Bridge structure anomaly detection method and system based on cavity convolution feature fusion

Abstract

The application discloses a bridge structure anomaly detection method and system based on cavity convolution feature fusion, and the method is as follows: collecting historical monitoring data of a bridge structure, extracting short-term, medium-term and long-term time sequence features in the historical monitoring data to obtain multi-branch feature maps; using cavity convolution with different expansion rates to extract features of the feature maps of each branch, calculating attention weights of each branch according to the features of the branch maps, weighting and fusing the features of each branch to obtain fused features; performing channel projection on the multi-branch feature maps to obtain residual features, adding the residual features and the fused features to obtain spliced features, inputting the spliced features into a time sequence convolution, generating a gating value through an activation function, element-by-element multiplying the gating value and the spliced features to obtain output features; remodeling the output features into a monitoring data prediction result of a time sequence, calculating a difference value between real-time monitoring data and the prediction result, and judging that the bridge structure is abnormal when the difference value exceeds a set difference value range.

Description

Technical Field

[0001] This invention relates to the field of bridge structure monitoring technology, specifically to a method and system for detecting bridge structural anomalies based on the fusion of void convolutional features. Background Technology

[0002] As vital transportation infrastructure, the structural health of bridges directly impacts public safety and transportation efficiency. In the field of bridge inspection, traditional methods primarily rely on regular manual inspections. This approach is not only time-consuming and labor-intensive but also susceptible to human error, making it difficult to detect potential structural anomalies in a timely and accurate manner.

[0003] In recent years, with the rapid development of IoT and AI technologies, intelligent bridge anomaly detection methods based on sensor data and deep learning have gradually become a research hotspot. Deep learning has made significant progress in the field of structural health monitoring, with various neural network models, such as convolutional neural networks (CNNs) and recurrent neural networks (RNNs), being widely applied to structural anomaly detection tasks. These models have improved the efficiency and accuracy of bridge anomaly detection to some extent.

[0004] However, traditional models still have some limitations when processing multi-source heterogeneous data on bridge structures. Existing technologies often extract features from bridge monitoring data by using fixed convolution kernels to extract features at a single time scale. This can easily lead to the omission of key information about the bridge structure at different time dimensions, and an inability to effectively balance local details and global correlations. As a result, the features are incomplete and fail to accurately reflect the true state of the bridge. Summary of the Invention

[0005] This invention proposes a bridge structure anomaly detection method and system based on dilated convolution feature fusion, which solves the problem of low accuracy of existing anomaly detection methods when processing multi-source heterogeneous data of bridge structures.

[0006] To address the aforementioned technical problems, this invention provides a bridge structural anomaly detection method based on dilated convolution feature fusion, comprising the following steps:

[0007] Step S1: Collect historical monitoring data of the bridge structure, extract short-term, medium-term and long-term time series features from the historical monitoring data, and obtain a multi-branch feature map;

[0008] Step S2: Extract the features of the feature map of each branch using dilated convolution with different dilation rates, calculate the attention weight of each branch based on the features of each branch map, and perform weighted fusion of the features of each branch based on the attention weight to obtain the fused features;

[0009] Step S3: Perform channel projection on the multi-branch feature map to obtain residual features. Add the residual features and the fused features to obtain the stitched features. Input the stitched features into a temporal convolution and generate a gate value through an activation function. Multiply the gate value element-wise with the stitched features to obtain the output features.

[0010] Step S4: Reshape the output features into a time series to obtain the prediction results of the monitoring data, calculate the difference between the real-time monitoring data and the prediction results, and determine the bridge structure as abnormal when the difference exceeds the set difference range.

[0011] Preferably, step S1 involves extracting short-term, medium-term, and long-term time-series features from the historical monitoring data, including the following steps:

[0012] Step S11: Perform a fast Fourier transform on the historical monitoring data to extract the period of the historical monitoring data, and reorganize the one-dimensional monitoring data into a two-dimensional matrix according to the period length;

[0013] Step S12: Extract short-term, medium-term, and long-term temporal features of the two-dimensional matrix using convolution kernels of different scales.

[0014] Preferably, the expression for reorganizing the one-dimensional monitoring data into a two-dimensional matrix according to the period length in step S11 is as follows:

[0015]

[0016] In the formula, X 2D [i] is a two-dimensional matrix; Reshape() represents the reshaping operation; x(t) is the monitoring data; T i For the i-th cycle; L is the total length of the monitoring data.

