A noise reduction method for vibration signals in structural health monitoring based on hybrid neural networks

Abstract

This invention proposes a method for denoising vibration signals from structural health monitoring based on a hybrid neural network. This method, implemented according to embodiments of the invention, eliminates the need for prior signal knowledge and manual parameter settings. It effectively removes various noise types from structural health monitoring vibration signals, significantly improves the signal-to-noise ratio, and reduces the root mean square error, achieving efficient and automated noise reduction of vibration signals.

Description

Technical Field

[0001] This invention relates to the field of vibration signal noise reduction technology for structural health monitoring, and in particular to a method for vibration signal noise reduction based on a hybrid neural network for structural health monitoring. Background Technology

[0002] Structural health monitoring (SHM), as a core technology for ensuring the safe operation of infrastructure, is widely used in civil engineering fields such as bridges and high-rise buildings. Among related technologies, vibration signal processing systems typically encompass key stages such as data acquisition, feature extraction, and damage assessment. Signal denoising is a fundamental task for ensuring the accuracy of modal analysis. Specifically, existing denoising techniques are mainly divided into time-domain, frequency-domain, and time-frequency-domain methods. Time-frequency-domain methods, such as wavelet transform, achieve noise separation through multi-scale analysis, but their effectiveness highly depends on manual parameter optimization such as wavelet basis function selection and decomposition level setting. With the development of deep learning technology, signal processing schemes based on models such as CNN and BiLSTM are gradually replacing traditional methods. CNN excels at local feature extraction, and BiLSTM possesses bidirectional time series modeling capabilities. However, existing hybrid architectures still suffer from technical bottlenecks such as insufficient modeling of long-range dependencies and weak global feature fusion capabilities.

[0003] However, existing signal denoising methods directly rely on manual experience for parameter selection without establishing an adaptive correlation mechanism between noise and structural features. This can lead to the loss of useful signal components or noise residue, or fluctuations in denoising performance when environmental parameters change dynamically. Specifically, frequency domain filtering methods require prior knowledge of the structural modal frequency range, but in actual engineering, the natural frequencies of structures drift with environmental factors such as temperature and load. Time-frequency domain methods are limited in their effectiveness for processing non-stationary signals due to the stability of the decomposition parameters, and the threshold selection process requires repeated adjustments. Therefore, while existing deep learning solutions can achieve parameter adaptation, the CNN-BiLSTM architecture struggles to simultaneously capture local temporal dependencies and global feature correlations when processing multi-scale vibration features, resulting in insufficient generalization ability in complex noisy environments. Furthermore, the vanishing gradient problem of traditional activation functions in nonlinear modeling and the limitations of single attention mechanisms in multi-channel feature fusion further restrict the performance improvement of denoising models. Summary of the Invention

[0004] The present invention aims to at least partially solve one of the technical problems in the related art.

[0005] Therefore, the first objective of this invention is to propose a method for noise reduction of vibration signals in structural health monitoring based on a hybrid neural network.

[0006] The second objective of this invention is to propose a noise reduction device for vibration signals in structural health monitoring based on a hybrid neural network.

[0007] To achieve the above objectives, the present invention provides a first method for denoising vibration signals in structural health monitoring based on a hybrid neural network, comprising: S1. Obtain the original signal of structural vibration response, and inject different levels of Gaussian white noise into the original signal to generate a noisy signal. Divide the noisy signal and the original signal into a training set, a validation set and a test set according to a preset ratio. S2, Construct a hybrid neural network model that includes a convolutional neural network with a self-attention mechanism and a bidirectional long short-term memory neural network, wherein the convolutional neural network uses alternating one-dimensional convolutional layers and average pooling layers, and the self-attention mechanism performs global weighted fusion of feature maps through a multi-head attention module; S3, input the training set and validation set into the hybrid neural network model for end-to-end training to minimize the mean square error loss function between the predicted denoised signal and the original signal, and obtain the optimal denoising model; S4. The optimal noise reduction model is used to denoise the noisy signals in the test set, and pink noise of different levels is injected into the same reference signal as the original signal to generate extended test data. The extended test data is then denoised using the optimal noise reduction model to verify the model's adaptability to various noise types.

[0008] In one embodiment of the present invention, S1 includes: S11 uses an accelerometer to collect the original signal of structural vibration response at a sampling frequency of 50Hz; S12, the dataset is divided into training, validation and test sets in a 2:1:1 ratio.

[0009] In one embodiment of the present invention, S2 includes: S21, The specific architecture of the self-attention layer is as follows: First, the input data is compressed using global average pooling. Then, the compressed data is multiplied by three different weight matrices to obtain the representations Q, K, and V. The self-attention weight H is calculated using the representations of Q and K, and multiplied by V to achieve a weighted average of the features at each position to obtain the output. S22, the number of heads in the multi-head self-attention mechanism is 8, and the multi-head self-attention mechanism is applied to each channel of the feature map in the neural network.

[0010] In one embodiment of the present invention, S3 includes: S31, after the convolution operation, the Leaky ReLU activation function is added, and its calculation formula is as follows: ; S32, the hybrid neural network model contains a dropout layer, and each neuron is dropped with an independent probability p=0.5 during training.

[0011] In one embodiment of the present invention, S4 includes: S41, the power spectral density (PSD) of pink noise is inversely proportional to the frequency f, and its mathematical expression is: ; S42 evaluates the noise reduction performance by calculating the root mean square error (RMSE) and signal-to-noise ratio (SNR) of the denoised signal. The formula for calculating RMSE is as follows: , The formula for calculating SNR is: .

[0012] In one embodiment of the present invention, it further includes: S5. During the training of the hybrid neural network model, a normalization layer is inserted between the convolutional layer and the self-attention layer. The normalization layer uses batch normalization technology to standardize the feature map in order to accelerate model convergence and improve training stability.

[0013] To achieve the above objectives, a second aspect of the present invention provides a vibration signal noise reduction device for structural health monitoring based on a hybrid neural network, comprising: The structural vibration signal acquisition and data partitioning module is used to acquire the original signal of structural vibration response, inject different levels of Gaussian white noise into the original signal to generate a noisy signal, and divide the noisy signal and the original signal into a training set, a validation set and a test set according to a preset ratio. A hybrid neural network model building module is used to build a hybrid neural network model that includes a convolutional neural network with a self-attention mechanism and a bidirectional long short-term memory neural network. The convolutional neural network uses alternating one-dimensional convolutional layers and average pooling layers, and the self-attention mechanism performs global weighted fusion of feature maps through a multi-head attention module. The end-to-end training and model optimization module is used to input the training set and validation set into the hybrid neural network model for end-to-end training, so as to minimize the mean square error loss function between the predicted denoised signal and the original signal and obtain the optimal denoising model. The noise reduction and adaptability verification module is used to perform noise reduction processing on the noisy signals in the test set using the optimal noise reduction model, and further inject pink noise of different levels into the same reference signal as the original signal to generate extended test data. The extended test data is then subjected to noise reduction processing by the optimal noise reduction model to verify the model's adaptability to multiple noise types.

