Structural health monitoring abnormal data reconstruction method based on submerged space diffusion model

Abstract

The invention discloses a structural health monitoring abnormal data reconstruction method based on a submerged space diffusion model, and belongs to the technical field of structural health monitoring. The method comprises the following steps: firstly, converting one-dimensional structure monitoring time sequence data into a two-dimensional image, mapping the two-dimensional image to a low-dimensional potential space by using a pre-trained variational auto-encoder, and obtaining a potential variable containing semantic information of an original signal; then, defining a forward diffusion process in the submerged space, and establishing an evolution path from the structured features to Gaussian noise; constructing a conditional diffusion U-Net network containing time step embedding and frequency domain conditional coding, and training the network through a mask region selectivity mechanism to learn noise inversion distribution; in a reverse generation stage, a known area forced alignment strategy and a bidirectional resampling mechanism are introduced, and a soft mask smoothing technology is combined. The invention aims to obtain a high-quality time sequence signal which can meet the requirements of subsequent modal parameter identification and long-term performance evaluation by using a generative probability inversion mechanism.

Description

Technical Field

[0001] This invention belongs to the fields of deep learning, data reconstruction, and structural health monitoring technology, and specifically relates to a method for reconstructing structural health monitoring data based on a latent space diffusion model. Background Technology

[0002] During the service life of large bridges and other civil engineering infrastructure, structural health monitoring systems typically require long-term, continuous acquisition of various response data, including acceleration, strain, and displacement, to support tasks such as modal parameter identification, damage early warning, and long-term performance evaluation. However, as a typical data-intensive system, structural health monitoring is inevitably affected by factors such as sensor aging and drift, acquisition equipment failure, communication link packet loss and jitter, external electromagnetic interference, and extreme environmental effects, leading to noise pollution, abnormal segments, and large continuous gaps in the monitoring sequence. On the one hand, engineering practice often employs anomaly removal and quality control strategies to avoid abnormal data interfering with subsequent analysis; but on the other hand, the data gaps after removal directly disrupt the temporal continuity and frequency domain integrity of the monitoring sequence, causing deviations in the identification and evaluation results based on long-term window statistical characteristics or frequency domain characteristics, thus affecting the reliability of structural condition assessment.

[0003] To address the problem of repairing missing and anomalous gaps in monitoring data, common solutions in existing technologies mainly include interpolation, sparse recovery, and matrix filling. These methods have certain advantages in handling sparse or locally smoothed signals. However, given the multi-scale, non-stationary, and nonlinear dynamic characteristics of structural responses, they struggle to simultaneously characterize long-range temporal dependencies, preserve high-frequency texture details, and restore complex dynamic consistency. This can easily lead to problems such as over-smoothing, spectral distortion, or loss of physical consistency, thus limiting their applicability and reconstruction accuracy in scenarios with high missing rates, segmented continuous missing data, and multiple sensor types.

[0004] In recent years, generative models have provided new pathways for modeling complex data distributions. In particular, diffusion models, through a probabilistic inversion mechanism of progressive noise addition and denoising, can learn data distributions incrementally and achieve high-fidelity generation and restoration. Simultaneously, placing the diffusion process in a low-dimensional latent space and combining it with conditional constraints promises to enhance the preservation of key features of structural responses while reducing computational complexity. However, several key challenges remain for the task of reconstructing missing structural health monitoring data: for example, how to perform targeted restoration of missing segments while ensuring that healthy segments remain undisturbed; how to suppress abrupt changes and artifacts at mask boundaries and achieve consistent reconstruction across scales; and how to establish a clearer constraint relationship between the inversion process and the global dynamic characteristics of the structural response to improve the engineering usability and physical plausibility of the reconstruction results. Summary of the Invention

[0005] This invention provides a method for reconstructing structural health monitoring anomaly data based on a latent space diffusion model. The aim is to use a generative probabilistic inversion mechanism to obtain high-quality time-series signals that can meet the requirements of subsequent modal parameter identification and long-term performance evaluation.

[0006] This invention is achieved through the following technical solution: A method for reconstructing structural health monitoring anomaly data based on a latent space diffusion model, the method comprising the following steps: Step 1: Convert the one-dimensional structure monitoring time series data into a two-dimensional image. Use a pre-trained variational autoencoder as a surrogate model to map the two-dimensional image to a low-dimensional latent space and obtain latent variables containing semantic information of the original data. Step 2: Define the forward diffusion process in the latent space, and gradually inject Gaussian noise into the latent variables through a noise scheduling sequence to establish a Markov chain evolution path from the true latent representation to the standard Gaussian noise distribution. Step 3: Construct a conditional diffusion U-Net network that includes time-step embedding and frequency-domain conditional coding. Introduce the frequency-domain conditional coding mechanism to impose explicit global dynamic constraints on the diffusion process, and combine the mask matrix and known region features as network inputs. Step 4: Employ a mask region selection mechanism to calculate noise prediction loss only in missing regions, and construct the total loss function by combining the latent space distribution regularization term to jointly optimize the model's noise inversion capability in abnormal regions and the statistical consistency of the latent space distribution. Step 5: Perform the back diffusion process, starting with random noise and gradually denoising using the trained conditional diffusion network; during the sampling process, introduce a known region forced alignment strategy and a bidirectional resampling mechanism to ensure that the latent variables of the known segments conform to the original distribution and coordinate the distribution of the boundary regions. Step 6: Input the complete latent variables generated by backdiffusion into the decoder of the variational autoencoder to restore them into a two-dimensional image, and then restore them into a one-dimensional structure monitoring time series signal through image-to-time backprojection.

