A Deep Learning-Based Method for Real-Time Monitoring and Damage Early Warning of Bridge Structures
By separating the elastic strain and plastic residual of bridge structures through deep learning networks and combining orientation angle consistency and adjacency relationship, the problem of untimely bridge damage identification and false detection in existing technologies is solved, and real-time and accurate damage early warning of bridge structures is realized.
Patent Information
- Application Number
- CN202611622043.2
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2026-10-20
- Publication Date
- 2026-11-17
AI Technical Summary
Existing bridge structural health monitoring methods cannot effectively distinguish between recoverable elastic response and irreversible plastic residue after unloading, resulting in untimely damage identification and a high false detection rate. In particular, it is difficult to obtain plastic residue benchmarks by resting under continuous vehicle loads.
A deep learning network is used to jointly process the strain time history of the loading segment and the vehicle load time history, separating the total strain increment into elastic strain components and plastic residual increment. The time series is constructed using the principal strain direction angle of the plastic residual strain, and the consistency of the direction angle is determined by circular statistics. Adjacency relationships are constructed by combining the spatial location of the measuring points and the force transmission path of the structure to form a connected region to confirm the damage location.
It improves the timeliness and accuracy of bridge damage early warning, reduces the false alarm rate, and enables real-time damage tracking and accurate early warning under continuous vehicle traffic conditions, without interrupting traffic or waiting for a static period.
Smart Images

Figure CN123115706A_ABST
Abstract
Description
Technical Field
[0001] This invention belongs to the field of bridge structural health monitoring technology, and more specifically, relates to a method for real-time monitoring and damage early warning of bridge structural health based on deep learning. Background Technology
[0002] Currently, bridge structural health monitoring typically involves deploying sensors at various cross-sections of the bridge to collect structural response data, and then triggering early warnings by comparing the monitored parameters with preset thresholds.
[0003] Existing technologies, such as the bridge structural health monitoring and early warning system disclosed in Chinese invention patent application number 202411239277.X, deploy multiple sensors to cover the key structure of the bridge and set personalized safety thresholds and early warning levels according to the actual situation of the bridge. Although personalized safety thresholds are set, they are essentially still threshold comparisons.
[0004] Existing technologies, such as the bridge structure health monitoring and early warning system disclosed in Chinese invention patent application No. 202610609180.6, further utilize the dynamic baseline table and real-time temperature and humidity data distributed from the cloud to dynamically correct the expected value of thermal strain and its error range, and trigger an early warning by whether the residual exceeds the error range, that is, upgrade the static preset threshold to the dynamically corrected threshold, but it still belongs to threshold comparison.
[0005] After temperature correction, the residual of the above threshold comparison method still contains two different components: the elastic response that can be fully recovered after unloading and the plastic residual of the permanent offset remaining after unloading. The former does not indicate damage, while the latter indicates that the structure has undergone irreversible plastic deformation. Furthermore, under continuous vehicle load, the amount of plastic residual generated by a single vehicle crossing event is extremely small, and the next vehicle load event follows immediately. It is impossible to make a judgment by waiting for the strain to fully recover after static loading.
[0006] In summary, the threshold comparison method only uses whether the residual amplitude exceeds the threshold as the criterion, which cannot distinguish between recoverable elastic response and irreversible plastic residue after unloading. Furthermore, it is difficult to obtain a plastic residue benchmark by resting under continuous vehicle load, thus failing to identify irreversible damage and provide early warning in a timely and reliable manner. Summary of the Invention
[0007] In view of this, in order to solve the above problems, a method for real-time monitoring of bridge structural health and early warning of damage based on deep learning is proposed.
[0008] The objective of this invention can be achieved through the following technical solution: This invention provides a method for real-time monitoring and damage early warning of bridge structural health based on deep learning. The method includes: acquiring the strain time history and vehicle load time history of each measuring point in each vehicle crossing event, and extracting the time period from the vehicle crossing the bridge to the first local strain peak point as the loading segment.
[0009] The strain time history of the loaded segment, the vehicle load time history, and the historical cumulative plastic residual at each measuring point are input into the pre-trained deep learning network. The network outputs the elastic strain components and plastic residual increments of the current vehicle crossing event and updates the historical cumulative plastic residual.
[0010] The direction angle of the principal strain of the plastic residual is calculated based on the plastic residual increment, and the time series of the direction angle of the measuring point is constructed. The ratio of the elastic strain component to the vehicle load is used as the compliance coefficient, and the time series of the compliance coefficient of the measuring point is constructed.
[0011] Perform circular statistics on the direction angle time series. If the statistical results meet the direction angle consistency condition, then the direction of the measuring point is determined to be locked.
[0012] The adjacency relationship is constructed based on the spatial location of the measuring points and the force transmission path of the structure. If the adjacent directions lock the measuring points to form a connected region, it is determined whether the time sequence of the compliance coefficient of each measuring point in the connected region changes in the same direction. If so, the region is regarded as the damage location and an early warning is issued.