[0014] In one embodiment of the present invention, the structural vibration signal acquisition and data segmentation module is further configured to: An accelerometer was used to acquire the raw signal of the structural vibration response at a sampling frequency of 50 Hz; The dataset is divided into training, validation, and test sets in a 2:1:1 ratio.

[0015] In one embodiment of the present invention, the hybrid neural network model building module is further configured to: The specific architecture of the self-attention layer is as follows: First, the input data is compressed using global average pooling. Then, the compressed data is multiplied by three different weight matrices to obtain the representations Q, K, and V. The self-attention weight H is calculated using the representations of Q and K, and then multiplied by V to achieve a weighted average of the features at each position to obtain the output. The number of heads in the multi-head self-attention mechanism is set to 8, and the multi-head self-attention mechanism is applied to each channel of the feature map in the neural network.

[0016] In one embodiment of the present invention, the end-to-end training and model optimization module is further configured to: After the convolution operation, the Leaky ReLU activation function is added, and its calculation formula is as follows: ; The hybrid neural network model includes a dropout layer, where each neuron is dropped with an independent probability p=0.5 during training.

[0017] The method and apparatus of this invention do not rely on prior knowledge of signals or manual parameter settings, effectively achieving automated noise reduction of various noise types in structural health monitoring vibration signals, significantly improving the signal-to-noise ratio and reducing the root mean square error, while preserving key vibration characteristic information of the structure.

[0018] Additional aspects and advantages of the invention will be set forth in part in the description which follows, and in part will be obvious from the description, or may be learned by practice of the invention. Attached Figure Description

[0019] The above and / or additional aspects and advantages of the present invention will become apparent and readily understood from the following description of the embodiments taken in conjunction with the accompanying drawings, wherein: Figure 1 This is a flowchart of a method for denoising vibration signals for structural health monitoring based on a hybrid neural network according to an embodiment of the present invention; Figure 2 This is an architecture diagram of a vibration signal noise reduction system for structural health monitoring based on a hybrid neural network according to an embodiment of the present invention; Figure 3This is a network architecture diagram of the self-attention layer in SACNN-BiLSTM according to an embodiment of the present invention; Figure 4 This is a diagram of the overall architecture of the SACNN-BiLSTM hybrid neural network according to an embodiment of the present invention; Figure 5 This is a comparison diagram of the original signal, the noisy signal, and the denoised signal according to an embodiment of the present invention. Figure 6 This is a structural diagram of a vibration signal noise reduction device for structural health monitoring based on a hybrid neural network according to an embodiment of the present invention. Detailed Implementation

[0020] It should be noted that, unless otherwise specified, the embodiments and features described in the present invention can be combined with each other. The present invention will now be described in detail with reference to the accompanying drawings and embodiments.

[0021] To enable those skilled in the art to better understand the present invention, the technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings of the embodiments of the present invention. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort should fall within the scope of protection of the present invention.

[0022] The following description, with reference to the accompanying drawings, describes a method for noise reduction of vibration signals in structural health monitoring based on a hybrid neural network, according to an embodiment of the present invention.

[0023] Example 1 Figure 1 This is a flowchart of a vibration signal noise reduction method for structural health monitoring based on a hybrid neural network according to an embodiment of the present invention, as shown below. Figure 1 As shown, it includes: S1. Obtain the original signal of structural vibration response, and inject different levels of Gaussian white noise into the original signal to generate a noisy signal. Divide the noisy signal and the original signal into a training set, a validation set and a test set according to a preset ratio.

[0024] Specifically, in step 1 of this invention, the raw vibration response signal of the structure under actual working conditions is first acquired using a high-precision accelerometer as a clean signal. This signal typically exists in the form of one-dimensional time-series data, with a sampling frequency set to 50Hz to ensure that the dynamic response characteristics of the structure in the low-to-mid-frequency range can be captured. In some implementations, the raw signal acquisition time is 48 hours to cover the response changes of the structure under different environmental and load conditions, thereby improving the diversity and representativeness of the data.

[0025] Subsequently, different levels of Gaussian white noise were injected into the original signal to simulate common noise pollution in real-world monitoring environments. The Gaussian white noise injection method involved superimposing a normally distributed random vector with a mean of 0 and a standard deviation identical to the original signal onto the original signal. The noise level was controlled by a parameter α; the larger α was, the higher the noise intensity. In this embodiment, the noise level α was set to 1-5 dB, generating five sets of noisy signals, thus constructing training samples with different levels of noise pollution.

[0026] Furthermore, the dataset is divided in a 2:1:1 ratio, with 70% of the data used for training, 15% for validation, and 15% for testing. This division aligns with standard practices for deep learning model training, ensuring the model has sufficient data for feature learning during training, while also exhibiting good generalization ability during validation and testing. The training set is used for updating model parameters, the validation set for adjusting hyperparameters and triggering early stopping mechanisms, and the test set for final performance evaluation.

[0027] Furthermore, this step provides a high-quality synthetic dataset for the subsequent training of the SACNN-BiLSTM hybrid neural network. By introducing controllable noise into the original signal, common noise interference in actual structural health monitoring is simulated, providing the model with rich training samples. This enables the model to learn the mapping relationship between signal features and noise at different noise levels, thereby improving the model's robustness and adaptability in real-world environments.

[0028] Furthermore, this step constructs a structured and standardized signal denoising training dataset, providing a good input-output pair for the deep learning model and effectively solving the problem of traditional denoising methods relying on prior knowledge and manual parameter adjustment. By introducing multi-level noise and reasonably partitioning the dataset, the model's generalization ability under different noise conditions is improved, laying a solid foundation for subsequent nonlinear mapping modeling and feature extraction.

[0029] Furthermore, S1 also includes: S11 uses an accelerometer to collect the original signal of the structural vibration response at a sampling frequency of 50Hz.

[0030] Specifically, in the structural health monitoring vibration signal noise reduction method of the present invention, an accelerometer is used to collect the original structural vibration response signal at a sampling frequency of 50Hz, which is the fundamental step for the entire system to acquire input data. This step is technically implemented based on the dynamic response acquisition capability of the sensor system, combined with signal injection and data partitioning strategies, to provide high-quality benchmark data for the subsequent training and validation of deep learning models.

[0031] Furthermore, the accelerometer is typically a triaxial MEMS (Micro-Electro-Mechanical Systems) sensor with a sampling frequency set to 50Hz, meaning it collects 50 vibration acceleration samples per second with a sampling period of 20ms. This frequency setting must satisfy the Nyquist sampling theorem, i.e., the sampling frequency should be at least twice the highest frequency component of the signal to ensure that key modal frequencies (such as the structure's natural frequencies, resonant frequencies, etc.) in the vibration signal can be completely captured. In this embodiment, the sensor collects the long-term vibration response of the structure under natural or artificial excitation, with a sampling duration of 48 hours to obtain sufficient time-series data for model training.

