Self-adaptive vibration signal noise reduction and modal recovery method and system

Abstract

The invention discloses a self-adaptive vibration signal noise reduction and modal recovery method and system, and relates to the technical field of civil engineering monitoring and artificial intelligence fusion application, and the method comprises the following steps: constructing a multi-source vibration signal collection channel, and obtaining time sequence response data of a structure in a service state; performing normalization and sliding slice preprocessing on the original signal; an optimal noise reduction strategy is dynamically selected based on a preset noise level, and self-adaptive adjustment of signal enhancement and noise suppression is realized; restoring a dominant frequency structure of the signal in combination with a spectrum analysis technology, and extracting structural modal parameters such as an eigenfrequency, a damping ratio and a vibration mode feature; a joint optimization model of signal de-noising and modal maintenance is constructed through a modal constraint loss function, and it is ensured that modal information is not damaged in the de-noising process. According to the adaptive vibration signal noise reduction and modal recovery method and the adaptive vibration signal noise reduction and modal recovery system, high-fidelity recovery of structural vibration signals and accurate extraction of key modal characteristics in a strong interference environment are realized.

Description

Technical Field

[0001] This invention relates to the field of integrated application technology of civil engineering monitoring and artificial intelligence, and in particular to an adaptive vibration signal noise reduction and modal recovery method and system. Background Technology

[0002] Structural health monitoring (SHM) is a crucial means of ensuring the long-term safety and reliable operation of major civil engineering structures such as bridges, high-rise buildings, and dams, and has been widely applied in various infrastructure projects. Vibration response signals, as one of the key raw inputs in SHM systems, contain coupled information on multiple dynamic parameters such as structural stiffness, mass, and damping, serving as the core basis for modal identification, damage location, condition assessment, and anomaly early warning. While the ability to acquire structural vibration data is continuously enhanced with improved sensor accuracy and increased deployment density, this also brings new challenges in complex data processing and high-reliability analysis.

[0003] In actual working conditions, structural vibration signals are often affected by multiple sources of interference, such as wind load, traffic load, equipment disturbance, and measurement noise, resulting in low signal-to-noise ratio and unclear or even missing modal information in the original signal. Traditional methods such as wavelet transform, Fourier filtering, and autoregressive analysis perform well when processing low-noise stationary signals, but their denoising and modal extraction effects are significantly insufficient when dealing with non-stationary, multi-frequency aliasing, and high-noise background structural response signals. In addition, although some deep learning methods can improve denoising capabilities, they lack constraint mechanisms in modal preservation, which easily leads to problems such as modal frequency shift and mode shape deformation, limiting their widespread application in practical engineering monitoring systems. Summary of the Invention

[0004] The purpose of this invention is to provide an adaptive vibration signal denoising and modal recovery method and system, which solves the problems of severe noise interference in vibration signals, unstable modal feature extraction, and lack of adaptive control capability in traditional processing procedures in existing structural monitoring. It achieves high-fidelity reconstruction of structural vibration response signals and accurate recovery of modal information under strong interference background, thereby improving the reliability of structural condition assessment and early damage identification.

[0005] To achieve the above objectives, the present invention provides an adaptive vibration signal noise reduction and modal recovery method, comprising the following steps: S1. Collect vibration response signals of the structure during its service life, build a database according to measurement points, channels and time periods, complete time synchronization and removal of abnormal segments, and record environmental and operating condition metadata for subsequent processing. Step S2: Preprocess the acquired structural vibration response signal by using a sliding window to slice the signal into samples of a uniform format to match the input format of the system's noise reduction module. Step S3: Based on the noise level and spectral characteristics of the input signal, dynamically select or weight and combine the optimal strategy from multiple noise reduction models preset in the system to perform adaptive noise reduction processing on the signal. Step S4: Perform spectrum analysis on the denoised signal, extract the dominant frequency component of the structural response, recover the intrinsic frequency, damping ratio and modal parameters, and output the set of modal information for health assessment. Step S5: Introduce a mode preservation mechanism during the noise reduction process. By constructing indicators such as modal frequency error, spectral structure similarity and mode shape matching degree, a joint optimization objective is formed. Step S6: Collect the noise reduction and modal recovery results, generate visual reports and interface data, and provide them to the structural status assessment, anomaly warning and life analysis modules for use in the monitoring platform.

