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

Abstract

The invention discloses a bridge structure anomaly detection method and system based on cavity convolution feature fusion, and the method comprises the steps: collecting the historical monitoring data of a bridge structure, extracting the short-term, medium-term and long-term time sequence features, and obtaining a multi-branch feature map; extracting the features of the feature map of each branch by adopting dilated convolution with different expansion rates, calculating the attention weight of each branch according to the features of each branch map, and performing weighted fusion on the features of each branch to obtain fused features; performing channel projection on the multi-branch feature map to obtain a residual feature, adding the residual feature and the fusion feature to obtain a splicing feature, inputting the splicing feature into time sequence convolution, generating a gating value through an activation function, and performing element-by-element multiplication on the gating value and the splicing feature to obtain an output feature; and remodeling the output features into a monitoring data prediction result of a time sequence, calculating a difference value between the real-time monitoring data and the prediction result, and when the difference value exceeds a set difference value range, judging that the bridge structure is abnormal.

Description

Technical Field

[0001] The present invention relates to the technical field of bridge structure monitoring, and in particular to a bridge structure anomaly detection method and system based on void convolution feature fusion. Background Art

[0002] As critical transportation infrastructure, the structural health of bridges is directly linked to public safety and transportation efficiency. Traditional bridge inspection methods rely primarily on regular manual inspections, a time-consuming and labor-intensive approach that is susceptible to human influence and struggles to accurately and promptly identify potential structural anomalies.

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

[0004] However, traditional models still have limitations when processing multi-source, heterogeneous data on bridge structures. Existing technologies often use fixed convolution kernels to extract features at a single time scale when extracting features from bridge monitoring data. This often misses key information about the bridge structure across different time dimensions and fails to effectively balance local details with global correlations. This results in incomplete features and makes it difficult to accurately reflect the bridge's true condition. Summary of the Invention

[0005] The present invention proposes a bridge structure anomaly detection method and system based on void 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 solve the above technical problems, the present invention provides a bridge structure anomaly detection method based on dilated convolution feature fusion, comprising the following steps:

[0007] Step S1: collecting historical monitoring data of the bridge structure, extracting short-term, medium-term, and long-term time series features from the historical monitoring data, and obtaining a multi-branch feature graph;

[0008] Step S2: Using dilated convolution with different expansion rates to extract the features of the feature map of each branch, calculating the attention weight of each branch according to the features of each branch map, and performing weighted fusion on the features of each branch according to the attention weight to obtain the fused feature;

[0009] Step S3: performing channel projection on the multi-branch feature map to obtain residual features, adding the residual features and the fusion features to obtain splicing features, inputting the splicing features into the temporal convolution, generating a gating value through an activation function, and performing element-wise multiplication of the gating value with the splicing features to obtain output features;

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

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

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

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

[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:

[0015]

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

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

[0018]

[0019] Where, F i is the feature of the feature map of the i-th branch; K is the upper limit of the number of elements of the hollow convolution kernel; x is the feature map; r is the dilation rate; k is the convolution kernel size; w(k) is the convolution kernel weight; x(i+r·k) represents the position in the feature map x with position i as the center, dilation rate r and convolution kernel size k as the offset.

[0020] Preferably, the step S2 of calculating the attention weight of each branch according to the features of each branch graph includes the following steps:

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

[0022] Step S22: Splice the spatial attention maps of multiple branches along the branch dimension, normalize all branch weights at each spatial position in the branch dimension, and obtain the attention weight of each branch at each spatial position.

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

[0024]

[0025] Where, is the spatial attention map of the i-th branch; DilatedConv i represents the dilated convolution operation; F i is the feature of the feature map of the i-th branch.

[0026] Preferably, the expression for normalizing all branch weights at each spatial position in the branch dimension in step S22 is:

[0027]

[0028] Where, ω i,h,w is the weight of the i-th branch at the spatial position (h, w); softmax() is the normalized exponential function; is the spatial attention map of the i-th branch.

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

[0030]

[0031] Where, F fused is the fusion feature; ω i,h,w is the weight of the i-th branch at the spatial position (h, w); n is the total number of branches; F i is the feature of the feature map of the i-th branch.