[0032] Furthermore, the sampling precision of the original signal is typically a 16-bit ADC (analog-to-digital converter), with a sampling range of ±2g (gravitational acceleration), and the sampling error controlled within ±0.5% to ensure data accuracy and stability. The noise injection process uses Gaussian white noise (GWN) with a mean of 0 and a standard deviation consistent with the original signal. The noise level is controlled by the signal-to-noise ratio (SNR), set to 1dB, 2dB, 3dB, 4dB, and 5dB respectively to simulate real monitoring environments under different levels of noise pollution. This process follows the specifications for vibration signal noise injection in ISO 16063-11 to ensure the comparability and scientific validity of the experimental data.

[0033] Furthermore, this step is applicable to long-term health monitoring systems for civil structures such as long-span bridges and high-rise buildings. By collecting structural vibration signals in laboratory or field environments and artificially adding noise, a representative training dataset is constructed to provide learning samples for the SACNN-BiLSTM model. The dataset is divided into training, validation, and test sets in a 2:1:1 ratio to ensure the model's generalization ability and robustness under different noise levels.

[0034] Furthermore, benchmark data of structural vibration signals and noise-contaminated samples are provided for the deep learning model, thereby constructing a denoising model with strong generalization ability. By precisely controlling the noise level and sampling parameters, it is ensured that the model can learn the complex nonlinear relationship between noise and the original signal during training, laying a solid foundation for the subsequent construction and optimization of the SACNN-BiLSTM hybrid neural network.

[0035] S12, the dataset is divided into training, validation and test sets in a 2:1:1 ratio.

[0036] Specifically, in step 1 of this invention, the partitioning of the dataset is a fundamental step in constructing the SACNN-BiLSTM hybrid neural network denoising model. Its technical implementation principle is based on data segmentation strategies in statistical learning theory, aiming to ensure that the model obtains a representative sample distribution during training, validation, and testing, thereby improving the model's generalization ability and robustness. The specific operation is as follows: First, a high-precision accelerometer is used to collect continuous vibration response signals of the structure under natural working conditions at a sampling frequency of 50Hz, serving as a clean signal. Then, a Gaussian white noise injection method is used to add noise at different signal-to-noise ratios (SNR) levels (1-5dB) to the original signal, generating the corresponding noisy signal. This noise injection process follows a standard normal distribution.

[0037] Furthermore, the dataset is divided into a training set, a validation set, and a test set in a 2:1:1 ratio. This division aligns with common data allocation strategies in deep learning, ensuring the training set has sufficient sample size (60%) to support adequate model parameter learning, the validation set (25%) for hyperparameter tuning and early stopping mechanism assessment, and the test set (15%) for final evaluation of the model's noise reduction performance on unseen data. In some implementations, stratified sampling or a sliding window for time series can be used to maintain consistency in the distribution of different noise levels and signal features across subsets.

[0038] This step plays a crucial role in the overall technical solution. On the one hand, reasonable dataset partitioning effectively prevents model overfitting, ensuring its good adaptability to unknown noise signals in real-world structural health monitoring scenarios. On the other hand, introducing synthetic data with different noise levels enhances the model's robustness to noise interference, providing diverse input samples for the subsequent training of the SACNN-BiLSTM model. Furthermore, this step provides a standardized testing framework for model performance evaluation. By calculating quantitative indicators such as the root mean square error (RMSE) and signal-to-noise ratio (SNR) of the denoised signals on the test set, it objectively reflects the model's denoising effect under different noise environments, providing data support for subsequent model optimization and engineering applications.

[0039] S2, Construct a hybrid neural network model that includes a convolutional neural network with a self-attention mechanism and a bidirectional long short-term memory neural network. The convolutional neural network uses alternating one-dimensional convolutional layers and average pooling layers, and the self-attention mechanism performs global weighted fusion of feature maps through a multi-head attention module.

[0040] Specifically, the SACNN-BiLSTM hybrid neural network model proposed in step 2 of this invention combines a self-attention convolutional neural network (SACNN) with a bidirectional long short-term memory neural network (BiLSTM) to achieve efficient noise reduction of vibration signals from structural health monitoring. This model integrates the advantages of CNN in local feature extraction with the bidirectional memory capability of BiLSTM in temporal modeling, while enhancing the model's ability to perceive and fuse global features through a self-attention mechanism.

[0041] Furthermore, the SACNN part employs alternating layers of one-dimensional convolutional layers (Conv1D) and average pooling layers (AvgPool) to progressively extract local temporal features from the vibration signal. Each convolutional layer has a kernel size of 32, a stride of 2, and channel numbers of 16, 32, and 64 respectively, to achieve layer-by-layer feature abstraction. After the convolution operation, the Leaky ReLU activation function is introduced, and its calculation formula is as follows: It is used to alleviate the gradient vanishing problem of traditional ReLU with negative input and enhance the nonlinear expressive power of the model.

[0042] The self-attention mechanism performs global weighted fusion of feature maps through a multi-head attention module. Each attention head independently learns the linear transformations of Query, Key, and Value, and finally generates the attention output through linear combination and concatenation operations. In this invention, the multi-head attention module is set to 8 heads, each with a dimension of 64. By calculating the attention weight matrix H, the weighted fusion of each channel in the feature map is achieved. The input of the self-attention layer is first compressed by global average pooling (GAP), and then multiplied by the three weight matrices respectively to generate Q, K, and V vectors. The attention weights are calculated through dot product and softmax operations, and finally multiplied by the V vector to obtain the weighted feature output, which is then fused with the original features to improve the model's ability to model long-range dependencies.

[0043] The BiLSTM part contains at least two layers of bidirectional long short-term memory structures, each with 128 hidden units. It employs tanh and sigmoid activation functions for gating control to capture the bidirectional dependence of vibration signals over time. By fusing local features extracted by SACNN with temporal features modeled by BiLSTM, the model can more comprehensively understand the time-frequency characteristics of the signal, thus achieving effective suppression of Gaussian white noise and pink noise without prior knowledge, significantly improving denoising accuracy and robustness.

[0044] Furthermore, S2 also includes: S21, the specific architecture of the self-attention layer is as follows: first, the input data is compressed using global average pooling; then, the compressed data is multiplied by three different weight matrices to obtain Q, K, and V representations; the self-attention weights H are calculated using the representations of Q and K, and multiplied by V to achieve a weighted average of the features at each position to obtain the output.

[0045] Specifically, the architecture design of the self-attention layer in this invention aims to enhance the ability of the SACNN-BiLSTM hybrid neural network to model global dependencies in vibration signals, thereby improving the noise reduction effect. In some implementations, this self-attention layer first uses Global Average Pooling (GAP) to compress the input feature map. The GAP operation compresses the input feature map of shape (C×T) (where C is the number of channels and T is the number of time steps) into a (C×1) vector by averaging each channel over the time dimension, thereby extracting the global feature representation of each channel. This compressed vector is then weighted by three different weight matrices. , , Multiplying these vectors generates three representation vectors: Query (Q), Key (K), and Value (V), typically with dimensions of (C×d), where d is the embedding dimension of the attention head.

