SHM data reconstruction method driven by compressed sensing and enhanced context encoder

Abstract

The invention discloses an SHM data reconstruction method driven by compressed sensing and an enhanced context encoder. The method comprises the steps that a CS-ECE model used for SHM data reconstruction is constructed based on the compressed sensing and the basic theory of the context encoder; the CS-ECE model comprises a generator G and a discriminator D; sHM data in an in-service basic structure or mechanical equipment are collected and preprocessed, the preprocessed data are used for training the CS-ECE model, a generator G is trained to map an original signal x to a compressed signal y, and the compressed signal y is used for generating a compressed signal y; a reconstructed high-quality recovery signal # imgabs0 # discriminator D is used for executing a classification task to distinguish an original signal x and a reconstructed signal # imgabs1 #, and the SHM data is reconstructed by using a trained CS-ECE model; according to the method, a closed-loop solution is provided for reconstruction and recovery of missing data so as to guarantee the data quality, and the method is used for guaranteeing the SHM data quality.

Description

Technical Field

[0001] The present invention relates to the technical field of structural health monitoring, and in particular to a SHM data reconstruction method driven by a compressed sensing and enhanced context encoder. Background Art

[0002] The collection and storage of massive data have prompted the research of prognostics health management (PHM) to enter the era of big data. The condition monitoring data of infrastructure and mechanical equipment contains rich health status information, which provides new opportunities for structural health monitoring (SHM). Due to the complex environment of data collection, transmission and storage in modern SHM, the acquired data has five major characteristics: fast speed, large scale, low quality, mixed types and low density, which makes it difficult to extract health status information. In particular, low-quality data significantly reduces the reliability of SHM results.

[0003] Therefore, data quality assurance has become a new challenge for SHM in the context of big data. Generally speaking, data quality includes four aspects: accuracy, completeness, timeliness and consistency. Accuracy indicates whether the monitoring data can accurately reflect the response signal of the target object; completeness measures the continuity of the data, including whether the data recording cycle is complete and whether the data value is missing; timeliness describes the effectiveness of data acquisition in timely capturing the current health status of the structural equipment; consistency evaluates the consistency of data sources, formats and structures. In engineering practice, the reasons for the decline in data quality can be summarized as follows: 1) The service environment is harsh, and the infrastructure and mechanical equipment are susceptible to various interferences. In addition, irregular data collection, such as incorrect sensor calibration or improper range setting, may cause outliers, noise and other abnormal data, thereby reducing data accuracy. 2) Failure of data acquisition equipment, such as sensor detachment, CPU damage, network interruption, etc., may lead to data loss. During this period, the data will be recorded as "Null value", "NaN" or values ​​close to / equal to zero, which will affect the integrity of the data. 3) Due to congestion in the transmission network and software and hardware memory management, time delays in the data transmission process may affect the timeliness of the data. 4) Data from various stimulus sources are directly stacked without any preprocessing or classification, and changes in sampling frequency during the test may affect the consistency of the data.

[0004] In summary, some inevitable interferences seriously reduce data quality and increase the challenges associated with SHM research. Driven by big data, deep learning-based algorithms have shown unique advantages and breakthrough progress in SHM research. However, these methods cannot evaluate data quality and have the disadvantage of "garbage in, garbage out". Conclusions drawn from low-quality SHM data may be incorrect or even misleading. Therefore, data quality assurance research is very important for obtaining reliable SHM results. Some scholars have pointed out that data cleaning and data reconstruction are important research to improve data quality. In previous studies, an adaptive abnormal data detection method was proposed to perform data cleaning tasks. However, this data cleaning strategy alone cannot obtain comprehensive health information because the missing content may destroy the temporal structure of SHM data.

