully display the data features. Therefore, global average pooling is improved to global maximum pooling, avoiding information loss caused by the positive and negative characteristics of the data, and thus enhancing the model's ability to express the key features of acceleration signals.

[0133] 3. When the sigmoid activation function is used, most of the channel weights will converge to 1 after network training, which is not conducive to comparing the weights of different sensor channels. Therefore, the activation function is improved to the prelu function, which is more conducive to linear expansion.

[0134] Neural Network Model Construction and Training:

[0135] The neural network model consists of an improved channel attention mechanism layer, a 1D-CNN layer, a global average pooling layer, and a linear output layer. The prepared dataset is input into the constructed neural network model for training. By minimizing the minimum value of the model output result and the data label until the number of training times reaches the set threshold or the value of the loss function reaches the set range and tends to be balanced, it can be considered that the model parameters have been trained, and the model results are saved. During the training process, the cross-entropy loss function is selected as the loss function.

[0136] The accuracy curves and loss curves before and after the improvement of the channel attention mechanism are as Figure 3 、 Figure 4 、 Figure 5 、 Figure 6 shown, and the accuracy and loss values are shown in Table 2:

[0137] Table 2 Comparison of Damage Recognition Accuracy and Loss before and after Improvement

[0138]

[0139] It can be seen from the comparison of the accuracy curve and the loss curve that the convergence speed of the improved model is significantly improved, and it can reach a higher accuracy faster. The experimental results show that the proposed method can effectively extract the features related to structural damage from the acceleration data, so as to accurately identify the damage location of the steel frame structure. The loss of the improved model on the test set decreased from 0.1130 to 0.0024, and the accuracy increased from 97.2% to 100%. This result verifies the effectiveness of the model improvement, indicating that the improved model can extract structural damage information more efficiently and significantly improve the recognition accuracy.

[0140] By visualizing the weights of the sensor channels of the trained model (as Figure 7 shown), the contribution differences of different sensor channels in damage recognition can be clearly observed. Based on the weight analysis results, the data of several sensor channels with larger weights were selected to retrain and predict the model again, and the results are as Figure 8 、 Figure 9 shown. The results show that an accuracy of 99.98% can be achieved by using only the data of 4 sensor channels with larger weights.

[0141] To fully verify the rationality of the method, the four sensors with the smallest weights and the four sensors with the largest weights were used for training and prediction respectively, and the results are as Figure 10 、 Figure 11 shown. The model using the data of the 4 sensors with the largest weights has a faster convergence extraction speed and a higher accuracy, verifying that the proposed method can effectively identify the structural damage information based on a small number of sensors. When the number of sensors of large structures is insufficient, this method can also provide a reference for the optimal layout of sensors.

[0142] Embodiment 3

[0143] The present invention also provides a structural damage recognition system based on an improved channel attention mechanism and a one-dimensional convolutional neural network. The system is used to implement any one of the methods described above. The system includes: an experimental model construction module, a data measurement module, a data segmentation module, a dataset construction module, a network model construction module, and a damage calculation module;

[0144] The experimental model construction module is used to establish experimental models of undamaged structures and structures with different damage locations;

[0145] The data measurement module is used to apply excitation to the structure to be measured, measure the acceleration time history data of each data acquisition point, and transmit it to the multi-channel sensor;

[0146] The data segmentation module is used to segment the multi-channel sensor data into acceleration time series segments with a fixed length;

[0147] The dataset construction module is used to construct a dataset based on the acceleration time series segments and divide the dataset into a training set and a test set;

[0148] The network model construction module is used to construct a convolutional neural network model based on the improved channel attention mechanism and the one-dimensional convolutional neural network, and train it using the training set;

[0149] The damage calculation module is used to input the data of the test set into the trained convolutional neural network model for calculation to obtain the structural damage identification result.

[0150] In this embodiment, the acceleration time history data received by the sensor is , where represents the acceleration time history data received by the c-th sensor, and the total number of acceleration sensors is C.

[0151] In this embodiment, the improved channel attention mechanism includes:

[0152] S41: Change the insertion position of the channel attention mechanism module from after the convolution operation to before the convolution operation, use the channel attention mechanism to weight the data of C sensors, and then judge the importance of the data at different measuring points;

[0153] S42: Improve the global average pooling to global max pooling;

[0154] S43: Channel expansion: Change the compression operation of the number of neurons in the middle layer to expand to the nearest power of 2;

[0155] S44: Improve the activation function to the PReLU function.

[0156] In this embodiment, the convolutional neural network model includes: an improved channel attention mechanism module, a first convolutional layer, a first max pooling layer, a second convolutional layer, a second max pooling layer, a global average pooling layer, and an output layer connected in sequence.

