Bridge monitoring method, system and equipment based on deep learning and medium

Abstract

The invention provides a bridge monitoring method, system and equipment based on deep learning and a medium, and the method comprises the following steps: collecting an original strain sequence and environmental parameters of each monitoring point of a road bridge, calculating a theoretical thermal strain according to the environmental parameters and a material thermal expansion coefficient, and carrying out the alignment subtraction of the theoretical thermal strain and an original data time sequence, so as to obtain a structural strain through correction; constructing a space-time strain matrix in a preset time window according to the monitoring point position and the acquisition time sequence, and uploading the space-time strain matrix to the cloud; inputting the matrix into a deployed convolutional neural network at a cloud end, extracting deep features by adopting multi-layer convolution comprising a time dimension convolution kernel and a space dimension convolution kernel, and then mapping the deep features into settlement state probability vectors through a full connection layer; and comparing the vector with a reference health state vector, and if the deviation exceeds a safety threshold, triggering settlement early warning. By implementing the technical scheme provided by the invention, environmental interference can be effectively eliminated, and the accuracy and real-time performance of bridge settlement monitoring are improved.

Description

Technical Field

[0001] This invention relates to the field of deep learning technology, and in particular to a bridge monitoring method, system, device and medium based on deep learning. Background Technology

[0002] With the rapid development of urban infrastructure construction and intelligent transportation systems, the coverage of road and bridge networks continues to expand, and the traffic load they bear is becoming increasingly heavy. How to collect, transmit, and deeply analyze the structural settlement, stress state, and environmental parameters of roads and bridges around the clock with high precision has become a core requirement for ensuring smooth traffic flow and extending the service life of infrastructure.

[0003] In existing technologies, health status monitoring of roads and bridges is typically achieved through alarm systems based on traditional discrete sensor networks and preset fixed thresholds. For example, an alarm is triggered when the instantaneous settlement or stress value at a monitoring point exceeds a fixed safety limit. However, in practical applications, existing technologies based on fixed thresholds and single physical quantity monitoring methods struggle to effectively distinguish between normal structural responses under complex and variable climatic environments (such as thermal expansion and contraction due to temperature changes) and abnormal settlement caused by early minor defects. This leads to risks such as high false alarm rates, insufficient accuracy in identifying minor deformations, and delayed early warning responses. Summary of the Invention

[0004] In view of this, this application provides a bridge monitoring method, system, device and medium based on deep learning to solve the above problems.

[0005] Firstly, a bridge monitoring method based on deep learning is provided, which includes:

[0006] The original strain monitoring data sequence of each monitoring point on the road and bridge was collected, as well as the environmental parameter data sequence of the area where each monitoring point is located;

[0007] The theoretical thermal strain numerical sequence of each monitoring point at different sampling times is calculated based on the environmental parameter data sequence and the preset material thermal expansion coefficient. The original strain monitoring data sequence and the theoretical thermal strain numerical sequence are then time-aligned and subtracted one by one to obtain the corrected structural strain sequence.

[0008] The corrected structural strain sequence is arranged according to the location of the monitoring points and the acquisition time to construct a spatiotemporal strain matrix within a preset time window;

[0009] The spatiotemporal strain matrix is ​​transmitted to the cloud analysis platform, and the spatiotemporal strain matrix is ​​input into the convolutional neural network model deployed on the cloud analysis platform;

[0010] Deep feature maps are generated by performing convolution operations on the spatiotemporal strain matrix using multi-layer convolutional kernels in a convolutional neural network model. The multi-layer convolutional kernels include temporal convolutional kernels and spatial convolutional kernels.

[0011] The deep feature map is input into the fully connected layer in the convolutional neural network model for numerical mapping to obtain the probability vector of the settlement state of the road and bridge at the current moment.

[0012] Calculate the numerical deviation between the settlement state probability vector and the preset baseline health state vector. If the numerical deviation exceeds the preset safety threshold, generate a settlement early warning command.

[0013] The above technical solution, by performing temperature compensation on the original strain data and constructing a spatiotemporal strain matrix, combined with joint learning of spatiotemporal features by a cloud-based convolutional neural network, can distinguish between normal thermal expansion and contraction caused by temperature changes and abnormal settlement caused by structural defects under complex environmental conditions. Compared with fixed threshold judgment methods, this method can adaptively learn the normal response patterns of bridges under different climatic and load conditions, thereby significantly reducing the false alarm rate, improving the accuracy of early minor settlement identification, and achieving real-time, intelligent early warning response.

[0014] Optionally, the theoretical thermal strain numerical sequence of each monitoring point at different sampling times is calculated based on the environmental parameter data sequence and the preset material thermal expansion coefficient, specifically including:

[0015] The environmental parameter data sequence is traversed through a sliding time window of preset length, and the sampling time located at the center of the sliding time window is determined as the target sampling time.

[0016] Based on the preset weight decay function, the time distance weight corresponding to each ambient temperature value is calculated according to the time interval between the sampling time of each ambient temperature value included in the sliding time window and the target sampling time. The time distance weight is negatively correlated with the time interval.