[0005] Traditional missing data recovery methods usually include regression, interpolation, fitting, and filling methods, which use the prior information of local data to extract the distribution pattern of global data and estimate the missing data based on some statistical features. However, these methods often need to solve a large number of high-dimensional equations, and even dimensional disasters may occur when processing big data, causing system crashes. In addition, these methods cannot analyze the coupling relationship between data structure and attributes, which may lead to poor recovery effects. Compression Sensing (CS) is a widely used accurate data reconstruction method that aims to extract important information of the original signal from a small number of randomly sampled samples. In other words, CS can reconstruct incomplete missing data into complete original signals. Therefore, CS has received widespread attention and application in SHM data analysis. Bao et al. conducted an in-depth study on the problem of accurate data recovery of civil structure acceleration signals based on CS. They discussed the key factors affecting data reconstruction performance through several practical cases and pointed out the problems that need to be solved in future research. For acoustic emission signals, Song et al. proposed an improved CS-based sparsity adaptive matching pursuit reconstruction method, which provides useful guidance for railway rail PHM. Cheng and Lu studied road roughness signals and developed two CS algorithms based on the improved simulated annealing algorithm and the golden section method for data reconstruction. For ultrasonic signals, Sawant et al. proposed a missing data reconstruction performance evaluation framework based on CS, which extended CS to the field of lossless PHM. In addition, Li et al. quantitatively demonstrated the effectiveness and robustness of CS in reconstructing electromechanical admittance signals at the theoretical and experimental levels. For wind speed signals in nature, Wan et al. conducted a CS-based PHM study on large-scale spatial structures using a time-shift strategy. It can be seen that CS has been successfully explored in various fields. However, the actual implementation of CS on real data is limited by strict randomness and sparsity assumptions and expensive reconstruction processes. Since the service environment of infrastructure and mechanical equipment is often harsh, the actual data is usually not sparse enough. The performance of traditional domain transformation methods is poor when performing signal sparse decomposition. In addition, most SHM data reconstruction scenarios cannot meet the requirements of random sampling because the missing data to be reconstructed are generally continuously distributed in the measured signal, rather than randomly distributed. In addition, the reconstruction process is usually completed by complex iterative calculations, which is very time-consuming. For example, it takes more than 100 seconds to reconstruct a signal with a length of 3000 points using the CS method. Therefore, the CS-based method cannot achieve accurate, fast and robust data reconstruction, especially for high-frequency SHM data.

[0006] Fortunately, due to the superior end-to-end learning ability of deep learning (DL), feature mapping patterns can be directly obtained from known low-resolution and high-resolution sample sets. This can eliminate the need to rely on the sparsity of unknown samples and eliminate the reliance on random sampling and sparsity. For example, Kulkarni et al. developed a DL-based ReconNet architecture that takes CS measurements as input and outputs reconstruction results in a non-iterative manner. Experiments show that ReconNet has better reconstruction quality and running efficiency than the CS algorithm. They further proposed a variant architecture of ReconNet, using adversarial loss to improve reconstruction quality. Similarly, Bora et al. proposed a CS scheme based on a generative model, which has better reconstruction performance than existing models. Among these deep learning networks, Generative Adversarial Networks (GANs) can obtain very detailed high-frequency textures, thereby achieving the best restoration effect. A typical example is Context Encoders (CE), which is a feature learning method based on contextual pixel prediction and performs well in restoring missing data. There are also some reconstruction models that have emerged in other fields, which can be considered as further applications and extensions based on CE. However, most of the above literature focuses on images, but does not address one-dimensional signals such as SHM data. In addition, methods used in other fields to recover one-dimensional signals may not be applicable to SHM data. One of the reasons is that these methods often only obtain good results in the time domain, while ignoring the importance of frequency domain information. Given the advantages of CE-based CS in data reconstruction in various fields, and their practical limitations in one-dimensional signals.

[0007] In summary, reconstructing and restoring missing data to improve data quality is of great significance for structural health monitoring (SHM). Existing missing data reconstruction methods are usually limited by strict mathematical assumptions, such as signal sparsity and random sampling process, coupled with high computational costs, making them poorly adaptable to SHM data reconstruction tasks. Summary of the invention

[0008] In order to solve the problems existing in the prior art, the purpose of the present invention is to provide a SHM data reconstruction method driven by a compressed sensing and enhanced context encoder. The present invention aims at reconstructing and recovering missing data to provide a closed-loop solution to the application-based problem of data quality assurance, which is used to ensure the quality of SHM data.

[0009] To achieve the above object, the technical solution adopted by the present invention is: a SHM data reconstruction method driven by a compressed sensing and enhanced context encoder, comprising the following steps:

[0010] Step 1: construct a CS-ECE model for SHM data reconstruction based on the basic theories of compressed sensing and context encoder; the CS-ECE model includes a generator G and a discriminator D;

[0011] Step 2: Collect SHM data from in-service infrastructure or mechanical equipment, preprocess the data, and use the preprocessed data to train the CS-ECE model, wherein the generator G is trained to map the original signal x to the compressed signal y to reconstruct a high-quality recovery signal The discriminator D is used to perform the classification task to distinguish the original signal x from the reconstructed signal

[0012] Step 3: Use the trained CS-ECE model to reconstruct the SHM data.

