CN-121997225-A - Bridge monitoring method, system, equipment and medium based on deep learning

CN121997225ACN 121997225 ACN121997225 ACN 121997225ACN-121997225-A

Abstract

The application provides a bridge monitoring method, system, equipment and medium based on deep learning, which comprises the following steps of collecting original strain sequences and environment parameters of monitoring points of a road bridge, calculating theoretical thermal strain according to the environment parameters and material thermal expansion coefficients, aligning and subtracting the theoretical thermal strain with original data time sequence to correct and obtain structural strain, constructing a space-time strain matrix in a preset time window according to the positions of the monitoring points and the collecting time sequence, uploading the space-time strain matrix to a cloud, inputting the matrix into a deployed convolutional neural network at the cloud, extracting deep features by adopting multi-layer convolution comprising a time dimension convolution kernel and a space dimension convolution kernel, mapping the deep features into a settlement state probability vector through a full connection layer, comparing the vector with a reference health state vector, and triggering settlement early warning if the deviation exceeds a safety threshold. By implementing the technical scheme provided by the application, the environmental interference can be effectively eliminated, and the accuracy and the instantaneity of bridge settlement monitoring are improved.

Inventors

LIU HONGBO
GUO RUI
GONG WENGANG
WANG PING

Assignees

陕西交通公路设计研究院有限公司

Dates

Publication Date: 20260508
Application Date: 20260115

Claims (10)