[0007] Furthermore, the reconstruction method further includes: Specifically, step one involves denoting the image representation corresponding to the monitored time-series signal as a tensor. The encoder using a variational autoencoder will Mapping to latent variables ; Step two specifically involves setting a noise scheduling sequence. Arbitrary time steps are generated through a forward diffusion process. Noisy latent variables ; Step three specifically involves constructing a conditional diffusion U-Net, with the network input consisting of noisy latent variables. Mask matrix and known regional latent variables It is assembled by splicing together the frequency domain feature vectors of the monitoring data and mapping them to latent space semantic embeddings. Inject into the network; Step four specifically involves generating training samples using the random stripe masking method and calculating the network prediction noise. With real noise The mask loss and KL divergence regularization terms are used to update the network parameters through backpropagation; Step five specifically involves using standard Gaussian noise. Begin reverse iterative sampling at each step. In the process, a known region forced alignment strategy is used to correct latent variables, and a bidirectional resampling mechanism is executed for boundary coordination until the desired result is generated. ; Step six specifically involves generating... The decoder of the input variational autoencoder obtains the reconstructed image. And convert it back to a timing signal.

[0008] Furthermore, the input structure and attention mechanism of the conditional diffusion denoising network are defined as follows: Multi-information joint input mechanism: Network input channel tensor Noisy latent variables at the current time step Binary mask matrix and known regional latent variables It is assembled from pieces and defined as:

[0009] Frequency domain conditional coding and cross-attention mechanism: Frequency domain feature mapping: mapping the input frequency domain feature vector Mapped to conditional embedding vector :

[0010] In the formula, For multilayer perceptron networks; cross-attention calculation: conditional embedding vectors are injected into U-Net features, and the attention formula is:

[0011] In the formula, the query matrix Features derived from latent space, key matrix Sum matrix From conditional embedding vectors Obtained by linear projection. For the attention dimension.

[0012] Furthermore, the loss function of the forward diffusion process and the conditional diffusion denoising network is defined as follows: During the forward diffusion process, at any time... latent variables Compared with the initial sample The relationship is:

[0013] In the formula, , This is the cumulative term of the noise attenuation coefficient; The mapping form of the frequency domain conditional coding mechanism is as follows:

[0014] In the formula, For the normalized logarithmic power spectrum, It is a multilayer perceptron network; Total loss function The formula is:

[0015] In the formula, The latent space binary mask matrix, For element-wise multiplication, To predict noise for the network, For the latent space distribution KL regularization term, and These are the weighting coefficients.

[0016] Furthermore, the strategy in the reverse generation and sampling reconstruction process is defined as follows: Step one is a known region forced alignment strategy, in which the latent variables are modified in the reverse sampling step:

[0017] In the formula, Latent variables generated for network prediction The sampled values ​​are obtained by adding noise to the known latent variables in the region; the second step is a bidirectional resampling mechanism, which generates... Then perform re-diffusion and reverse update:

[0018] In the formula, The step length, The noise is random; step three is a soft masking smoothing mechanism, which uses a soft mask in the later stages of sampling. Alternate binary mask :

[0019] In the formula, It is a two-dimensional Gaussian filter operator.

[0020] Furthermore, the formula for the fusion update of latent variables is adjusted as follows:

[0021] This mechanism is enabled only during the low-noise phase of backsampling, eliminating the sharp boundaries produced at the edges by traditional hard masks and improving the smoothness and structural continuity of the generated results.

[0022] Furthermore, the specific steps of the random stripe masking method are as follows: Given an image size... Occlusion ratio and the ratio of single strip width The mask matrix is ​​generated as follows: Determine the target width of a single missing band. Randomly select the starting column index. , making the first The set of column indexes for missing bands is The generated mask matrix satisfies:

[0023] The mask remains consistent in the vertical direction of the image, simulating signal loss over continuous time intervals in structure monitoring.