[0017] Preferably, the expression for extracting the features of each branch's feature map using dilated convolutions with different dilation rates in step S2 is:

[0018]

[0019] In the formula, F i Let x be the feature map of the i-th branch; K is the upper limit of the number of elements in the dilated convolution kernel; x is the feature map; r is the dilation rate; k is the kernel size; w(k) is the kernel weight; x(i+r·k) represents the position in the feature map x after offset from position i with dilation rate r and kernel size k.

[0020] Preferably, step S2, which involves calculating the attention weight of each branch based on the characteristics of each branch graph, includes the following steps:

[0021] Step S21: Generate a spatial attention map for each branch using dilated convolution;

[0022] Step S22: Concatenate the spatial attention maps of multiple branches along the branch dimension, and normalize the weights of all branches at each spatial location along the branch dimension to obtain the attention weights of each branch at each spatial location.

[0023] Preferably, the expression for generating a spatial attention map for each branch using dilated convolution in step S21 is:

[0024]

[0025] In the formula, This is the spatial attention graph for the i-th branch; DilatedConv i Indicates dilated convolution operation; F i The features of the feature map of the i-th branch.

[0026] Preferably, the expression for normalizing the weights of all branches at each spatial location in the branch dimension in step S22 is as follows:

[0027]

[0028] In the formula, ω i,h,w is the weight of the i-th branch at spatial location (h, w); softmax() is the normalized exponential function; Let be the spatial attention graph of the i-th branch.

[0029] Preferably, the expression for the fusion feature in step S2 is:

[0030]

[0031] In the formula, F fused For fusion features; ω i,h,w Let F be the weight of the i-th branch at spatial location (h, w); n is the total number of branches; F i The features of the feature map of the i-th branch.

[0032] Preferably, the expression for the output feature in step S3 is:

[0033] F final =(F fused +F residual )⊙Gate;

[0034] F residual =ChannelProjection(x);

[0035] Gate = σ(TemporalConv(F)fused +F residual ));

[0036] In the above formula, F final For output features; F fused For fusion features; F residual is the residual feature; represents element-wise multiplication; Gate is the gate value; ChannelProjection() represents the channel projection operation; x is the feature map; σ is the Sigmoid activation function; TemporalConv() represents the convolution operation along the time axis.

[0037] This invention also provides a bridge structure anomaly detection system based on dilated convolution feature fusion, which is based on the above-mentioned bridge structure anomaly detection method based on dilated convolution feature fusion, and includes: a data acquisition module, a temporal feature extraction module, a multi-scale feature learning module, an anomaly diagnosis module, and a visualization early warning module;

[0038] The data acquisition module collects monitoring data of the bridge structure.

[0039] The time-series feature extraction module converts historical monitoring data of the bridge structure into frequency domain signals, extracts periodic frequency components, filters the top k frequency components with the highest power to obtain the period length, and reorganizes the monitoring data into a two-dimensional tensor based on the period length.

[0040] The multi-scale feature learning module extracts short-term, medium-term, and long-term features from a two-dimensional tensor using different convolutional kernels, expands the receptive field by adjusting the dilation rate, captures the relationships between multi-scale features, dynamically generates weights to achieve adaptive feature fusion, uses temporal convolution and activation functions to generate gating values, strengthens key anomalous temporal features, and outputs predicted values.

[0041] The anomaly diagnosis module calculates the difference between the predicted value and the real-time collected monitoring data. When the difference exceeds the set difference range, it determines that the bridge structure is abnormal, determines the area where the anomaly occurs by combining the sensor location, traces back the starting time of the anomaly through time-series characteristics, and classifies the anomaly level according to the difference magnitude.

[0042] The visualization and early warning module dynamically displays the raw data of each sensor and anomaly point markers, visualizes the spatiotemporal distribution of abnormal events, generates structural health scores and trend analysis charts, and pushes abnormal information via SMS, email, or API interface.

[0043] The advantages of this invention include at least the following:

[0044] 1. By extracting short-term, medium-term, and long-term time-series features, the dynamic changes of the bridge structure can be fully captured from different time scales, avoiding feature omissions caused by analysis at a single time scale; by using dilated convolutions with different dilation rates, the receptive field of the convolution kernel is effectively expanded, while considering local details and global correlations, thus improving the comprehensiveness of feature extraction.

[0045] 2. By dynamically allocating weights based on the importance of each branch feature, key features can be highlighted, redundant information can be suppressed, the fused features can be made more representative, and the model's ability to identify abnormal features can be improved. The gating mechanism enables the model to adaptively adjust features and dynamically determine the degree of feature retention or suppression based on data characteristics, thereby improving adaptability to complex working conditions and noise. Attached Figure Description

[0046] Figure 1 This is a schematic diagram of the method flow according to an embodiment of the present invention;

[0047] Figure 2 This is a schematic diagram of the system hierarchy according to an embodiment of the present invention;

[0048] Figure 3 This is a schematic diagram of the structure of the multi-scale feature learning module in an embodiment of the present invention. Detailed Implementation

[0049] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. 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 are within the protection scope of the present invention.

