Deep learning method for structural health monitoring acceleration data restoration

Abstract

The invention belongs to the field of structural health monitoring and data processing, and aims to solve the problem that manual intervention is always needed in the aspects of abnormal segment detection and elimination in traditional data restoration. The invention provides a deep learning method for structural health monitoring acceleration data restoration, and the method comprises a structural health monitoring acceleration abnormal data set simulation method and a deep learning method for health monitoring abnormal data restoration. Wherein the structural health monitoring acceleration abnormal data simulation method is used for manually adding various types of abnormal signals and making a data set of a deep learning model, and the health monitoring abnormal data restoration deep learning method is used for restoring various types of actually measured abnormal data into normal data. According to the invention, various types of abnormal features of the health monitoring acceleration data can be learned, the data directly collected by the acceleration sensor is used as the input of the network, and the repaired data is directly output without manual intervention in the use stage.

Description

Technical Field

[0001] The present invention belongs to the field of structural health monitoring and data processing, and in particular relates to a deep learning method for structural health monitoring acceleration data repair. Background Art

[0002] With the improvement of global economic level and the advancement of engineering technology, large-scale engineering structures are constantly developing towards larger spans and larger scales. To ensure the safe operation of these structures, health monitoring systems are widely used, covering hundreds or thousands of sensors, and continuously generating massive monitoring data every day. While these data are multidimensional, complex and random, they are also inevitably subject to abnormal interference caused by sensor failures, signal transmission problems and external environmental influences. When using unprocessed monitoring data for structural status assessment, it may lead to a decrease in assessment accuracy and even lead to false alarms or omissions of structural accidents. A large amount of abnormal monitoring data that is directly eliminated will also lead to serious data loss, which seriously affects the actual application effect of monitoring data. Therefore, in order to ensure the integrity and availability of structural monitoring data, how to repair abnormal monitoring data has become an important issue in the field.

[0003] Traditional data repair methods usually target specific abnormal data types and often use data interpolation, fitting, and filtering techniques to remove or replace abnormal data. These methods are simple and easy to implement in the case of a single sensor, but it is difficult to process data with long vacant sequences. With the development of artificial intelligence technology, deep learning technology has gradually been applied to health monitoring data processing. Fan et al. proposed a random loss data recovery method based on U-Net, which can use multiple sensor data to repair the lost data of a single sensor. Lei et al. introduced adversarial loss into the network, which can effectively reconstruct high-frequency information in the signal. Jeong et al. proposed a signal reconstruction method based on bidirectional recurrent neural network (BRNN) and successfully used it for modal analysis of structures. Tang et al. further combined compressed sensing with convolutional networks and used deep learning instead of traditional methods to solve the pathological problems in signal reconstruction. Ju Hanwen et al. proposed a structural monitoring data repair method based on gated recurrent neural network (GRU) and successfully applied it to the repair of strain, crack and temperature monitoring data of an ancient city wall.

[0004] In general, the current traditional data repair methods are based on relatively complete mathematical models and theoretical foundations. However, they are often only effective for specific types of anomalies and require a lot of manual intervention, so they are not suitable for processing massive monitoring data. Deep learning-based methods generally treat abnormal data segments as lost data, and often still require manual intervention in the detection and removal of abnormal segments, which limits the repair efficiency. Current deep learning methods mainly use the correlation between multiple sensors as the basis for repair, and the research on the loss of abnormal data of a single sensor is relatively limited. In health monitoring systems, such as the acceleration monitoring data of the hanger on a suspension bridge, the correlation between sensors is poor, so the multi-sensor repair method is not applicable. Summary of the invention

[0005] In order to solve at least one of the above-mentioned technical problems existing in the prior art, the present invention provides a deep learning method for repairing acceleration data of structural health monitoring.

