Building structure health monitoring missing data repair method based on CNN-BiLSTM-attention optimized

Abstract

The building structure health monitoring missing data repairing method based on AO optimization of CNN-BiLSTM-Attention can effectively capture the space-time characteristics of structure health monitoring data, and realize high-precision structure health monitoring missing data repairing.The specific implementation process is as follows: A. pre-collecting and pre-processing structure health monitoring complete data, and constructing training and verification data sets; B. constructing a CNN-SE framework to extract the spatial characteristics of monitoring data; C. constructing a BiLSTM-TA framework to extract the time characteristics of monitoring data; D. constructing a CNN-BiLSTM-Attention model, and adopting an AO algorithm to optimize the model super parameters; E. inputting the data collected by normal sensors, and repairing the missing data of faulty sensors.The method disclosed by the application can realize high-precision repairing of structure health monitoring missing data, and provide support for structure state diagnosis and prediction.

Description

TECHNICAL FIELD

[0001] The application relates to a building structure health monitoring missing data repairing method, which combines CNN and SE attention mechanism, combines BiLSTM and TA attention mechanism, and adopts an AO optimization algorithm to optimize model hyperparameters, so that missing data of structure health monitoring is repaired with high precision, and belongs to the technical field of structure health monitoring. BACKGROUND

[0002] Structure health monitoring technology can evaluate the performance state of a building structure in real time, and plays an important role in ensuring the safe operation of civil engineering structures. The effectiveness of a structure health monitoring system depends on reliable sensor data. However, due to the harsh environment (such as strong wind, earthquake, and sharp temperature change) of high-rise buildings, sensor problems such as battery failure and communication interruption often occur, resulting in missing monitoring data. These missing data will affect the accuracy of structure state evaluation. Therefore, it is of great engineering application value to develop an effective missing data repairing method.

[0003] Current data repairing methods are mainly based on deep learning models, which repair missing monitoring data by modeling the correlation between complete and missing data. However, these methods usually only focus on local features of data, ignoring the complex internal correlation of data in time and space dimensions, and do not consider the importance difference between different features. Therefore, it is necessary to develop a high-precision data repairing method that can effectively capture the spatiotemporal correlation between different sensor data and consider the importance of different feature channels and key time steps. SUMMARY

[0004] In view of the deficiencies of the prior art, the application provides a building structure health monitoring missing data repairing method based on AO optimization CNN-BiLSTM-Attention to improve the repairing precision of structure health monitoring missing data. The specific content includes:

[0005] The building structure health monitoring missing data repairing method based on AO optimization CNN-BiLSTM-Attention comprises the following steps:

[0006] A. Pre-collect and pre-process the complete structure health monitoring data to construct training and verification data sets;

[0007] B. Construct a CNN-SE framework to extract spatial features of monitoring data;

[0008] C. Construct a BiLSTM-TA framework to extract time features of monitoring data;

[0009] D. Construct a CNN-BiLSTM-Attention model, and adopt an AO algorithm to optimize model hyperparameters;

[0010] E. Input the data collected by normal sensors and repair the missing data of faulty sensors.

[0011] Further, step A specifically includes:

[0012] A1. Pre-collect complete structural health monitoring data;

[0013] A2. Process the pre-collected signal data using the max-min normalization method to scale the data to the range of [-1, 1], thereby reducing the influence on the generalization ability of the network model;

[0014] A3. Divide the normalized data to construct training and validation data sets.