[0017] The effective temperature value corresponding to the target sampling time is obtained by using the time distance weight to calculate the weighted average of each ambient temperature value contained in the sliding time window, and the sequence composed of multiple effective temperature values ​​is determined as the effective temperature sequence.

[0018] The effective temperature sequence is resampled using a preset interpolation algorithm to generate an aligned temperature sequence corresponding to the timestamp of the original strain monitoring data sequence.

[0019] Subtract the preset reference temperature value from each temperature value in the aligned temperature sequence to obtain the temperature difference sequence;

[0020] By multiplying each value in the temperature difference sequence by the material's thermal expansion coefficient, a theoretical thermal strain numerical sequence is obtained.

[0021] The above technical solution obtains an effective temperature sequence by using a sliding time window and a weighted calculation method based on time distance, and aligns it to the strain sampling time. This allows for accurate estimation of the thermal strain contribution at each moment when the temperature changes drastically. The theoretical thermal strain is calculated using the coefficient of thermal expansion and subtracted from the original data, ensuring that the structural strain data is unaffected by climate temperature fluctuations. This significantly reduces false alarms of settlement caused by temperature errors and improves the authenticity and reliability of the data.

[0022] Optionally, the corrected structural strain sequence is arranged according to the monitoring point location and acquisition time sequence to construct a spatiotemporal strain matrix within a preset time window, specifically including:

[0023] A sliding slice operation is performed on the corrected structural strain sequence according to a preset time step and a preset window length, and the difference sequence of the corrected structural strain sequence on the time axis after the slice operation is calculated to obtain the strain rate subsequence.

[0024] Based on the preset correlation calculation model, the correlation coefficient between the target monitoring point and each adjacent monitoring point is calculated, and the correlation coefficient is compared with the preset spatial correlation threshold to determine the set of valid monitoring points with a correlation coefficient greater than the spatial correlation threshold.

[0025] The strain rate subsequence corresponding to the effective monitoring point set is used as the first feature channel, and the corrected structural strain sequence corresponding to the effective monitoring point set is used as the second feature channel. The first feature channel and the second feature channel are combined to obtain a multi-channel matrix.

[0026] The multi-channel matrix is ​​normalized to generate a spatiotemporal strain matrix.

[0027] The above technical solution constructs a spatiotemporal strain matrix containing multi-channel features by sliding slices and differencing the corrected structural strain sequence on the time axis and combining it with spatial correlation screening between monitoring points. This matrix can simultaneously capture the continuous deformation patterns of the bridge in terms of both temporal variation and spatial distribution. This matrix integrates local and global structural response features, providing a high-dimensional, low-noise input foundation for subsequent models, thereby improving the spatial resolution and global consistency of settlement identification.

[0028] Optionally, based on a preset correlation calculation model, the correlation coefficient between the target monitoring point and each adjacent monitoring point is calculated, specifically including:

[0029] Extract the first numerical sequence of the target monitoring point within a preset time window from the corrected structural strain sequence, and extract the second numerical sequence of the adjacent monitoring point within a preset time window.

[0030] Within a preset phase offset search range, multiple discrete time lags are set, and the data elements in the second numerical sequence are offset according to each time lag, generating multiple sets of candidate test sequences corresponding to different time lags.

[0031] Calculate the standard cross-correlation values ​​between the first numerical sequence and each candidate test sequence to obtain the cross-correlation value set;

[0032] The value with the largest absolute value is selected from the set of cross-correlation values, and this value is determined as the correlation coefficient value.

[0033] The above technical solution identifies the hysteresis correlation between the target monitoring point and its adjacent points through phase offset cross-correlation calculation, which can accurately capture the spatiotemporal characteristics of structural deformation propagation. Compared with simple spatial distance correlation methods, this method can reflect the true coupling degree of the structural dynamic response, thereby optimizing the spatial feature extraction results, reducing false identification caused by local anomalies or noise, and improving the stability of the model under complex structural conditions.

[0034] Optionally, a deep feature map can be generated by performing convolution operations on the spatiotemporal strain matrix using multiple convolutional kernels in a convolutional neural network model, specifically including:

[0035] The temporal dimension of the spatiotemporal strain matrix is ​​extracted using a temporal convolution kernel to obtain temporal feature components;

[0036] Spatial dimension features are extracted from the spatiotemporal strain matrix using spatial dimension convolution kernels to obtain spatial feature components;

[0037] The temporal and spatial feature components are fused and adaptively calibrated to obtain the calibrated feature tensor.

[0038] The dimensionality of the calibrated feature tensor is reduced by using the max pooling layer in the convolutional neural network model to generate a deep feature map.

[0039] The above technical solution extracts features from different dimensions by setting convolutional kernels in the temporal and spatial dimensions, and generates deep feature maps through pooling operations. This allows for the simultaneous capture of the dynamic evolution and spatial diffusion characteristics of bridge deformation at multiple scales. This multi-dimensional convolutional structure reflects the overall health status of the structure better than a single monitoring indicator, thereby improving the sensitivity and detection accuracy of minute settlement changes.