[0046] Furthermore, the self-attention weight matrix It is obtained by calculating the dot product between the query and the key and then performing Softmax normalization, i.e. The weight matrix reflects the correlation between positions in the input sequence, thus achieving a weighted summation of the Value vector, ultimately outputting a weighted feature map of shape (C×T). In this invention, each channel independently applies a multi-head attention mechanism, with 8 heads, each head having a dimension of d=64, resulting in a total output dimension of 512. Through linear transformation and concatenation operations, the outputs of each head are merged to form the final attention output.

[0047] Furthermore, this self-attention layer is embedded between the convolutional layers and BiLSTM layers of the SACNN-BiLSTM network to enhance the interaction between the local features extracted by the CNN and the time-series features processed by the BiLSTM. By introducing the self-attention mechanism, the model can automatically focus on key features in the signal that are less correlated with noise, improving its ability to preserve weak damage features in structural vibration signals. This step significantly expands the receptive field of the CNN, enabling it to capture dependencies over longer distances, thereby achieving more robust noise reduction performance in complex noisy environments.

[0048] S22, the number of heads in the multi-head self-attention mechanism is 8, and the multi-head self-attention mechanism is applied to each channel of the feature map in the neural network.

[0049] Specifically, in the SACNN-BiLSTM hybrid neural network of this invention, the number of heads in the multi-head self-attention mechanism is set to 8. This mechanism is applied to each channel of the feature map in the neural network to enhance the model's ability to model global dependencies in vibration signals. The self-attention mechanism achieves dynamic weighted aggregation of features by calculating the similarity of query, key, and value between positions in the input feature map, thereby extracting more discriminative global features. In the multi-head self-attention mechanism, each attention head independently learns a set of trainable weight matrices and performs linear transformations on the input features to generate corresponding query, key, and value vectors. Specifically, each channel of the input feature map is first linearly projected into multiple subspaces, each subspace corresponding to one attention head. Each head performs standard self-attention calculations, i.e., by calculating the dot product between query and key and performing softmax normalization to obtain the attention weight matrix H, and then multiplying it by the value matrix to obtain the output feature of that head. Finally, the output features of all 8 attention heads are merged through a concatenation operation and integrated through a learnable linear transformation layer to output the fused feature map.

[0050] Furthermore, the input feature map of the self-attention layer is first compressed in the channel dimension using global average pooling (GAP) to generate channel-level feature representations. Subsequently, these compressed features are multiplied by three different weight matrices to obtain the representations of query, key, and value.

[0051] Furthermore, in this invention, the number of heads in the multi-head self-attention mechanism is fixed at 8. This parameter is chosen based on a balance between experimental verification and model complexity. Too many heads increase the computational burden, while too few heads may limit the model's ability to model multi-dimensional feature relationships. Setting up 8 heads can effectively improve the expressive power of the feature map while ensuring model performance. In addition, the dimension of each attention head is usually set to 1 / 8 of the number of feature map channels to maintain the controllability of parameter scale.

[0052] Furthermore, vibration signals often exhibit non-stationarity, multi-scale characteristics, and long-range dependencies. Traditional CNNs, limited by their local receptive field, struggle to capture dependencies at long distances within the signal. By applying a multi-head self-attention mechanism to each channel, the SACNN-BiLSTM model significantly expands its receptive range, enhancing its understanding of the global signal structure. This allows it to retain more useful structure-related information during denoising, improving both denoising accuracy and signal fidelity. This step is one of the core technologies for achieving efficient and intelligent denoising in this invention, providing higher-quality feature inputs for subsequent BiLSTM layers to handle temporal dependencies, and playing a crucial role in improving the overall model performance.

[0053] S3. Input the training set and validation set into the hybrid neural network model for end-to-end training to minimize the mean square error loss function between the predicted denoised signal and the original signal, and obtain the optimal denoising model.

[0054] Specifically, in step 3 of this invention, the training set and validation set are input into the constructed hybrid neural network model to achieve end-to-end vibration signal denoising training. The core of this step lies in optimizing the model parameters and obtaining the optimal denoising model by minimizing the mean squared error (MSE) loss function between the predicted denoised signal and the original signal.

[0055] Furthermore, the training process employs supervised learning, using noisy signals as model input and the original clean signal as target output. The model calculates the predicted signal through forward propagation and iteratively updates the network parameters using backpropagation combined with stochastic gradient descent (SGD) or the Adam optimizer. Specifically, the SACNN-BiLSTM model consists of multiple convolutional layers (Conv1D), self-attention layers, average pooling layers (AvgPool), bidirectional long short-term memory (BiLSTM) layers, dropout layers, and fully connected (dense) layers. Each layer participates in the joint modeling of feature extraction and noise suppression. Among them, the self-attention mechanism uses a multi-head structure (8 attention heads) to globally weight the feature maps, enhancing the model's ability to model long-range dependencies, thereby improving noise reduction accuracy.

[0056] Furthermore, this training process is applicable to vibration signals collected in structural health monitoring systems, such as acceleration response data of infrastructure like bridges and high-rise buildings. By collecting raw signals in laboratory or actual engineering environments and artificially injecting Gaussian white noise or pink noise with different signal-to-noise ratios (1-5dB), representative training and validation datasets are constructed, thereby training a general noise reduction model suitable for various noise environments.

[0057] Furthermore, through an end-to-end training mechanism, the model can automatically learn the complex nonlinear mapping relationship between noise and the original signal, without relying on the manual setting of wavelet basis functions, decomposition levels, or thresholds as in traditional methods. The final optimal denoising model shows a significant improvement in denoising performance on the test set, as shown in Table 1. The root mean square error (RMSE) of the denoised signal is reduced by about 40%, and the signal-to-noise ratio (SNR) is improved by about 150%, fully verifying the technical value of this invention in achieving efficient and intelligent signal denoising in the field of structural health monitoring.

[0058] Furthermore, S3 also includes: S31, after the convolution operation, the Leaky ReLU activation function is added, and its calculation formula is as follows: .

[0059] Specifically, in the SACNN-BiLSTM hybrid neural network constructed in this invention, the introduction of the Leaky ReLU activation function after convolution is a key step in achieving nonlinear feature mapping and enhancing the model's expressive power. This step is technically based on the fundamental principles of activation functions in deep learning, aiming to enhance the network's ability to model complex noise and signal relationships.

[0060] Furthermore, the convolutional layer employs a one-dimensional convolution (Conv1D) structure to extract local temporal features from the input vibration signal. After the convolution operation, elements in each feature map are input into the Leaky ReLU activation function for nonlinear transformation. The mathematical expression for Leaky ReLU is: .