[0006] Preferably, in S2, a fixed-length sliding window is used to slice the continuous signal. The window length of the sliding window is 1 / 4 to 1 / 8 of the total length of the original signal, and the step size is 25% to 50% of the window length.

[0007] Preferably, the noise reduction module in S3 includes wavelet threshold denoising, autoencoder network, and spectral domain filter bank method. The system automatically switches or weights the strategy according to the current noise energy distribution.

[0008] Preferably, the spectral analysis in S4 uses a combination of Fast Fourier Transform (FFT) and modal stability judgment to output multiple sets of modal parameters and filter the main modal solution according to the stability score.

[0009] Preferably, mode preservation in S5 includes a modal frequency deviation term, a spectral structure similarity term, and a mode shape consistency term. These three terms are weighted and adjusted to form a joint loss function that participates in model training or parameter tuning.

[0010] Preferably, the joint optimization model in S5 consists of the U-Net framework, Res2Net, and CBAM.

[0011] An adaptive vibration signal denoising and modal recovery system includes a data acquisition module, a signal preprocessing module, an adaptive denoising module, a modal analysis module, a result display module, and an interface communication module.

[0012] Therefore, the present invention adopts the above-mentioned adaptive vibration signal denoising and modal recovery method and system, and through the joint optimization of multi-source vibration data modeling, dynamic denoising strategy selection, spectrum analysis and modal parameter extraction, and modal preservation constraint mechanism, it achieves high-fidelity recovery of structural vibration signals and accurate extraction of key modal features under strong interference environment.

[0013] The technical solution of the present invention will be further described in detail below with reference to the accompanying drawings and embodiments. Attached Figure Description

[0014] Figure 1 This is a flowchart of an adaptive vibration signal noise reduction and modal recovery method according to the present invention; Figure 2 This is a flowchart of the signal preprocessing process of the present invention; Figure 3 This is a block diagram showing the selection of adaptive noise reduction strategies in this invention; Figure 4 This is a diagram of the modal preservation constraint structure of the present invention; Figure 5 This is a comparison chart of the modal recovery effects of the present invention. Detailed Implementation

[0015] The technical solution of the present invention will be further described below with reference to the accompanying drawings and embodiments.

[0016] Unless otherwise defined, the technical or scientific terms used in this invention shall have the ordinary meaning understood by one of ordinary skill in the art to which this invention pertains. The terms "first," "second," and similar terms used in this invention do not indicate any order, quantity, or importance, but are merely used to distinguish different components. Terms such as "comprising" or "including" mean that the element or object preceding the word encompasses the elements or objects listed following the word and their equivalents, without excluding other elements or objects. Terms such as "connected" or "linked" are not limited to physical or mechanical connections, but can include electrical connections, whether direct or indirect. Terms such as "upper," "lower," "left," and "right" are used only to indicate relative positional relationships; when the absolute position of the described object changes, the relative positional relationship may also change accordingly.

[0017] Example Please see Figures 1-5 This invention provides an adaptive vibration signal noise reduction and modal recovery method, comprising the following steps: S1. Collect vibration response signals of the structure under service conditions and complete the initial signal processing, including timing synchronization, abnormal segment removal and metadata annotation, to build a basic data set suitable for subsequent processing.

[0018] Structural vibration response signals are acquired using an array of accelerometers deployed on bridges or high-rise buildings. The data includes time-series responses from multiple measurement points, channels, and continuous time periods. After acquisition, time synchronization, outlier removal, and metadata binding are performed to construct a standardized dataset for noise reduction and modal analysis.