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

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

[0034] F residual =ChannelProjection(x);

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

[0036] In the above formula, F final is the output feature; F fused is the fusion feature; F residual is the residual feature; represents element-by-element multiplication; Gate is the gating 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] The embodiment of the present invention also provides a bridge structure anomaly detection system based on void convolution feature fusion, which is implemented based on the above-mentioned bridge structure anomaly detection method based on void convolution feature fusion, and includes: a data acquisition module, a time series feature extraction module, a multi-scale feature learning module, an anomaly diagnosis module and a visual warning module;

[0038] The data acquisition module is used to collect monitoring data of the bridge structure;

[0039] The time series feature extraction module converts the historical monitoring data of the bridge structure into a frequency domain signal, extracts the periodic frequency components, selects the first k frequency components with the largest power, obtains the period length, and reorganizes the monitoring data into a two-dimensional tensor according to the period length;

[0040] The multi-scale feature learning module uses different convolution kernels to extract short-term, medium-term, and long-term features from a two-dimensional tensor. It adjusts the dilation rate to expand the receptive field, captures the relationship between multi-scale features, dynamically generates weights to achieve adaptive feature fusion, and uses time-domain convolution and activation functions to generate gating values, strengthen key abnormal time series features, and output prediction values.

[0041] The abnormality diagnosis module calculates the difference between the predicted value and the real monitoring data collected in real time. When the difference exceeds the set difference range, it determines that the bridge structure is abnormal. The abnormality occurrence area is determined based on the sensor location. The abnormality starting point is traced back through the time series characteristics, and the abnormality level is classified according to the difference amplitude.

[0042] The visual warning module dynamically displays the raw data and abnormal point marks of each sensor, 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 benefits of the present invention include at least:

[0044] 1. By extracting short-term, medium-term, and long-term temporal features, the dynamic changes of bridge structures can be fully captured from different time scales, avoiding feature omissions caused by single-time scale analysis. The use of dilated convolution with different dilation rates effectively expands the receptive field of the convolution kernel, while considering local details and global correlations, improving the comprehensiveness of feature extraction.

[0045] 2. Dynamically assigning weights based on the importance of each branch feature can highlight key features, suppress redundant information, make the fused features more representative, and improve the model's ability to identify abnormal features. The gating mechanism enables the model to adaptively control features, dynamically determining the degree of feature retention or suppression based on data characteristics, and improving adaptability to complex working conditions and noise. BRIEF DESCRIPTION OF THE DRAWINGS

[0046] Figure 1 Schematic diagram of a method flow in an embodiment of the present invention;

[0047] Figure 2 A schematic diagram of the system hierarchy of an embodiment of the present invention;

[0048] Figure 3 Schematic diagram of the structure of a multi-scale feature learning module according to an embodiment of the present invention. DETAILED DESCRIPTION

[0049] The following is a clear and complete description of the technical solutions in the embodiments of the present invention, in conjunction with the accompanying drawings. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. All other embodiments obtained by ordinary technicians in this field based on the embodiments of the present invention without making any creative efforts are within the scope of protection of the present invention.

[0050] Although existing bridge structure anomaly detection technologies have been applied in practice, they still face many challenges in achieving accurate and timely anomaly detection. Bridge structure data is multi-source and heterogeneous, covering various types of data such as vibration data, strain data, and displacement data. These data have complex spatiotemporal correlations, making it difficult for traditional anomaly detection methods to effectively integrate and analyze them. Bridge structure data is easily affected by factors such as environmental noise and sensor failure, which reduces the reliability of anomaly detection. In addition, structural data often exhibits strong temporal characteristics and periodic changes. These characteristics are closely related to environmental factors such as temperature and traffic volume, further increasing the difficulty of anomaly pattern recognition. Traditional anomaly detection methods rely too much on threshold setting 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, and are prone to false positives or false negatives in anomaly detection.

[0051] At the same time, bridge structural anomaly detection requires a system that can respond quickly and identify potential risks promptly. Existing deep learning models have high computational overhead when processing long sequences of structural data, making them difficult to meet the requirements of real-time anomaly detection. Furthermore, existing anomaly detection methods have high data quality requirements and lack robustness against interference.