[0013] As a further improvement of the present invention, the generator G includes an encoder having multiple stacked convolutional layers and a decoder having multiple corresponding transposed convolutional layers, a channel fully connected layer is arranged between the encoder and the decoder, and an activation function layer and a batch normalization layer are arranged after each convolution or transposed convolution layer; the stacked convolutional layer of the encoder is used to extract high-level abstract features from the input compressed signal y, and then establish the features of the predicted object through the combined decoder; the channel fully connected layer is used to enable the convolutional layer to only establish connections between all feature maps, but not to link all nodes within a given feature map, and to propagate information across channels in the activation of each feature map with fewer parameters.

[0014] As a further improvement of the present invention, the discriminator D comprises five convolutional layers, followed by a flattening layer and two fully connected layers; the discriminator D receives the reconstructed signal Or the original signal x as input, using the output with Sigmoid activation for binary classification to determine whether each input is a reconstructed signal generated by G It is still the real signal x.

[0015] As a further improvement of the present invention, in step 2, preprocessing the data includes outlier cleaning, cutting, place-taking and masking.

[0016] As a further improvement of the present invention, in the CS-ECE model, the loss function of the discriminator D is the cross entropy between the original signal x and the restored signal G(y):

[0017]

[0018] By training the discriminator G, we generate the reconstructed signal The loss function is:

[0019]

[0020] The mean square error MSE is used as an additional loss function of the discriminator G to improve the reconstruction accuracy as follows:

[0021]

[0022] A loss function that can evaluate frequency characteristics is introduced, called focal frequency loss FFL. FFL allows the weights corresponding to hard frequencies to be adaptively enhanced through the weight vector ω, as follows:

[0023]

[0024] The Hadamard product is used to combine the weight vector ω and the differentiable Euclidean distance with x in the frequency domain, and the FFL expression shown below is defined:

[0025]

[0026] Where α is an adjustable scaling factor, and f() is the Fourier transform;

[0027] The final joint loss function of the discriminator G is obtained as follows:

[0028]

[0029] Among them, λ1, λ2, and λ3 are the weights corresponding to different loss terms.

[0030] As a further improvement of the present invention, it also includes:

[0031] Use gradient descent method to alternately maximize the loss function L D And minimize the loss function L G To optimize the CS-ECE model.

[0032] The beneficial effects of the present invention are:

[0033] 1. In order to improve the quality of SHM data, a missing data reconstruction method based on CS and context encoder (CS-ECE) is proposed. It can reconstruct the detailed features of the target signal in the time and frequency domains without being restricted by the mathematical assumptions of sparsity and random sampling.

[0034] 2. The proposed CS-ECE can directly learn the feature mapping relationship from the compressed signal and the original signal samples in an adversarial manner. It is a completely feed-forward algorithm that does not require any iteration and optimization calculation, greatly speeding up the CS reconstruction process.

[0035] 3. A joint loss function consisting of reconstruction loss for CS tasks, time domain waveform loss for SHM data, focal frequency domain loss and adversarial loss is designed, so that the CS-ECE model can provide more realistic and accurate restoration results. BRIEF DESCRIPTION OF THE DRAWINGS

[0036] Figure 1 This is a CS-ECE model framework diagram in an embodiment of the present invention;

[0037] Figure 2 The network architecture and training diagram of the CS-ECE model in an embodiment of the present invention;

[0038] Figure 3 It is a schematic diagram of a field test in an embodiment of the present invention;

[0039] Figure 4 A schematic diagram of a portion of axle box acceleration (ABA) and wheel roughness data collected in an embodiment of the present invention;

[0040] Figure 5 It is a time-frequency domain schematic diagram of a sample data of axle box acceleration in an embodiment of the present invention;

[0041] Figure 6 It is a characteristic curve diagram of each index D and G with the number of training rounds in the embodiment of the present invention;

[0042] Figure 7 This is a schematic diagram of the time domain reconstruction results generated by G under different numbers of training rounds in an embodiment of the present invention;