[0061] Specifically, compared to the traditional ReLU function, Leaky ReLU assigns a non-zero gradient when the input is negative, thereby effectively alleviating the "neuron death" problem, avoiding the gradient vanishing phenomenon, and improving the stability and convergence speed of the model during training.

[0062] Furthermore, this step is applicable to feature extraction and nonlinear mapping modeling of noisy vibration signals in structural health monitoring systems. Since vibration signals acquired in actual engineering projects are often subject to various noise interferences, their signal-noise relationship is highly nonlinear. The use of Leaky ReLU can enhance the model's ability to fit this complex relationship, thereby improving noise reduction accuracy.

[0063] Furthermore, the introduction of Leaky ReLU significantly enhances the nonlinear expressive power of the SACNN-BiLSTM model, enabling the network to more effectively learn the mapping relationship between noise and the original signal. Combined with the subsequent self-attention mechanism and BiLSTM structure, this activation function provides the model with stronger generalization ability and robustness, which is an important technical support for achieving high-quality vibration signal denoising.

[0064] S32, the hybrid neural network model contains a dropout layer, and each neuron is dropped with an independent probability p=0.5 during training.

[0065] Specifically, in the SACNN-BiLSTM hybrid neural network model constructed in this invention, a dropout layer is introduced to enhance the model's generalization ability and alleviate overfitting. During training, each neuron is randomly dropped with an independent probability p=0.5; that is, during each forward propagation, the neuron's output is set to zero with a 50% probability, and retained and multiplied by 2 with a 50% probability as compensation. This mechanism forces the network to operate independently of any single neuron during training, thereby improving the model's robustness to noise and input perturbations. In some implementations, the dropout layer is deployed before the fully connected layer of the network to prevent that layer from overfitting noisy patterns in the training data due to excessive parameters.

[0066] Furthermore, p=0.5 is the optimal parameter setting verified through experiments. This value is widely considered in the field of deep learning as a recommended value for balancing the model's generalization ability and training efficiency. In this invention, the selection of this parameter is based on the results of multiple cross-validations on vibration signals with different noise levels (1-5dB), ensuring that the model can maintain high denoising accuracy and stability when facing high noise interference. In addition, the introduction of randomness in the dropout layer makes the network structure slightly different in each training process, thereby simulating the effect of ensemble learning to a certain extent and improving the average performance of the model.

[0067] Furthermore, this dropout mechanism is applicable to vibration signals acquired in structural health monitoring systems, especially in complex on-site environments with diverse noise sources (such as Gaussian white noise and pink noise). By introducing a random dropout strategy during the training phase, the model can better adapt to signal characteristics under different noise distributions, thus exhibiting stronger generalization ability during the testing phase.

[0068] Furthermore, this step effectively suppresses the overfitting of the model to noise signals during training, improving its denoising performance in unknown noise environments. Combining the global feature extraction capability of SACNN and the time series modeling advantages of BiLSTM, the introduction of the dropout layer further enhances the stability and reliability of the entire hybrid neural network, providing a solid technical guarantee for intelligent denoising of vibration signals in structural health monitoring.

[0069] S4. The optimal noise reduction model is used to denoise the noisy signals in the test set, and pink noise of different levels is injected into the same reference signal as the original signal to generate extended test data. The extended test data is then denoised using the optimal noise reduction model to verify the model's adaptability to various noise types.

[0070] Specifically, this step aims to verify the adaptability of the SACNN-BiLSTM hybrid neural network denoising model to various noise types, particularly its generalization ability in non-white noise scenarios. Specifically, this step first uses the optimal denoising model to denoise noisy signals in the test set that have already had Gaussian white noise added, to evaluate the model's performance under standard noise conditions. Then, different levels of pink noise are injected into the same time series as the original signal to generate an extended test dataset, and the optimal model is used again for denoising, thereby verifying its adaptability to noise with different spectral characteristics.

[0071] Furthermore, the extended test data generation uses the same original signal as in S1 as a reference, injecting pink noise into it by adjusting the noise level (e.g., 1-5 dB). The power spectral density (PSD) of pink noise is inversely proportional to the frequency f, i.e., PSD ∝ 1 / f, and its spectral characteristics are closer to actual environmental noise, such as wind noise and mechanical vibration noise. This noise injection process is implemented using digital signal processing tools, typically employing frequency domain synthesis methods. First, white noise is generated, and then its spectrum is shaped using a 1 / f filter to ensure it conforms to the spectral distribution characteristics of pink noise.

[0072] Furthermore, the noise level of the extended test data was kept consistent with that of the Gaussian white noise test set to ensure the fairness of the comparative experiments. When processing the extended test data, the denoising model used a 1D vibration time-series signal of length N as input, with a sampling frequency of 50Hz, and outputted a denoised signal of the same length as the input. The model's loss function remained the root mean square error (MSE), and parameters were updated using the Adam optimizer with a learning rate of 0.001 and a batch size of 64.

[0073] This step is of great significance in practical applications, especially in structural health monitoring systems. Vibration signals collected on-site are often subject to various noise contaminations, such as environmental noise and equipment noise, and their spectral characteristics may differ significantly from Gaussian white noise. By introducing pink noise testing, actual working conditions can be simulated more realistically, verifying the robustness and generalization ability of the model in non-ideal noise environments.

[0074] Furthermore, this step not only evaluated the model's denoising performance under standard noise but also verified its ability to handle noise with uneven spectral distribution. By calculating the signal-to-noise ratio (SNR) and root mean square error (RMSE) of the denoised signal, the model's denoising effect under different noise types can be quantified. Experimental results show that the model's SNR improvement under pink noise conditions is similar to that under Gaussian white noise, indicating that it has good adaptability across noise types and provides strong support for practical engineering applications.

[0075] S41, the power spectral density (PSD) of pink noise is inversely proportional to the frequency f, and its mathematical expression is: .

[0076] Specifically, in step 4 of this invention, for the evaluation of the noise reduction performance of pink noise, the same original vibration signal as in step 1 is used as the basis, and pink noise of different noise levels is superimposed on it to construct a more challenging test dataset. Pink noise is a type of noise with specific energy distribution characteristics in the frequency domain. Its power spectral density (PSD) is inversely proportional to the frequency f, and the mathematical expression is: .

[0077] This characteristic indicates that pink noise has a higher energy density in the low-frequency band, while its energy gradually decays in the high-frequency band. It is highly similar to environmental noise commonly found in actual structural health monitoring (such as wind vibration and traffic noise). Therefore, the evaluation of its noise reduction effect is more engineering-significant and practically representative.

[0078] Furthermore, the spectral characteristics of the original signal need to be analyzed using Fast Fourier Transform (FFT) to ensure that the superimposed pink noise satisfies the aforementioned inverse relationship in the frequency domain. The noise signal can be generated using either filtering or frequency domain synthesis, with frequency domain synthesis being more accurate. First, a white noise signal with a flat power spectrum is generated. Then, a 1 / f filter is designed to perform frequency domain weighting on the white noise, ensuring that its power spectral density satisfies the characteristic of being inversely proportional to the frequency at each frequency point.