[0050] While existing bridge structural anomaly detection technologies have been applied in practice, they still face numerous challenges in achieving accurate and timely anomaly detection. Bridge structural data is multi-source and heterogeneous, encompassing various types such as vibration, strain, and displacement data. These data exhibit complex spatiotemporal relationships, making it difficult for traditional anomaly detection methods to effectively integrate and analyze them. Bridge structural data is susceptible to environmental noise and sensor malfunctions, reducing the reliability of anomaly detection. Furthermore, structural data often exhibits strong temporal characteristics and periodic variations, which are closely related to environmental factors such as temperature and traffic flow, further increasing the difficulty of anomaly pattern recognition. Traditional anomaly detection methods rely heavily on threshold settings and expert experience, making it difficult to accurately identify minor structural anomalies. Conventional machine learning methods perform poorly when processing high-dimensional nonlinear data of bridge structures, easily leading to false positives or false negatives in anomaly detection.

[0051] Meanwhile, bridge structural anomaly detection requires a system capable of rapid response and timely detection of potential risks. Existing deep learning models incur high computational costs when processing long sequence structural data, making them unsuitable for real-time anomaly detection. Furthermore, existing anomaly detection methods have high data quality requirements and insufficient robustness against interference.

[0052] To overcome the aforementioned problems, this invention proposes a bridge structural anomaly detection method based on dilated convolution feature fusion. Periodic analysis is performed using Fast Fourier Transform (FFT) to reconstruct a two-dimensional tensor representation. A multi-scale convolutional structure is employed to simultaneously capture the short, medium, and long-term behavioral features of the bridge. Dilated convolution-based feature fusion (DCFF) technology is used to capture implicit correlations among features, and a temporal activation gating (TAG) mechanism is applied to focus on key anomaly events and suppress noise. Compared to traditional sequence models such as RNNs or LSTMs, the improved time series prediction model TimesNet replaces recursive computation with two-dimensional convolution operations, significantly reducing computational complexity and thus greatly improving the speed of anomaly detection.

[0053] like Figure 1 As shown, the method of this embodiment of the invention includes the following steps:

[0054] Step S1: Collect historical monitoring data of the bridge structure, extract short-term, medium-term and long-term time series features from the historical monitoring data, and obtain a multi-branch feature map.

[0055] Specifically, to achieve comprehensive monitoring of the bridge structure, this embodiment of the invention deploys a sensor network at key locations on the bridge to collect multi-source, heterogeneous monitoring data. These sensors encompass various types, including temperature and humidity sensors, strain sensors, and displacement sensors, and are installed on the main components of the bridge, such as main beams, piers, and supports, to ensure real-time data acquisition from various key parts of the bridge.

[0056] During data acquisition, a precise time synchronization mechanism ensures that data collected by all sensors remains consistent in time, avoiding data mismatch issues caused by time deviations. Simultaneously, advanced data acquisition equipment and transmission technologies are employed to ensure that data is not lost or interfered with during acquisition and transmission, guaranteeing data integrity and reliability.

[0057] In addition, to further improve data quality, a series of preprocessing operations were performed on the collected raw data, including noise reduction to remove potential environmental noise and other interference signals, and data imputation to ensure data continuity and integrity. These preprocessing steps provide a high-quality data foundation for subsequent data analysis and processing, thereby enabling a more accurate assessment of the bridge's structural health.

[0058] After data preprocessing, time-series features are extracted from historical monitoring data. Periodicity analysis aims to identify the main periodic patterns in time-series data, thereby revealing the underlying periodic regularities. To achieve this goal, this embodiment of the invention employs Fast Fourier Transform (FFT) to perform spectral analysis on each monitoring variable, converting the time-domain signal into a frequency-domain signal. The Fast Fourier Transform obtains the spectral representation through the following integral formula:

[0059] X(f)=∫x(t)e -j2πft dt;

[0060] In the formula, X(f) represents the spectral coefficient at frequency f, which reflects the intensity of the frequency component in the original signal; x(t) is the monitoring data at time t; and j is the imaginary unit.

[0061] The power distribution of each frequency component is evaluated by calculating the power spectral density (PSD), and the k frequency components with the highest power {f1, f2, ..., fk} are selected from the PSD. k The corresponding period is {T1,T2,...T}. k},in These major periods reflect the most significant periodic variations in the data, providing a foundation for subsequent feature extraction and modeling. The expression for calculating the power spectral density is:

[0062] PSD(f)=|X(f)| 2 .