[0019] S2. The original signal is normalized, detrended, and bandpass filtered to eliminate unstructured drift and background noise. Then, a sliding window method is used to slice the signal, generating time-series samples of uniform size to match the input format of the system's noise reduction module.

[0020] The sliding window has a window length of 1 / 4 to 1 / 8 of the total length of the original signal, and a step size of 25% to 50% of the window length. The window overlap setting is used to improve the stability of time-frequency analysis. In this embodiment, the slice length is 1024 points and the step size is 512 points. Non-overlapping or semi-overlapping slices form a unified input sample for subsequent adaptive noise reduction module.

[0021] Figure 2 This is a flowchart illustrating the signal preprocessing process in an embodiment of the present invention, showing the processing of the original structural response signal from acquisition to slice input. The process includes steps such as outlier removal, normalization, detrending, bandpass filtering, and sliding window slicing, forming samples in a uniform format for subsequent processing.

[0022] The deep learning network model refers to an improved U-shaped noise reduction model based on a dual attention mechanism and a residual multi-level network. Its basic architecture is the U-Net framework. This model is not a typical stacked autoencoder, but rather developed based on the U-Net architecture, featuring a typical "U"-shaped symmetrical structure, including an encoder and a decoder. The decoder is an upsampling path, and the encoder is a downsampling path. Skip connections are used to directly pass features from the encoding layer to the decoding layer, which helps preserve high-frequency details in the signal and addresses the problem of deep networks easily losing detailed information. Deconvolution techniques are used for upsampling, gradually restoring the time-series dimension of the signal.

[0023] The model incorporates a dual attention mechanism and a residual multi-level network on top of U-Net. The dual attention mechanism is a spatial-channel dual attention mechanism. Channel attention automatically identifies which feature channels contain more useful structural vibration information and which channels are mainly noise, assigning different weights accordingly. Spatial attention, in the time series dimension, focuses on key fluctuation regions of the signal and suppresses background noise regions. This enables dynamic adjustment of feature weights, focusing on core structural response features amidst complex interference.

[0024] The residual multilevel network embeds residual blocks within the layers of U-Net. By introducing residual connections, it solves the gradient vanishing problem in deep network training, allowing the network to be built deeper for multi-scale feature extraction. The model can simultaneously capture both macroscopic trends (low frequencies) and microscopic details (high frequencies) of a signal.

[0025] During its development, this model was benchmarked against the standard U-Net and DenseNet models. This deep learning model is an improved U-Net network integrating a spatial-channel dual attention mechanism and residual modules. It utilizes skip connections and deconvolution techniques to specifically extract multi-scale features and adaptively suppress noise from vibration signals of civil engineering structures.

[0026] S3. Based on the noise level and spectral characteristics of the input signal, dynamically select or weightedly combine the optimal strategy from multiple preset noise reduction models in the system to perform adaptive noise reduction processing on the signal, achieving noise suppression and feature enhancement. For example... Figure 3 The adaptive noise reduction strategy selection block diagram shown illustrates how the system adaptively calls different noise reduction models or combinations of strategies based on noise level and spectral characteristics after receiving signals with different noise intensities. This achieves automatic scheduling capability for noise reduction paths under different noise levels, improving the adaptability and accuracy of signal processing.

[0027] The noise reduction model includes wavelet thresholding, autoencoders, and spectral filter banks. The strategy selection module classifies and judges noise based on a preset noise identification model and calls the corresponding processing path to achieve adaptive signal enhancement and noise suppression.

[0028] The dynamic selection process operates in two dimensions: channel and time. The dynamic selection in the channel dimension aims to address the question of which features to focus on. Since the feature map output by the convolutional layer contains multiple channels, some channels may contain crucial structural damage information, while others may mainly contain background noise.

[0029] Compression begins by using a global average pooling operation to compress the spatial information of each channel into a real number, thus representing the global distribution of that channel.