[0024] A structural health monitoring anomaly data reconstruction system based on a latent space diffusion model is provided. The reconstruction system uses the aforementioned structural health monitoring anomaly data reconstruction method based on a latent space diffusion model. The reconstruction system includes: Data preprocessing and latent space mapping module: Converts one-dimensional structure monitoring time series data into two-dimensional images, uses a pre-trained variational autoencoder as a surrogate model, maps the two-dimensional images to a low-dimensional latent space, and obtains latent variables containing semantic information of the original data; Constructing a forward diffusion module in the latent space: Define the forward diffusion process in the latent space, gradually inject Gaussian noise into the latent variables through a noise scheduling sequence, and establish a Markov chain evolution path from the true latent representation to the standard Gaussian noise distribution; Constructing a conditional diffusion denoising network module: Constructing a conditional diffusion U-Net network that includes time-step embedding and frequency-domain conditional coding, introducing a frequency-domain conditional coding mechanism to impose explicit global dynamic constraints on the diffusion process, and combining the mask matrix and known region features as network inputs; Model training module: It adopts a mask region selection mechanism to calculate the noise prediction loss only in the missing region, and combines the latent space distribution regularization term to construct the total loss function, jointly optimizing the model's noise inversion ability in the abnormal region and the statistical consistency of the latent space distribution. The reverse generation and sampling reconstruction module performs a reverse diffusion process, starting with random noise and gradually denoising using a trained conditional diffusion network. During the sampling process, a known region forced alignment strategy and a bidirectional resampling mechanism are introduced to ensure that the latent variables of the known segments conform to the original distribution and to coordinate the distribution of the boundary regions. Data decoding and restoration module: The complete latent variables generated by backdiffusion are input into the decoder of the variational autoencoder to restore them into a two-dimensional image, and then restored to a one-dimensional structure monitoring time series signal through image-to-time backprojection.

[0025] A computer device includes a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor, when executing the computer program, implements the method described above.

[0026] A computer-readable storage medium storing a computer program that, when executed by a processor, implements the method described above.

[0027] The beneficial effects of this invention are: This invention breaks through the limitations of traditional interpolation methods: Based on the latent space diffusion model, this invention uses a generative probabilistic inversion mechanism to model the dynamic laws of structural response. Compared with traditional interpolation and sparse recovery methods, it can better capture long-range temporal dependencies, effectively restore complex nonlinear dynamic characteristics, and avoid smoothing artifacts in the reconstructed signal.

[0028] This invention improves computational efficiency and feature stability by using a pre-trained variational autoencoder to compress high-dimensional monitoring images into a low-dimensional latent space for diffusion sampling. Compared to performing diffusion directly in the pixel domain, latent space operations significantly reduce computational complexity, while leveraging the high-level semantic features of the latent space to enhance the model's stability at the structural feature level.

[0029] This invention enhances global dynamic consistency by innovatively introducing a frequency-domain conditional coding mechanism into the network architecture, using frequency-domain features as strong constraints in the denoising process. This enables the model to infer reasonable vibration frequencies and energy distributions based on global spectral distribution patterns when recovering missing waveforms, ensuring that the reconstructed results maintain a high degree of consistency with the real structure in terms of dynamic characteristics.

[0030] This invention eliminates boundary artifacts and achieves smooth transitions: Addressing the discontinuity issue at splicing points, this invention proposes a forced alignment strategy for known regions and a bidirectional resampling mechanism during the sampling phase. The forced alignment strategy prevents known regions from being interfered with by noise, achieving separation between structure preservation and anomaly repair; the bidirectional resampling mechanism, through a round-trip process of "re-diffusion-reverse update," statistically better integrates boundary regions, eliminating the sharp boundaries produced at edges by traditional hard masks.

[0031] This invention has excellent robustness and versatility: Experiments show that this invention is not only applicable to various types of sensor data such as acceleration, GPS displacement, and strain, but also adaptable to different missing rates and complex missing patterns such as fragmentation and long gaps, and can still maintain a low reconstruction error under extreme working conditions. Attached Figure Description

[0032] Figure 1 This is a schematic diagram of the method flow of the present invention.

[0033] Figure 2 This is a schematic diagram of the latent space diffusion model structure of the present invention.

[0034] Figure 3 This is a schematic diagram of the conditional diffusion U-Net network structure of the present invention.

[0035] Figure 4 This is a schematic diagram of the reconstruction results of the actual bridge acceleration monitoring data under different missing working conditions.

[0036] Figure 5 This is a schematic diagram illustrating the reconstruction results of GPS displacement monitoring data of a real bridge under different missing working conditions.

[0037] Figure 6 This is a schematic diagram of the reconstruction results of the actual bridge strain monitoring data under different missing working conditions according to the present invention.

[0038] Figure 7 This is a schematic diagram comparing the reconstruction result of the actual bridge acceleration signal under typical working conditions with the original sequence (time domain-frequency domain), where (a) is the time domain result of acceleration signal sample 1, (b) is the frequency domain result of acceleration signal sample 1, (c) is the time domain result of acceleration signal sample 2, and (d) is the frequency domain result of acceleration signal sample 2.

[0039] Figure 8 This is a schematic diagram comparing the reconstruction results of the actual bridge GPS displacement signal under typical working conditions with the original sequence, where (a) is the time domain result of GPS signal sample 1 and (b) is the time domain result of GPS signal sample 2.