[0052] To overcome the above problems, an embodiment of the present invention proposes a bridge structure anomaly detection method based on dilated convolution feature fusion. Fast Fourier transform (FFT) is used to perform periodic analysis and reconstruct a two-dimensional tensor representation. A multi-scale convolution structure is used to simultaneously capture the short-, medium-, and long-term behavioral characteristics of the bridge. The dilated convolution-based feature fusion (DCFF) technology is used to capture the implicit correlation of features, and the time domain activation gating (TAG) mechanism is applied to focus on key abnormal events and suppress noise. Compared with traditional sequence models such as RNN or LSTM, the improved time series prediction model TimesNet replaces recursive calculations with two-dimensional convolution operations, significantly reducing the computational complexity, thereby greatly improving the speed of anomaly detection.

[0053] like Figure 1 As shown, the method of the embodiment of the present 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 graph.

[0055] Specifically, to achieve comprehensive monitoring of bridge structures, the present invention deploys a sensor network at key locations across the bridge to collect heterogeneous monitoring data from multiple sources. These sensors, including temperature and humidity sensors, strain sensors, and displacement sensors, are installed on key bridge components such as girders, piers, and supports, ensuring real-time data capture from every critical bridge location.

[0056] During the data collection process, a precise time synchronization mechanism ensures that the data collected by all sensors remains consistent in time, avoiding data mismatches caused by time deviations. At the same time, the use of advanced data collection equipment and transmission technology ensures that data is not lost or interfered with during the collection and transmission process, ensuring data integrity and reliability.

[0057] To further improve data quality, the collected raw data undergoes a series of preprocessing operations, including noise reduction to remove potential environmental noise and other interfering signals, and filling in missing data to ensure data continuity and integrity. These preprocessing steps provide a high-quality data foundation for subsequent data analysis and processing, enabling a more accurate assessment of the bridge's structural health.

[0058] After data preprocessing, the time series features in the historical monitoring data are extracted. Periodicity analysis aims to identify the main periodic patterns in time series data, thereby revealing the periodic laws behind the data. To achieve this goal, the embodiment of the present invention uses a 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] Where 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 an imaginary unit.

[0061] The power distribution of each frequency component is evaluated by calculating the power spectral density (PSD), and the first k frequency components with the largest power {f1, f2, ..., f k}, the corresponding period is {T1,T2,...T k},in These main cycles reflect the most significant periodic changes in the data and provide a basis for subsequent feature extraction and modeling. The expression for calculating power spectral density is:

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

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

[0064]

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

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

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

[0068] 1×1 convolution branch: Uses a small convolution kernel to extract point features and capture independent information at each time step, which is used to analyze transient signals such as local strain of bridges.

[0069] 3×3 convolutional branch: With a medium-sized receptive field, it extracts local correlations and identifies the relationship between adjacent time steps for analyzing the transient vibration characteristics of the bridge.

[0070] 5×5 and larger convolution branches: Capture long-range dependencies with the help of expanded receptive fields and analyze the response patterns of bridges to long-term factors such as temperature changes.

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

[0072] Step S2: Use dilated convolution with different expansion rates to extract the features of the feature map of each branch, calculate the attention weight of each branch according to the features of each branch map, and perform weighted fusion of the features of each branch according to the attention weight to obtain the fused features.

[0073] Specifically, the complexity of bridge structure monitoring data is reflected in the interweaving of multi-scale features. On the one hand, features of different scales exist independently, such as instantaneous vibration, short-term strain and long-term deformation; on the other hand, these features influence each other through complex physical processes. For example, instantaneous vibration may amplify local strain anomalies, while long-term deformation may change the vibration mode. Traditional feature fusion methods mainly focus on explicit physical scales and rely on convolution kernels of different sizes to extract features, but ignore the implicit associations between features. Therefore, an embodiment of the present invention proposes a feature fusion (DCFF) method based on dilated convolution. By adjusting the expansion rate, dilated convolution can capture a wider range of contextual information without increasing the number of parameters. By adopting a two-layer design of explicit physical scale and implicit feature association, it adapts to the needs of bridge monitoring and improves 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 calculation formula of dilated convolution is:

[0075]

[0076] Where, F iis the feature of the feature map of the i-th branch; K is the upper limit of the number of elements of the hollow convolution kernel, which serves as the boundary of the summation range and is used to determine the sampling width of feature extraction; x is the input feature map; r is the dilation rate, which is used to control the size of the receptive field; k is the convolution kernel size; w(k) is the convolution kernel weight; x(i+r·k) represents the position in the feature map x with position i as the center, dilation rate r and convolution kernel size k offset, which is used to extract the feature value of the corresponding position, that is, i+r·k is the offset index calculated by dilation rate r and convolution kernel index k, which determines 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 implicit connections between features. Even though the multi-scale convolution stage extracts explicit features, their interactions remain 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 caused by bridge structural damage.