[0063] After completing the periodic analysis, the one-dimensional time series data is converted into a two-dimensional tensor representation. For each identified major period, the original time series is reorganized according to the period length to generate a two-dimensional matrix:

[0064]

[0065] In the formula, X 2D [i] is a two-dimensional matrix; Reshape() represents the reshaping operation; x(t) is the monitoring data at time t; T i For the i-th cycle; L is the total length of the monitoring data. This indicates the number of columns after the reorganization.

[0066] Reconstructed matrix X 2D Each row of [i] corresponds to a time step within a complete period, capturing subtle changes within the period; each column corresponds to changes between different periods, reflecting the consistency or trend of change between periods. Through this reconstruction, the periodic features in the time series are represented in a structured way, providing rich structural information for subsequent multi-scale feature learning.

[0067] Bridge structure monitoring data exhibits significant multi-scale characteristics, encompassing physical phenomena at various time scales, from millisecond-level vibration responses to diurnal temperature fluctuations. To effectively capture these multi-scale temporal features, this invention employs a multi-scale convolutional structure. This structure contains multiple convolutional branches, each equipped with a convolutional kernel of different sizes, ranging from 1×1 to (2N+1)×(2N+1). The functions of each branch are as follows:

[0068] 1×1 convolutional branch: Utilizes small convolutional kernels to extract point-state features, capturing independent information at each time step, and is used to analyze instantaneous signals such as local strain of bridges.

[0069] 3×3 convolution branch: Extracts local correlations with a moderately sized receptive field and identifies relationships between neighboring time steps, which is used to analyze the instantaneous vibration characteristics of bridges.

[0070] 5×5 and larger convolutional branches: By leveraging the extended receptive field, long-range dependencies are captured, and the response patterns of bridges under long-term factors such as temperature changes are analyzed.

[0071] Each branch generates an output feature map through a convolution operation, and these feature maps will be fused in subsequent steps.

[0072] Step S2: Extract the features of each branch's feature map using dilated convolutions with different dilation rates, calculate the attention weights of each branch based on the features of each branch map, and perform weighted fusion of the features of each branch based on the attention weights to obtain the fused features.

[0073] Specifically, the complexity of bridge structure monitoring data lies in the interweaving of multi-scale features. On the one hand, features at different scales exist independently, such as instantaneous vibration, short-term strain, and long-term deformation; on the other hand, these features interact through complex physical processes, such as instantaneous vibration potentially amplifying local strain anomalies, while long-term deformation may alter vibration patterns. Traditional feature fusion methods primarily focus on explicit physical scales, relying on convolutional kernels of different sizes to extract features, but neglecting the implicit correlations between features. Therefore, this invention proposes a feature fusion method based on dilated convolution (DCFF). By adjusting the dilation rate, dilated convolution can capture a wider range of contextual information without increasing the number of parameters. By employing a two-layer design of explicit physical scales and implicit feature correlations, it adapts to the needs of bridge monitoring, improving the accuracy and robustness of anomaly detection.

[0074] The specific steps are as follows: after multi-branch feature extraction, dilated convolution is applied to further process these feature maps. The formula for calculating dilated convolution is:

[0075]

[0076] In the formula, F idenoted as , where is the feature map of the i-th branch; K is the upper limit of the number of elements in the dilated convolution kernel, serving as the boundary of the summation range and used to determine the sampling width for feature extraction; x is the input feature map; r is the dilation rate, used to control the size of the receptive field; k is the kernel size; w(k) is the kernel weight; x(i+r·k) represents the position in feature map x centered at position i, offset by dilation rate r and kernel size k, used to extract the feature value at the corresponding position. That is, i+r·k is the offset index calculated by dilation rate r and kernel index k, determining the specific position to be accessed in x, and x(i+r·k) is the value at this position.

[0077] After multi-branch feature extraction, dilated convolution is used to further explore the implicit connections between features. Even though explicit features have been extracted in the multi-scale convolution stage, their interactions still need to be fully explored. Dilated convolution can establish long-range connections, allowing features at different scales to influence each other, which is crucial for addressing the multi-frequency coupling problem caused by bridge structural damage.

[0078] For each branch, a spatial attention map is generated using the corresponding dilated convolution:

[0079]

[0080] In the formula, This is the spatial attention graph for the i-th branch; DilatedConv i This indicates a dilated convolution operation.