[0043] Figure 8 This is a schematic diagram of the frequency domain reconstruction results generated by G under different numbers of training rounds in an embodiment of the present invention;

[0044] Fig. 9 It is a schematic diagram of improving the quality of axle box acceleration data of the CS-ECE model in an embodiment of the present invention;

[0045] Fig.10 Schematic diagram of experimental results of CS-ECE ablation experiment in an embodiment of the present invention;

[0046] Fig.11 is a schematic diagram of vehicle body vibration acceleration data collected under normal and interference conditions in an embodiment of the present invention;

[0047] Fig.12 The time domain and frequency domain diagrams of the reconstructed signal in the embodiment of the present invention;

[0048] Fig.13 Schematic diagram of data reconstruction performance based on a classic model in an embodiment of the present invention. DETAILED DESCRIPTION

[0049] The embodiments of the present invention are described in detail below with reference to the accompanying drawings.

[0050] Example

[0051] A SHM data reconstruction method driven by a compressed sensing and enhanced context encoder, comprising:

[0052] 1. Basic theory of compressed sensing:

[0053] The following formula can illustrate the basic theory of CS

[0054] y=Φx=ΦΨs=As (1)

[0055] in, is the original signal to be reconstructed, which satisfies the classic Nyquist-Shannon sampling theorem. To measure the signal. They are the measurement matrix, sparse transformation matrix and sensing matrix respectively. Represents sparse coefficients. According to CS theory, if the signal x is sparse, or the transformed signal s obtained by applying the transformation basis Ψ to x has sufficient sparsity, the original signal x can be downsampled by the measurement matrix Φ first, and then the original signal x can be reconstructed as accurately as possible by solving the optimization problem with the measured signal y, as shown below:

[0056] min||x||0 st y=Φx=As (2)

[0057] Where ||·||0 is the 0-norm, that is, the number of non-zero terms. Theoretically, solving the above equation is an NP-hard problem. At present, enumeration is often used to solve such equations, that is, exhausting all possible non-zero value combinations in the original signal. It can be seen that the amount of calculation is huge and it is not suitable for direct calculation. Therefore, this type of problem is gradually transformed into an approximate equivalent problem of finding the minimum 1-norm, as shown below:

[0058] min||x||1 st y=Φx=As (3)

[0059] Since the above equation is a convex optimization problem, it can be solved directly. It should be noted that the sensing matrix A must satisfy the restricted isometry property condition. However, in the actual SHM missing data reconstruction, it is difficult to meet this requirement.

[0060] 2. Basic theory of context encoder:

[0061] The overall architecture of the classic CE is an intuitive encoder-decoder structure. The encoder takes the incomplete signal as input data and outputs the potential feature information of the incomplete signal. The decoder accepts the feature information and generates the missing area.

[0062] The encoder is derived from the famous AlexNet architecture. This encoder network is specifically constructed for context prediction tasks rather than the traditional ImageNet classification task. In addition, to ensure effective propagation of information, a channel-wise fully connected layer is utilized, allowing direct connections between all activations.

[0063] Channel fully connected layer This layer is designed to facilitate the propagation of information in each feature map activation. Unlike the fully connected layer, it only passes information in the graph without linking parameters from different feature maps. Therefore, unlike the fully connected layer, 2 p 4 Compared with the number of parameters, the number of parameters in this channel form is op 4 (o feature maps of size p×p).

[0064] The decoder contains several upconvolutional layers, each with feature learning filters and a ReLU activation function. Upconvolution operations can produce higher dimensions or resolutions. These upconvolution operations and nonlinear weighted upsampling can make the feature maps close to the original target dimensions.

[0065] 3. SHM data reconstruction model framework:

[0066] Figure 1 The framework of the CS-ECE model is demonstrated, which has better reconstruction accuracy and operation efficiency compared with the classic DL-based and sparsity-based CS methods.

[0067] 3.1 Data preprocessing:

[0068] The SHM data input to CS-ECE are usually collected from in-service infrastructure or mechanical equipment. The data volume is usually large, and various information is coupled and related to each other. Direct analysis of these data will cause the operation calculation to crash. Therefore, data preprocessing is essential. The CS-ECE preprocessing strategy for SHM data quality assurance mainly includes outlier cleaning, cutting, placeholders and masks.

[0069] Cutting: A sliding window of 1024 data points is used to divide the outlier-cleaned data into a series of segments. It should be noted that the window length can be adjusted adaptively according to different requirements.