[0006] The present invention is implemented by the following technical solution: A deep learning method for structural health monitoring acceleration data repair comprises the following steps:

[0007] Acquire normal acceleration data in structural health monitoring, and add different abnormal signal features to the normal data to obtain various types of abnormal data; the addition of the abnormal signal features is based on the generation of multiple simulated abnormal data, and the simulated situations include data jump point simulation, data gain simulation, trend item simulation, data drift simulation, precision degradation simulation and data loss simulation;

[0008] Build a deep learning model based on acceleration data repair, use the data set composed of normal data and abnormal data as the input of the model, use the normal data matching the input data as the output of the model, and train and verify the model;

[0009] The acceleration data for structural health monitoring is collected and input into a deep learning model based on acceleration data repair. For normal data in the acceleration data, the model reconstructs the data and keeps the original data output. For abnormal data in the acceleration data, the model detects the abnormal type and location, corrects the abnormal data, reconstructs it into normal data, and then outputs it.

[0010] Preferably, the data jump point simulation is to randomly select a number of data points on a data segment of unit length, and introduce random noise into the data points, which is set to n1 times the amplitude, and 1.5<n1<8;

[0011] The data gain simulation includes four cases: overall amplification, overall reduction, local amplification, and local reduction. For overall amplification, a gain coefficient of n2 times is randomly generated, and 1.5<n2<10; for overall reduction, a gain coefficient of n3 times is randomly generated, and 0.02<n3<0.3; for local gain, data of n4 times the length of the data segment is randomly selected for gain, and

[0012] 0.2<n4<0.9.

[0013] Preferably, the trend item is simulated as randomly dividing a data segment of unit length into segments, and adding a linear height difference with an amplitude of n5 times the amplitude to each segment, including four cases of increasing, decreasing, and random positive and negative signs; a sinusoidal function with a random period, a random initial phase, and a random amplitude is further superimposed on each segment, wherein the period ranges from 0.5π to 2.5π, the random initial phase is between 0 and 2π, the random amplitude is between n5 times the amplitude, and 0.3<n5<2.5.

[0014] Preferably, the data drift simulation is to randomly select a data segment of n6 times the data length, and 0.1<n6<0.4; and generate a peak point height of n7 times the amplitude, and 0.5<n7<2.5; randomly determine the number of fitting functions of the local "trend term" and superimpose it on the original data;

[0015] The precision reduction is simulated as adding random noise with a signal-to-noise ratio between -5 and 0 to the original signal;

[0016] The data loss simulation is a randomly generated signal loss rate between 0.2 and 0.7, and is divided into two situations: continuous whole-segment data loss and discrete point data loss.

[0017] Preferably, the simulation also includes normal data simulation, and the normal data simulation is to keep the normal data unchanged.

[0018] Preferably, the deep learning model for acceleration data restoration includes four parts: a block embedding module, an encoder module, a decoder module, and a loss function;

[0019] The block embedding module includes a reshaping operation module, a two-dimensional convolution module and a position encoding module, which is used to divide the input data into fixed-size data segments, expand them into a series of vector representations and perform linear projection transformations; the input data is reshaped and convolved in the block embedding module to become features of standardized size, and then sine-cosine position encoding is used to generate an invariant position information tensor, and the position information is attached to the feature by direct addition;

[0020] The encoder module is composed of several Transformer modules, which include regularization layer, multi-head attention mechanism, residual structure and fully connected layer. The output of the block embedding module is layer-normalized and input to the multi-head attention mechanism. In the multi-head attention mechanism, the number of heads is set to k, and the features of the normalized size are passed through the fully connected layer and reshape operation to generate k groups of query, key and numerical features. The k groups of query, key and numerical features are respectively input to the scaled dot product attention module to obtain k groups of features. The output of each attention head is then connected to a two-layer feedforward neural network. The residual structure is nested in each sublayer, including the attention mechanism layer and the multi-layer perceptron layer.

[0021] The decoder module is composed of several Transformer modules. The output of the encoder module is first mapped by the fully connected layer, and then the sine-cosine position encoding is used to generate an invariant position information tensor, and the position information is attached to the feature; the decoder output feature is again normalized by the layer, fully connected layer and reshaped, and the reconstructed data is output.