[0040] Figure 9 This is a schematic diagram comparing the reconstruction results of the strain signal of the real bridge under typical working conditions with the original sequence, where (a) is the time domain result of strain signal sample 1 and (b) is the time domain result of strain signal sample 2. Detailed Implementation

[0041] In the following description, specific details such as particular system architectures and techniques are set forth for illustrative purposes and not for limitation, in order to provide a thorough understanding of the embodiments of this application. However, those skilled in the art will understand that this application may also be implemented in other embodiments without these specific details. In other instances, detailed descriptions of well-known systems, apparatuses, circuits, and methods are omitted so as not to obscure the description of this application with unnecessary detail.

[0042] It should be understood that, when used in this specification and the appended claims, the term "comprising" indicates the presence of the described features, integrals, steps, operations, elements and / or components, but does not exclude the presence or addition of one or more other features, integrals, steps, operations, elements, components and / or collections thereof.

[0043] It should also be understood that the terminology used in this application specification is for the purpose of describing particular embodiments only and is not intended to limit the application. As used in this application specification and the appended claims, the singular forms “a,” “an,” and “the” are intended to include the plural forms unless the context clearly indicates otherwise.

[0044] The technical solutions in the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only a part of the embodiments of this application, and not all of the embodiments. Based on the embodiments of this application, all other embodiments obtained by those of ordinary skill in the art without creative effort are within the scope of protection of this application.

[0045] Many specific details are set forth in the following description in order to provide a full understanding of this application. However, this application may also be implemented in other ways different from those described herein. Those skilled in the art can make similar extensions without departing from the spirit of this application. Therefore, this application is not limited to the specific embodiments disclosed below.

[0046] Implementation Method 1 This embodiment provides a method for reconstructing structural health monitoring anomaly data based on a latent space diffusion model, such as... Figure 1 As shown, this method mainly includes three stages: data preprocessing and latent space mapping, diffusion model training, and outlier data reconstruction. The reconstruction method includes the following steps: First, the one-dimensional structure monitoring time-series data is converted into two-dimensional images, and then mapped to a low-dimensional latent space using a pre-trained variational autoencoder to obtain latent variables containing semantic information of the original signal. Subsequently, a forward diffusion process is defined in the latent space to establish an evolutionary path from structured features to Gaussian noise. A conditional diffusion U-Net network, incorporating time-step embedding and frequency-domain conditional coding, is constructed, and the network is trained to learn the noise inversion distribution through a mask region selectivity mechanism. In the reverse generation stage, a known region forced alignment strategy and a bidirectional resampling mechanism are introduced, combined with soft mask smoothing techniques.

[0047] Step 1: Convert the one-dimensional structure monitoring time series data into a two-dimensional image. Use a pre-trained variational autoencoder as a surrogate model to map the two-dimensional image to a low-dimensional latent space and obtain latent variables containing semantic information of the original data. Step 2: Define the forward diffusion process in the latent space, and gradually inject Gaussian noise into the latent variables through a noise scheduling sequence to establish a Markov chain evolution path from the true latent representation to the standard Gaussian noise distribution. Step 3: Construct a conditional diffusion U-Net network that includes time-step embedding and frequency-domain conditional coding. Introduce the frequency-domain conditional coding mechanism to impose explicit global dynamic constraints on the diffusion process, and combine the mask matrix and known region features as network inputs. Step 4: Employ a mask region selection mechanism to calculate noise prediction loss only in missing regions, and construct the total loss function by combining the latent space distribution regularization term to jointly optimize the model's noise inversion capability in abnormal regions and the statistical consistency of the latent space distribution. Step 5: Perform the back diffusion process, starting with random noise and gradually denoising using the trained conditional diffusion network; during the sampling process, introduce a known region forced alignment strategy and a bidirectional resampling mechanism to ensure that the latent variables of the known segments conform to the original distribution and coordinate the distribution of the boundary regions. Step Six: The complete latent variables generated by backdiffusion are input into the decoder of the variational autoencoder to restore a two-dimensional image, and then the image is restored to a one-dimensional structural monitoring time-series signal through temporal backprojection. The aim is to establish a dynamic law model of structural response through a generative probabilistic inversion mechanism, thereby enabling high-fidelity reconstruction of semantically consistent and dynamically reasonable continuous missing or abnormal regions while ensuring that the health monitoring data segments remain undisturbed.

[0048] The reconstruction method further includes: Specifically, step one involves denoting the image representation corresponding to the monitored time-series signal as a tensor. The encoder using a variational autoencoder will Mapping to latent variables ; Step two specifically involves setting a noise scheduling sequence. Arbitrary time steps are generated through a forward diffusion process. Noisy latent variables ; Step three specifically involves constructing a conditional diffusion U-Net, with the network input consisting of noisy latent variables. Mask matrix and known regional latent variables It is assembled by splicing together the frequency domain feature vectors of the monitoring data and mapping them to latent space semantic embeddings. Inject into the network; Step four specifically involves generating training samples using the random stripe masking method and calculating the network prediction noise. With real noise The mask loss and KL divergence regularization terms are used to update the network parameters through backpropagation; Step five specifically involves using standard Gaussian noise. Begin reverse iterative sampling at each step. In the process, a known region forced alignment strategy is used to correct latent variables, and a bidirectional resampling mechanism is executed for boundary coordination until the desired result is generated. ; Step six specifically involves generating... The decoder of the input variational autoencoder obtains the reconstructed image. And convert it back to a timing signal.