[0015] Further, step B specifically includes:

[0016] B1. Set a convolutional neural network (CNN) network containing two convolutional layers, use ReLU activation function after each convolutional layer, obtain the convolutional feature map of the monitoring data, and pass these feature maps to the dot product layer;

[0017] B2. After the first convolutional operation, the Squeeze-and-Excitation (SE) mechanism first performs global average pooling on the features of each channel to obtain the statistical information of the channel global features, which can be represented as:

[0018]

[0019] In the formula: H and W represent the width and height of the features respectively; u c is the input data; z c is the output value;

[0020] B3. The SE attention mechanism inputs z c into a fully connected layer, processes it through a ReLU activation function, and then enters another fully connected layer, which is activated by a sigmoid function to obtain the weight value of each channel. The weight value of each channel is passed to the dot product layer, which can be represented as:

[0021] S SE = σ(W2δ(W1z)) (2)

[0022] In the formula: W1 and W2 represent the weights of the two fully connected layers respectively; S SE is the output value; δ is the ReLU function; σ is the sigmoid function;

[0023] B4. The generated channel weight S SEThe convolution feature map is point multiplied with each channel to realize adaptive weighting of different channels, and is flattened into a vector by a flatten layer and transmitted to a subsequent BiLSTM layer.

[0024] Further, step C specifically comprises:

[0025] C1. A bidirectional long short-term memory network (BiLSTM) is established to generate a hidden state for each time step and transmit it to a dot product layer;

[0026] C2. After the data passes through the first BiLSTM layer, a temporal attention mechanism (TA) adopts a fully connected layer and a tanh function to perform nonlinear mapping on the input data, and converts the mapping result into an attention score, which can be expressed as:

[0027] e t = v T tanh(W3h t + b) (3)

[0028] In the formula: e t is the attention score; W3 is the weight matrix; v is the weight vector; h t is the hidden state of the t-th time step; b is the bias vector;

[0029] C3. The attention score is normalized by a softmax function to obtain the attention weight of each time step, which is transmitted to a dot product layer, which can be expressed as:

[0030]

[0031] In the formula: α t is the attention weight of the t-th time step;

[0032] C4. The hidden state of each time step is point multiplied with the attention weight α t to obtain the weighted time information, which is then processed by a BiLSTM layer, a fully connected layer and a regression layer to output the final prediction result of the model.

[0033] Further, step D specifically comprises:

[0034] D1. For the initial learning rate, the L2 regularization coefficient, the learning rate reduction factor, and the number of hidden units in the BiLSTM layer and the fully connected layer, the Aquila Optimization (AO) algorithm generates a certain number of random solutions, each solution corresponding to a combination of hyperparameters to be optimized.

[0035] D2. For each hyperparameter combination, establish a CNN-BiLSTM-Attention model, train the model, and calculate the fitness value (RMSE) with the prediction results of the validation set, which can be expressed as:

[0036]

[0037] In the formula: k is the number of data points; y s is the true value of the data; is the repaired value of the data;

[0038] D3. Adopt the AO optimization algorithm to minimize the fitness value, obtain the optimal hyperparameter combination, and establish the optimal CNN-BiLSTM-Attention model.

[0039] The advantages of the present application are:

[0040] (1) The present application combines CNN with SE attention mechanism, effectively extracts the spatial features of monitoring data through CNN, obtains the weights of different spatial feature channels through SE attention mechanism, and effectively extracts the internal correlation in the spatial dimension of data.

[0041] (2) The present application combines BiLSTM with TA attention mechanism, effectively extracts the time features of sensor data through BiLSTM, obtains the weights of key time steps through TA attention mechanism, and effectively captures the time features of data.

[0042] (3) The present application uses the AO optimization algorithm to optimize the hyperparameters of the neural network, effectively improving the accuracy of the model in repairing missing data. BRIEF DESCRIPTION OF DRAWINGS

[0043] Figure 1 is a flowchart of the method of the present application;

[0044] Figure 2 is a strain measurement point layout diagram of the South Taihu CBD Mansion connecting truss of the present application;

[0045] Figure 3 is a CNN-BiLSTM-Attention framework schematic diagram of the present application;

[0046] Figure 4 is a missing data repair result diagram of the present application;

[0047] Figure 5 is a missing data repair result comparison diagram of different methods;

[0048] Figure 6 is a missing data repair index comparison diagram of different methods; DETAILED DESCRIPTION

[0049] The application will be further described in detail below in combination with the drawings and an example.