[0079] In this invention, the noise level is controlled by the signal-to-noise ratio (SNR), specifically set to 1-5 dB, to simulate noise interference of varying intensities in actual monitoring environments. The superimposed noisy signal is used as a test set and input into the trained SACNN-BiLSTM denoising model. The model, through its multi-level feature extraction and time-series modeling capabilities, effectively suppresses the noise components in the signal and outputs the denoised vibration signal. This step not only verifies the model's adaptability to non-uniform spectral noise but also further demonstrates the robustness and generalization ability of SACNN-BiLSTM in complex noisy environments without prior knowledge.

[0080] S42 evaluates the noise reduction performance by calculating the root mean square error (RMSE) and signal-to-noise ratio (SNR) of the denoised signal. The formula for calculating RMSE is as follows: , The formula for calculating SNR is: .

[0081] Specifically, this invention quantitatively evaluates the performance of the proposed SACNN-BiLSTM hybrid neural network in denoising vibration signals for structural health monitoring by calculating the root mean square error (RMSE) and signal-to-noise ratio (SNR) of the denoised signal. This step employs standard signal quality assessment methods at the technical implementation level, combined with actual test data, to quantitatively analyze the difference between the denoised signal output by the model and the original clean signal, thereby verifying the model's denoising effect.

[0082] Furthermore, the formula for calculating RMSE is: , in, Indicates the first The original signal values ​​of each sampling point The total number of sampling points for the signal. RMSE reflects the overall error level between the denoised signal and the original signal; a smaller value indicates a better denoising effect. In this invention, RMSE is used to measure the denoising accuracy of the model at different noise levels (1-5dB), especially the performance difference when dealing with Gaussian white noise and pink noise.

[0083] The formula for calculating the signal-to-noise ratio (SNR) is: , This formula evaluates the relative proportion of effective and noise components in a signal by comparing the energy of the original signal with the energy of the error between the denoised signal and the original signal. A higher SNR value indicates that more of the original signal components are preserved in the denoised signal, and the noise is suppressed more thoroughly. In embodiments of this invention, SNR is used to evaluate the denoising capability of a model under different noise types (such as Gaussian white noise and pink noise), particularly whether the model can effectively recover the original characteristics of the signal in the absence of prior knowledge.

[0084] Furthermore, this invention employs a uniform number of sampling points. Calculations are performed to ensure comparability between different test samples. Simultaneously, noise levels are considered. Gradient injection was performed with step sizes ranging from 0.1 to 0.5 to simulate noise interference of varying intensities in real-world engineering scenarios. After model training, RMSE and SNR were calculated on the test set and pink noise test data, respectively, as key indicators of the model's generalization ability and robustness.

[0085] This step plays a crucial role in the evaluation and verification of the entire technical solution. By introducing two complementary evaluation metrics, RMSE and SNR, we can not only measure the model's accuracy from an error perspective but also assess its noise suppression capability from a signal energy perspective. This multi-dimensional performance evaluation method helps to comprehensively verify the practicality and superiority of the SACNN-BiLSTM model in denoising vibration signals for structural health monitoring, providing reliable data support for subsequent model optimization and engineering deployment.

[0086] The vibration signal noise reduction method for structural health monitoring based on hybrid neural networks in this invention can automatically extract noise-insensitive features from vibration signals, effectively remove various noise interferences, improve the signal-to-noise ratio and noise reduction accuracy, and achieve efficient and intelligent noise reduction of structural health monitoring signals.

[0087] Example 2 The purpose of this invention is to propose a method for denoising vibration signals of structural health monitoring based on SACNN-BiLSTM hybrid neural network. This method can achieve efficient and automated denoising of SHM vibration signals without relying on prior knowledge of the target signal or manual parameter setting.

[0088] To achieve the objectives of this invention, the invention is implemented through the following technical solutions, such as... Figure 2 As shown: In one embodiment of the present invention, a method for denoising vibration signals for structural health monitoring based on a SACNN-BiLSTM hybrid neural network is characterized by comprising the following steps: Specifically, S101 uses sensors to collect the raw signal of the structural vibration response as a clean signal. Adding different levels of Gaussian white noise to the original signal yields a noisy signal. The dataset is divided into training, validation and test sets according to the appropriate proportions. Specifically, in S102, the following will be introduced: Figure 3 The self-attention mechanism convolutional neural network shown is combined with a bidirectional long short-term memory neural network to construct a system as follows: Figure 4 The SACNN-BiLSTM hybrid neural network noise reduction model shown; Specifically, S103, such as Figure 5 As shown, the noisy signal As the input signal, the original signal As the target signal, noise is labeled accordingly, and a system is built. and mapping relationship This generates a noise-reduced signal. This enables the prediction of clean signals. Mapping relationship. The iterative optimization objective is to minimize the predicted denoised signal. With the corresponding original signal The difference between them is:

[0089] The SACNN-BiLSTM hybrid neural network denoising model is trained using the training set and validation set to obtain the optimal model. Then, the optimal model is used to denoise the test set data to obtain the denoised vibration signal. Specifically, in S104, to further evaluate the applicability of the proposed method in practice, a set of test data is generated by adding different levels of pink noise to the original signal, which is the same as in step 1, and the noise reduction model trained in S103 is used to perform noise reduction processing on it.

[0090] Furthermore, the method for denoising vibration signals for structural health monitoring based on a SACNN-BiLSTM hybrid neural network is characterized by the following: This patent proposes a novel deep learning model—SACNN-BiLSTM—and applies it to denoising vibration signals for structural health monitoring. The specific architecture of the constructed SACNN-BiLSTM hybrid neural network includes, in sequence, an input layer, a convolutional layer, a self-attention layer, a pooling layer, a convolutional layer, a self-attention layer, a pooling layer, a recurrent layer, a recurrent layer, a dropout layer, a fully connected layer, and an output layer. The convolutional layer is Conv1D (one-dimensional convolution), the pooling layer is Avgpool (average pooling), and the recurrent layer is BiLSTM (bidirectional long short-term memory). After the convolutional layer, the activation function is Leaky ReLU (leakage rectified linear unit), and the loss function is MSE (root mean square error).

[0091] Furthermore, the aforementioned method for denoising vibration signals for structural health monitoring based on a SACNN-BiLSTM hybrid neural network is characterized by the fact that the relationship between noise and the original signal is undoubtedly complex, and linear functions are insufficient to model this relationship. Therefore, an activation function is added after the convolution operation to transform the mapping from linear to nonlinear, enabling SACNN-BiLSTM to model this complex nonlinear relationship. Leaky ReLU, as an improved version of ReLU, is a high-level activation function, and its specific calculation formula is as follows: , The Leaky ReLU function has a wider output range than the ReLU function. When the input is negative, it gives the input a very small slope, which solves the zero gradient problem in the case of negative input and effectively alleviates the vanishing gradient problem.