[0013] Compared with the prior art, the beneficial effects of the present invention are as follows: (1) The present invention uses a deep learning network to jointly process the strain time history of the loading segment and the vehicle load, and separates the total strain increment of the loading segment into elastic strain components and plastic residual increment. This solves the problem that the existing threshold comparison method cannot distinguish between recoverable elastic response and irreversible plastic residual, and avoids the interference of elastic response fluctuation and temperature residual on the damage criterion, thereby further improving the timeliness and accuracy of bridge damage early warning.
[0014] (2) This invention utilizes the principal strain direction angles of plastic residuals under multiple vehicle crossing events to construct a time sequence. It uses circular statistics to determine whether the direction angles tend to be consistent and uses the consistency of direction to replace the amplitude exceeding the limit as the judgment criterion. This allows the irreversible plastic deformation direction to be identified even under the condition that the plastic residuals of a single event are minimal. It does not need to rely on the plastic residuals of a single event, thus realizing the continuous tracking of the irreversible plastic deformation direction under the condition of continuous vehicle passage. At the same time, it does not need to interrupt traffic or wait for a static period, thus ensuring the timeliness of tracking.
[0015] (3) Based on the construction of adjacency relationship according to the spatial location of the measuring point and the force transmission path of the structure, and after locking the measuring points in adjacent directions to form a connected region, the invention eliminates non-damaging factors such as local stress concentration or sensor drift in single-point directional consistency by verifying whether the flexibility coefficient in the region increases in the same direction. This realizes the expansion of single-point directional consistency into a dual confirmation of spatial connectivity and continuous stiffness degradation, thereby effectively reducing the false alarm rate of early warning. Attached Figure Description
[0016] Figure 1 This is a schematic diagram of the overall implementation process of the method of the present invention;
[0017] Figure 2 This is a schematic diagram of the measurement point direction locking determination process of the present invention;
[0018] Figure 3 This is a schematic diagram of the loading segment division of the present invention;
[0019] Figure 4 This is a schematic diagram of the deep learning network structure of the present invention;
[0020] Figure 5 This is a schematic diagram of the construction of connected regions in this invention.
[0021] Figure labels: 1. First local strain peak point; 2. Loading segment. Detailed Implementation
[0022] 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 scope of protection of the present invention.
[0023] Currently, bridge structural health monitoring typically compares the temperature-corrected residual amplitude with a preset threshold. However, due to the continuous action of vehicle loads, the plastic residual generated by a single event is extremely small, and the next vehicle load arrives immediately afterward, before the elastic response of the previous event has fully recovered. Therefore, the residual contains a mixture of elastic response and plastic residual. Existing methods, which only use whether the residual amplitude exceeds the threshold as a criterion, may misclassify recoverable elastic responses as damage events, leading to a high false detection rate.
[0024] Based on this, this invention discloses a method for real-time monitoring of bridge structural health and damage early warning based on deep learning. By separating the total strain increment of the loaded segment into elastic strain components and plastic residual increment, and replacing the residual amplitude exceeding the threshold with the temporal consistency of the principal strain direction angle of the plastic residual as the damage criterion, the accuracy of bridge damage early warning can be effectively improved.
[0025] Please refer to the details. Figure 1 As shown, the present invention provides a method for real-time monitoring of bridge structural health and damage early warning based on deep learning. The method includes: S1, obtaining the strain time history and vehicle load time history of each measuring point in each vehicle crossing event, and extracting the time period from the vehicle crossing the bridge to the first local strain peak point as the loading segment.
[0026] The measuring points are distributed across various monitoring sections of the bridge. Each measuring point is equipped with a triaxial strain gauge to collect the strain time history at that point. The triaxial strain gauge is preferably a vibrating wire strain gauge or a fiber optic strain gauge, and is equipped with a temperature-compensated strain gauge. Before data collection for each vehicle crossing the bridge event, zero-point calibration is performed on each measuring point. When the zero-point drift exceeds a preset drift threshold (e.g., 5% of full range), a sensor maintenance alarm is triggered.
[0027] Vehicle load time history can be acquired synchronously through methods such as weighing sensors, axle detectors, or video recognition. It includes at least one parameter such as axle load, wheelbase, vehicle speed, and lane information. Each parameter can be acquired by its corresponding detection device; for example, axle load is acquired by a weighing sensor, wheelbase and vehicle speed by an axle detector, and lane information by a camera mounted on the bridge. The vehicle load parameters are arranged chronologically within the loading segment to form the vehicle load time history.
[0028] It should be noted that the period from when the vehicle hits the bridge to the first local strain peak corresponds to the stage from when the vehicle load begins to act on the bridge until the structural response reaches its first peak. During this stage, the structural response is primarily dominated by the vehicle load loading process; the elastic response and plastic residual are not yet fully mixed, facilitating subsequent deep learning networks to separate the elastic strain components from the plastic residual increments. For details, please refer to [link to relevant documentation]. Figure 3 As shown in the figure, the strain gradually increases after the vehicle enters the bridge, reaching the first local peak point. The period before this peak point is the loading phase. (The figure shows...) Indicates the time when the vehicle enters the bridge. This indicates the first local strain peak point.