[0040] Optionally, the temporal and spatial feature components are fused and adaptively calibrated to obtain a calibrated feature tensor, specifically including:

[0041] The temporal and spatial feature components are concatenated to construct a fused feature tensor.

[0042] Global average pooling is performed on the fused feature tensor to obtain a one-dimensional weight vector describing the statistical information between feature channels;

[0043] A one-dimensional weight vector is input into a preset weight generation sub-network to calculate the weight coefficient sequence corresponding to each feature channel. The weight generation sub-network is constructed from multiple fully connected layers.

[0044] Each weight coefficient in the weight coefficient sequence is multiplied by the corresponding feature channel in the fused feature tensor to obtain the calibrated feature tensor.

[0045] The above technical solution employs adaptive feature channel weighting through global average pooling and a weight generation subnetwork, which automatically adjusts the importance of temporal and spatial features based on real-time input data. This mechanism enables the model to maintain a dynamic balance of feature responses under different climate, load, or structural conditions, thereby improving the robustness of identification under complex environmental conditions and reducing false positives and false negatives.

[0046] Optionally, the deep feature map is input into the fully connected layer of the convolutional neural network model for numerical mapping to obtain the settlement state probability vector of the road and bridge at the current time, specifically including:

[0047] The multidimensional deep feature map is reshaped into a one-dimensional feature vector using a flattening function;

[0048] A matrix multiplication operation is performed between the one-dimensional feature vector and the preset weight matrix in the fully connected layer, and a preset bias vector is superimposed to obtain a logical numerical sequence.

[0049] Input the logical numerical sequence into the preset probability normalization function to calculate the probability distribution value of each settlement state category and generate a settlement state probability vector.

[0050] The above technical solution, through the mapping between fully connected layers and probability normalization functions, can transform high-dimensional convolutional features into settlement state probability vectors in real time, and automatically trigger early warning commands based on deviation thresholds. This approach transforms early warning from fixed threshold triggering to intelligent decision-making based on probability assessment, enabling rapid response to early, minute settlement changes, continuous monitoring, and immediate alarms, significantly improving the problem of lag in traditional systems.

[0051] Secondly, a bridge monitoring system based on deep learning is provided, the system comprising:

[0052] The data acquisition module is configured to collect raw strain monitoring data sequences from various monitoring points on roads and bridges, and to collect environmental parameter data sequences from the areas where each monitoring point is located.

[0053] The data processing module is configured to calculate the theoretical thermal strain numerical sequence of each monitoring point at different sampling times based on the environmental parameter data sequence and the preset material thermal expansion coefficient, and to time-align the original strain monitoring data sequence with the theoretical thermal strain numerical sequence and subtract them one by one to obtain the corrected structural strain sequence.

[0054] The matrix construction module is configured to arrange the corrected structural strain sequence according to the monitoring point location and acquisition time order, and construct the spatiotemporal strain matrix within a preset time window;

[0055] The data transmission module is configured to transmit the spatiotemporal strain matrix to the cloud analysis platform and input the spatiotemporal strain matrix into a convolutional neural network model deployed on the cloud analysis platform.

[0056] The feature extraction module is configured to perform convolution operations on the spatiotemporal strain matrix through multiple convolutional kernels in the convolutional neural network model to generate deep feature maps. The multiple convolutional kernels include temporal convolutional kernels and spatial convolutional kernels.

[0057] The state analysis module is configured to input deep feature maps into the fully connected layer of the convolutional neural network model for numerical mapping, and obtain the settlement state probability vector of the road and bridge at the current moment.

[0058] The settlement early warning module is configured to calculate the numerical deviation between the settlement state probability vector and the preset benchmark health state vector. If the numerical deviation exceeds the preset safety threshold, a settlement early warning command is generated.

[0059] Thirdly, an electronic device is provided, including a processor, a memory, a user interface, and a network interface, wherein the memory is used to store instructions, the user interface and the network interface are both used to communicate with other devices, and the processor is used to execute the instructions stored in the memory to cause the electronic device to perform the method as described in any of the above.

[0060] Fourthly, a computer-readable storage medium is provided, the computer-readable storage medium storing instructions that, when executed, perform the method as described in any of the preceding claims.

[0061] In summary, implementing one or more technical solutions provided in this application has at least the following technical effects or advantages:

[0062] By introducing a deep learning-based multidimensional feature analysis framework into bridge health monitoring, the entire process, from raw strain acquisition, temperature compensation, and spatiotemporal feature modeling to settlement identification and early warning, is fully automated. This solution does not rely on manually set thresholds or single monitoring indicators and can continuously learn and adaptively optimize under the dynamic updating of large-scale sensor data. Through a cloud-deployed convolutional neural network model, monitoring data from multiple bridges and time periods can be centrally processed while ensuring computational efficiency, enabling unified management and remote analysis of monitoring resources. Furthermore, this invention maintains stable identification accuracy and response speed under different environmental, seasonal, and load conditions, significantly improving the automation level, data processing efficiency, and long-term operational reliability of bridge structural health monitoring, providing an intelligent, low-manual-reliance solution for bridge safety operation and maintenance. Attached Figure Description

[0063] Figure 1 This is an exemplary system architecture diagram of a bridge monitoring method or a bridge monitoring system based on deep learning, which applies the present application.