[0022] Preferably, the loss function of the overall network consists of two parts, namely, the average relative error of the reconstructed data And the correlation coefficient between the reconstructed data and the original data Where y is the original input data, a is the actual output of the network, N is the length of each segment of data, a is the mean of the actual output data of the network; y is the mean of the original input data; i represents the i-th, and n represents the number of samples;

[0023] Finally, the overall loss function of the network is defined as:

[0024] Preferably, in the normal data in structural health monitoring, the moving window scanning method is used to intercept the data, the time series length and the moving step length are set to 1024, and the initial training set and the test set are constructed; the mean and variance of the monitoring signals of each sensor in the initial training set are statistically calculated, and the data in the initial training set and the initial test set are preprocessed using the Z-score normalization method; for the initial training set and the initial test set, the generation of several kinds of abnormal data is randomly simulated to produce the input data of the deep learning model for acceleration data repair, and the final training set and the test set are obtained.

[0025] Compared with the prior art, the present invention has the following beneficial effects:

[0026] The device developed by the present invention can be widely used in the preprocessing of acceleration data for structural health monitoring. It can complete the clear processing of large-scale acceleration monitoring data in a short time without manual intervention:

[0027] 1. It can realize the rapid and clear abnormal data of structural health monitoring data. Compared with the traditional method: the algorithm proposed in the present invention does not require manual intervention to judge the abnormal type, adjust the threshold and other coefficients, directly takes the original data as input, and completely relies on the high-dimensional feature mapping extracted by the convolutional network to realize the correction of various abnormal types of data. It has a wide range of use and a high degree of automation.

[0028] 2. There is no need to worry about information loss caused by normal data input into the network. After normal data is input into the network, the input data can be directly reconstructed with high precision. The average relative error and average correlation coefficient of the output data and output are 0.065 and 0.999 respectively. BRIEF DESCRIPTION OF THE DRAWINGS

[0029] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the drawings required for use in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying creative work.

[0030] Figure 1 It is the overall technical roadmap of the present invention;

[0031] Figure 2 It is a jump point data simulation diagram in the present invention;

[0032] Figure 3 is the data gain simulation in the present invention;

[0033] Figure 4 is the trend term and drift simulation in the present invention;

[0034] Figure 5 It is the precision reduction simulation in the present invention;

[0035] Figure 6 It is the data loss simulation in the present invention;

[0036] Figure 7 It is the Transformer structure diagram in the present invention;

[0037] Figure 8 is the network convergence curve in the present invention Figure 1 ;

[0038] Fig. 9 is the network convergence curve in the present invention Figure 2 ;

[0039] Fig.10 It is the output effect of the normal data input network in the present invention;

[0040] Fig.11This is the repair effect of the jump point data input model in the present invention;

[0041] Fig.12 is the repair effect of the gain data input model in the present invention;

[0042] Fig.13 It is the repair effect of the drift data input model in the present invention;

[0043] Fig.14 It is the repair effect of the data input model with different signal-to-noise ratios in the present invention;

[0044] Fig.15 This is the repair effect of the different loss rate data input models in the present invention. DETAILED DESCRIPTION

[0045] In conjunction with the drawings in the embodiments of the present invention, the technical solutions in the embodiments of the present invention are clearly and completely described. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other implementation methods obtained by ordinary technicians in this field without making creative work are within the scope of protection of the present invention.

[0046] It should be noted that the structures, proportions, sizes, etc. illustrated in the drawings of this specification are only used to match the contents disclosed in the specification so that people familiar with this technology can understand and read them. They are not used to limit the conditions under which the present invention can be implemented, and therefore have no substantive technical significance. Any structural modification, change in proportional relationship or adjustment of size should fall within the scope of the technical contents disclosed in the present invention without affecting the effects and purposes that can be achieved by the present invention. It should be noted that in this specification, relational terms such as first and second are only used to distinguish one entity from several other entities, and do not necessarily require or imply any actual relationship or order between these entities.