[0029] For example, when obtaining the strain time history, strain data at each measuring point can be collected synchronously at a set sampling frequency using a triaxial strain gauge during the vehicle's crossing of the bridge. The strain data at the same moment can be aligned with the vehicle load data by timestamp to form the strain time history and vehicle load time history within the loading segment.
[0030] It should be understood that the first local strain peak point refers to the first local maximum value point that appears after the vehicle enters the bridge during the strain time history. This point corresponds to the first peak value reached during the loading process. Afterward, continued vehicle movement may cause the strain to change again. Extracting this period as the loading segment is mainly used to focus on the main loading stage of the vehicle load and avoid interference from the subsequent unloading and recovery processes on the identification of plastic residues.
[0031] S2. Input the strain time history of the loading segment, the vehicle load time history, and the historical cumulative plastic residual of each measuring point into the pre-trained deep learning network, output the elastic strain component and plastic residual increment of the current vehicle crossing the bridge event, and update the historical cumulative plastic residual.
[0032] As a preferred embodiment, please refer to Figure 4 As shown, the pre-trained deep learning network includes a strain load feature extraction layer, an elastic response output branch, and a plastic residual output branch. The strain load feature extraction layer receives the strain time series of the loading segment and the vehicle load, outputs a shared feature vector, and its output is connected to both the elastic response output branch and the plastic residual output branch. The plastic residual output branch also receives the historical cumulative plastic residual at each measuring point. The elastic response output branch and the plastic residual output branch output the elastic strain component and the plastic residual increment of the current vehicle crossing event, respectively.
[0033] In one specific embodiment, the strain load feature extraction layer employs a structure of a one-dimensional convolutional neural network and a long short-term memory network cascaded together. The one-dimensional convolutional neural network includes two one-dimensional convolutional layers: the first layer has 32 kernels, a kernel size of 3, a stride of 1, and uses ReLU activation; the second layer has 64 kernels, a kernel size of 3, a stride of 1, and uses ReLU activation. Each convolutional layer is followed by a max-pooling layer with a pooling window size of 2. The long short-term memory network includes one long short-term memory layer with a hidden layer dimension of 128, outputting a fixed-length shared feature vector of 128 dimensions. This shared feature vector is simultaneously input to both the elastic response output branch and the plastic residual output branch.
[0034] Furthermore, both the elastic response output branch and the plastic residual output branch adopt a three-layer fully connected network. The first fully connected layer has 64 neurons and the activation function is ReLU. The second fully connected layer has 32 neurons and the activation function is ReLU. The third fully connected layer has 1 neuron and no activation function is set in the third layer.
[0035] It should be added that the input of the plastic residual output branch is a concatenated vector of the shared feature vector and the historical cumulative plastic residual of each measurement point.
[0036] It should also be noted that the plastic residual increment is the scalar value of the plastic residual increment generated at this measuring point in the current vehicle crossing the bridge event. Furthermore, the splicing method involves expanding the scalar value of the historically accumulated plastic residual into a vector of the same dimension as the shared feature vector, and then splicing it with the shared feature vector by channel.
[0037] Furthermore, the plastic residual increment is added to the historical cumulative plastic residual before the update to obtain the updated historical cumulative plastic residual. Specifically, the plastic residual increment output by the current vehicle crossing event is added to the historical cumulative plastic residual before the update to obtain the updated historical cumulative plastic residual, which is then used as the historical cumulative plastic residual input to the plastic residual output branch in the next vehicle crossing event.
[0038] It is also important to note that, considering the network output accuracy may not meet requirements in the early stages of training, if the historical cumulative plastic residual is updated by accumulating the network output, the error will accumulate with each bridge crossing event and be fed back into the input of the next training sample, leading to instability in the training process. However, in the inference phase, the network parameters have been sufficiently trained and fixed, and their output accuracy has been verified by independent test sets to meet the strain measurement resolution requirements. At this point, accumulating the network output ensures that the cumulative error remains within an acceptable range. Therefore, during the training phase, the historical cumulative plastic residual at each measurement point is updated by accumulating the plastic residual increment label to ensure the accuracy of the training input data; during the inference phase, the historical cumulative plastic residual at each measurement point is updated by accumulating the plastic residual increment of the network output to achieve real-time continuous tracking of the irreversible plastic deformation direction.
[0039] In another example, the pre-trained deep learning network is trained in the following manner: the loading segment strain sequence, vehicle load, and historical cumulative plastic residual at each measuring point before the historical vehicle crossing the bridge event are used as training inputs, and the elastic strain component labels and plastic residual increment labels of the historical vehicle crossing the bridge event are used as training labels to supervise the training of the deep learning network, and the historical cumulative plastic residual at each measuring point is updated by accumulating the plastic residual increment labels.