[0050] As Figure 1 , the embodiment relates to a building structure health monitoring missing data repair method based on CNN-BiLSTM-Attention of AO optimization, and specifically comprises the following steps:

[0051] A. Pre-acquisition and preprocessing of structure health monitoring complete data, construction of training and verification data set, specifically including:

[0052] A1. The South Taihu CBD Building is installed with a structure health monitoring system, and 12 strain sensors based on fiber Bragg gratings are deployed in the connecting truss part, with a sampling frequency of 1 Hz, continuous acquisition for 120 hours, a total of 432000 data, and the measuring point arrangement is as shown in Figure 2 ;

[0053] A2. The original data is averaged every 60 groups, 7200 processed data are obtained, and the maximum-minimum normalization method is used to process the pre-acquired signal data, so that the data is scaled to the range of [-1, 1];

[0054] A3. The normalized data is divided, and 5760 data from 0 to 72 hours and 96 to 120 hours are used as the training set of the model, and 1440 data from 72 to 96 hours are used as the verification set of the model.

[0055] B. Construction of CNN-SE framework to extract spatial features of monitoring data, specifically including:

[0056] B1. Set a CNN network containing two convolutional layers, use a ReLU activation function after each convolutional layer, obtain the convolutional feature map of the monitoring data, and pass it to the dot product layer;

[0057] B2. After the first BiLSTM layer, the SE attention mechanism first performs global average pooling on the features of each channel to obtain the statistical information z c :

[0058] B3. The SE attention mechanism inputs z c into the full connection layer, and after ReLU activation function processing, it enters another full connection layer, is activated by a sigmoid function, obtains the weight value S SE of each channel, and is passed to the dot product layer;

[0059] B4. The generated channel weight S SEThe convolutional feature map is point-wise multiplied with each channel to realize adaptive weighting of different channels, and is flattened into a vector by a flatten layer and passed to a subsequent BiLSTM layer.

[0060] C. Constructing a BiLSTM-TA framework to extract time features of monitoring data, specifically including:

[0061] C1. Establishing a BiLSTM layer to generate a hidden state for each time step and passing it to a dot product layer;

[0062] C2. After the data passes through the first BiLSTM layer, the TA attention mechanism uses a fully connected layer and a tanh function to perform nonlinear mapping on the input data, and converts the mapping result into attention scores e t ;

[0063] C3. Normalizing the attention scores by a softmax function to obtain attention weights a t for each time step and passing them to the dot product layer;

[0064] C4. Dot product operation is performed on the hidden state of each time step and the attention weight a t to obtain weighted time information, which is then processed by a BiLSTM layer, a fully connected layer, and a regression layer in turn to output the final prediction result of the model;

[0065] D. Constructing a CNN-BiLSTM-Attention model as shown in Figure 3 , which uses an AO optimization algorithm to optimize the model hyperparameters, specifically including:

[0066] D1. For the initial learning rate, L2 regularization coefficient, learning rate reduction factor, and the number of hidden units in the BiLSTM layer and fully connected layer, the AO optimization algorithm generates a certain number of random solutions, each solution corresponding to a combination of hyperparameters to be optimized;

[0067] D2. For each combination of hyperparameters, a CNN-BiLSTM-Attention model is established and trained, and the fitness value is calculated based on the prediction performance of the validation set;

[0068] D3. The AO optimization algorithm is used to minimize the fitness value to obtain the optimal combination of hyperparameters, and the optimal CNN-BiLSTM-Attention model is established.

[0069] E. Input data collected by normal sensors to repair missing data of faulty sensors. The repair effect of missing data by the CNN-BiLSTM-Attention model is shown in Figure 4 , and the comparison with other methods is shown in Figure 5 , and the repair effect indicators are as followsFigure 6 as shown.

[0070] The person skilled in the art understands that the embodiments described in the specification are only a list of implementation forms of the inventive concept, and the protection scope of the present application should not be regarded as being limited to the specific forms described in the embodiments. The protection scope of the present application also extends to equivalent technical means that can be thought of by the person skilled in the art according to the inventive concept.