[0047] The present invention provides an embodiment:

[0048] A deep learning method for repairing acceleration data for structural health monitoring comprises the following steps: obtaining normal data of acceleration in structural health monitoring, and adding different abnormal signal features to the normal data to obtain various types of abnormal data; the addition of the abnormal signal features is based on the generation of multiple simulated abnormal data, and the simulated situations include data jump point simulation, data gain simulation, trend item simulation, data drift simulation, precision reduction simulation and data loss simulation; constructing a deep learning model based on the repair of acceleration data, taking a data set composed of normal data and abnormal data as the input of the model, taking normal data matching the input data as the output of the model, and training and verifying the model; collecting acceleration data in structural health monitoring, and inputting the deep learning model based on the repair of acceleration data; for normal data in the acceleration data, the model reconstructs the data and keeps the original data output, and for abnormal data in the acceleration data, the model detects the abnormal type and position, and corrects and reconstructs the abnormal data into normal data for output.

[0049] In the normal data of structural health monitoring, the moving window scanning method is used to intercept the data, and the time series length and the moving step are set to 1024 to build the initial training set and test set; the mean and variance of the monitoring signals of each sensor in the initial training set are statistically calculated, and the Z-score normalization method is used to preprocess the data in the initial training set and the initial test set; for the initial training set and the initial test set, the generation of several kinds of abnormal data is randomly simulated to produce the input data of the deep learning model for acceleration data repair, and the final training set and test set are obtained.

[0050] In this embodiment, the simulation also includes normal data simulation, and the normal data simulation is to keep the normal data unchanged.

[0051] The data jump point simulation is to randomly select 4 to 8 data points on each data segment of length 1024, and introduce random noise into the data points, which is set to n1 times of the amplitude, and 1.5<n1<8;

[0052] The data gain simulation includes four cases: overall enlargement, overall reduction, local enlargement, and local reduction. For the case of overall enlargement, a gain coefficient of n2 times is randomly generated, and 1.5<n2<10; for the case of overall reduction, a gain coefficient of n3 times is randomly generated, and 0.02<n3<0.3; for the case of local gain, data of n4 times the length of the data segment is randomly selected for gain, and 0.2<n4<0.9.

[0053] The trend item simulation is to randomly divide the data segment of length 1024 into segments (1 to 4 segments), and add a linear height difference with an amplitude of n5 times the amplitude to each segment, including four cases of increasing, decreasing and random positive and negative signs; and further superimpose a sinusoidal function of random period, random initial phase and random amplitude on each segment, where the period ranges from 0.5π to 2.5π, the random initial phase is between 0 and 2π, the random amplitude is between n5 times the amplitude, and 0.3<n5<2.5.

[0054] The data drift simulation is to randomly select a data segment of n6 times the data length, and 0.1<n6<0.4; and generate a peak point height of n7 times the amplitude, and 0.5<n7<2.5; randomly determine the number of fitting functions of the local "trend term" and superimpose it on the original data;

[0055] The accuracy reduction is simulated by adding random noise with a signal-to-noise ratio (SNR) between -5 and 0 to the original signal;

[0056] The data loss simulation is a randomly generated signal loss rate between 0.2 and 0.7, and is divided into two situations: continuous whole-segment data loss and discrete point data loss.

[0057] The deep learning model for acceleration data repair consists of four parts: a block embedding module, an encoder module, a decoder module, and a loss function. The input of the model is a time series monitoring signal of length 1×1024, and the output is a repaired monitoring signal of 1×1024. The block embedding module divides the input signal into fixed-size signal segments, expands them into a series of vector representations, and performs linear projection transformations. It mainly consists of reshaping operations, two-dimensional convolutions, and position encoding. The 1×1024 input data is first reshaped to a size of 1×8×128. In the two-dimensional convolution, the convolution kernel size used is 1×16, and the convolution step size is (1,16). After the convolution operation, the input signal becomes a 1×64×256 feature. Subsequently, the sine-cosine position encoding is used to generate an invariant position information tensor, and the position information is attached to the feature by direct addition.