[0157] In this embodiment, the loss function adopted in the convolutional neural network model is the cross-entropy loss function, and the formula is:

[0158]

[0159] In the formula, represents the loss function, represents the number of categories, represents the sign function. If the true category of the sample is equal to take 1, otherwise take 0; represents the observed sample belongs to the category the predicted probability of, is the number of samples.

[0160] The embodiments described above are only descriptions of the preferred embodiments of the present invention and do not limit the scope of the present invention. Without departing from the spirit of the present invention design, various deformations and improvements made by those of ordinary skill in the art to the technical solutions of the present invention shall fall within the protection scope determined by the claims of the present invention.

Claims

1. A structural damage identification method based on an improved channel attention mechanism and a one-dimensional convolutional neural network, characterized in that The method includes: Establishing experimental models of the undamaged structure and structures with different damage locations; Applying excitation to the structure to be measured, measuring the acceleration time history data of each data acquisition point, and transmitting it to the multi-channel sensor; Segmenting the multi-channel sensor data into acceleration time series segments of a fixed length; Constructing a data set based on the acceleration time series segments and dividing the data set into a training set and a test set; Constructing a convolutional neural network model based on the improved channel attention mechanism and the one-dimensional convolutional neural network, and training it using the training set; Inputting the data of the test set into the trained convolutional neural network model for calculation to obtain the structural damage identification result.

2. The method according to claim 1, wherein The acceleration time history data received by the sensor is , where represents the acceleration time history data received by the c-th sensor, and the total number of acceleration sensors is C.

3. The method according to claim 1, wherein The improved channel attention mechanism includes: S41: Changing the insertion position of the channel attention mechanism module from after the convolution operation to before the convolution operation, weighting the C sensor data using the channel attention mechanism, and then judging the importance of the data at different measurement points; S42: Changing the global average pooling to global max pooling; S43: Changing the compression of the number of neurons in the middle layer to appropriate expansion; S44: Changing the activation function to the PReLU function.

4. The method according to claim 1, wherein The convolutional neural network model includes: an improved channel attention mechanism module, a first convolutional layer, a first max pooling layer, a second convolutional layer, a second max pooling layer, a global average pooling layer, and an output layer connected in sequence.

5. The method according to claim 1, characterized in that The loss function adopted in the convolutional neural network model is the cross-entropy loss function, and the formula is: Wherein, represents the loss function, represents the number of categories, represents the sign function. If the true category of the sample is equal to take 1, otherwise take 0; represents the observed sample belongs to the category predicted probability, is the number of samples.

6. A structural damage identification system based on an improved channel attention mechanism and a one-dimensional convolutional neural network, the system is used to implement the method described in any one of claims 1-5, and is characterized in that The system includes: an experimental model construction module, a data measurement module, a data segmentation module, a data set construction module, a network model construction module, and a damage calculation module; The experimental model construction module is used to establish experimental models of the undamaged structure and structures with different damage locations; The data measurement module is used to apply excitation to the structure to be measured, measure the acceleration time history data of each data acquisition point, and transmit it to the multi-channel sensor; The data segmentation module is used to segment the multi-channel sensor data into acceleration time series segments of a fixed length; The data set construction module is used to construct a data set based on the acceleration time series segments and divide the data set into a training set and a test set; The network model construction module is used to construct a convolutional neural network model based on the improved channel attention mechanism and the one-dimensional convolutional neural network, and train it using the training set; The damage calculation module is used to input the data of the test set into the trained convolutional neural network model for calculation to obtain the structural damage identification result.

7. The system according to claim 6, wherein The acceleration time history data received by the sensor is , where represents the acceleration time history data received by the c-th sensor, and the total number of acceleration sensors is C.

8. The system according to claim 6, wherein The improved channel attention mechanism includes: S41: Changing the insertion position of the channel attention mechanism module from after the convolution operation to before the convolution operation, weighting the C sensor data using the channel attention mechanism, and then judging the importance of the data at different measurement points; S42: Changing the global average pooling to global max pooling; S43: Changing the compression of the number of neurons in the middle layer to appropriate expansion; S44: Changing the activation function to the PReLU function.

9. The system according to claim 6, wherein The convolutional neural network model includes: an improved channel attention mechanism module, a first convolutional layer, a first max pooling layer, a second convolutional layer, a second max pooling layer, a global average pooling layer, and an output layer connected in sequence.

10. The system according to claim 6, wherein The loss function adopted in the convolutional neural network model is the cross-entropy loss function, and the formula is as follows: Wherein, represents the loss function, represents the number of categories, represents the sign function. If the true category of the sample is equal to take 1, otherwise take 0; represents the observed sample belongs to the category the predicted probability of, is the number of samples.

Images