[0040] Specifically, during training, for each historical vehicle crossing event, the strain time series of the loading segment and the vehicle load are input into the strain-load feature extraction layer, and the historical cumulative plastic residual is input into the plastic residual output branch. The network outputs elastic strain components and plastic residual increments, and calculates the loss with the corresponding elastic strain component labels and plastic residual increment labels. The network parameters are updated through backpropagation. For example, the loss functions for both the elastic response output branch and the plastic residual output branch are mean squared error losses, and the total network loss is the sum of the losses of the two branches. The mean squared error loss is calculated by dividing the sum of the squares of the differences between the network output elastic strain components and their corresponding labels by the number of training samples, and the sum of the squares of the differences between the network output plastic residual increments and their corresponding labels by the number of training samples. The sum of these two results in the total network loss. Training uses the Adam optimizer with an initial learning rate of 0.001, a batch size of 32, and 200 training epochs. Training terminates when the validation set loss does not decrease for 10 consecutive epochs.
[0041] It should be added that the training samples are historical vehicle bridge crossing event data accumulated after the monitoring system was put into operation.
[0042] After training, the network performance is validated on an independent test set consisting of historical vehicle bridge-crossing events that were not part of the training. Training is considered successful when the root mean square error of both the elastic response output branch and the plastic residual output branch on the test set does not exceed the strain measurement resolution at that point, and the network parameters are fixed for the inference phase. If the test accuracy is insufficient, historical vehicle bridge-crossing events are accumulated and retrained until the aforementioned accuracy condition is met.
[0043] Furthermore, in this embodiment of the invention, the training labels are generated in the following manner: A1, vehicle crossing events with a time interval greater than a first preset duration after the vehicle leaves the bridge are selected as training samples.
[0044] Understandably, the first preset duration is used to ensure that the elastic response generated by the previous vehicle crossing the bridge has been fully recovered, so that the strain value before the vehicle crosses the bridge only contains the irrecoverable historical cumulative plastic residue. The first preset duration is determined based on the first-order natural vibration period of the bridge, which is 30 to 60 times the first-order natural vibration period and not less than 5 minutes. The first-order natural vibration period is obtained by performing spectral analysis on the strain time history after the vehicle crosses the bridge.
[0045] It should be noted that this selection criterion is only used during the training phase to generate training labels to ensure their accuracy. During the online inference phase after training is complete, this selection criterion is not required; the system can output the plastic residual increment in real time for each consecutive vehicle crossing the bridge event and continuously track changes in the direction angle.
[0046] A2. For the selected training samples, the strain value corresponding to the moment when the vehicle leaves the bridge and the second preset time has elapsed is taken as the residual strain value of the measuring point. The strain value of the measuring point before the vehicle crosses the bridge in this vehicle crossing event is taken as the initial strain value. The difference between the residual strain value and the initial strain value is taken as the plastic residual increment label of the measuring point.
[0047] A2. For the selected training samples, the strain value corresponding to the moment when the vehicle leaves and the second preset time is reached is taken as the residual strain value of the measuring point. The strain value of the measuring point before the vehicle enters the bridge in this vehicle crossing event is taken as the initial strain value. Since the training samples have been selected as events where the time interval between the vehicle leaving and the next vehicle entering the bridge is greater than the first preset time, the elastic response of the measuring point before the vehicle enters the bridge has been fully recovered. The strain value of the measuring point before the vehicle enters the bridge is taken as the strain value of the most recent sampling point before the vehicle enters the bridge.
[0048] Understandably, the second preset duration is used to ensure that the elastic response generated by the vehicle crossing the bridge has been fully recovered, so that the residual strain value only includes the sum of the historical cumulative plastic residual and the current plastic residual increment. The second preset duration is determined based on the fact that the change in strain at the measuring point within a number of consecutive sampling points after the vehicle leaves is less than the strain measurement resolution. The time between the earliest time that meets this condition and the time when the vehicle leaves is taken as the second preset duration, wherein there are no fewer than three sampling points.
[0049] Optionally, for bridges that are under heavy traffic and have continuous vehicle passage, making it impossible to obtain a static period of the second preset duration after a vehicle leaves the bridge and before the next vehicle enters, the following alternative method is used to determine the residual strain value: Calculate the absolute value of the strain difference between adjacent sampling points during the strain time history of the measuring point from the departure of a vehicle to the entry of the next vehicle. Take the strain value of the next sampling point corresponding to the adjacent sampling point with the smallest absolute value of the difference as the residual strain value. Calculate the range of the strain within three consecutive sampling points of the adjacent sampling point. When the range is less than the strain measurement resolution, it is determined that the residual strain value can be used for label generation. The difference between the residual strain value and the initial strain value of the measuring point before the vehicle enters the bridge in this vehicle crossing event is used as the plastic residual increment label of the measuring point. If the range is greater than or equal to the strain measurement resolution, the training sample corresponding to this vehicle crossing event is removed and does not participate in network training.
[0050] It should be noted that, since the sample selection criteria ensure that the time elapsed between the vehicle's entry onto the bridge and its previous departure is greater than a first preset time, the initial strain value only includes historically accumulated plastic residuals. The residual strain value, however, includes the sum of historically accumulated plastic residuals and the current increase in plastic residuals. Therefore, the difference between the residual strain value and the initial strain value is the increase in plastic residuals generated by the current vehicle crossing the bridge event, with historically accumulated plastic residuals being offset within this difference.