[0064] Figure 2 This is a flowchart illustrating a bridge monitoring method based on deep learning disclosed in this application;

[0065] Figure 3 This is a schematic diagram of a bridge monitoring system based on deep learning disclosed in this application;

[0066] Figure 4 This is a schematic diagram of the structure of an electronic device disclosed in this application.

[0067] Explanation of reference numerals in the attached figures: 100, System architecture; 101, First terminal device; 102, Second terminal device; 103, Third terminal device; 104, Network; 105, Server; 301, Data acquisition module; 302, Data processing module; 303, Matrix construction module; 304, Data transmission module; 305, Feature extraction module; 306, Status analysis module; 307, Settlement early warning module; 401, Processor; 402, Communication bus; 403, User interface; 404, Network interface; 405, Memory. Detailed Implementation

[0068] To enable those skilled in the art to better understand the technical solutions in this specification, the technical solutions in the embodiments of this specification will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, and not all embodiments.

[0069] In the description of the embodiments of this application, the words "for example" or "for instance" are used to indicate examples, illustrations, or explanations. Any embodiment or design that is described as "for example" or "for instance" in the embodiments of this application should not be construed as being more preferred or advantageous than other embodiments or design options. Rather, the use of the words "for example" or "for instance" is intended to present the relevant concepts in a specific manner.

[0070] In the description of the embodiments of this application, the term "multiple" means two or more. For example, multiple systems means two or more systems, and multiple screen terminals means two or more screen terminals. Furthermore, the terms "first" and "second" are used for descriptive purposes only and should not be construed as indicating or implying relative importance or implicitly specifying the indicated technical features. Thus, a feature defined with "first" or "second" may explicitly or implicitly include one or more of that feature. The terms "comprising," "including," "having," and variations thereof all mean "including but not limited to," unless otherwise specifically emphasized.

[0071] Figure 1 An exemplary system architecture diagram is shown, illustrating an embodiment of a deep learning-based bridge monitoring method or a deep learning-based bridge monitoring system to which this application can be applied.

[0072] like Figure 1 As shown, the system architecture 100 may include a first terminal device 101, a second terminal device 102, a third terminal device 103, a network 104, and a server 105. The network 104 is used as a medium to provide communication links between the terminal devices 101, 102, 103, and the server 105. The network 104 may include various connection types, such as wired or wireless communication links or fiber optic cables, etc.

[0073] Users can use terminal devices 101, 102, and 103 to interact with server 105 via network 104 to receive or send messages, etc. Various communication client applications can be installed on terminal devices 101, 102, and 103, such as model training applications, video recognition applications, web browser applications, social platform software, etc.

[0074] Terminal devices 101, 102, and 103 can be either hardware or software. When terminal devices 101, 102, and 103 are hardware, they can be various electronic devices with displays, including but not limited to smartphones, tablets, e-book readers, MP3 (Moving Picture Experts Group Audio Layer III) players, MP4 (Moving Picture Experts Group Audio Layer IV) players, laptops, and desktop computers, etc. When terminal devices 101, 102, and 103 are software, they can be installed in the aforementioned electronic devices. They can be implemented as multiple software programs or software modules (e.g., multiple software programs or software modules used to provide distributed services) or as a single software program or software module. No specific limitations are imposed here.

[0075] When terminals 101, 102, and 103 are hardware devices, video capture devices can also be installed on them. These video capture devices can be various devices capable of capturing video, such as cameras, sensors, etc. Users can use the video capture devices on terminals 101, 102, and 103 to capture video.

[0076] Server 105 can be a server that provides various services, such as a backend server for processing data displayed on terminal devices 101, 102, and 103. The backend server can analyze and process the received data and can feed back the processing results (such as recognition results) to the terminal devices.

[0077] It should be noted that a server can be either hardware or software. When the server is hardware, it can be implemented as a distributed server cluster consisting of multiple servers, or as a single server. When the server is software, it can be implemented as multiple software programs or software modules (e.g., multiple software programs or software modules used to provide distributed services), or as a single software program or software module. No specific limitations are made here.

[0078] It should be understood that Figure 1 The number of terminal devices, networks, and servers shown is merely illustrative. Depending on implementation needs, any number of terminal devices, networks, and servers can be included. In particular, if the target data does not need to be obtained remotely, the above system architecture may exclude the network and include only terminal devices or servers.

[0079] Figure 2This is a flowchart illustrating a bridge monitoring method based on deep learning, as described in this application. This method can be implemented using a computer program, a microcontroller, or run on a bridge monitoring system based on deep learning. The computer program can be integrated into an application or run as a standalone utility application. The specific steps of the deep learning-based bridge monitoring method are described in detail below.

[0080] S201: Collect the original strain monitoring data sequence of each monitoring point on the road and bridge, and collect the environmental parameter data sequence of the area where each monitoring point is located.