[0030] The incentive mechanism utilizes the nonlinear relationships between channels learned in the fully connected layer to generate channel weight vectors.

[0031] In the selection process, channels with weighting coefficients close to 1 are considered important features and are retained, while channels close to 0 are considered noise and are suppressed.

[0032] The dynamic selection of spatial dimensions aims to address the question of which time period to focus on. For one-dimensional vibration signals, effective information is often concentrated in specific time segments, while the rest is mostly invalid random noise.

[0033] Feature aggregation involves the network compressing or convolving features along the channel axis to extract spatially salient features.

[0034] Generate a mask, and use the Sigmoid activation function to generate a spatial attention map to map the features to the [0, 1] interval.

[0035] In this selection process, signal peak regions receive high weights, while flat noise regions receive low weights. The weights corresponding to peak positions (signal) are close to 1, while the weights corresponding to flat positions (background noise) are close to 0. Multiplying the image with the original waveform automatically removes the noise regions. The final dynamic selection is achieved by element-wise multiplying the input feature map with the two attention weights generated above.

[0036] S4. Perform spectrum analysis on the denoised signal, extract the dominant frequency component of the structural response, recover the modal parameters such as intrinsic frequency, damping ratio and mode shape, and output a set of modal information for health assessment.

[0037] The module first performs a Fast Fourier Transform (FFT) to extract the dominant frequency component of the signal, and then extracts the first to third order modal frequencies, damping ratios, and mode shape parameters based on frequency stability assessment. To ensure the physical rationality of the modal parameters, the system performs stability scoring and repeatability analysis on the extraction results, and outputs modal identification results and confidence labels.

[0038] S5. A mode preservation mechanism is introduced during the noise reduction process to ensure that the modal characteristics are not distorted. By constructing indicators such as modal frequency error, spectral structure similarity, and mode shape matching, a joint optimization objective is formed to achieve integrated constraint optimization of noise reduction and mode preservation. Figure 4 The diagram showing the modal preservation loss constraint structure illustrates how modal frequency difference, spectral structure similarity, and mode shape error are introduced as joint optimization objectives during model training to ensure the consistency of modal features in noise reduction.

[0039] The joint optimization model refers to an end-to-end overall training architecture based on an improved U-shaped network using a dual attention mechanism and a residual multi-level network. The joint optimization model is deeply integrated from three core components, and the parameters of these components are optimized simultaneously during training to minimize the error between the denoised signal and the true signal. The U-Net framework provides a basic encoder-decoder structure. The encoder is responsible for compressing the signal and extracting features, while the decoder is responsible for restoring the signal size through deconvolution.

[0040] Res2Net replaces or enhances the convolutional layers in a standard U-Net. It extracts multi-scale vibration features through fine-grained feature grouping without significantly increasing computational cost. During optimization, Res2Net learns how to capture vibration details across different frequency ranges, preventing high-frequency, minute vibrations from being mistakenly filtered out during denoising.

[0041] CBAM incorporates channel attention and spatial attention. Weights are dynamically adjusted across network layers.

[0042] The modality preservation mechanism constructs a new mechanism through a specific combination to solve the modality loss problem, that is, traditional noise reduction often treats high-frequency modes as noise to be filtered out.

[0043] Multi-scale capture (the role of Res2Net): By utilizing the multi-receptive field characteristics of Res2Net, the mechanism ensures that the model can simultaneously "see" low-frequency large-amplitude vibrations (low-order modes) and high-frequency small vibrations (high-order modes). This is modal information capture at the physical level.

[0044] Adaptive filtering (the role of CBAM): Using an attention mechanism, the model automatically identifies which frequency components belong to structural modes (giving them high weights) and which belong to environmental noise (giving them low weights).

[0045] Skip connections: U-Net's original skip connections directly pass the encoder's high-resolution features to the decoder, which in principle avoids the loss of high-frequency modal details caused by deep networks.