[0081] The spatial attention maps of multiple branches are concatenated along the branch dimensions. The weights of all branches at each spatial location are normalized along the branch dimensions to ensure that the sum of the weights at each spatial location (h, w) is 1, thus obtaining the attention weights of each branch at each spatial location:

[0082]

[0083] In the formula, ω i,h,w is the weight of the i-th branch at spatial location (h, w); softmax() is the normalized exponential function.

[0084] Finally, the fusion feature is obtained by weighted summation:

[0085]

[0086] In the formula, F fused is the fusion feature; n is the total number of branches.

[0087] This dynamic fusion mechanism enables the model to adaptively focus on key frequency components in bridge structure monitoring data. For example, it enhances the weights of low-frequency branches in temperature fluctuation scenarios and strengthens high-frequency branch features in vibration anomalies caused by traffic loads, thereby accurately identifying different types of bridge anomalies. Since the weights are learned through dilated convolution, they not only reflect the independent importance of features at each scale but also implicitly imply the interactions between features.

[0088] Step S3: Perform channel projection on the multi-branch feature map to obtain residual features. Add the residual features and the fused features to obtain the concatenated features. Input the concatenated features into a temporal convolution and generate a gate value through an activation function. Multiply the gate value element-wise with the concatenated features to obtain the output features.

[0089] Specifically, to ensure consistency between input and output channels, this embodiment of the invention also introduces channel projection of 1×1 convolution:

[0090] F residual =ChannelProjection(x);

[0091] In the formula, F residual As residual features, they are obtained by channel projection and fusion feature F. fused Align the number of channels to ensure dimensional consistency in gating calculations; ChannelProjection() represents the channel projection operation.

[0092] In bridge monitoring, identifying sparse anomalous signals against a high-noise background is a critical problem. To address this, this invention proposes a Temporal Activation Gating (TAG) mechanism. Furthermore, by adding the input signal to the fused features, low-frequency features, such as the slow degradation trend of bridge stiffness, are preserved, effectively solving the problem of traditional deep networks easily losing low-frequency information. The expression for the TAG mechanism is:

[0093] Gate = σ(TemporalConv(F) fused +F residual ));

[0094] F final =(F fused +F residual )⊙Gate;

[0095] In the above formula, F final represents the output feature; ⊙ indicates element-wise multiplication; Gate is the gate value; σ is the Sigmoid activation function; TemporalConv() represents a 3×1 convolution operation along the time axis.

[0096] This mechanism, based on temporal convolution, can accurately identify temporal patterns in data. Specifically, temporal convolution, through its local connectivity properties, is particularly suitable for capturing short-term anomalies in bridge structures, such as strain jumps caused by crack propagation. Simultaneously, the sigmoid activation function transforms the identified temporal patterns into gating values ​​between 0 and 1. In bridge structure monitoring, structural damage typically induces characteristic temporal patterns, such as abrupt changes in vibration frequency or strain peaks. These patterns generate gating values ​​close to 1, thus being effectively preserved. Conversely, random environmental noise such as wind-induced vibration, lacking consistent temporal patterns, generates gating values ​​close to 0, and is therefore suppressed.

[0097] Through the TAG mechanism, the model in this embodiment of the invention can achieve fine-grained feature selection in the time dimension, automatically highlighting key abnormal events in the bridge structure while effectively suppressing the interference of environmental noise. This enhanced temporal feature representation capability provides more accurate input for subsequent anomaly detection based on prediction differences, significantly improving the model's sensitivity to structural anomalies and enhancing its robustness, thus providing more reliable technical support for bridge health monitoring.

[0098] Step S4: Reshape the output features into a time series to obtain the prediction results of the monitoring data. Calculate the difference between the real-time collected monitoring data and the prediction results. When the difference exceeds the set difference range, determine that the bridge structure is abnormal.

[0099] Specifically, after feature extraction via the temporal activation mechanism, the processed two-dimensional features are reshaped into a time series format to restore their original temporal dependencies and obtain the prediction results of the monitoring data. In particular, the feature tensor is converted into a time series through an inverse operation to ensure the continuity of the temporal dimension. Let X be the two-dimensional feature tensor after feature extraction. 2D Its shape is [B, M, N], where B is the batch size, M is the feature dimension, and N is the number of time steps. The reshaped time series is x. ts Its shape is [B,T], and the total length of the time series is T = M × N. This reshaping process can be represented as:

[0100] x ts [b] = Reshape -1 (X 2D [b],[T]);

[0101] In the formula, b is the sample index in the batch; Reshape -1 () indicates the inverse reshaping operation, which converts a two-dimensional feature tensor back into a one-dimensional time series, thereby restoring the continuity of the time dimension.