[0078] For each branch, the corresponding dilated convolution is used to generate a spatial attention map:

[0079]

[0080] Where, is the spatial attention map of the i-th branch; DilatedConv i Represents a dilated convolution operation.

[0081] The spatial attention maps of multiple branches are spliced ​​along the branch dimension, and all branch weights of each spatial position are normalized in the branch dimension to ensure that the sum of the weights at each spatial position (h, w) is 1. The attention weight of each branch at each spatial position is obtained:

[0082]

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

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

[0085]

[0086] Where, 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 can enhance the weights of low-frequency branches in temperature fluctuation scenarios and strengthen high-frequency branch features in vibration anomalies caused by traffic loads, thereby accurately identifying different types of bridge anomalies. Because the weights are learned through dilated convolution, they not only reflect the independent importance of features at each scale, but also implicitly reflect the interactions between features.

[0088] Step S3: Perform channel projection on the multi-branch feature map to obtain the residual feature, add the residual feature and the fusion feature to obtain the splicing feature, input the splicing feature into the temporal convolution, generate the gating value through the activation function, multiply the gating value with the splicing feature element by element to obtain the output feature.

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

[0090] F residual =ChannelProjection(x);

[0091] Where, F residual It is the residual feature, which is projected through the channel and fused with the feature F fused Align the number of channels to ensure dimensional consistency of gating calculations; ChannelProjection() represents the channel projection operation.

[0092] In bridge monitoring, identifying sparse abnormal signals in a high-noise background is an urgent problem. To this end, the present 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 retained, effectively solving the problem that traditional deep networks are prone to losing low-frequency information. The expression of the temporal activation gating 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 is the output feature; ⊙ represents element-by-element 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 is based on time-domain convolution operations and can accurately identify temporal patterns in data. Specifically, time-domain convolution, due to its local connection characteristics, is particularly suitable for capturing short-term abnormal phenomena in bridge structures, such as strain jumps generated at the moment of crack expansion. At the same time, the Sigmoid activation function converts the identified temporal patterns into gating values ​​between 0 and 1. In bridge structure monitoring, structural damage usually causes characteristic temporal patterns, such as sudden changes in vibration frequency or strain peaks. These patterns generate gating values ​​close to 1 and are thus effectively retained. In contrast, random environmental noise such as wind vibration will produce gating values ​​close to 0 due to the lack of consistent temporal patterns and is suppressed.

[0097] Through the TAG mechanism, the model in this embodiment of the present invention enables fine-grained feature selection in the temporal dimension, automatically highlighting key abnormal events in the bridge structure while effectively suppressing interference from ambient noise. This enhanced temporal feature representation capability provides more accurate input for subsequent anomaly detection based on prediction differences. This not only significantly improves the model's sensitivity to structural anomalies but also enhances its robustness, 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 monitoring data and the prediction results, and when the difference exceeds the set difference range, judge that the bridge structure is abnormal.

[0099] Specifically, after the feature extraction of the time domain activation mechanism is completed, the processed two-dimensional features are reshaped into a time series form to restore their original time dependency and obtain the prediction results of the monitoring data. Specifically, the feature tensor is converted into a time series through the inverse operation to ensure the continuity of the time dimension. Let the two-dimensional feature tensor after feature extraction be X 2D , whose 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. The reshaping process can be expressed as:

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

[0101] Where b is the sample index in the batch; Reshape -1 () represents the inverse reshape operation, which transforms the two-dimensional feature tensor back into a one-dimensional time series, thus restoring the time dimension continuity.