[0092] Furthermore, the aforementioned method for denoising vibration signals in structural health monitoring based on a SACNN-BiLSTM hybrid neural network is characterized by the introduction of a self-attention mechanism within the SACNN-BiLSTM hybrid neural network. This self-attention mechanism comprises three key components: query, key, and value, all of which are homologous. Each position in the input feature map is treated as a query, and the correlation between the query and other positions is calculated by weighting the similarity between the query and the key to obtain the value of the query relative to other positions. This approach can capture the dependencies within the input sequence. The mathematical expression of the self-attention mechanism is shown below:

[0093] In the formula, Q, K, and V represent query, key, and value, respectively, and T represents transpose. It is generally a relatively small positive number.

[0094] Multi-head self-attention is an extension of single-head self-attention by using multiple independent attention heads to process the input sequence simultaneously. Each attention head learns its own linear transformations of Q, K, and V, and then performs independent self-attention computation. Finally, the outputs of the multiple attention heads are combined through linear transformations and concatenation to obtain the final output. In this case, multi-head self-attention is applied to each channel of the feature map in the neural network, with a maximum of 8 heads.

[0095] The specific architecture of the self-attention layer is as follows: First, the input data is compressed using GAP (Global Average Pooling). Then, the compressed data is compared with three different weight matrices. , , Multiply to get The representation, then using , The representation calculates the self-attention weight H and combines it with... Multiplication performs a weighted average of the features at each location to obtain the output. Finally, the weighted features of all channels are fused with the original features to generate a new feature map, which is then passed to subsequent network layers. The self-attention layer allows the entire input vector to be processed at once, which greatly expands the receptive field of CNNs (Convolutional Neural Networks) limited by the kernel size. SACNN-BiLSTM embeds the self-attention mechanism into the computation process to extract richer global features from the data.

[0096] Furthermore, the proposed method for denoising vibration signals for structural health monitoring based on a SACNN-BiLSTM hybrid neural network is characterized by the use of dropout technology, an emerging tool for addressing overfitting. When training with different batches of samples, a portion of neurons are randomly selected and discarded, resulting in more robust training of the remaining neurons and enhanced generalization ability of the network. Each neuron has an independent probability p for being discarded, where p=0.5 is a suggested parameter value used for training the proposed SACNN-BiLSTM network. Furthermore, the aforementioned method for denoising vibration signals for structural health monitoring based on a SACNN-BiLSTM hybrid neural network is characterized by evaluating the denoising performance of the denoising model by calculating the root mean square error (RMSE) and signal-to-noise ratio (SNR) of the denoised signal.

[0097]

[0098] In the formula, n represents the total number of sampling points. Indicates the noise reduction signal. This represents the original signal.

[0099] Furthermore, the method for denoising vibration signals for structural health monitoring based on a SACNN-BiLSTM hybrid neural network is characterized in that, in step 1, different levels of Gaussian white noise are injected into the original signal according to the following formula:

[0100] In the formula, noise is a normally distributed random vector with a mean of zero and a standard deviation identical to that of the original signal. It refers to the noise level.

[0101] Furthermore, unlike Gaussian white noise whose energy intensity is uniformly distributed across the entire frequency band, the power spectral density (PSD) of pink noise is inversely proportional to the frequency f, and can be expressed as:

[0102] Furthermore, the aforementioned method for denoising vibration signals in structural health monitoring based on a SACNN-BiLSTM hybrid neural network is characterized by combining the advantages of CNN (Convolutional Neural Network) in handling noise interference and extracting weak signal features with the advantages of BiLSTM (Bidirectional Long Short-Term Memory Neural Network) in processing time series. It also introduces a self-attention mechanism into the network. Self-attention is a powerful modeling tool in deep learning; by calculating the attention weights between elements, it can better handle long-range dependencies and capture global information, learning the importance and correlation between extracted multi-level features, thereby further improving model performance.

[0103] The present invention also has the following technical effects Specifically, the SACNN-BiLSTM denoising model combines the advantages of both models by using a CNN with a self-attention mechanism to extract local and global features from the input data and a BiLSTM layer to capture the temporal dependencies between these features. This allows the model to better capture complex patterns and dependencies in time-series data while avoiding the limitations of using each model alone. Furthermore, using Leaky ReLU as the activation function improves the model's non-linear modeling ability, solves the gradient vanishing and exploding problems, reduces training difficulty, improves the model's generalization performance, and ensures the good performance of the intelligent denoising model. Furthermore, the SACNN-BiLSTM denoising model directly processes one-dimensional vibration signals, automatically extracting noise-insensitive features from the measurement signal to achieve denoising. This effectively avoids the manual selection and fine-tuning of relevant parameters in traditional denoising methods, requiring no prior knowledge of the vibration signal or human intervention. The process is simple to operate, achieving automated and intelligent denoising of vibration signals. Moreover, the SACNN-BiLSTM denoising model can effectively eliminate different types of noise. In summary, this invention has the advantages of simple operation, high signal noise reduction efficiency, and good noise reduction quality. It can be used for noise reduction of signals from different sources and with different characteristics, and has broad application prospects in engineering practice.

[0104] Example 3 To achieve the above embodiments, such as Figure 6 As shown, this embodiment also provides a structural health monitoring vibration signal noise reduction device 10 based on a hybrid neural network, comprising: The structural vibration signal acquisition and data division module 100 is used to acquire the original signal of structural vibration response, inject different levels of Gaussian white noise into the original signal to generate a noisy signal, and divide the noisy signal and the original signal into a training set, a validation set and a test set according to a preset ratio. The hybrid neural network model building module 200 is used to build a hybrid neural network model that includes a convolutional neural network with a self-attention mechanism and a bidirectional long short-term memory neural network. The convolutional neural network uses alternating one-dimensional convolutional layers and average pooling layers, and the self-attention mechanism performs global weighted fusion of feature maps through a multi-head attention module. The end-to-end training and model optimization module 300 is used to input the training set and validation set into the hybrid neural network model for end-to-end training, so as to minimize the mean square error loss function between the predicted denoised signal and the original signal and obtain the optimal denoising model. The noise reduction and adaptability verification module 400 is used to perform noise reduction processing on the noisy signals in the test set using the optimal noise reduction model, and further inject pink noise of different levels into the same reference signal as the original signal to generate extended test data. The extended test data is then subjected to noise reduction processing by the optimal noise reduction model to verify the model's adaptability to multiple noise types.

[0105] Furthermore, the structural vibration signal acquisition and data segmentation module 100 is also used for: An accelerometer was used to acquire the raw signal of the structural vibration response at a sampling frequency of 50 Hz; The dataset is divided into training, validation, and test sets in a 2:1:1 ratio.

[0106] Furthermore, the hybrid neural network model building module 200 is also used for: The specific architecture of the self-attention layer is as follows: First, the input data is compressed using global average pooling. Then, the compressed data is multiplied by three different weight matrices to obtain the representations Q, K, and V. The self-attention weight H is calculated using the representations of Q and K, and then multiplied by V to achieve a weighted average of the features at each position to obtain the output. The number of heads in the multi-head self-attention mechanism is set to 8, and the multi-head self-attention mechanism is applied to each channel of the feature map in the neural network.