[0049] The forward diffusion process is a process of gradually injecting noise into the latent variable. For example... Figure 2 As shown, let the real samples in the latent space be... The noise scheduling sequence is Then, a forward-diffusion Markov chain is defined as:

[0050] In the formula, For the first Latent variables of the step, The unit covariance matrix; after iterative expansion, at any time... Noisy latent variables Can be directly from the initial sample and standard Gaussian noise The sampling yielded:

[0051] In the formula, , This is the cumulative term of the noise attenuation coefficient. This process simulates the degradation and evolution of real sensor signals under noise pollution.

[0052] The conditional diffusion U-Net network, such as Figure 3 As shown, it includes a multi-information joint input mechanism and a frequency domain conditional coding mechanism.

[0053] Furthermore, the input structure and attention mechanism of the conditional diffusion denoising network are defined as follows: Multi-information joint input mechanism: Network input channel tensor Noisy latent variables at the current time step Binary mask matrix and known regional latent variables It is assembled from pieces and defined as:

[0054] Frequency domain conditional coding and cross-attention mechanism: Frequency domain feature mapping: mapping the input frequency domain feature vector Mapped to conditional embedding vector :

[0055] In the formula, For multilayer perceptron networks; cross-attention calculation: conditional embedding vectors are injected into U-Net features, and the attention formula is:

[0056] In the formula, the query matrix Features derived from latent space, key matrix Sum matrix From conditional embedding vectors Obtained by linear projection. For the attention dimension.

[0057] Furthermore, the training process employs a mask region selection mechanism, calculating the noise prediction loss only in missing regions. A binary mask matrix is ​​defined. The missing region is represented by 1, and the known region by 0. The total loss function of the model. Regression loss due to mask noise and KL regularization terms composition:

[0058]

[0059] In the formula, Represents element-wise product. For network prediction noise, and These are the weighting coefficients. Ensure that the model gradient only propagates in abnormal regions, so that the model can focus on learning the noise inversion ability of missing regions; Used to constrain the latent space distribution and prevent distribution drift.

[0060] Furthermore, during the backsampling phase, to ensure that healthy regions are not corrupted, a known region enforcement strategy is employed. This is done at each backsampling time step. First, the network predicts and generates the latent variables of the entire graph. At the same time, real data from known areas are utilized. Forward noise generation correspondence The final combination of latent variables is as follows:

[0061] In the formula, By using real samples The sampled data satisfies This strategy ensures that the health latent variables always conform to the original distribution.

[0062] Specific Implementation Method Six: This implementation method supplements step five of Specific Implementation Method One. To further coordinate the boundary distribution, a bidirectional resampling mechanism is introduced. This is achieved by taking a small step back to the previous method via "re-diffusion". This involves re-coordinating the latent space distribution of the boundaries. Specifically, in the generation... Then, perform a short-boost forward spread, and then perform a reverse update again:

[0063] In the formula, The step length, This is the newly sampled random noise. Within the latent space, this process is equivalent to a small-step time-stochastic differential equation (SDE) round-trip process, gradually harmonizing the distribution of anomalous and healthy regions within the latent space.

[0064] Furthermore, in the later stages of backsampling, to avoid abrupt changes or artifacts at the boundaries, a soft mask smoothing mechanism is used. This involves applying Gaussian blur to the mask to make the mask boundaries change continuously. The soft mask uses a continuous weight matrix. The calculation is as follows:

[0065] In the formula, This represents the convolution operation. The standard deviation is The two-dimensional Gaussian filter operator. The latent variable fusion update formula is adjusted as follows:

[0066] This mechanism is enabled only during the low-noise phase of backsampling, eliminating the sharp boundaries produced at the edges by traditional hard masks and improving the smoothness and structural continuity of the generated results.

[0067] Furthermore, during the model training phase, a controllable random stripe masking method was designed to simulate sensor anomalies or missing data. Given an image size... By randomly generating the mask matrix Define the area to be repaired. The generation of the mask matrix must meet the preset occlusion ratio and strip width. Column index set with missing bands Defined as:

[0068] In the formula, The starting column index is randomly selected. The width is the stripe width. This is the generated mask matrix. satisfy:

[0069] This striped random mask maintains horizontal continuity, which is beneficial for simulating segmented gaps common in structured signals. This embodiment applies the structural health monitoring anomaly data reconstruction method based on the latent space diffusion model proposed in this invention to a real health monitoring system of an in-service long-span cable-stayed bridge to verify the engineering applicability of the method. Three typical sensor data types from this long-span bridge monitoring system are selected as the research objects: accelerometer data, GPS displacement data, and strain sensor data. Masking rates are set from 5% to 40%, covering various anomaly types such as fragmented missing data and long-gap continuous missing data.