[0102] The predicted difference Δ between the monitored data and the model's output is calculated. If Δ exceeds the range [μΔ - 3σΔ, μΔ + 3σΔ], the structural state is considered abnormal. μΔ is the mean predicted difference in the dataset over the initial one-month period of normal system operation, and σΔ is the corresponding standard deviation. This statistically based anomaly detection method effectively identifies abnormal patterns in structural behavior and promptly identifies potential structural issues. When an anomaly is detected, the system issues an early warning signal and provides detailed information such as the anomaly's location, time, and severity, supporting further diagnostic analysis.

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

[0104] Data acquisition module: Through the sensor network deployed at key locations of the bridge, multi-source heterogeneous monitoring data is collected, and the original data is pre-processed by noise reduction, filling, etc. to ensure data quality.

[0105] Time series feature extraction module: Based on the improved TimesNet model, the monitoring data is modeled and features are extracted. The monitoring data is converted into frequency domain signals, the periodic frequency components are extracted, and the top k frequency components with the largest power are screened to obtain the period length. The monitoring data is reorganized into a two-dimensional tensor based on the period length.

[0106] like Figure 3 The figure shows a hierarchical structure diagram of the multi-scale feature learning module of an embodiment of the present invention. The module uses different convolution kernels to extract short-term, medium-term, and long-term features in a two-dimensional tensor, 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 time domain convolution and activation functions to generate gating values, strengthens key abnormal time series features, and outputs predicted values.

[0107] Abnormal diagnosis module: Calculates the difference between the predicted value and the actual monitoring data. When the difference exceeds the set difference range, the bridge structure is judged to be abnormal. The abnormality area is determined based on the sensor location. The abnormality starting time point is traced back through the time series characteristics, and the abnormality level is divided according to the difference amplitude.

[0108] Visual early warning module: Dynamically displays the raw data of each sensor and abnormal point marks, 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 present invention were conducted on a high-performance computing platform equipped with an NVIDIA RTX 3090 graphics card and 72GB of memory, using the PyTorch 2.4.1 deep learning framework. The experiments used a real-world bridge health monitoring sensor dataset as the evaluation target and rationally set network training parameters to verify the effectiveness of the bridge anomaly data detection method in this embodiment of the present invention.

[0110] To comprehensively evaluate the model's performance, the model proposed in this embodiment of the present invention was compared with various time series prediction models, including a long short-term memory (LSTM) network, an Informer long series prediction model, an Autoformer long series prediction model based on a deep decomposition architecture and autocorrelation mechanism, and a Transformer-based long series prediction model, PatchTST. Mean squared error (MSE) and mean absolute error (MAE) were used as evaluation metrics, and tests were performed at prediction step sizes of 72, 144, and 288, respectively. 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 the present embodiment achieves the best MSE and MAE metrics at each prediction step size. When the prediction step size is 72, the MSE of the model in the present embodiment is 0.124, a 34.0% reduction compared to the second-best performing PatchTST model; when the prediction step size is 144, the MSE of the model in the present embodiment is 0.144, a 33.9% reduction compared to the PatchTST model; and when the prediction step size is 288, the MSE of the model in the present embodiment is 0.178, a 24.9% reduction compared to the Autoformer model. Comparative analysis shows that traditional LSTM models perform poorly when handling long sequence prediction tasks, especially at longer prediction step sizes, where their MSE and MAE metrics are significantly higher than other models. While Transformer-based models, such as the Informer, offer significant improvements over LSTM, they still have limitations in capturing multi-scale temporal features. While PatchTST improves prediction performance by introducing a patch embedding mechanism for time series, it still cannot effectively handle the complex temporal patterns and periodic features found in bridge monitoring data. The improved TimesNet-based model in this embodiment of the present invention effectively captures short-term fluctuations, medium-term trends, and long-term periodic variations in bridge vibration signals through multi-scale feature extraction after two-dimensional time series reconstruction, thereby improving the ability to identify abnormal bridge structural conditions.

[0115] To further validate the effectiveness of the Dilated Convolutional Feature Fusion (DCFF) and Temporal Activation Gating (TAG) modules proposed in this embodiment of the present invention, we conducted an ablation experiment. The experiment compared the performance of the baseline TimesNet, the DCFF module alone, the TAG module alone, and the complete solution combining DCFF and TAG at different prediction step sizes. The results are shown in Table 2.