[0102] The system calculates the difference Δ between the monitoring data and the predicted data output by the model. If Δ exceeds the range of [μΔ-3σΔ, μΔ+3σΔ], the structure is considered abnormal. Here, μΔ is the mean difference between the predicted data and the predicted data for the first month of normal system operation, and σΔ is the corresponding standard deviation. This statistical anomaly detection method effectively identifies abnormal patterns in structural behavior and promptly detects potential structural problems. When an anomaly is detected, the system issues an early warning signal, providing detailed information such as the location, time, and severity of the anomaly to support further diagnostic analysis.

[0103] like Figure 2 As shown, this embodiment of the invention also provides a bridge structure anomaly detection system based on dilated convolution feature fusion, which is based on the above-mentioned bridge structure anomaly detection method based on dilated convolution feature fusion, and includes: a data acquisition module, a temporal feature extraction module, a multi-scale feature learning module, an anomaly diagnosis module, and a visualization early warning module.

[0104] Data acquisition module: Collects multi-source heterogeneous monitoring data through a sensor network deployed at key locations on the bridge. At the same time, it performs preprocessing such as noise reduction and data filling on the raw data to ensure data quality.

[0105] The time-series feature extraction module is based on an improved TimesNet model. It models and extracts features from the monitoring data, converts the monitoring data into a frequency domain signal, extracts the periodic frequency components, selects the k frequency components with the highest power, obtains the period length, and reassembles the monitoring data into a two-dimensional tensor based on the period length.

[0106] like Figure 3 The diagram shows the hierarchical structure of the multi-scale feature learning module in an embodiment of the present invention. This module uses different convolution kernels to extract short-term, medium-term, and long-term features from a two-dimensional tensor. It expands the receptive field by adjusting the dilation rate, captures the relationship between multi-scale features, dynamically generates weights to achieve adaptive feature fusion, uses temporal convolution and activation functions to generate gating values, strengthens key anomalous temporal features, and outputs predicted values.

[0107] Anomaly Diagnosis Module: Calculates the difference between the predicted value and the actual monitoring data. When the difference exceeds the set difference range, it determines that the bridge structure is abnormal. It determines the area where the anomaly occurs by combining the sensor location, traces back the starting time of the anomaly through time-series characteristics, and classifies the anomaly level according to the magnitude of the difference.

[0108] Visualized early warning module: dynamically displays raw data from each sensor and anomaly markers, visualizes the spatiotemporal distribution of abnormal events, generates structural health scores and trend analysis charts, and pushes abnormal information via SMS, email, or API interface.

[0109] The experiments in this embodiment of the invention were conducted on a high-performance computing platform equipped with an NVIDIA RTX 3090 graphics card and 72GB of RAM, and implemented using the PyTorch 2.4.1 deep learning framework. The experiments used a real-world bridge health monitoring sensor dataset as the evaluation object, and appropriately set the network training parameters to verify the effectiveness of the bridge anomaly data detection method of this embodiment.

[0110] To comprehensively evaluate the model performance, the model proposed in this embodiment of the invention is compared with various time series prediction models, including Long Short-Term Memory (LSTM) network, Informer long-sequence prediction model, Autoformer long-sequence prediction model based on deep decomposition architecture and autocorrelation mechanism, and PatchTST long-sequence prediction model based on Transformer. Mean squared error (MSE) and mean absolute error (MAE) are used as evaluation metrics, and tests are conducted at prediction step sizes of 72, 144, and 288. Table 1 shows the prediction results of each model on the bridge dataset.

[0111] Table 1 Comparison of prediction results of each model

[0112]

[0113]

[0114] As shown in Table 1, the improved TimesNet model proposed in this embodiment achieves the best MSE and MAE indices for all prediction step sizes. When the prediction step size is 72, the MSE of the model in this embodiment is 0.124, which is 34.0% lower than the second-best performing PatchTST model; when the prediction step size is 144, the MSE of the model in this embodiment is 0.144, which is 33.9% lower than the PatchTST model; and when the prediction step size is 288, the MSE of the model in this embodiment is 0.178, which is 24.9% lower than the Autoformer model. Comparative analysis shows that the traditional LSTM model performs poorly in handling long sequence prediction tasks, especially at longer prediction step sizes, where its MSE and MAE indices are significantly higher than other models. Although models based on the Transformer structure, represented by Informer, show significant improvements compared to LSTM, they still have limitations in capturing multi-scale temporal features. While PatchTST improves prediction performance by introducing a time-series patch embedding mechanism, it still cannot effectively handle the complex temporal patterns and periodic features in bridge monitoring data. The improved TimesNet-based model in this invention, through multi-scale feature extraction after temporal two-dimensional reconstruction, effectively captures short-term fluctuations, medium-term trends, and long-term periodic changes in bridge vibration signals, thereby improving the ability to identify abnormal bridge structural states.