1. A bridge monitoring method based on deep learning, the method comprising: collecting original strain monitoring data sequences of monitoring points of roads and bridges, and collecting environmental parameter data sequences of areas where the monitoring points are located; Calculating theoretical thermal strain numerical value sequences of all monitoring points at different sampling moments according to the environmental parameter data sequences and preset material thermal expansion coefficients, and carrying out time sequence alignment and one-by-one subtraction on the original strain monitoring data sequences and the theoretical thermal strain numerical value sequences to obtain a corrected structural strain sequence; arranging the corrected structure strain sequence according to the positions of the monitoring points and the acquisition time sequence, and constructing a space-time strain matrix in a preset time window; Transmitting the space-time strain matrix to a cloud analysis platform, and inputting the space-time strain matrix into a convolutional neural network model deployed on the cloud analysis platform; performing convolution operation on the space-time strain matrix through a multi-layer convolution kernel in the convolution neural network model to generate a deep feature map, wherein the multi-layer convolution kernel comprises a time dimension convolution kernel and a space dimension convolution kernel; Inputting the deep feature map into a full-connection layer in the convolutional neural network model for numerical mapping to obtain a settlement state probability vector of the road and bridge at the current moment; and calculating the numerical deviation between the sedimentation state probability vector and a preset reference health state vector, and if the numerical deviation exceeds a preset safety threshold, generating a sedimentation early warning instruction.
2. The method according to claim 1, wherein calculating the theoretical thermal strain value sequence of each monitoring point at different sampling moments according to the environmental parameter data sequence and a preset thermal expansion coefficient of the material specifically comprises: traversing the environment parameter data sequence through a sliding time window with a preset length, and determining a sampling time positioned at the center of the sliding time window as a target sampling time; Based on a preset weight attenuation function, respectively calculating time distance weights corresponding to the environment temperature values according to time intervals between sampling moments of the environment temperature values and the target sampling moments contained in the sliding time window, wherein the time distance weights and the time intervals are in negative correlation; Performing weighted average calculation on each environmental temperature value contained in the sliding time window by using the time distance weight to obtain an effective temperature value corresponding to the target sampling moment, and determining a sequence formed by a plurality of effective temperature values as an effective temperature sequence; Resampling the effective temperature sequence through a preset interpolation algorithm to generate an aligned temperature sequence corresponding to the timestamp of the original strain monitoring data sequence; Subtracting preset reference temperature values from each temperature value in the aligned temperature sequence to obtain a temperature difference sequence; and multiplying each numerical value in the temperature difference value sequence by the thermal expansion coefficient of the material to obtain the theoretical thermal strain numerical value sequence.
3. The method according to claim 1, wherein the arranging the modified structural strain sequence according to the monitoring point position and the acquisition time sequence, and constructing a space-time strain matrix within a preset time window specifically comprises: Performing sliding slicing operation on the corrected structural strain sequence according to a preset time step and a preset window length, and calculating a differential sequence of the corrected structural strain sequence after the slicing operation on a time axis to obtain a strain rate subsequence; Calculating a correlation coefficient value between a target monitoring point and each adjacent monitoring point based on a preset correlation calculation model, comparing the correlation coefficient value with a preset spatial correlation threshold value, and determining an effective monitoring point set with the correlation coefficient value larger than the spatial correlation threshold value; taking a strain rate subsequence corresponding to the effective monitoring point set as a first characteristic channel, taking a modified structural strain sequence corresponding to the effective monitoring point set as a second characteristic channel, and combining the first characteristic channel and the second characteristic channel to obtain a multi-channel matrix; And carrying out normalization processing on the multichannel matrix to generate the space-time strain matrix.
4. A method according to claim 3, wherein calculating a correlation coefficient value between the target monitoring point and each adjacent monitoring point based on the preset correlation calculation model specifically includes: Extracting a first numerical sequence of the target monitoring point in the preset time window from the corrected structural strain sequence, and extracting a second numerical sequence of the adjacent monitoring point in the preset time window; Setting a plurality of discrete time lag amounts in a preset phase deviation searching range, and respectively carrying out position deviation processing on data elements in the second numerical sequence according to each time lag amount to generate a plurality of groups of candidate sequences to be detected which respectively correspond to different time lag amounts; Respectively calculating standard cross-correlation values between the first numerical sequence and each candidate sequence to be detected to obtain a cross-correlation value set; and screening out the value with the largest absolute value from the cross-correlation value set, and determining the value with the largest absolute value as the correlation coefficient value.
5. The method according to claim 1, wherein the convolving the space-time strain matrix by a multi-layer convolution check in the convolutional neural network model to generate a deep feature map, specifically comprising: Performing feature extraction of time dimension by using the time dimension convolution check to the space-time strain matrix to obtain a time feature component; Performing feature extraction of space dimension by utilizing the space dimension convolution check to the space-time strain matrix to obtain a space feature component; Performing feature fusion and self-adaptive calibration on the time feature component and the space feature component to obtain a calibrated feature tensor; And performing dimension reduction processing on the calibrated feature tensor through a maximum pooling layer in the convolutional neural network model to generate the deep feature map.
6. The method according to claim 5, wherein the performing feature fusion and adaptive calibration on the temporal feature component and the spatial feature component to obtain a calibrated feature tensor specifically includes: Channel splicing is carried out on the time characteristic component and the space characteristic component, and a fusion characteristic tensor is constructed; global average pooling is carried out on the fusion feature tensors, and one-dimensional weight vectors describing statistical information among feature channels are obtained; Inputting the one-dimensional weight vector into a preset weight generation sub-network, and calculating to obtain a weight coefficient sequence corresponding to each characteristic channel, wherein the weight generation sub-network is constructed by a plurality of full-connection layers; And multiplying each weight coefficient in the weight coefficient sequence with a corresponding characteristic channel in the fusion characteristic tensor to obtain the calibrated characteristic tensor.
7. The method of claim 1, wherein the inputting the deep feature map into the full-connection layer in the convolutional neural network model performs numerical mapping to obtain a settlement state probability vector of the road bridge at the current moment, and specifically comprises: remodelling the multidimensional deep feature map into a one-dimensional feature vector by using a flattening function; performing matrix multiplication operation on the one-dimensional feature vector and a preset weight matrix in the full-connection layer, and superposing a preset offset vector to obtain a logic value sequence; and inputting the logic numerical value sequence into a preset probability normalization function, calculating probability distribution values of all sedimentation state categories, and generating the sedimentation state probability vector.
8. A bridge monitoring system based on deep learning, the system comprising: The data acquisition module is configured to acquire original strain monitoring data sequences of all monitoring points of the road and bridge and acquire environment parameter data sequences of areas where the monitoring points are located; The data processing module is configured to calculate theoretical thermal strain value sequences of the monitoring points at different sampling moments according to the environmental parameter data sequences and preset material thermal expansion coefficients, and to align the original strain monitoring data sequences with the theoretical thermal strain value sequences in time sequence and subtract the original strain monitoring data sequences one by one to obtain a corrected structural strain sequence; The matrix construction module is configured to arrange the modified structure strain sequence according to the positions of the monitoring points and the acquisition time sequence to construct a space-time strain matrix in a preset time window; The data transmission module is configured to transmit the space-time strain matrix to a cloud analysis platform and input the space-time strain matrix into a convolutional neural network model deployed on the cloud analysis platform; the feature extraction module is configured to perform convolution operation on the space-time strain matrix through a multi-layer convolution kernel in the convolution neural network model to generate a deep feature map, wherein the multi-layer convolution kernel comprises a time dimension convolution kernel and a space dimension convolution kernel; The state analysis module is configured to input the deep feature map into a full-connection layer in the convolutional neural network model to carry out numerical mapping so as to obtain a settlement state probability vector of the road and bridge at the current moment; and the sedimentation early warning module is configured to calculate the numerical deviation between the sedimentation state probability vector and a preset reference health state vector, and if the numerical deviation exceeds a preset safety threshold, a sedimentation early warning instruction is generated.
9. An electronic device comprising a processor, a memory, a user interface, and a network interface, the memory for storing instructions, the user interface and the network interface each for communicating with other devices, the processor for executing instructions stored in the memory to cause the electronic device to perform the method of any of claims 1-7.
10. A computer readable storage medium storing instructions which, when executed, perform the method of any one of claims 1-7.