[0070] Placeholder: It is known that there is cleaned data, that is, compressed signal y, which is a part of the original signal x (y<<x), and the length of y is determined by the compression ratio. It is necessary to fill y with appropriate placeholders to ensure that y and x meet the same scale. Here, zero-valued data can be added to the corresponding positions of the cleaned data, which helps CS-ECE to flexibly handle different data characteristics.

[0071] Mask: CS-ECE may over-learn placeholder features, so it is important to emphasize to the model the distribution of the placeholder regions in y. Therefore, a Boolean mask sequence is set and attached to CS-ECE as constraint information. In this sequence, any null value index means that the feature information at the corresponding position should be ignored during the calculation process.

[0072] 3.2 CS-ECE Architecture

[0073] After preprocessing the data, the structured annoying dataset can be provided to CS-ECE for training. The architecture of CS-ECE consists of two DL networks: the generator G and the discriminator D. G is trained to map the original signal x to the compressed signal y to reconstruct a high-quality restored signal Instead, D performs a classification task to distinguish between the original signal x and the reconstructed signal The training process of CS-ECE is as follows Figure 2 As shown. The generator G is an improved enhanced CE. It consists of two modules: an encoder with five stacked convolutional layers and a decoder with five corresponding transposed convolutional layers. An activation function layer and a batch normalization layer are set after each convolution or transposed convolution layer. The architecture of G is designed to extract high-level abstract features from the input y using the stacked convolutional layers of the encoder, and then establish the features of the predicted object by combining the decoder. The channel fully connected layer has two purposes. First, as a link between G and D, it can overcome the limitation of the convolutional layer, that is, the convolutional layer only establishes connections between all feature maps, but cannot link all nodes within a given feature map. Second, it can propagate information across channels in the activations of each feature map with fewer parameters. In addition, unlike the classic CE, the proposed G generates The dimension of G and D is consistent with that of x, and it does not only output data about the missing area. This greatly enhances the continuity of the final reconstructed signal. Table 1 summarizes the parameter settings of G and D. Figure 2 The structure of the discriminator D is also shown. Specifically, D contains five convolutional layers, followed by a flatten layer and two fully connected layers. D receives the reconstructed signal Or the original signal x as input, using the output with Sigmoid activation for binary classification to determine whether each input is a reconstructed signal generated by G It is still the real signal x. Table 1 gives the specific structure and parameters of G and D.

[0074] Table 1 Structure and parameter description of G and D

[0075]

[0076] Conv: convolution; L-Re: Leaky ReLU activation function; BN: batch normalization; Re: ReLU activation function. TConv: transposed convolution

[0077] 3.3. Joint loss function:

[0078] After defining the architecture of G and D, this embodiment proposes a joint loss function. In CS-ECE, the loss function of D is the cross entropy between the original signal x and the restored signal G(y).

[0079]

[0080] By training G, we can generate a reconstructed signal that is realistic enough to be mistaken for a fake. Even D can not distinguish The difference between the real signal x.

[0081]

[0082] Although the above formula can give a reconstruction result that looks more realistic, it may not generate a signal with sufficient accuracy. The error between and x helps improve the accuracy. Therefore, the mean square error (MSE) is used as an additional loss function of G to improve the reconstruction accuracy, as shown below

[0083]

[0084] Due to its simple calculation and easy implementation, MSE is often used in regression tasks in deep learning. By minimizing the MSE, G is forced to generate a reconstructed signal that is close to the true signal x. until For SHM data, equation (6) can be used to evaluate the error in the time domain or spatial domain, but it ignores important frequency domain features. On the other hand, due to the inherent properties of the DL model, G tends to avoid frequency components that are difficult to fit, namely hard frequencies, and converges to a suboptimal point (non-global optimal point). To solve this problem, this embodiment introduces a loss function that can evaluate frequency characteristics, called focal frequency loss (FFL), which is inspired by focal loss and hard sample mining. FFL allows the weights corresponding to hard frequencies to be adaptively enhanced through the weight vector ω, as defined in equation (7), which is adaptively determined based on the characteristic loss of each frequency during the training process. It can force the generator G to quickly focus on the hard frequency and gradually enhance the frequency details of the generated signal to obtain better recovery effects. The Hadamard product is used to combine the weight vector ω and the differentiable Euclidean distance with x in the frequency domain. This embodiment defines the FFL expression shown in equation (8).