[0115] To further verify the effectiveness of the dilated convolutional feature fusion (DCFF) and temporal activation gating (TAG) modules proposed in this embodiment of the invention, ablation experiments were also designed. The experiments compared the performance of the baseline TimesNet, the scheme with only the DCFF module, the scheme with only the TAG module, and the complete scheme using a combination of DCFF and TAG at different prediction strides. The results are shown in Table 2.

[0116] Table 2 Ablation Experiment Results of the Improved Module

[0117]

[0118]

[0119] Ablation experiments show that the DCFF and TAG modules significantly improve model performance. In short-term predictions, both the DCFF and TAG modules effectively reduce prediction errors when used alone, with the DCFF module slightly outperforming the TAG module. In medium-term predictions, the performance improvement from using either the DCFF or TAG module alone is limited, but their combined use shows a significant synergistic effect, reducing the MSE from 0.156 to 0.144, a decrease of 7.7%. In long-term predictions, the DCFF module shows a strong performance improvement, while the combination of DCFF and TAG is the most effective, reducing the error by 4.8% compared to the baseline. This series of experiments demonstrates the complementarity of the DCFF and TAG modules at different time scales. DCFF effectively enhances the model's ability to capture implicit correlations between features through a dynamic feature fusion mechanism; TAG enhances the model's ability to capture key local features through temporal attention gating.

[0120] It is important to note that while this invention aims at detecting bridge anomaly data, performance evaluation through time-series prediction is a crucial step and technological foundation for achieving anomaly detection. In the system proposed in this invention, the core of anomaly detection lies in identifying the deviation between actual observations and the expected normal state. An accurate time-series prediction model can establish the behavioral patterns of the bridge structure under normal conditions. When a significant difference appears between the actual observed data and the model's predicted values, it can be identified as a potential anomaly. Therefore, the higher the prediction accuracy of the model, the more accurately it can capture subtle anomaly signals, improving the sensitivity and accuracy of anomaly detection. The excellent performance of the model provided in this invention on prediction tasks directly demonstrates its superior performance in bridge anomaly data detection.

[0121] This invention improves upon the simple averaging fusion method of the original TimesNet by proposing a feature fusion strategy based on dilated convolution. This strategy utilizes dilated convolution to capture the implicit correlations between features at different scales and generates dynamic weights, thus achieving a two-layer design that combines explicit physical scale with implicit feature correlations. This design can adaptively focus on the most diagnostically valuable frequency components in the bridge structure, enabling accurate identification of different types of bridge anomalies.

[0122] An innovative temporal activation gating structure is introduced, which selectively strengthens key anomalous features by learning the importance distribution of temporal features, while effectively suppressing environmental noise. Complementing the temporal activation gating structure, it can highlight potential anomalous patterns and mitigate the interference of environmental factors such as temperature changes and traffic flow fluctuations, thereby significantly improving the robustness of the model in complex environments.

[0123] Based on an improved TimesNet model, an optimized framework specifically designed for anomaly detection in bridge structures was constructed. By specifically adjusting network parameters and optimizing the feature extraction process, this framework makes the system more sensitive to anomaly patterns in bridge structures, such as sudden damage and fatigue crack development, thereby improving the accuracy and timeliness of anomaly detection.

[0124] The technical features of the above embodiments can be combined arbitrarily. For the sake of brevity, not all possible combinations of the technical features in the above embodiments are described; only preferred embodiments of the present invention are illustrated. The descriptions are relatively specific and detailed, but they should not be construed as limiting the scope of the present invention. As long as the combination of these technical features does not contradict each other, it should be considered within the scope of this specification.

[0125] It should be noted that those skilled in the art can make various modifications and improvements without departing from the inventive concept, and these all fall within the scope of protection of this invention. Therefore, the scope of protection of this invention should be determined by the appended claims.