[0046] S6. Bind the recovered modal information with the original measurement point information, output the structural response characteristics and modal recognition results, generate a visualization report, and upload it to the structural health monitoring platform for storage and retrieval through the interface.

[0047] The system output includes results such as the denoised signal, modal parameters, modal extraction confidence scores, and frequency domain reconstruction comparison charts. For example... Figure 5 The comparison diagram of modal extraction results shown compares the original noisy signal and the signal processed by the method of this invention in terms of spectral structure and main modal frequency extraction results, verifying the modal recovery accuracy and robustness.

[0048] Therefore, the present invention adopts the above-mentioned adaptive vibration signal denoising and modal recovery method and system, and through the joint optimization of multi-source vibration data modeling, dynamic denoising strategy selection, spectrum analysis and modal parameter extraction, and modal preservation constraint mechanism, it achieves high-fidelity recovery of structural vibration signals and accurate extraction of key modal features under strong interference environment.

[0049] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and not to limit them. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can still be made to the technical solutions of the present invention, and these modifications or equivalent substitutions cannot cause the modified technical solutions to deviate from the spirit and scope of the technical solutions of the present invention.

Claims

1. An adaptive vibration signal noise reduction and modal recovery method, characterized in that, Includes the following steps: S1. Collect vibration response signals of the structure during its service life, build a database according to measurement points, channels and time periods, complete time synchronization and removal of abnormal segments, and record environmental and operating condition metadata. Step S2: Preprocess the acquired structural vibration response signal by using a sliding window to slice the signal into samples of a uniform format to match the input format of the system's noise reduction module. Step S3: Based on the noise level and spectral characteristics of the input signal, dynamically select or weight and combine the optimal strategy from multiple noise reduction models preset by the system to perform adaptive noise reduction processing on the signal. Step S4: Perform spectrum analysis on the denoised signal, extract the dominant frequency component of the structural response, recover the intrinsic frequency, damping ratio and modal parameters, and output the set of modal information for health assessment. Step S5: Introduce a mode preservation mechanism during the noise reduction process. By constructing indicators such as modal frequency error, spectral structure similarity and mode shape matching degree, a joint optimization objective is formed. Step S6: Collect the noise reduction and modal recovery results, generate visual reports and interface data, and provide them to the structural status assessment, anomaly warning and life analysis modules for use in the monitoring platform.

2. The adaptive vibration signal denoising and modal recovery method according to claim 1, characterized in that: In S2, a fixed-length sliding window is used to slice the continuous signal. The window length of the sliding window is 1 / 4 to 1 / 8 of the total length of the original signal, and the step size is 25% to 50% of the window length.

3. The adaptive vibration signal noise reduction and modal recovery method according to claim 2, characterized in that: The noise reduction module in S3 includes wavelet thresholding, autoencoder network, and spectral domain filter bank method. The system automatically switches or weights the strategy based on the current noise energy distribution.

4. The adaptive vibration signal noise reduction and modal recovery method according to claim 3, characterized in that: The spectral analysis in S4 uses a combination of Fast Fourier Transform (FFT) and modal stability assessment to output multiple sets of modal parameters and filter the master mode solution according to the stability score.

5. The adaptive vibration signal noise reduction and modal recovery method according to claim 4, characterized in that: In S5, mode preservation includes modal frequency deviation, spectral structure similarity, and mode shape consistency. These three terms are weighted and adjusted to form a joint loss function that is used in model training or parameter tuning.

6. The adaptive vibration signal noise reduction and modal recovery method according to claim 5, characterized in that: The joint optimization model in S5 consists of the U-Net framework, Res2Net, and CBAM.

7. An adaptive vibration signal denoising and modal recovery system, comprising a data acquisition module, a signal preprocessing module, an adaptive denoising module, a modal analysis module, a result display module, and an interface communication module, characterized in that, The system is used to perform the method as described in any one of claims 1 to 6.

Images