[0085]

[0086] Where α is an adjustable scaling factor, which is recommended to be set to 1. f() is the Fourier transform. In order to obtain reliable results, both the amplitude and phase information of the Fourier transformed data should be taken into account. Otherwise, the generated signal may be accompanied by irregular textures or noise interference.

[0087] The final joint loss function of G is:

[0088]

[0089] Where λ1, λ2, and λ3 are weights corresponding to different loss terms. This embodiment adopts the following settings: λ1 = 0.01, λ2 = 0.9, λ3 = 0.09, so that the sizes of different loss terms have similar scales.

[0090] 3.4、CS-ECE adversarial training:

[0091] Use gradient descent method to alternately maximize the loss function L D And minimize the loss function L G To optimize CS-ECE, RMSProp with a learning rate of 0.0002 was used to speed up the convergence during training.

[0092] 4. Practical application cases:

[0093] Due to the complex service environment and harsh operating conditions, the quality of high-speed rail monitoring data often declines, which poses a major challenge to high-speed rail health monitoring. Therefore, the feasibility and superiority of the proposed method are verified by taking the SHM data of high-speed trains as an example. The tested high-speed rail runs on a line from Chengdu to Xingwen in China. Figure 3 The schematic diagram of the field test is shown in Figure 1. Figure 3 Among them, (a) is the Chengdu-Xingwen high-speed railway line, (b) is the axle box vibration acceleration test, (c) is the wheel roughness test, (d) is the vehicle body vibration acceleration test, and (e) is the vehicle body vibration acceleration test interference.

[0094] 4.1. Reconstruction of missing data of axle box vibration acceleration:

[0095] Figure 4 A portion of the collected axle box acceleration (ABA) and wheel roughness data is shown, where (a) is the measured data of axle box acceleration, (b) is the polar coordinates of wheel roughness, and (c) is the measured data in order coordinates.

[0096] It can be seen that there are some abnormal data in ABA, and the 24th and 25th order wheel polygons are worn significantly. The measured ABA data is divided into a sample set of 10000*1024. 8000 samples are used for training, and the remaining 2000 samples (including abnormal data) are used for testing. Figure 5 As shown, Figure 5 In the figure, (a) is the time domain and (b) is the frequency domain. It can be seen that the ABA signal is not strictly sparse in both the time domain and the frequency domain. This is a key factor that limits the practical use of traditional sparse-based CS models. For the MHM data quality assurance study, the abnormal data detection method is first used to clean up the outliers. Then, the cleaned data (missing data) is obtained, that is, the compressed signal y. It should be noted that the specific manifestation of the abnormal data is random, that is, in the SHM data, it does not appear at a fixed position and scale. In order to simulate this situation, this embodiment constructs a structural element containing 205 zero values ​​and randomly writes it into each sample set x to obtain a data cleaning result y with a compression rate of 20%. Therefore, by reconstructing the signal By comparing with the original signal x, the performance of the proposed CS-ECE can be quantitatively evaluated. Figure 6 The curves of the various items in D and G as the number of training rounds is shown. It describes that the loss of the generator G and the discriminator D has undergone a fluctuating adversarial evolution, indicating that there is a continuous competition between D and G. In addition to the loss term, this embodiment uses D acc To describe the discriminator D and x, and use the R-squared coefficient (R 2 ) to evaluate The reconstruction quality between and x is defined as:

[0097]

[0098] in is the average value of x, R 2The closer it is to 1, the higher the degree of fit and the more accurate the recovery result. The ideal training process is: each loss term gradually decreases until it approaches a minimum approximation value; D acc Converges to about 0.5, that is, the samples generated by G are too real and D cannot correctly identify them; R 2 It keeps approaching 1, which means that the generated signal is gradually consistent with the real signal. Figure 6 The R2 in the final convergence is 0.976. Figure 7 The output of G at different training stages is shown for a sample. Initially, G generates a completely meaningless noise. As training progresses, the reconstructed signal Gradually approach the target signal x, and finally obtain true reconstruction under sufficient training conditions. Figure 8 Describes Figure 7 The frequency characteristics of the signal shown in Figure 2 show that there is not only a clear difference in the dominant frequency between the original signal x and the compressed signal y, but also a significant difference in the corresponding energy. G-FFL Advantages of reconstructing signals The frequency characteristics of gradually approach the target signal x, ensuring its authenticity and accuracy. Figure 4 Abnormal data in . Fig. 9 is the recovery result of an abnormal sample. It can be seen that not only the time domain waveform of the measured ABA signal has a significant deviation from the normal level, but also its frequency characteristics cannot reflect the 24th and 25th order wheel polygon features. Fortunately, the trained CS-ECE can significantly improve the quality of the ABA signal. This method eliminates the distortion reflected in the time domain and can accurately derive the order information of the wheel polygon based on the main frequency information.