Claims

1. A method for detecting bridge structural anomalies based on dilated convolution feature fusion, characterized in that, Includes the following steps: Step S1: Collect historical monitoring data of the bridge structure, extract short-term, medium-term and long-term time series features from the historical monitoring data, and obtain a multi-branch feature map; Step S2: Extract the features of the feature map of each branch using dilated convolution with different dilation rates, calculate the attention weight of each branch based on the features of each branch map, and perform weighted fusion of the features of each branch based on the attention weight to obtain the fused features; The expression for extracting the features of each branch's feature map using dilated convolutions with different dilation rates is as follows: ； In the formula, The features of the feature map of the i-th branch; This represents the upper limit of the number of elements in the dilated convolution kernel; For feature maps; Expansion rate; The kernel size; These are the kernel weights; In the feature map In the middle, based on position Centered on, expansion rate and convolution kernel size is The offset position; The step of calculating the attention weight of each branch based on the characteristics of each branch graph includes the following steps: Step S21: Generate a spatial attention map for each branch using dilated convolution: ； In the formula, This is the spatial attention graph for the i-th branch; This represents the dilated convolution operation; The features of the feature map of the i-th branch; Step S22: Concatenate the spatial attention maps of multiple branches along the branch dimension, and normalize the weights of all branches at each spatial location along the branch dimension to obtain the attention weights of each branch at each spatial location. Step S3: Perform channel projection on the multi-branch feature map to obtain residual features, add the residual features and the fused features to obtain the concatenated features, input the concatenated features into temporal convolution, generate a gate value through an activation function, and multiply the gate value with the concatenated features element by element to obtain the output features; Step S4: Reshape the output features into a time series to obtain the prediction results of the monitoring data, calculate the difference between the real-time monitoring data and the prediction results, and determine the bridge structure as abnormal when the difference exceeds the set difference range.

2. The bridge structural anomaly detection method based on dilated convolution feature fusion according to claim 1, characterized in that: Step S1 involves extracting short-term, medium-term, and long-term time-series features from the historical monitoring data, including the following steps: Step S11: Perform a fast Fourier transform on the historical monitoring data to extract the period of the historical monitoring data, and reorganize the one-dimensional monitoring data into a two-dimensional matrix according to the period length; Step S12: Extract short-term, medium-term, and long-term temporal features of the two-dimensional matrix using convolution kernels of different scales.

3. The bridge structural anomaly detection method based on dilated convolution feature fusion according to claim 2, characterized in that: The expression for reorganizing the one-dimensional monitoring data into a two-dimensional matrix according to the period length in step S11 is as follows: ； In the formula, It is a two-dimensional matrix; Indicates a reshaping operation; The monitoring data is at time t; For the i-th period; The total length of the monitoring data.

4. The bridge structural anomaly detection method based on dilated convolution feature fusion according to claim 1, characterized in that: The expression for normalizing the weights of all branches at each spatial location along the branch dimension, as described in step S22, is: ； In the formula, For the i-th branch in spatial position The weights; It is a normalized exponential function; Let be the spatial attention graph of the i-th branch.

5. The bridge structural anomaly detection method based on dilated convolution feature fusion according to claim 4, characterized in that: The expression for the fusion feature mentioned in step S2 is: ； In the formula, Features of fusion; For the i-th branch in spatial position The weights; Total number of branches; The features of the feature map of the i-th branch.

6. The bridge structural anomaly detection method based on dilated convolution feature fusion according to claim 1, characterized in that: The expression for the output feature in step S3 is: ； ； ； In the above formula, For output features; Features of fusion; Residual characteristics; This indicates element-wise multiplication; This is the gate value; Indicates channel projection operation; For feature maps; Use the Sigmoid activation function; This indicates a convolution operation along the time axis.

7. A bridge structural anomaly detection system based on dilated convolution feature fusion, implemented based on the bridge structural anomaly detection method based on dilated convolution feature fusion as described in any one of claims 1 to 6, characterized in that, include: The module includes a data acquisition module, a time-series feature extraction module, a multi-scale feature learning module, an anomaly diagnosis module, and a visualization and early warning module. The data acquisition module collects monitoring data of the bridge structure. The time-series feature extraction module converts historical monitoring data of the bridge structure into frequency domain signals, extracts periodic frequency components, filters the top k frequency components with the highest power to obtain the period length, and reorganizes the monitoring data into a two-dimensional tensor based on the period length. The multi-scale feature learning module extracts short-term, medium-term, and long-term features from a two-dimensional tensor using different convolutional kernels, expands the receptive field by adjusting the dilation rate, captures the relationships between multi-scale features, dynamically generates weights to achieve adaptive feature fusion, uses temporal convolution and activation functions to generate gating values, strengthens key anomalous temporal features, and outputs predicted values. The anomaly diagnosis module calculates the difference between the predicted value and the real-time collected monitoring data. When the difference exceeds the set difference range, it determines that the bridge structure is abnormal, determines the area where the anomaly occurs by combining the sensor location, traces back the starting time of the anomaly through time-series characteristics, and classifies the anomaly level according to the difference magnitude. The visualization and early warning module dynamically displays the raw data of each sensor and anomaly point markers, visualizes the spatiotemporal distribution of abnormal events, generates structural health scores and trend analysis charts, and pushes abnormal information via SMS, email, or API interface.

Images