Claims

1. A building structure health monitoring missing data repair method based on AO-optimized CNN-BiLSTM-Attention, comprising the following steps: A. Pre-acquiring and pre-processing complete structure health monitoring data to construct a training set and a validation data set; B. Constructing a CNN-SE framework to extract spatial features of the monitoring data, specifically including: B1. Setting a CNN network containing two convolutional layers, using a ReLU activation function after each convolutional layer, obtaining convolutional feature maps of the monitoring data, and passing these feature maps to a dot product layer; B2. After the first convolutional operation, the SE attention mechanism first performs global average pooling on the features of each channel to obtain statistical information of the channel global features, which can be represented as: where: H and W represent the width and height of the feature, respectively; u c is the input data; z c is the output value; B3. SE attention mechanism will z c input into a fully connected layer, after ReLU activation function processing, into another fully connected layer, through sigmoid function activation, to obtain the weight value S of each channel SE , the weight value of each channel is transmitted to the dot product layer, which can be represented as: S SE = σ(W2δ(W1z)) (2) wherein: W1 and W2 represent the weights of two fully connected layers, respectively; S SE is the output value; δ is the ReLU function; and σ is the sigmoid function. B4. Generated channel weights S SE The second convolutional layer feature map is point-wise multiplied with the first convolutional layer feature map, and adaptive weighting of different channels is realized. The flattened layer is used to flatten the vector, which is transmitted to the subsequent BiLSTM layer. C. Constructing a BiLSTM-TA framework to extract time features of the monitoring data, specifically including: C1. Establishing a first BiLSTM layer to generate hidden states for each time step and passing them to a dot product layer; C2. After the data passes through the first BiLSTM layer, the TA attention mechanism uses a fully connected layer and a tanh function to perform nonlinear mapping on the input data, and converts the mapping results into attention scores, which can be represented as: where: e t is the attention score; W3is a weight matrix; v is a weight vector; h t is the hidden state at the tth time step; b is a bias vector; C3. Normalizing the attention scores by a softmax function to obtain the attention weights of each time step and passing them to a dot product layer, which can be represented as: In the formula, α t is the attention weight for the t-th time step; C4. Hidden state of each time step and attention weight a t The dot product operation is performed to obtain the weighted time information, which is then processed by a second BiLSTM layer, a fully connected layer, and a regression layer to output the final prediction result of the model. D. Constructing a CNN-BiLSTM-Attention model and optimizing the model hyperparameters using the AO optimization algorithm; E. Inputting data collected by normal sensors to repair missing data of faulty sensors.

2. The CNN-BiLSTM-Attention based on AO optimization method for building structure health monitoring missing data repair according to claim 1, wherein, Step A specifically includes: A1. Pre-acquiring complete building structure health monitoring data; A2. Using the max-min normalization method to process the pre-acquired signal data, scaling the data to the range [-1, 1] to reduce the impact on the generalization ability of the network model; A3. Dividing the normalized data to construct a training set and a validation data set.

3. The CNN-BiLSTM-Attention based on AO optimization method for building structure health monitoring missing data repair according to claim 1, wherein, Step D specifically includes: D1. For the initial learning rate, L2 regularization coefficient, learning rate decay factor, and the number of hidden units in the BiLSTM layer and fully connected layer, the AO optimization algorithm generates a certain number of random solutions, each solution corresponds to a combination of hyperparameters to be optimized; D2. For each hyperparameter combination, establish a CNN-BiLSTM-Attention model and train it, calculate the fitness value (RMSE) based on the prediction performance of the validation set, which can be represented as: wherein: k is the number of data points; y s is the true value of the data; is the repaired value of the data; D3. Minimize the fitness value using the AO optimization algorithm to obtain the optimal hyperparameter combination and establish the optimal CNN-BiLSTM-Attention model.

Images