Description

Bridge monitoring method, system, equipment and medium based on deep learning Technical Field The invention relates to the technical field of deep learning, in particular to a bridge monitoring method, system, equipment and medium based on deep learning. Background With the rapid development of urban infrastructure construction and intelligent traffic systems, the coverage area of road and bridge networks is continuously enlarged, and the carried traffic load is increasingly heavy. How to collect, transmit and deeply analyze the structural settlement, stress state and environmental parameters of roads and bridges in all weather and high precision has become a core requirement for guaranteeing traffic safety smoothness and prolonging the service life of infrastructure. In the prior art, the health status monitoring of roads and bridges is generally realized through an alarm system based on a traditional discrete sensor network and a preset fixed threshold value. For example, an alarm is triggered when the instantaneous settlement or stress value of the monitoring point exceeds a fixed safety limit. However, in the prior art, based on a fixed threshold value and a single physical quantity monitoring method in practical application, effective distinction between normal structural response under complex and changeable climatic environments (such as thermal expansion and cold contraction caused by temperature change) and abnormal settlement caused by early-stage micro diseases is difficult, and risks of high false alarm rate, insufficient micro deformation recognition precision, early warning response lag and the like exist. Disclosure of Invention In view of the above, the present application provides a method, system, device and medium for monitoring a bridge based on deep learning to solve the above problems. In a first aspect, a bridge monitoring method based on deep learning is provided, the method comprising: Collecting original strain monitoring data sequences of all monitoring points of a road bridge, and collecting environment parameter data sequences of areas where all monitoring points are located; Calculating theoretical thermal strain numerical sequences of all monitoring points at different sampling moments according to the environmental parameter data sequences and the preset material thermal expansion coefficients, and carrying out time sequence alignment and one-by-one subtraction on the original strain monitoring data sequences and the theoretical thermal strain numerical sequences to obtain a corrected structural strain sequence; arranging the corrected structure strain sequence according to the positions of the monitoring points and the acquisition time sequence, and constructing a space-time strain matrix in a preset time window; transmitting the space-time strain matrix to a cloud analysis platform, and inputting the space-time strain matrix into a convolutional neural network model deployed on the cloud analysis platform; Performing convolution operation on the space-time strain matrix through a multi-layer convolution kernel in the convolution neural network model to generate a deep feature map, wherein the multi-layer convolution kernel comprises a time dimension convolution kernel and a space dimension convolution kernel; inputting the deep feature map into a full-connection layer in a convolutional neural network model for numerical mapping to obtain a settlement state probability vector of a road and a bridge at the current moment; And calculating the numerical deviation between the sedimentation state probability vector and a preset reference health state vector, and if the numerical deviation exceeds a preset safety threshold, generating a sedimentation early warning instruction. According to the technical scheme, the temperature compensation is carried out on the original strain data, the space-time strain matrix is constructed, and the combination learning of the cloud convolutional neural network on the space-time characteristics is combined, so that the normal thermal expansion and contraction caused by temperature change and the abnormal settlement caused by structural diseases can be distinguished under the complex environmental condition. Compared with a fixed threshold judgment mode, the method can adaptively learn the normal response mode of the bridge under different climates and loads, thereby remarkably reducing the false alarm rate, improving the recognition precision of early micro-settlement and realizing real-time and intelligent early warning response. Optionally, calculating a theoretical thermal strain numerical sequence of each monitoring point at different sampling moments according to the environmental parameter data sequence and a preset material thermal expansion coefficient, and specifically including: traversing the environmental parameter data sequence through a sliding time window with a preset length, and determining the sampling time positioned at the c