[0070] like Figure 4 and Figure 7 As shown, the acceleration signal exhibits significant high-frequency broadband characteristics and its amplitude fluctuates wildly. Experimental results demonstrate that despite environmental background noise, the method of this invention successfully recovers the amplitude envelope and phase structure of the signal. Even with a relatively high missing rate of 30%, the structural similarity index (SSIM) of the reconstructed results remains above 0.9, proving that the model recovers crucial vibrational modal information.

[0071] like Figure 5 and Figure 8 As shown, GPS signals are primarily dominated by low-frequency trend terms. The method of this invention exhibits optimal reconstruction performance for such signals, accurately capturing the slow drift trend of structures and the effects of solar radiation and temperature. Even under extreme conditions of 40% missing data, the SSIM index of the reconstructed signal remains above 0.958, demonstrating extremely strong long-range time-series dependency capture capabilities.

[0072] like Figure 6 and Figure 9 As shown, the strain data has a small amplitude and is extremely sensitive to minute disturbances. Experiments show that the method of this invention can accurately reconstruct the direction and relative amplitude of change within the missing segment, maintaining the phase and periodic structure of local fluctuations consistent with the real data.

[0073] As the amount of information lost increases, the model's reconstruction accuracy shows a gradual downward trend without any precipitous failure. The boundary gradient consistency (BGC) index remains at a low level under all operating conditions, indicating that the reconstructed waveform remains smooth at the splicing points without significant bends or artifacts.

[0074] In summary, the method of the present invention is compatible with the frequency characteristics and noise levels of different physical quantities, and can achieve stable, reliable and high-fidelity data reconstruction under the conditions of high noise, non-stationary and multimodal response in real engineering scenarios.

[0075] Implementation Method 2 This embodiment provides a structural health monitoring anomaly data reconstruction system based on a latent space diffusion model. The reconstruction system uses the structural health monitoring anomaly data reconstruction method based on a latent space diffusion model as described in Embodiment 1. The reconstruction system includes: Data preprocessing and latent space mapping module: Converts one-dimensional structure monitoring time series data into two-dimensional images, uses a pre-trained variational autoencoder as a surrogate model, maps the two-dimensional images to a low-dimensional latent space, and obtains latent variables containing semantic information of the original data; Constructing a forward diffusion module in the latent space: Define the forward diffusion process in the latent space, gradually inject Gaussian noise into the latent variables through a noise scheduling sequence, and establish a Markov chain evolution path from the true latent representation to the standard Gaussian noise distribution; Constructing a conditional diffusion denoising network module: Constructing a conditional diffusion U-Net network that includes time-step embedding and frequency-domain conditional coding, introducing a frequency-domain conditional coding mechanism to impose explicit global dynamic constraints on the diffusion process, and combining the mask matrix and known region features as network inputs; Model training module: It adopts a mask region selection mechanism to calculate the noise prediction loss only in the missing region, and combines the latent space distribution regularization term to construct the total loss function, jointly optimizing the model's noise inversion ability in the abnormal region and the statistical consistency of the latent space distribution. The reverse generation and sampling reconstruction module performs a reverse diffusion process, starting with random noise and gradually denoising using a trained conditional diffusion network. During the sampling process, a known region forced alignment strategy and a bidirectional resampling mechanism are introduced to ensure that the latent variables of the known segments conform to the original distribution and to coordinate the distribution of the boundary regions. Data decoding and restoration module: The complete latent variables generated by backdiffusion are input into the decoder of the variational autoencoder to restore them into a two-dimensional image, and then restored to a one-dimensional structure monitoring time series signal through image-to-time backprojection.

[0076] Implementation Method 3 This invention provides an electronic device including a memory, a processor, and a computer program stored in the memory and executable on the processor. The memory stores software programs and modules, and the processor executes various functional applications and data processing by running the software programs and modules stored in the memory. The memory and processor are connected via a bus. Specifically, the processor implements any step in Embodiment 1 by running the computer program stored in the memory.

[0077] It should be understood that, in the embodiments of the present invention, the processor may be a Central Processing Unit (CPU), but it may also be other general-purpose processors, digital signal processors (DSPs), application-specific integrated circuits (ASICs), field-programmable gate arrays (FPGAs), or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc. A general-purpose processor may be a microprocessor or any conventional processor.

[0078] Memory may include read-only memory, flash memory, and random access memory, and provides instructions and data to the processor. Some or all of the memory may also include non-volatile random access memory.