[0107] Furthermore, the end-to-end training and model optimization module 300 is also used for: After the convolution operation, the Leaky ReLU activation function is added, and its calculation formula is as follows: ; The hybrid neural network model includes a dropout layer, where each neuron is dropped with an independent probability p=0.5 during training.

[0108] This invention discloses a method for denoising vibration signals in structural health monitoring based on a hybrid neural network. This method does not rely on prior knowledge of the signal or manual parameter setting, and effectively achieves automated denoising of various noise types in structural health monitoring vibration signals. It significantly improves the signal-to-noise ratio and reduces the root mean square error, while preserving key vibration characteristic information of the structure.

[0109] In the description of this specification, the references to terms such as "one embodiment," "some embodiments," "example," "specific example," or "some examples," etc., refer to specific features, structures, materials, or characteristics described in connection with that embodiment or example, which are included in at least one embodiment or example of the present invention. In this specification, the illustrative expressions of the above terms do not necessarily refer to the same embodiment or example. Furthermore, the specific features, structures, materials, or characteristics described may be combined in any suitable manner in one or more embodiments or examples. Moreover, without contradiction, those skilled in the art can combine and integrate the different embodiments or examples described in this specification, as well as the features of different embodiments or examples.

[0110] Furthermore, the terms "first" and "second" are used for descriptive purposes only and should not be construed as indicating or implying relative importance or implicitly specifying the number of technical features indicated. Thus, a feature defined as "first" or "second" may explicitly or implicitly include at least one of that feature. In the description of this invention, "a plurality of" means at least two, such as two, three, etc., unless otherwise explicitly specified.

Claims

1. A method for denoising vibration signals in structural health monitoring based on hybrid neural networks, characterized in that, include: S1. Obtain the original signal of structural vibration response, and inject different levels of Gaussian white noise into the original signal to generate a noisy signal. Divide the noisy signal and the original signal into a training set, a validation set and a test set according to a preset ratio. S2, Construct a hybrid neural network model that includes a convolutional neural network with a self-attention mechanism and a bidirectional long short-term memory neural network, wherein the convolutional neural network uses alternating one-dimensional convolutional layers and average pooling layers, and the self-attention mechanism performs global weighted fusion of feature maps through a multi-head attention module; S3, input the training set and validation set into the hybrid neural network model for end-to-end training to minimize the mean square error loss function between the predicted denoised signal and the original signal, and obtain the optimal denoising model; S4. The optimal noise reduction model is used to denoise the noisy signals in the test set, and pink noise of different levels is injected into the same reference signal as the original signal to generate extended test data. The extended test data is then denoised using the optimal noise reduction model to verify the model's adaptability to various noise types.

2. The method as described in claim 1, characterized in that, S1 includes: S11 uses an accelerometer to collect the original signal of the structural vibration response at a sampling frequency of 50Hz; S12, the dataset is divided into training, validation and test sets in a 2:1:1 ratio.

3. The method as described in claim 1, characterized in that, S2 includes: S21, The specific architecture of the self-attention layer is as follows: First, the input data is compressed using global average pooling. Then, the compressed data is multiplied by three different weight matrices to obtain the representations Q, K, and V. The self-attention weight H is calculated using the representations of Q and K, and multiplied by V to achieve a weighted average of the features at each position to obtain the output. S22, the number of heads in the multi-head self-attention mechanism is 8, and the multi-head self-attention mechanism is applied to each channel of the feature map in the neural network.

4. The method as described in claim 1, characterized in that, The S3 includes: S31, after the convolution operation, the Leaky ReLU activation function is added, and its calculation formula is as follows: ； S32, the hybrid neural network model contains a dropout layer, and each neuron is dropped with an independent probability p=0.5 during training.

5. The method as described in claim 1, characterized in that, The S4 includes: S41, the power spectral density (PSD) of pink noise is inversely proportional to the frequency f, and its mathematical expression is: ； S42 evaluates the noise reduction performance by calculating the root mean square error (RMSE) and signal-to-noise ratio (SNR) of the denoised signal. The formula for calculating RMSE is as follows: ， The formula for calculating SNR is: 。 6. The method as described in claim 1, characterized in that, include: S5. During the training of the hybrid neural network model, a normalization layer is inserted between the convolutional layer and the self-attention layer. The normalization layer uses batch normalization technology to standardize the feature map in order to accelerate model convergence and improve training stability.

7. A vibration signal noise reduction device for structural health monitoring based on a hybrid neural network, characterized in that, include: The structural vibration signal acquisition and data partitioning module is used to acquire the original signal of structural vibration response, inject different levels of Gaussian white noise into the original signal to generate a noisy signal, and divide the noisy signal and the original signal into a training set, a validation set and a test set according to a preset ratio. A hybrid neural network model building module is used to build a hybrid neural network model that includes a convolutional neural network with a self-attention mechanism and a bidirectional long short-term memory neural network. The convolutional neural network uses alternating one-dimensional convolutional layers and average pooling layers, and the self-attention mechanism performs global weighted fusion of feature maps through a multi-head attention module. The end-to-end training and model optimization module is used to input the training set and validation set into the hybrid neural network model for end-to-end training, so as to minimize the mean square error loss function between the predicted denoised signal and the original signal and obtain the optimal denoising model. The noise reduction and adaptability verification module is used to perform noise reduction processing on the noisy signals in the test set using the optimal noise reduction model, and further inject pink noise of different levels into the same reference signal as the original signal to generate extended test data. The extended test data is then subjected to noise reduction processing by the optimal noise reduction model to verify the model's adaptability to multiple noise types.

8. The apparatus as claimed in claim 7, characterized in that, The structural vibration signal acquisition and data segmentation module is also used for: An accelerometer was used to acquire the raw signal of the structural vibration response at a sampling frequency of 50 Hz; The dataset is divided into training, validation, and test sets in a 2:1:1 ratio.

9. The apparatus as claimed in claim 7, characterized in that, The hybrid neural network model building module is also used for: The specific architecture of the self-attention layer is as follows: First, the input data is compressed using global average pooling. Then, the compressed data is multiplied by three different weight matrices to obtain the representations Q, K, and V. The self-attention weight H is calculated using the representations of Q and K, and then multiplied by V to achieve a weighted average of the features at each position to obtain the output. The number of heads in the multi-head self-attention mechanism is set to 8, and the multi-head self-attention mechanism is applied to each channel of the feature map in the neural network.

10. The apparatus as claimed in claim 7, characterized in that, The end-to-end training and model optimization module is also used for: After the convolution operation, the Leaky ReLU activation function is added, and its calculation formula is as follows: ； The hybrid neural network model includes a dropout layer, where each neuron is dropped with an independent probability p=0.5 during training.

Images