[0116] Table 2 Ablation experiment results of improved modules

[0117]

[0118]

[0119] Ablation experiment results show that the DCFF and TAG modules have a significant effect on improving model performance. In short-term predictions, the use of the DCFF module and the TAG module alone can effectively reduce the prediction error, with the DCFF module slightly outperforming the TAG module. In medium-term predictions, the use of the DCFF or TAG module alone has limited performance improvement, but when used in combination, a significant synergistic effect is shown, with the MSE dropping from the baseline of 0.156 to 0.144, a decrease of 7.7%. In long-term predictions, the DCFF module shows a strong performance improvement, and the combination of DCFF and TAG is the best, with a 4.8% reduction compared to the baseline. This series of experiments demonstrates the complementarity of the DCFF and TAG modules at different time scales. DCFF effectively improves the model's ability to capture implicit correlations between features through a dynamic feature fusion mechanism; TAG enhances the model's ability to capture local key features through temporal attention gating.

[0120] It should be noted that although the embodiment of the present invention aims at detecting abnormal data of bridges, the key link and technical basis for achieving abnormality detection is through time series prediction performance evaluation. In the system proposed in the embodiment of the present invention, the core of abnormality detection lies in identifying the deviation between the actual observation value and the expected normal state. An accurate time series prediction model can establish the behavior pattern of the bridge structure under normal conditions. When there is a significant difference between the actual observation data and the model prediction value, it can be determined as a potential abnormal state. Therefore, the higher the prediction accuracy of the model, the more accurately it can capture tiny abnormal signals and improve the sensitivity and accuracy of abnormality detection. The outstanding performance of the model provided by the embodiment of the present invention in the prediction task directly proves its superior performance in detecting abnormal data of bridges.

[0121] This embodiment of the present invention improves on TimesNet's existing simple average fusion approach and proposes a feature fusion strategy based on dilated convolution. This strategy uses dilated convolution to capture implicit associations between features at different scales and generates dynamic weights, thereby implementing a two-layer design that associates explicit physical scales with implicit features. This design adaptively focuses on the most diagnostically valuable frequency components in bridge structures, enabling accurate identification of different types of bridge anomalies.

[0122] The innovative introduction of a temporal activation gating structure selectively enhances key anomaly features by learning the importance distribution of temporal features while effectively suppressing environmental noise. Complementing the temporal activation gating structure, it can highlight potential anomaly patterns and mitigate interference from environmental factors such as temperature changes and traffic volume fluctuations, significantly improving the model's robustness in complex environments.

[0123] Based on the improved TimesNet model, an optimized framework specifically for bridge structural anomaly detection was constructed. By specifically adjusting network parameters and optimizing the feature extraction process, the framework makes the system more sensitive to abnormal 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 may be combined in any manner. To simplify the description, not all possible combinations of the technical features in the above embodiments are described. Only preferred embodiments of the present invention are presented. While the description is relatively specific and detailed, it should not be construed as limiting the scope of the present invention. As long as there are no conflicts in the combination of these technical features, they should be considered to be within the scope of this specification.

[0125] It should be noted that, for those skilled in the art, various modifications and improvements can be made without departing from the scope of the present invention, and these modifications and improvements fall within the scope of protection of the present invention. Therefore, the scope of protection of the present invention shall be determined by the appended claims.

Claims

1. A bridge structure anomaly detection method based on dilated convolution feature fusion, characterized in that: The following steps are involved: Step S1: collecting historical monitoring data of the bridge structure, extracting short-term, medium-term, and long-term time series features from the historical monitoring data, and obtaining a multi-branch feature graph; Step S2: Using dilated convolution with different expansion rates to extract the features of the feature map of each branch, calculating the attention weight of each branch according to the features of each branch map, and performing weighted fusion on the features of each branch according to the attention weight to obtain the fused feature; Step S3: performing channel projection on the multi-branch feature map to obtain residual features, adding the residual features and the fusion features to obtain splicing features, inputting the splicing features into the temporal convolution, generating a gating value through an activation function, and performing element-wise multiplication of the gating value with the splicing features to obtain output features; Step S4: reshape the output features into a time series, obtain the prediction results of the monitoring data, calculate the difference between the real-time monitoring data and the prediction results, and when the difference exceeds the set difference range, determine that the bridge structure is abnormal.