[0079] It should be understood that if the integrated modules / units described above are implemented as software functional units and sold or used as independent products, they can be stored in a computer-readable storage medium. Based on this understanding, all or part of the processes in the methods described above can also be implemented by a computer program instructing related hardware. This computer program can be stored in a computer-readable storage medium, and when executed by a processor, it can implement the steps of the various method embodiments described above. The computer program includes computer program code, which can be in the form of source code, object code, executable files, or certain intermediate forms. The computer-readable medium can include: any entity or device capable of carrying the computer program code, recording media, USB flash drives, portable hard drives, magnetic disks, optical disks, computer memory, read-only memory (ROM), random access memory (RAM), electrical carrier signals, telecommunication signals, and software distribution media, etc. It should be noted that the content contained in the computer-readable storage medium can be appropriately increased or decreased according to the requirements of legislation and patent practice in the jurisdiction.

[0080] The above description of the disclosed embodiments enables those skilled in the art to make or use the invention. Various modifications to these embodiments will be readily apparent to those skilled in the art, and the general principles defined herein may be implemented in other embodiments without departing from the spirit or scope of the invention. Therefore, the invention is not to be limited to the embodiments shown herein, but is to be accorded the widest scope consistent with the principles and novel features disclosed herein.

[0081] Those skilled in the art will clearly understand that, for the sake of convenience and brevity, the above-described division of functional units and modules is merely an example. In practical applications, the above functions can be assigned to different functional units and modules as needed, that is, the internal structure of the above device can be divided into different functional units or modules to complete all or part of the functions described above. The functional units and modules in the embodiments can be integrated into one processing unit, or each unit can exist physically separately, or two or more units can be integrated into one unit. The integrated unit can be implemented in hardware or as a software functional unit. Furthermore, the specific names of the functional units and modules are only for easy differentiation and are not intended to limit the scope of protection of this invention. The specific working process of the units and modules in the above system can be referred to the corresponding process in the foregoing method embodiments, and will not be repeated here.

[0082] It should be noted that the methods and detailed examples provided in the above embodiments can be incorporated into the apparatus and devices provided in the embodiments for mutual reference, and will not be repeated here.

[0083] Those skilled in the art will recognize that the units and algorithm steps of the various examples described in conjunction with the embodiments disclosed herein can be implemented in electronic hardware, or a combination of computer software and electronic hardware. Whether these functions are implemented in hardware or software depends on the specific application and design constraints of the technical solution. Those skilled in the art can use different methods to implement the described functions for each specific application, but such implementations should not be considered beyond the scope of this invention.

[0084] In the embodiments provided by this invention, it should be understood that the disclosed apparatus / terminal devices and methods can be implemented in other ways. For example, the apparatus / device embodiments described above are merely illustrative. For instance, the division of modules or units described above is merely a logical functional division, and in actual implementation, it can be divided in other ways. For example, multiple units or components can be combined or integrated into another system, or some features can be ignored or not executed.

[0085] The above embodiments are only used to illustrate the technical solutions of the present invention, and are not intended to limit them. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention, and should all be included within the protection scope of the present invention.

Claims

1. A method for reconstructing structural health monitoring anomaly data based on a latent space diffusion model, characterized in that, The reconstruction method involves first converting one-dimensional structure monitoring time-series data into a two-dimensional image, and then using a pre-trained variational autoencoder to map it to a low-dimensional latent space to obtain latent variables containing semantic information of the original signal. Subsequently, a forward diffusion process is defined in the latent space to establish an evolution path from structured features to Gaussian noise. A conditional diffusion U-Net network containing time-step embedding and frequency-domain conditional coding is constructed, and the network is trained to learn the noise inversion distribution through a mask region selectivity mechanism. In the reverse generation stage, a known region forced alignment strategy and a bidirectional resampling mechanism are introduced, combined with soft mask smoothing technology, to input the complete latent variables generated by the reverse diffusion into the decoder of the variational autoencoder to restore the two-dimensional image, and then recovering the one-dimensional structure monitoring time-series signal through image-to-time backprojection.

2. The method for reconstructing abnormal data from structural health monitoring according to claim 1, characterized in that, The reconstruction method includes the following steps: Step 1: Denote the image representation corresponding to the monitored time-series signal as a tensor. The encoder using a variational autoencoder will Mapping to latent variables ; Step 2: Set the noise scheduling sequence Arbitrary time steps are generated through a forward diffusion process. Noisy latent variables ; Step 3: Construct a conditional diffusion U-Net, with the network input consisting of noisy latent variables. Mask matrix and known regional latent variables It is assembled by splicing together the frequency domain feature vectors of the monitoring data and mapping them to latent space semantic embeddings. Inject into the network; Step 4: Generate training samples using the random stripe masking method and calculate the network prediction noise. With real noise The mask loss and KL divergence regularization terms are used to update the network parameters through backpropagation; Step 5, from standard Gaussian noise Begin reverse iterative sampling at each step. In the process, a known region forced alignment strategy is used to correct latent variables, and a bidirectional resampling mechanism is executed for boundary coordination until the desired result is generated. ; Step six, generate The decoder of the input variational autoencoder obtains the reconstructed image. And convert it back to a timing signal.