[0081] In this embodiment, the original strain monitoring data sequence refers to a set of values ​​arranged in chronological order, continuously collected by strain sensing devices installed on the surface of the bridge structure at a preset sampling frequency. This set characterizes the amount of geometric deformation of the bridge structure at different times due to stress or environmental effects. For example, this data sequence can be represented as a set of time-series data collected by a fiber optic grating sensor at a frequency of 100Hz, containing specific timestamps and corresponding micro-strain values.

[0082] Specifically, to ensure the comprehensiveness of the collected data and the feasibility of the project, the following strategies can be adopted during the implementation of this step: Hardware selection and deployment: Distributed fiber optic sensor networks or high-precision wireless resistance strain gauge arrays will be used for data acquisition. The deployment of monitoring points will vary depending on the bridge type: for example, for suspension bridges, the focus will be on monitoring the strain in the main cable anchorage zone, mid-span hangers, and the base of the bridge towers; for beam bridges, the focus will be on monitoring the shear zone near the lower edge of the main beam mid-span and the supports. Considering that strain sensors and environmental sensors (such as thermometers and anemometers) may belong to different subsystems, to ensure strict alignment of the data sequence on the time axis, the system will adopt a timing mechanism based on Network Time Protocol (NTP) or Precision Time Protocol (PTP) to ensure that the timestamp error of all acquisition terminals is controlled within milliseconds (e.g., ±10ms), preventing correlation calculation errors caused by time misalignment. During the data acquisition process, if the sensor is subjected to transient electromagnetic interference that generates outlier noise, the system will use the "3σ criterion" or "median filtering" algorithm to remove abnormal jump values ​​in real time, and use linear interpolation between the previous and next time points to fill in the gaps, so as to ensure the continuity and physical authenticity of the original sequence.

[0083] Furthermore, to obtain comprehensive and accurate bridge condition data, the system needs to deploy monitoring points at key stress locations such as the mid-span of the main girder, the top of the piers, and the anchorage zone of the cable towers, based on the bridge's structural mechanical characteristics. High-precision strain acquisition devices are installed at these monitoring points. During the acquisition process, these devices sense the minute expansion and contraction deformations of the structural materials in real time, converting them into digital signals and adding time stamps to generate a continuous sequence of raw strain monitoring data. Simultaneously, considering that environmental factors (especially temperature) can significantly affect the thermal expansion and contraction of structural materials, thus interfering with the actual stress-strain analysis, environmental sensing sensors (such as patch thermometers and air temperature and humidity meters) deployed in local areas of each monitoring point must be activated concurrently with strain data acquisition. These environmental sensors are synchronized with the strain acquisition devices, recording a sequence of environmental parameter data in real time. This environmental parameter data sequence mainly consists of temperature values ​​corresponding one-to-one with the timestamps of the strain data, and may also include auxiliary parameters such as humidity or wind speed. The acquired data of both types are aggregated and stored, providing a complete data foundation for subsequent calculations of theoretical thermal strain and the elimination of environmental interference.

[0084] S202: Calculate the theoretical thermal strain numerical sequence of each monitoring point at different sampling times based on the environmental parameter data sequence and the preset material thermal expansion coefficient, and then align the original strain monitoring data sequence with the theoretical thermal strain numerical sequence in time and subtract them one by one to obtain the corrected structural strain sequence.

[0085] For example, this step aims to eliminate the interference of environmental thermal effects on structural health monitoring through a physical mechanism-driven data cleaning approach. Due to the inherent thermal expansion and contraction characteristics of bridge materials, the raw signals collected by sensors often contain a large number of non-structural thermal strain components. The system establishes a quantitative physical mapping relationship between ambient temperature and material deformation, accurately calculates the current theoretical thermal background, and uses the differential principle to separate it from the total monitoring data. This decouples the actual mechanical response (i.e., the corrected structural strain) caused only by traffic loads, wind loads, or potential structural damage, ensuring that the input data for subsequent deep learning model analysis has a high signal-to-noise ratio and physical purity.

[0086] In one possible implementation, the theoretical thermal strain numerical sequence of each monitoring point at different sampling times is calculated based on the environmental parameter data sequence and a preset material thermal expansion coefficient. Specifically, this includes: traversing the environmental parameter data sequence through a sliding time window of a preset length, and determining the sampling time at the center of the sliding time window as the target sampling time; calculating the time distance weight corresponding to each environmental temperature value based on a preset weight decay function and the time interval between the sampling time of each environmental temperature value included in the sliding time window and the target sampling time, where the time distance weight is negatively correlated with the time interval; performing a weighted average calculation on each environmental temperature value included in the sliding time window using the time distance weight to obtain the effective temperature value corresponding to the target sampling time, and determining the sequence composed of multiple effective temperature values ​​as the effective temperature sequence; resampling the effective temperature sequence using a preset interpolation algorithm to generate an aligned temperature sequence corresponding to the timestamp of the original strain monitoring data sequence; subtracting a preset reference temperature value from each temperature value in the aligned temperature sequence to obtain a temperature difference sequence; and multiplying each value in the temperature difference sequence by the material thermal expansion coefficient to obtain the theoretical thermal strain numerical sequence.