[0099] 4.2 Ablation test:

[0100] To further illustrate the superiority of the proposed CS-ECE, three pseudo-CS-ECE models are constructed and three sets of ablation tests are performed to evaluate the advantages of the proposed joint loss function. Table 2 shows the main details and differences of the three pseudo-CS-ECE models. 2 To evaluate the error level and restoration quality of the reconstructed data, each ablation study was performed 50 times to minimize random interference. Fig.10 The black dots in the figure represent the MSE and R of the CS-ECE and non-pseudo-CS-ECE models. 2 Therefore, the range of the horizontal and vertical axes is [0, 1]. The significance of the red line (coordinate (0,0)(1,1)) is to intuitively show the performance difference between CS-ECE and the comparison model. 2 The black dots in the figure are distributed above the red line, which means that the R 2The value is greater than the R of the comparison model 2 Conversely, if the point in the MSE graph is below the red line, it means that the MSE value of CS-ECE is smaller than the MSE value of the comparison model. As can be seen from Figure 10, most points are significantly higher than (R 2 ) or below the (MSE) red line. Fig.10 The distances between the points in v2 and the red line are relatively close, which indicates that L G-FFL It plays a significant role in reconstructing the details of MHM data. Ablation studies have shown that the proposed CS-ECE model with a joint loss function of time and frequency domain information is very beneficial to the data reconstruction task.

[0101] Table 2 Three pseudo CS-ECE models

[0102]

[0103] 4.3. Reconstruction of missing data of vehicle body vibration acceleration:

[0104] The performance of this method in practical applications is further verified through specific experiments. Figure 3 In (c) and (d), accelerometer A is installed on the floor of the vehicle body to collect the acceleration signal of the vehicle body. During the test, the sensor connector is separated and then quickly reconnected to simulate a sudden disturbance. At the same time, another accelerometer B is installed next to sensor A and records data as a benchmark during the entire test. Fig.11 (a) and (b) show the recorded data of the two sensors under normal and interference conditions, respectively. The waveforms of the two sensors look almost the same under normal conditions, which shows the feasibility of using sensor B data to verify the recovery effect of sensor A data. Fig.12 Shows Fig.11 The data recovery result of sensor A in (b) is shown in Figure 2. The data processed by the CS-ECE model is consistent with the reference signal, and the characteristic frequency of about 1 Hz can be observed. The experiment proves that the proposed data quality assurance method can be applied to practical engineering problems.

[0105] 4.4, Comparative analysis:

[0106] At present, the most popular reconstruction methods in the field of compressed sensing research can be divided into two categories: greedy iterative algorithms and convex optimization algorithms. The former reconstructs the original signal by iteratively searching the support set of sparse vectors and using least squares estimation. Its classic representative algorithm is orthogonal matching pursuit (OMP), and its derivative algorithms such as sparsity adaptive matching pursuit (SAMP) and regularized orthogonal matching pursuit (ROMP); gradient descent method is often used to solve constrained optimal problems in convex optimization, among which basis pursuit (BP) is the most typical method. The above four algorithms are used to process vehicle acceleration signals. Fig.13The results shown show that among the greedy iterative methods, SAMP has the highest restoration quality, but it also consumes the most time. ROMP is more efficient than OMP, but the reconstruction quality is poor. It should be noted that the time cost of BP is very high, but the results obtained are not the best. The reason is that the computational complexity of convex optimization is much higher than that of the greedy iterative algorithm under the same conditions. In addition to the classical methods, the current five representative state-of-the-art CS algorithms are also analyzed. Table 3 shows the corresponding results. Comparative analysis proves that the CS-ECE proposed in this embodiment can obtain better results, especially in terms of time cost, and is at least two orders of magnitude faster than the classical method. This has great advantages for the timeliness of online structural health monitoring. (All calculations in this embodiment are based on a computing platform. Hardware: Intel(R) Xeon(R) W-2265CPU@3.5GHz, RAM 64GB, GPU NVIDIARTX A5000. Software: Windows 10, Anaconda 2020.11, conda 4.9.2, Python 3.8.5, Tensorflow-GPU 2.5.0)