3. The method for reconstructing structural health monitoring anomaly data according to claim 1 or 2, characterized in that, The input structure and attention mechanism of the conditional diffusion denoising network are defined as follows: Multi-information joint input mechanism: Network input channel tensor Noisy latent variables at the current time step Binary mask matrix and known regional latent variables It is assembled from pieces and defined as: Frequency domain conditional coding and cross-attention mechanism: Frequency domain feature mapping: mapping the input frequency domain feature vector Mapped to conditional embedding vector : In the formula, For multilayer perceptron networks; cross-attention calculation: conditional embedding vectors are injected into U-Net features, and the attention formula is: In the formula, the query matrix Features derived from latent space, key matrix Sum matrix From conditional embedding vectors Obtained by linear projection. For the attention dimension.

4. The method for reconstructing abnormal data from structural health monitoring according to claim 3, characterized in that, The forward diffusion process and the loss function of the conditional diffusion denoising network are defined as follows: During the forward diffusion process, at any time... latent variables Compared with the initial sample The relationship is: In the formula, , This is the cumulative term of the noise attenuation coefficient; The mapping form of the frequency domain conditional coding mechanism is as follows: In the formula, For the normalized logarithmic power spectrum, It is a multilayer perceptron network; Total loss function The formula is: In the formula, The latent space binary mask matrix, For element-wise multiplication, To predict noise for the network, For the latent space distribution KL regularization term, and These are the weighting coefficients.

5. The method for reconstructing structural health monitoring anomaly data according to claim 4, characterized in that, The strategy defined in the reverse generation and sampling reconstruction process is as follows: Step one is to use a known region forced alignment strategy and modify the latent variables in the reverse sampling step: In the formula, Latent variables generated for network prediction The sampled values ​​are obtained by adding noise to the known latent variables in the region; the second step is a bidirectional resampling mechanism, which generates... Then perform re-diffusion and reverse update: In the formula, The step length, The noise is random; step three is a soft masking smoothing mechanism, which uses a soft mask in the later stages of sampling. Alternate binary mask : In the formula, It is a two-dimensional Gaussian filter operator.

6. The method for reconstructing abnormal data from structural health monitoring according to claim 5, characterized in that, The formula for the fusion update of latent variables is adjusted as follows: This mechanism is enabled only during the low-noise phase of backsampling, eliminating the sharp boundaries produced at the edges by traditional hard masks and improving the smoothness and structural continuity of the generated results.

7. The method for reconstructing abnormal data from structural health monitoring according to claim 5, characterized in that, The specific steps of the random stripe masking method are as follows: Given an image size Occlusion ratio and the ratio of single strip width The mask matrix is ​​generated as follows: Determine the target width of a single missing band. Randomly select the starting column index. , making the first The set of column indexes for missing bands is The generated mask matrix satisfies: The mask remains consistent in the vertical direction of the image, simulating signal loss in continuous time intervals during structural monitoring.

8. A structural health monitoring anomaly data reconstruction system based on a latent space diffusion model, characterized in that, The reconstruction system uses the structural health monitoring anomaly data reconstruction method based on the latent space diffusion model as described in any one of claims 1-7, and the reconstruction system includes: Data preprocessing and latent space mapping module: Converts one-dimensional structure monitoring time series data into two-dimensional images, uses a pre-trained variational autoencoder as a surrogate model, maps the two-dimensional images to a low-dimensional latent space, and obtains latent variables containing semantic information of the original data; Constructing a forward diffusion module in the latent space: Define the forward diffusion process in the latent space, gradually inject Gaussian noise into the latent variables through a noise scheduling sequence, and establish a Markov chain evolution path from the true latent representation to the standard Gaussian noise distribution; Constructing a conditional diffusion denoising network module: Constructing a conditional diffusion U-Net network that includes time-step embedding and frequency-domain conditional coding, introducing a frequency-domain conditional coding mechanism to impose explicit global dynamic constraints on the diffusion process, and combining the mask matrix and known region features as network inputs; Model training module: It adopts a mask region selection mechanism to calculate the noise prediction loss only in the missing region, and combines the latent space distribution regularization term to construct the total loss function, jointly optimizing the model's noise inversion ability in the abnormal region and the statistical consistency of the latent space distribution. The reverse generation and sampling reconstruction module performs a reverse diffusion process, starting with random noise and gradually denoising using a trained conditional diffusion network. During the sampling process, a known region forced alignment strategy and a bidirectional resampling mechanism are introduced to ensure that the latent variables of the known segments conform to the original distribution and to coordinate the distribution of the boundary regions. Data decoding and restoration module: The complete latent variables generated by backdiffusion are input into the decoder of the variational autoencoder to restore them into a two-dimensional image, and then restored to a one-dimensional structure monitoring time series signal through image-to-time backprojection.

9. A computer device, characterized in that, It includes a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein when the processor executes the computer program, it implements the method as described in any one of claims 1-7.

10. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores a computer program that, when executed by a processor, implements the method as described in any one of claims 1-7.

Images