[0051] A3. Calculate the difference between the measured strain value at the peak of the loading segment and the initial strain value, as the total strain increment. Use the difference between the total strain increment and the plastic residual increment label as the elastic strain component label. The total strain increment reflects the increment of the total structural response during the current vehicle crossing event, and it is composed of the elastic strain component and the plastic residual increment. Subtracting the plastic residual increment label from the total strain increment yields the elastic strain component label.
[0052] Therefore, the training labels are generated independently from the measured strain values after the vehicle leaves, which are two preset time intervals from the departure time. They do not depend on the cumulative update of the plastic residual increment of the network output for the current vehicle crossing the bridge event after training is completed. The training input, training labels and network output are consistent.
[0053] S3. Calculate the direction angle of the principal plastic strain based on the plastic residual increment, construct the time series of the direction angle of the measuring point, and use the ratio of the elastic strain component to the vehicle load as the compliance coefficient to construct the time series of the compliance coefficient of the measuring point.
[0054] Specifically, the method for calculating the principal strain direction angle based on the plastic residual increment is as follows: For each measurement point, obtain the plastic residual increment in each direction of the triaxial strain rosette output by the deep learning network, and calculate the plastic residual principal strain direction angle of each measurement point according to the formula for calculating the principal strain direction of the triaxial strain rosette.
[0055] It should be added that the formula for calculating the principal strain direction of the triaxial strain rosette is a conventional formula in this field. The principal strain direction angle of the plastic residual can be obtained by calculating the angle of the plastic residual increment in each direction of the triaxial strain rosette. For example, for triaxial strain rosettes of 0°, 45°, and 90°, let the plastic residual increments in the three directions be respectively , , Then the direction angle of the plastic residual principal strain satisfy: .
[0056] In another specific instance, the ratio of the elastic strain component to the vehicle load is used as the compliance coefficient. The vehicle load is taken as the peak value of the vehicle load time history within the loading segment or the equivalent uniformly distributed load value. The equivalent uniformly distributed load value is taken as the integral of the vehicle load time history within the loading segment with respect to time divided by the duration of the loading segment.
[0057] In addition, the orientation angle timing and compliance coefficient timing are constructed separately on a per-measurement-point basis for subsequent orientation locking determination and damage confirmation.
[0058] S4. Perform circular statistics on the direction angle time sequence. If the statistical results meet the direction angle consistency condition, then the direction of the measuring point is determined to be locked.
[0059] Considering that the direction of residual plasticity caused by damage is determined by the location and form of damage, it should remain stable across multiple vehicle crossing events. Furthermore, the direction of vehicle load in bridge design is typically along the longitudinal direction of the lanes, and the load direction itself remains largely consistent across multiple crossing events. Strain caused by temperature effects is largely recoverable after unloading and does not generate residual plasticity components. Therefore, under the premise of relatively stable load direction and temperature conditions, the direction angle of residual plasticity caused by non-damage factors such as noise should exhibit random fluctuations. If the direction angle of the principal residual plasticity strain at a certain measuring point shows a statistically significant concentration trend across multiple valid events, it indicates the existence of a stable direction source dominated by irreversible structural deformation, i.e., the direction of residual plasticity caused by damage.
[0060] Based on this, please refer to Figure 2As shown, the present invention determines whether there is a stable dominant direction by performing circular statistics on the direction angle time sequence. The specific process is as follows: obtain the vehicle load and plastic residual increment of each vehicle crossing the bridge event, remove events where the vehicle load does not reach the set load level and the plastic residual increment is less than the preset increment threshold, and record the remaining events after removal as valid events.
[0061] Understandably, the set load level is determined based on the lowest vehicle load corresponding to the event that produces plastic residual increment in the historical monitoring data of the measuring point. Specifically, the minimum value of the vehicle load corresponding to all events that produce plastic residual increment during the monitoring period of the measuring point is taken as the set load level. In the early stage of monitoring, the set load level is taken as 60% of the standard value of bridge design load.
[0062] It is also understandable that the preset incremental threshold is determined based on the noise level of the strain sensor at the measuring point. Specifically, the strain data at the measuring point is collected during periods when no vehicles are crossing the bridge, the standard deviation is calculated, and three times the standard deviation is taken as the preset incremental threshold.
[0063] Obtain the plastic residual principal strain direction angle and plastic residual increment of each measuring point under each effective event from the measuring point direction angle time series.
[0064] The direction angle of the plastic residual principal strain is doubled and mapped, that is... Change to The cosine and sine values of the plastic residual principal strain direction angle after mapping are calculated. The double-angle mapping is used to eliminate the periodic effect of the direction angle, so that 0° and 180° are mapped to the same direction, avoiding periodic jump error.
[0065] The weight of each valid event is determined by the proportion of the plastic residual increment of each valid event to the sum of the plastic residual increments of all valid events. Based on the weight, the cosine and sine values of each valid event are weighted and summed to obtain the weighted cosine sum and the weighted sine sum.