[0087] In this embodiment, the theoretical thermal strain numerical sequence refers to a time series calculated using physical constitutive equations based on the thermophysical properties of bridge building materials (such as the coefficient of thermal expansion) and real-time monitored ambient temperature change data. This series characterizes the theoretical deformation of the structure caused solely by temperature changes under no external load. For example, when an increase in ambient temperature causes concrete material to expand, the value at the corresponding moment in this sequence is the calculated positive micro-strain value, which is used to subsequently separate the temperature component from the total monitored strain.

[0088] Specifically, to accurately eliminate interference from ambient temperature on structural monitoring, the system uses a sliding time window of a preset length (e.g., 30 minutes or 1 hour) to progressively traverse the environmental parameter data sequence along the time axis. During each slide, the time point at the center of the window is locked and marked as the target sampling time. Considering the hysteresis and thermal inertia of heat conduction, a simple arithmetic mean cannot accurately reflect the true thermal state at the center time. Therefore, a preset weight decay function (e.g., a Gaussian kernel function) is invoked to calculate the corresponding time distance weight based on the absolute time difference (i.e., time interval) between the sampling times of each ambient temperature value within the window and the target sampling time. This calculation follows the negative correlation principle that closer distances result in greater weights, thus giving higher reliability to data near the center time. The calculated weights are then used to perform a weighted average operation on each ambient temperature value within the window, smoothing out instantaneous noise and calculating the effective temperature value at the target sampling time. A continuous effective temperature sequence is then generated as the window slides. Since the temperature acquisition frequency is typically lower than the strain acquisition frequency, a preset interpolation algorithm (such as linear interpolation or cubic spline interpolation) is then used to perform high-frequency resampling of the effective temperature sequence, significantly improving its temporal resolution and generating an aligned temperature sequence that is completely synchronized with the original strain monitoring data sequence on the time axis. A preset reference temperature value (usually the initial state temperature of the structure) is subtracted from each temperature point in this aligned temperature sequence, thereby quantifying the temperature difference sequence relative to the reference state. Based on the principles of thermoelasticity, each difference data point in this temperature difference sequence is multiplied by a preset material thermal expansion coefficient to obtain a theoretical thermal strain numerical sequence reflecting the pure temperature effect.

[0089] Furthermore, in the above embodiments, regarding the selection and parameter setting of the preset weight decay function, this application provides a more refined control strategy: In addition to using the standard Gaussian kernel function, the preset weight decay function can also be an exponential decay function or an inverse distance weighting function, selected based on the heat conduction characteristics of the actual monitoring environment. The environmental parameter data sequence here includes not only temperature but can also be extended to include solar radiation intensity and wind speed. In this case, the weighting function is upgraded to a multivariate coupling function. Specifically, the multivariate coupling function adopts a dynamic variance model based on wind speed correction. The system sets the basic Gaussian kernel bandwidth to σ. base Real-time wind speed data v t The dynamic bandwidth is calculated using the linear correction formula:

[0090]

[0091] Where k is a preset wind speed sensitivity coefficient (e.g., a value of 0.05). When the wind speed v t When it increases, the calculated σ dynamicReducing the size of the Gaussian kernel function narrows its distribution curve, which is mathematically equivalent to shortening the effective integration range of the sliding time window (i.e., reducing the weight of distant historical data), thus objectively matching the physical characteristics of the bridge structure's faster thermal response under strong convection conditions.

[0092] S203: Arrange the corrected structural strain sequence according to the location of the monitoring points and the order of acquisition time to construct a spatiotemporal strain matrix within a preset time window.

[0093] For example, this step primarily completes the structured reorganization and feature enhancement of the data. The system does not view the data from individual monitoring points in isolation, but rather transforms the discrete one-dimensional monitoring sequence into a high-dimensional tensor that can be directly analyzed by a deep learning model by introducing a spatiotemporal coupling perspective. In this process, the system not only retains the original static strain amplitude but also introduces dynamic strain rate information through differential calculations and filters out physically closely related spatial neighborhoods based on signal synchronicity. This constructs a standard input format that includes both local spatial topological relationships and short-term temporal evolution patterns, preparing the data for efficient feature extraction by subsequent convolutional neural networks.

[0094] In one possible implementation, the corrected structural strain sequence is arranged according to the monitoring point location and acquisition time sequence to construct a spatiotemporal strain matrix within a preset time window. Specifically, this includes: performing a sliding slice operation on the corrected structural strain sequence according to a preset time step and a preset window length, and calculating the difference sequence of the corrected structural strain sequence on the time axis after the slice operation to obtain a strain rate subsequence; calculating the correlation coefficient between the target monitoring point and each adjacent monitoring point based on a preset correlation calculation model, and comparing the correlation coefficient with a preset spatial correlation threshold to determine the set of valid monitoring points with correlation coefficients greater than the spatial correlation threshold; using the strain rate subsequence corresponding to the set of valid monitoring points as the first feature channel, and the corrected structural strain sequence corresponding to the set of valid monitoring points as the second feature channel, and combining the first feature channel and the second feature channel to obtain a multi-channel matrix; and normalizing the multi-channel matrix to generate a spatiotemporal strain matrix.