[0058] The position encoding formula is: Where, i∈(0,dim / 2), PE (pos,i) Represents the value of the posth position and the i-th dimension in the position encoding matrix, and dim represents the dimension of the model embedding vector;

[0059] The encoder consists of 12 Transformer modules, which take as input 1×64×256 features and output 1×64×256 features. The Transformer module mainly consists of a regularization layer, a multi-head attention mechanism, a residual structure, and a fully connected layer. The 1×64×256 features are first normalized by the layer and then input into the multi-head attention mechanism. In the multi-head attention mechanism, the number of heads is set to 8, and the 1×64×256 standardized features are passed through a fully connected layer and a reshape operation to generate 8 sets of query (Q), key (K), and value (V) features, all of which are 1×64×32 in size. The 8 sets of Q, K, and V features are respectively input into the scaled dot product attention module to obtain eight sets of features of 1×64×32 in size. The output of each attention head is then connected to a two-layer feedforward neural network, in which the dimension of the hidden layer is set to 128. The residual structure is nested in each sublayer, including the attention mechanism layer and the multi-layer perceptron layer.

[0060] The encoder output 1×64×256 features are first mapped to 1×64×128 features through the fully connected layer; again using the formula The sine and cosine position encoding generates an invariant position information tensor and attaches the position information to the feature; the decoder consists of 4 Transformer modules, with an input of 1×64×128 features and an output of 1×64×128 features. Finally, the decoder output features are again normalized, fully connected, and reshaped to output reconstructed data of 1×1024;

[0061] The loss function of the overall network consists of two parts, namely the average relative error of the reconstructed data And the correlation coefficient between the reconstructed data and the original data Where y is the original input data, a is the actual output of the network, N is the length of each segment of data, a is the mean of the actual output data of the network; y is the mean of the original input data; i represents the i-th, and n represents the number of samples;

[0062] Finally, the overall loss function of the network is defined as:

[0063] For the trained model, the acceleration data collected by the structural health monitoring system can be directly used as input with a length of 1×1024; for normal data input, the model can directly reconstruct the input data without worrying about information loss; for abnormal data input, the model can automatically detect the abnormal type and abnormal location, and output high-precision repaired data.

[0064] The present invention is further explained below in conjunction with the accompanying drawings and specific embodiments:

[0065] Step 1: On the original data, the moving window scanning method is used to intercept the data set, and the time series length and moving step size are set to 1024. The mean and variance of the monitoring signals of each sensor in the initial training set are counted, and the data of the initial training set and the initial test set are preprocessed using the Z-score normalization method to obtain a training set containing 200,000 data and a test set containing 50,000 data, which are also the target true values ​​of the data; for the training set and the test set, seven situations are randomly simulated to produce the input data of the network, specifically: normal data simulation, data jump point simulation ( Figure 2 )、Data gain simulation( Figure 3 ), trend items and data drift simulation ( Figure 4 )、Accuracy reduction simulation( Figure 5 ) and data loss simulation ( Figure 6 ), and finally we get the training set and test set with various types of anomalies randomly added.

[0066] Step 2: Build a deep learning model ( Figure 1 ), the network input size is 1×1024 data points, and the output is 1×1024 repaired data; the model includes a block embedding module, an encoder module, a decoder module and a loss function; the block embedding module divides the input signal into fixed-size signal segments, linearly projects them into a series of vectors, and performs reshaping and two-dimensional convolution operations to convert 1×1024 data into 1×8×128 features, and then adds position information through sine and cosine position encoding; the encoder module consists of 12 Transformer modules ( Figure 7 ), including layer normalization, multi-head attention mechanism (8 heads), residual structure and feedforward neural network (hidden layer dimension is 128); the encoder output is mapped through a fully connected layer, sine and cosine position encoding is added, and then converted into reconstructed data through 4 Transformer modules, and finally the repaired signal is output through normalization and reshaping operations. The Adam algorithm is used to update the network parameters, so that the gap between the model output and the target true value is gradually reduced until the network converges.