[0066] Understandably, events with larger increases in plastic residual are more representative of the true plastic residual direction of the structure by their orientation angles, and therefore are given higher weight.
[0067] The orientation angle concentration at each measuring point is calculated based on the weighted cosine sum and the weighted sine sum, where the orientation angle concentration corresponds to the resultant vector length. Specifically, the orientation angle concentration can be calculated using the following formula: ,in, Indicates the concentration of the orientation angle. This represents the weighted cosine sum. This represents the weighted sine sum. A larger value indicates a more concentrated range of directional angles for the valid events. The smaller the value, the more dispersed the distribution.
[0068] If the concentration of the direction angle at a certain measuring point is greater than or equal to the concentration threshold, the direction of the measuring point is determined to be locked; otherwise, the measuring direction is determined to be unlocked.
[0069] Understandably, the concentration threshold is determined based on the direction angle concentration of each valid vehicle crossing event within one month after the measurement point is confirmed to be undamaged by manual inspection or load test, and the maximum value of the direction angle concentration of each valid event within that month is taken as the concentration threshold.
[0070] S5. Construct adjacency relationships based on the spatial location of the measuring points and the structural force transmission path. If adjacent directions lock the measuring points to form a connected region, determine whether the flexibility coefficient of each measuring point in the connected region changes in the same direction. If so, take the region as the damage location and issue an early warning.
[0071] Specifically, the process of constructing the adjacency relationship based on the spatial location of the measuring points and the structural force transmission path is as follows: obtain the bridge design drawings, and pre-mark the pairs of measuring points adjacent to each measuring point along the force transmission path according to the structural force transmission path in the bridge design drawings.
[0072] For pre-marked pairs of measuring points, if the spatial distance between them is less than a preset distance threshold, it is determined that the two measuring points are adjacent. The preset distance threshold is set according to the bridge span; the larger the bridge span, the larger the value. For conventional bridges, the preset distance threshold is usually 1 / 10 to 1 / 5 of the bridge span. For example, for a 30m span bridge, the preset distance threshold is 3 to 6m. For a 100m span bridge, the preset distance threshold is 10 to 20m.
[0073] It should be noted that when adjacency is determined solely by spatial distance, measuring points that are spatially close but have unrelated force transmission paths may be misjudged as adjacent, such as measuring points on the left and right sides of a bridge that are adjacent but experience independent forces. This invention introduces structural force transmission paths, ensuring that adjacency relationships align with actual force transmission, and resulting in more accurate location of connected areas.
[0074] It should also be noted that the actual spatial distance can be obtained through the coordinates of the measuring points in the design drawings or through on-site measurement.
[0075] After establishing adjacency relationships, further determine the connected regions. For example, please refer to [reference needed]. Figure 5 As shown, the connected region is defined according to the following rules: a direction-locked measurement point graph is constructed using direction-locked measurement points as nodes and adjacency relationships as edges. Connected components are searched within the direction-locked measurement point graph; each connected component is a set of measurement points consisting of mutually adjacent direction-locked measurement points.
[0076] A connected component containing at least two directional locking measurement points is defined as a connected region.
[0077] Understandably, the search for the connected components can be implemented using either depth-first search or breadth-first search, and these search methods are existing technologies, so the search process will not be described in detail here.
[0078] It should be noted that the directional locking of a single measuring point may be caused by local stress concentration or sensor drift, and is insufficient to confirm a damage area. However, the simultaneous directional locking of multiple adjacent measuring points indicates that the irreversible deformation has spatial continuity, which is closer to the actual damage. Therefore, by identifying connected regions, this embodiment of the invention can combine spatially related directional-locked measuring points, thereby providing a spatial range for subsequent judgment of the same-direction change of the compliance coefficient and avoiding misjudgment of isolated single points.
[0079] After determining the connected region, it is further determined whether the time series of the compliance coefficients of each measuring point within the connected region changes in the same direction. For example, the specific process of the determination is as follows: the time series of compliance coefficients of each measuring point are processed by sliding window averaging to generate a smoothed compliance sequence for each measuring point.
[0080] Within the same interval of vehicle crossing event numbers, using the event sequence number as the independent variable, the linear regression slope of the compliance smoothing sequence at each measuring point is calculated to obtain the compliance change slope at each measuring point.
[0081] Determine whether the slope of the compliance change at each measuring point is positive and greater than a preset slope threshold. If it is positive and greater than the preset slope threshold, then determine that the compliance coefficients of each measuring point within the connected region change in the same direction over time.
[0082] Understandably, the window length of the sliding window is determined based on the average daily number of vehicle crossing events on the bridge, taking 1 / 4 to 1 / 2 of that average daily number of events. For example, if the average daily number of vehicle crossing events at this measuring point is 20, then the window length is 5 to 10 events. The preset slope threshold is determined based on the maximum absolute value of the flexibility coefficient slope at this measuring point during the initial monitoring period, and this maximum value is taken as the preset slope threshold.