[0095] In this embodiment, the spatiotemporal strain matrix refers to a multidimensional data structure generated by reorganizing one-dimensional time-series data from monitoring points scattered at different physical locations on a bridge according to spatial topological relationships and time dimensions, and used for feature extraction by deep learning models. For example, the matrix can be represented as a three-dimensional tensor of the form (T, N, C), where T represents the length of the time window, N represents the number of spatially associated monitoring points (Nodes) selected, and C represents the number of feature channels (Channels) containing strain values ​​and strain rates.

[0096] Specifically, to construct a high-dimensional feature input that simultaneously reflects the local spatial correlation and temporal dynamic evolution of the structure, the system performs a sliding slicing operation on the one-dimensional corrected structural strain sequence according to a preset time step and a preset window length, truncating the long sequence into multiple fixed-length time segments. To capture the rate of strain change over time, a first-order difference operation is performed on the sliced ​​sequence along the time axis to calculate the strain rate subsequence. To focus on the spatial region with actual physical significance for the evaluation target point, a preset correlation calculation model (such as Pearson correlation coefficient or cross-correlation function) is used to quantify the signal synchronization degree between the target monitoring point and its neighboring monitoring points, i.e., the correlation coefficient value. By rigorously comparing this value with a preset spatial correlation threshold, interfering nodes with weak correlation are eliminated, and a group of nodes with highly synchronized physical responses is selected as the effective monitoring point set. It is important to note that the spatial correlation threshold is not a fixed value, but is determined using an adaptive threshold algorithm based on neighborhood statistical distribution. The system first calculates the correlation coefficient set C={r1, r2, ..., r...} between the target monitoring point and all candidate adjacent points (e.g., M points within a radius R). M}. Calculate the arithmetic mean μ of this set. r and standard deviation σ r Using the formula Th adaptive =μ r +λ·σ r Dynamically generate the filtering threshold for the current time window, where λ is the sensitivity coefficient (e.g., a value of 1.0). Only retain those satisfying r. i >Th adaptive monitoring points, r i This represents the correlation coefficient between the target monitoring point and a specific candidate adjacent point. This dynamic threshold mechanism can automatically adapt to the overall response level of the bridge under different traffic flow conditions, ensuring that the selected effective monitoring points are always strongly correlated nodes in the current environment, rather than strongly correlated in absolute value.

[0097] Furthermore, a construction method similar to that used in image processing (such as RGB channels) is adopted. The strain rate subsequence corresponding to the effective monitoring point set is used as the first feature channel describing the dynamic trend, and the corrected structural strain sequence (i.e., the original slice values) corresponding to the same set is used as the second feature channel describing the deformation amplitude. These two feature channels are stacked or combined in the depth direction to form a multi-channel matrix containing rich physical information. In order to eliminate the scale difference between data of different dimensions and accelerate the convergence of the neural network, normalization processing (such as standard deviation normalization or min-max normalization) is performed on the multi-channel matrix, finally generating a normalized spatiotemporal strain matrix.

[0098] Furthermore, in constructing the spatiotemporal strain matrix, to improve the convergence speed and robustness of model training, this step also includes the following key processing details: When normalizing the multi-channel matrix, given the significant differences in physical dimensions and numerical ranges between the strain rate subsequence (first channel) and the strain numerical sequence (second channel) (the former may be 10), -6 The order of magnitude, the latter being 10 -3 (At the order of magnitude), mixed normalization can mask the numerical contribution of small-value features by high-value features. Therefore, the system calculates the mean and standard deviation of each channel separately for Z-score standardization, adjusting the data distribution of each channel to a standard normal distribution with a mean of 0 and a variance of 1, ensuring that the model is equally sensitive to dynamic trends and static amplitudes. To improve the generalization ability of the convolutional neural network model and prevent overfitting, augmentation techniques specifically for spatiotemporal data are introduced during the matrix construction stage. For example, random time-axis masking (randomly zeroing a certain period of time) or node dropout (simulating sensor failure) is applied to the spatiotemporal matrix of the input model. This mechanism forces the model to learn to reconstruct the missing state using the spatial context information of the remaining nodes, thereby significantly improving the robustness of the system in the case of partial sensor failure.

[0099] In one possible implementation, based on a preset correlation calculation model, the correlation coefficient between the target monitoring point and each adjacent monitoring point is calculated. Specifically, this includes: extracting a first numerical sequence of the target monitoring point within a preset time window from the corrected structural strain sequence, and extracting a second numerical sequence of the adjacent monitoring points within the preset time window; setting multiple discrete time lags within a preset phase offset search range, and performing position offset processing on the data elements in the second numerical sequence according to each time lag, generating multiple sets of candidate test sequences corresponding to different time lags; calculating the standard cross-correlation values ​​between the first numerical sequence and each candidate test sequence to obtain a set of cross-correlation values; selecting the value with the largest absolute value from the set of cross-correlation values, and determining the value with the largest absolute value as the correlation coefficient value.