[0107] Table 3 Comparison with some of the most advanced algorithms

[0108]

[0109] In order to obtain more reliable SHM results, this embodiment develops an algorithm based on compressed sensing and enhanced context encoder (CS-ECE) to reconstruct missing data. Whether in the time domain or the frequency domain, CS-ECE can reconstruct the detailed features of the missing signal. The effectiveness of the proposed method is demonstrated using measured data collected on high-speed trains.

[0110] The above-mentioned embodiments only express the specific implementation of the present invention, and the description thereof is relatively specific and detailed, but it cannot be understood as limiting the scope of the present invention. It should be pointed out that, for ordinary technicians in this field, several variations and improvements can be made without departing from the concept of the present invention, which all belong to the protection scope of the present invention.

Claims

1. A SHM data reconstruction method driven by compressed sensing and enhanced context encoder, characterized in that: The following steps are involved: Step 1: construct a CS-ECE model for SHM data reconstruction based on the basic theories of compressed sensing and context encoder; the CS-ECE model includes a generator G and a discriminator D; Step 2: Collect SHM data from in-service infrastructure or mechanical equipment, preprocess the data, and use the preprocessed data to train the CS-ECE model, wherein the generator G is trained to map the original signal x to the compressed signal y to reconstruct a high-quality recovery signal. The discriminator D is used to perform the classification task to distinguish the original signal x from the reconstructed signal Step 3: Use the trained CS-ECE model to reconstruct the SHM data.

2. The SHM data reconstruction method driven by compressed sensing and enhanced context encoder according to claim 1, characterized in that: The generator G includes an encoder having multiple stacked convolutional layers and a decoder having multiple corresponding transposed convolutional layers, a channel fully connected layer is arranged between the encoder and the decoder, and an activation function layer and a batch normalization layer are arranged after each convolution or transposed convolution layer; the stacked convolutional layer of the encoder is used to extract high-level abstract features from the input compressed signal y, and then establish the features of the predicted object through the combined decoder; the channel fully connected layer is used to enable the convolutional layer to only establish connections between all feature maps, but not to link all nodes within a given feature map, and to propagate information across channels in the activation of each feature map with fewer parameters.

3. The SHM data reconstruction method driven by compressed sensing and enhanced context encoder according to claim 2 is characterized in that: The discriminator D includes five convolutional layers, followed by a flattening layer and two fully connected layers; the discriminator D receives the reconstructed signal Or the original signal x as input, using the output with Sigmoid activation for binary classification to determine whether each input is a reconstructed signal generated by G It is still the real signal x.

4. The SHM data reconstruction method driven by compressed sensing and enhanced context encoder according to claim 1, characterized in that: In step 2, the data is preprocessed including outlier cleaning, cutting, placeholders, and masks.

5. The SHM data reconstruction method driven by compressed sensing and enhanced context encoder according to claim 3 is characterized in that: In the CS-ECE model, the loss function of the discriminator D is the cross entropy between the original signal x and the restored signal G(y): By training the discriminator G, we generate the reconstructed signal The loss function is: The mean square error MSE is used as an additional loss function of the discriminator G to improve the reconstruction accuracy as follows: A loss function that can evaluate frequency characteristics is introduced, called focal frequency loss FFL. FFL allows the weights corresponding to hard frequencies to be adaptively enhanced through the weight vector ω, as follows: The Hadamard product is used to combine the weight vector ω and the differentiable Euclidean distance with x in the frequency domain, and the FFL expression shown below is defined: Where α is an adjustable scaling factor, and f() is the Fourier transform; The final joint loss function of the discriminator G is obtained as follows: Among them, λ1, λ2, and λ3 are the weights corresponding to different loss terms.

6. The SHM data reconstruction method driven by compressed sensing and enhanced context encoder according to claim 5, characterized in that: Also includes: Use gradient descent method to alternately maximize the loss function L D And minimize the loss function L G To optimize the CS-ECE model.

Images