[0083] It should be noted that the essence of structural damage is a decrease in stiffness and an increase in flexibility. The flexibility coefficients at each measuring point show a significant increasing trend over time, indicating that stiffness degradation has occurred at each measuring point in the connected region, which can be identified as the damaged area.
[0084] Optionally, in this embodiment of the invention, the method further includes: if the direction-locked measuring points do not form a connected region, then the component positions where each direction-locked measuring point is located are output as suspected damage locations, and encrypted monitoring is triggered. The encrypted monitoring includes increasing the data acquisition frequency of the measuring point and adjacent measuring points to obtain more data for subsequent judgment.
[0085] It should be noted that the directional locking of a single measuring point that has not formed a connected region may be an early sign of damage or may be caused by local factors. Outputting it as a suspected damage location and triggering encrypted monitoring avoids directly triggering damage warnings to prevent false alarms from single points, while retaining the ability to detect early damage signals.
[0086] This invention, through constructing adjacency relationships based on the spatial location of measurement points and the structural force transmission path, searching for connected regions using direction-locked measurement points as nodes, and judging the temporal changes of flexibility coefficients within connected regions in the same direction, achieves step-by-step confirmation from single-point direction locking to spatial connected regions, and then to continuous stiffness degradation, ultimately outputting the damage location and issuing an early warning. For single-point direction locking that has not formed a connected region, the suspected damage location is output and encrypted monitoring is triggered, balancing the reliability of early warning and the ability to detect early damage.
[0087] In this embodiment, direction locking, connected region determination, and compliance coefficient temporal change judgment are executed sequentially. Direction locking is used to extract the stable direction dominated by irreversible plastic deformation from continuous bridge crossing events; connected region determination is used to exclude local stress concentration or sensor drift at isolated single points; and compliance coefficient temporal change judgment is used to confirm continuous stiffness degradation within the connected region. The output of the previous stage serves as the basis for the judgment in the next stage. The direction locking result provides nodes for the connected region, the connected region provides the spatial range for the compliance coefficient temporal judgment, and the compliance coefficient temporal change confirms continuous stiffness degradation within this spatial range. Therefore, damage localization can be achieved without waiting for a static period under continuous vehicle traffic conditions.
[0088] The above content is merely an example and illustration of the concept of the present invention. Those skilled in the art can make various modifications or additions to the specific embodiments described, or use similar methods to replace them, as long as they do not deviate from the concept of the invention or exceed the scope defined by the present invention, and all such modifications and additions should fall within the protection scope of the present invention.
Claims
1. A method for real-time monitoring of bridge structural health and early warning of damage based on deep learning, characterized in that, The method includes: Obtain the strain time history and vehicle load time history of each measuring point for each vehicle crossing the bridge event, and extract the time period from when the vehicle goes onto the bridge to the first local strain peak point as the loading segment; The strain time history of the loaded segment, the vehicle load time history, and the historical cumulative plastic residual at each measuring point are input into the pre-trained deep learning network. The network outputs the elastic strain components and plastic residual increments of the current vehicle crossing event and updates the historical cumulative plastic residual. The direction angle of the principal strain of the plastic residual is calculated based on the plastic residual increment, and the time series of the direction angle of the measuring point is constructed. The ratio of the elastic strain component to the vehicle load is used as the compliance coefficient, and the time series of the compliance coefficient of the measuring point is constructed. Perform circular statistics on the direction angle time series. If the statistical results meet the direction angle consistency condition, then the direction of the measuring point is determined to be locked. The adjacency relationship is constructed based on the spatial location of the measuring points and the force transmission path of the structure. If the adjacent directions lock the measuring points to form a connected region, it is determined whether the time sequence of the compliance coefficient of each measuring point in the connected region changes in the same direction. If so, the region is regarded as the damage location and an early warning is issued.
2. The method for real-time monitoring and damage early warning of bridge structural health based on deep learning as described in claim 1, characterized in that: The pre-trained deep learning network includes a strain load feature extraction layer, an elastic response output branch, and a plastic residual output branch, wherein: The strain load feature extraction layer receives the strain time sequence of the loading segment and the vehicle load, outputs a shared feature vector, and its output is connected to the elastic response output branch and the plastic residual output branch respectively. The plastic residual output branch also receives the historical cumulative plastic residual of each measuring point. The elastic response output branch and the plastic residual output branch respectively output the elastic strain component and the plastic residual increment of the current vehicle crossing the bridge event. The plastic residual increment is used to add with the historical cumulative plastic residual before the update to obtain the updated historical cumulative plastic residual. The plastic residual increment is used to accumulate and update the historical cumulative plastic residual at each measuring point.
3. The method for real-time monitoring and damage early warning of bridge structural health based on deep learning as described in claim 2, characterized in that: The pre-trained deep learning network is trained in the following way: The deep learning network is trained by using the strain timing of the loading segment, vehicle load, and historical cumulative plastic residual at each measuring point before the historical vehicle crossing the bridge event as training inputs, and the elastic strain component label and plastic residual increment label of the historical vehicle crossing the bridge event as training labels. During training, the historical cumulative plastic residual of each measurement point is updated by accumulating the plastic residual increment label to generate the input for the next training sample; After training is completed, the historical cumulative plastic residual of each measurement point is updated by accumulating the plastic residual increment output by the network during the inference phase.