[0067] Step 3. After the model is trained, the data collected by the health monitoring system is directly input into the network. The length of the monitoring data collected by the structural health monitoring system is much longer than the length of the network input data. The sliding window scanning detection method is used to scan each sensor to collect data. Each scanning detection window is staggered in the direction of the monitoring data length (1024) in the collected data. The segmented data is used as the input of the network in turn, and the corrected data is directly output.

[0068] The present invention is further described below by means of specific examples, but the protection scope of the present invention should not be limited thereto.

[0069] Taking the 17-week monitoring data of 8 acceleration sensors on the publicly available Dowling Hall pedestrian bridge as an example, the data from the first 12 weeks are used to construct the initial training set, and the data from the last 5 weeks are used to construct the initial test set. We then further construct training sets and test sets containing various types of abnormal data. The training set contains 200,000 data segments, and the test set contains 50,000 data segments.

[0070] Training the data repair proposed by the present invention, the model training process is as follows Figure 8 , 9 As shown in the figure, the training converged and there was no overfitting. The training loss and test loss in the initial stage of training showed a clear downward trend. After training for 20 epochs, the improvement of network performance gradually slowed down, and the best test result was obtained at the 96th epoch. As the number of training times increases, the relative error between the repaired data and the true value gradually decreases, and the average correlation coefficient on the test set can reach 0.95.

[0071] Fig.10 It shows the output effect of the network with normal data input, and the output data and input data almost completely overlap.

[0072] Fig.11 The repair effect of inputting jump point data into the model is demonstrated. The model can automatically locate the jump point and repair it.

[0073] Fig.12 The repair effect of the gain data input model is demonstrated, and the model can repair the amplitude of the data gain with high accuracy.

[0074] Fig.13 The repair effect of the drift data input model is demonstrated. The model can locate the abnormal segment with high accuracy and repair it.

[0075] Fig.14 The repair effect of the model with different signal-to-noise ratio data input is demonstrated, and the model can greatly improve the signal-to-noise ratio of the noisy signal.

[0076] Fig.15 The repair effect of the model with different loss rate data input is demonstrated. Even at a loss rate of 40%, the lost data can be repaired with high accuracy.

[0077] The above is only a preferred specific embodiment of the present invention, but the protection scope of the present invention is not limited thereto. Any changes or substitutions that can be easily thought of by a person skilled in the art within the technical scope disclosed by the present invention should be included in the protection scope of the present invention. Therefore, the protection scope of the present invention shall be based on the protection scope of the claims.

Claims

1. A deep learning method for structural health monitoring acceleration data repair, characterized in that: The following steps are involved: Acquire normal acceleration data in structural health monitoring, and add different abnormal signal features to the normal data to obtain various types of abnormal data; the addition of the abnormal signal features is based on the generation of multiple simulated abnormal data, and the simulated situations include data jump point simulation, data gain simulation, trend item simulation, data drift simulation, precision degradation simulation and data loss simulation; Build a deep learning model based on acceleration data repair, use the data set composed of normal data and abnormal data as the input of the model, use the normal data matching the input data as the output of the model, and train and verify the model; Collect acceleration data from structural health monitoring and input it into a deep learning model for acceleration data-based repair; For normal data in the acceleration data, the model reconstructs the data and keeps the original data output. For abnormal data in the acceleration data, the model detects the abnormal type and location, corrects the abnormal data, reconstructs it into normal data, and then outputs it.

2. The deep learning method for structural health monitoring acceleration data repair according to claim 1, characterized in that: The data jump point simulation is to randomly select a number of data points on a data segment of unit length, and introduce random noise into the data points, which is set to n1 times of the amplitude, and 1.5<n1<8; The data gain simulation includes four cases: overall enlargement, overall reduction, local enlargement, and local reduction. For the case of overall enlargement, a gain coefficient of n2 times is randomly generated, and 1.5<n2<10; for the case of overall reduction, a gain coefficient of n3 times is randomly generated, and 0.02<n3<0.3; for the case of local gain, data of n4 times the length of the data segment is randomly selected for gain, and 0.2<n4<0.