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

3. The bridge structure anomaly detection method based on dilated convolution feature fusion according to claim 2 is characterized by: The expression for reorganizing the one-dimensional monitoring data into a two-dimensional matrix according to the cycle length in step S11 is: Where, X 2D [i] is a two-dimensional matrix; Reshape() represents the reshaping operation; x(t) is the monitoring data at time t; T i is the i-th cycle; L is the total length of the monitoring data.

4. The bridge structure anomaly detection method based on dilated convolution feature fusion according to claim 1 is characterized by: The expression for extracting the features of the feature map of each branch using dilated convolution with different dilation rates in step S2 is: Where, F i is the feature of the feature map of the i-th branch; K is the upper limit of the number of elements of the hollow convolution kernel; x is the feature map; r is the dilation rate; k is the convolution kernel size; w(k) is the convolution kernel weight; x(i+r·k) represents the position in the feature map x with position i as the center, dilation rate r and convolution kernel size k after offset.

5. The bridge structure anomaly detection method based on dilated convolution feature fusion according to claim 1 is characterized by: Calculating the attention weight of each branch according to the features of each branch graph in step S2 includes the following steps: Step S21: Generate a spatial attention map for each branch through dilated convolution; Step S22: Splice the spatial attention maps of multiple branches along the branch dimension, normalize all branch weights at each spatial position in the branch dimension, and obtain the attention weight of each branch at each spatial position.

6. The bridge structure anomaly detection method based on dilated convolution feature fusion according to claim 5 is characterized by: The expression for generating the spatial attention map for each branch through the dilated convolution in step S21 is: Where, is the spatial attention map of the i-th branch; DilatedConv i represents the dilated convolution operation; F i is the feature of the feature map of the i-th branch.

7. The bridge structure anomaly detection method based on dilated convolution feature fusion according to claim 5 is characterized by: The expression for normalizing all branch weights at each spatial position in the branch dimension in step S22 is: Where, ω i,h,w is the weight of the i-th branch at the spatial position (h, w); softmax() is a normalized exponential function; is the spatial attention map of the i-th branch.

8. The bridge structure anomaly detection method based on dilated convolution feature fusion according to claim 7 is characterized by: The expression of the fusion feature in step S2 is: Where, F fused It is a fusion feature; ω i,h,w is the weight of the i-th branch at the spatial position (h, w); n is the total number of branches; F i is the feature of the feature map of the i-th branch.

9. The bridge structure anomaly detection method based on dilated convolution feature fusion according to claim 1 is characterized by: The expression of the output feature in step S3 is: F final =(F fused +F residual )⊙Gate; F residual =ChannelProjection(x); Gate=σ(TemporalConv(F fused +F residual )); In the above formula, F final is the output feature; F fused is the fusion feature; F residual is the residual feature; ⊙ represents element-by-element multiplication; Gate is the gating 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.

10. A bridge structure anomaly detection system based on dilated convolution feature fusion, implemented based on the bridge structure anomaly detection method based on dilated convolution feature fusion according to any one of claims 1 to 9, characterized in that: include: Data acquisition module, time series feature extraction module, multi-scale feature learning module, anomaly diagnosis module and visual warning module; The data acquisition module is used to collect monitoring data of the bridge structure; The time series feature extraction module converts the historical monitoring data of the bridge structure into a frequency domain signal, extracts the periodic frequency components, selects the first k frequency components with the largest power, obtains the period length, and reorganizes the monitoring data into a two-dimensional tensor according to the period length; The multi-scale feature learning module uses different convolution kernels to extract short-term, medium-term, and long-term features from a two-dimensional tensor. It adjusts the dilation rate to expand the receptive field, captures the relationship between multi-scale features, dynamically generates weights to achieve adaptive feature fusion, and uses time-domain convolution and activation functions to generate gating values, strengthen key abnormal time series features, and output prediction values. The abnormality diagnosis module calculates the difference between the predicted value and the real monitoring data collected in real time. When the difference exceeds the set difference range, it determines that the bridge structure is abnormal. The abnormality occurrence area is determined based on the sensor location. The abnormality starting point is traced back through the time series characteristics, and the abnormality level is classified according to the difference amplitude. The visual warning module dynamically displays the raw data and abnormal point marks of each sensor, 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