4. The method for real-time monitoring and damage early warning of bridge structural health based on deep learning as described in claim 3, characterized in that: The training labels are generated in the following way: Vehicle crossing events where the time interval between the departure of a vehicle and the next vehicle crossing the bridge is greater than a first preset time are selected as training samples. For the selected training samples, the strain value corresponding to the moment when the vehicle leaves the bridge and the second preset time has elapsed is taken as the residual strain value of the measuring point. The strain value of the measuring point before the vehicle crosses the bridge in this vehicle crossing event is taken as the initial strain value. The difference between the residual strain value and the initial strain value is taken as the plastic residual increment label of the measuring point. The difference between the measured strain value at the peak of the loading segment and the initial strain value is calculated as the total strain increment, and the difference between the total strain increment and the plastic residual increment label is used as the elastic strain component label.
5. The method for real-time monitoring and damage early warning of bridge structural health based on deep learning as described in claim 1, characterized in that: The specific method for calculating the principal strain direction angle based on the plastic residual increment is as follows: For each measurement point, obtain the plastic residual increment in each direction of the triaxial strain rosette output by the deep learning network, and calculate the plastic residual principal strain direction angle of each measurement point according to the formula for calculating the principal strain direction of the triaxial strain rosette.
6. The method for real-time monitoring and damage early warning of bridge structural health based on deep learning as described in claim 1, characterized in that: Perform circular statistics on the orientation angle time series. If the statistical results meet the orientation angle consistency condition, the orientation of the measuring point is determined to be locked. Specifically, this includes: Obtain the vehicle load and plastic residual increment for each vehicle crossing the bridge event, remove events where the vehicle load does not reach the set load level and the plastic residual increment is less than the preset increment threshold, and record the remaining events as valid events. Obtain the plastic residual principal strain direction angle and plastic residual increment of each measuring point under each effective event from the measuring point direction angle time series; The plastic residual principal strain direction angle is doubled and mapped, and the cosine and sine values of the mapped plastic residual principal strain direction angle are calculated. The weight of each effective event is determined by the proportion of the plastic residual increment of each effective event to the sum of the plastic residual increments of all effective events. Based on the weight, the cosine and sine values of each effective event are weighted and summed to obtain the weighted cosine sum and the weighted sine sum. The concentration of the orientation angle at each measuring point is calculated based on the weighted cosine sum and the weighted sine sum. If the concentration of the direction angle at a certain measuring point is greater than or equal to the concentration threshold, then the direction of the measuring point is determined to be locked.
7. The method for real-time monitoring and damage early warning of bridge structural health based on deep learning as described in claim 1, characterized in that: The specific process for constructing the adjacency relationship based on the spatial location of the measuring point and the structural force transmission path is as follows: Obtain bridge design drawings and, based on the structural force transmission path in the bridge design drawings, pre-mark adjacent pairs of measuring points along the force transmission path at each measuring point. For a pair of pre-marked measurement points, if the spatial distance between them is less than a preset distance threshold, it is determined that there is an adjacency relationship between the two measurement points.
8. The method for real-time monitoring and damage early warning of bridge structural health based on deep learning as described in claim 1, characterized in that: The connected regions are defined according to the following rules: Construct a direction-locked measurement point graph using direction-locked measurement points as nodes and adjacency relationships as edges; Search for connected components in the orientation-locked measurement point map, where the connected components are a set of measurement points consisting of mutually adjacent orientation-locked measurement points; A connected component containing at least two directional locking measurement points is defined as a connected region.
9. The method for real-time monitoring and damage early warning of bridge structural health based on deep learning as described in claim 1, characterized in that: The specific process for determining whether the time sequence of the compliance coefficients of each measuring point in the connected region changes in the same direction is as follows: The time series of compliance coefficients at each measuring point are processed by sliding window mean to generate a smoothed compliance sequence for each measuring point. Within the same interval of vehicle crossing event numbers, the linear regression slope of the compliance smoothing sequence at each measuring point is calculated using the event sequence number as the independent variable, thus obtaining the compliance change slope at each measuring point. Determine whether the slope of the compliance change at each measuring point is positive and greater than the preset slope threshold. If all values are positive and greater than the preset slope threshold, then it is determined that the flexibility coefficients of each measuring point in the connected region change in the same direction over time.
10. The method for real-time monitoring and damage early warning of bridge structural health based on deep learning as described in claim 1, characterized in that: The method also includes: If the direction-locked measuring points do not form a connected region, the component locations where each direction-locked measuring point is located will be output as suspected damage locations, and encrypted monitoring will be triggered.
Citation Information
Patent Citations
Bridge structure health monitoring and early warning system
CN118758377A
Bridge structure health monitoring and early warning system
CN122468200A