9.

3. The deep learning method for structural health monitoring acceleration data repair according to claim 2, characterized in that: The trend item simulation is to randomly divide a data segment of unit length into segments, and add a linear height difference with an amplitude of n5 times the amplitude to each segment, including four cases of increasing, decreasing, and random positive and negative signs; a sinusoidal function with a random period, a random initial phase, and a random amplitude is further superimposed on each segment, where the period ranges from 0.5π to 2.5π, the random initial phase is between 0 and 2π, the random amplitude is between n5 times the amplitude, and 0.3<n5<2.

5.

4. The deep learning method for structural health monitoring acceleration data repair according to claim 3, characterized in that: The data drift simulation is to randomly select a data segment of n6 times the data length, and 0.1<n6<0.4; and generate a peak point height of n7 times the amplitude, and 0.5<n7<2.5; randomly determine the number of fitting functions of the local "trend term" and superimpose it on the original data; The precision reduction is simulated as adding random noise with a signal-to-noise ratio between -5 and 0 to the original signal; The data loss simulation is a randomly generated signal loss rate between 0.2 and 0.7, and is divided into two situations: continuous whole-segment data loss and discrete point data loss.

5. The deep learning method for structural health monitoring acceleration data repair according to claim 4, characterized in that: The simulation situation also includes normal data simulation, which means keeping the normal data unchanged.

6. The deep learning method for structural health monitoring acceleration data repair according to claim 1, characterized in that: The deep learning model for acceleration data restoration consists of four parts: block embedding module, encoder module, decoder module, and loss function; The block embedding module includes a reshaping operation module, a two-dimensional convolution module and a position encoding module, which is used to divide the input data into fixed-size data segments, expand them into a series of vector representations and perform linear projection transformations; the input data is reshaped and convolved in the block embedding module to become features of standardized size, and then sine-cosine position encoding is used to generate an invariant position information tensor, and the position information is attached to the feature by direct addition; The encoder module is composed of several Transformer modules, which include regularization layer, multi-head attention mechanism, residual structure and fully connected layer. The output of the block embedding module is layer-normalized and input into the multi-head attention mechanism. In the multi-head attention mechanism, the number of heads is set to k, and the features of the normalized size are passed through the fully connected layer and reshape operation to generate k groups of query, key and numerical features. The k groups of query, key and numerical features are respectively input into the scaled dot product attention module to obtain k groups of features. The output of each attention head is then connected to a two-layer feedforward neural network; the residual structure is nested in each sub-layer, including the attention mechanism layer and the multi-layer perceptron layer; The decoder module is composed of several Transformer modules. The output of the encoder module is first mapped by the fully connected layer, and then the sine and cosine position encoding is used to generate the invariant position information tensor, and the position information is attached to the feature. The decoder output features are again subjected to layer normalization, fully connected layers, and reshaping operations to output reconstructed data.

7. The deep learning method for structural health monitoring acceleration data repair according to claim 6, characterized in that: The loss function of the overall network consists of two parts, namely the average relative error of the reconstructed data And the correlation coefficient between the reconstructed data and the original data Where y is the original input data, a is the actual output of the network, and N is the length of each segment of data. is the mean of the actual output data of the network; is the mean of the original input data; i represents the i-th one, and n represents the number of samples; Finally, the overall loss function of the network is defined as:

8. The deep learning method for structural health monitoring acceleration data repair according to claim 1, characterized in that: In the normal data of structural health monitoring, the moving window scanning method is used to intercept the data, and the time series length and moving step length are set to 1024 to build the initial training set and test set; the mean and variance of the monitoring signals of each sensor in the initial training set are counted, and the Z-score normalization method is used to preprocess the data in the initial training set and the initial test set; For the initial training set and the initial test set, we randomly simulated the generation of several abnormal data to produce the input data of the deep learning model for acceleration data repair, and obtained the final training set and test set.

Images