[0100] In this embodiment, the correlation coefficient is a statistical index used to quantify the similarity between the strain change trends of the target monitoring point and adjacent monitoring points within a specific time period under optimal time alignment, taking into full account the propagation delay characteristics of vibration waves in the bridge structure. For example, this value can be a floating-point number between -1 and 1, where a value of 0.95 indicates that the waveforms of the two monitoring points have extremely high consistency after correcting for a specific time lag.

[0101] Specifically, to accurately capture the asynchronous signal correlation characteristics between monitoring points at different spatial locations caused by physical distance and medium properties, the system precisely extracts data segments from the target monitoring point within a preset time window from the temperature-corrected structural strain sequence, based on the current analysis period, as the first numerical sequence. Simultaneously, it extracts data segments from adjacent monitoring points within the same time window as the second numerical sequence for comparison. Considering the phase difference in vibration signal transmission within large structures, zero-delay comparison alone is insufficient. Therefore, a series of discrete time lags need to be set within a preset phase offset search range. The phase offset search range is calculated based on the physical wave velocity constraints of the bridge structure, obtaining the straight-line physical distance L between the target monitoring point and adjacent monitoring points. dist And the theoretical minimum velocity v of stress wave propagation in bridge materials (such as concrete or steel). min (For example, approximately 2000 m / s in concrete). Calculate the maximum physical delay time:

[0102]

[0103] Combined with sampling frequency f s Determine the range of discrete time lag steps. The system sets the search range to [-K, +K] sampling points, where K is a positive integer. This range setting based on physical constraints avoids blind mathematical searches, significantly reducing the amount of computation, and also eliminates spurious high-correlation matches that are impossible in terms of causality (i.e., eliminates spurious correlations that exceed the speed of light or violate the common sense of wave speed).

[0104] Furthermore, using these time lags as sliding parameters, the data elements in the second numerical sequence are shifted left or right or cyclically along the time axis to construct multiple sets of candidate test sequences representing different delay assumption scenarios. A standardized cross-correlation algorithm is employed to calculate the standard cross-correlation value between the baseline first numerical sequence and each set of candidate test sequences, thereby eliminating the influence of signal amplitude differences and obtaining a set of cross-correlation values ​​containing multiple calculation results. This set is then traversed and searched to select the value with the largest absolute value, which represents the maximum similarity under optimal time lag compensation. This value is ultimately determined as the correlation coefficient describing the relationship between the two points.

[0105] S204: Transmit the spatiotemporal strain matrix to the cloud analysis platform and input the spatiotemporal strain matrix into the convolutional neural network model deployed on the cloud analysis platform.

[0106] In the embodiments of this application, the convolutional neural network model refers to a deep learning network architecture specifically designed to process data with a grid structure (such as spatiotemporal matrices or image data). It automatically extracts local features and high-dimensional abstract patterns from the data by simulating biological visual perception mechanisms and utilizing a combination of convolutional layers, pooling layers, and activation functions. For example, this model could be an end-to-end neural network model based on an improved deep residual network (ResNet) containing multiple layers of spatiotemporal convolutional kernels, capable of identifying settlement trends from complex strain matrices.

[0107] Specifically, to fully utilize the high-performance parallel computing capabilities of remote server clusters (such as GPU (Graphics Processing Unit) clusters), the system uses high-speed communication networks (such as 5G mobile networks or fiber optic broadband) and encrypted transmission protocols (such as HTTPS (Hypertext Transfer Protocol Secure) or MQTT (Message Queuing Telemetry Transport)) to upload the locally constructed spatiotemporal strain matrix in real time to a pre-built cloud analysis platform. This cloud analysis platform, acting as the central hub for data processing and model inference, immediately performs format verification and tensor transformation on the spatiotemporal strain matrix upon receiving the data, ensuring its data dimensions match the model's input layer interface requirements (e.g., by adding batch processing dimensions). It then calls the inference interface of the deep learning framework, loading the formatted spatiotemporal strain matrix as an input tensor and inputting it into the convolutional neural network model deployed on the cloud analysis platform, thereby triggering the model's forward propagation operation and providing the basic input for subsequent feature extraction and state discrimination.

[0108] S205: The spatiotemporal strain matrix is ​​convolved using multi-layer convolutional kernels in the convolutional neural network model to generate a deep feature map. The multi-layer convolutional kernels include temporal convolutional kernels and spatial convolutional kernels.

[0109] For example, this step aims to leverage the powerful nonlinear feature extraction capabilities of deep neural networks to perform multi-dimensional feature decoupling and reconstruction on the input spatiotemporal strain matrix. By deploying convolutional kernels targeting the time and spatial dimensions respectively, the model can simulate the spatiotemporal divide-and-conquer analytical logic. On the one hand, it captures the strain evolution pattern of a single monitoring point over time (i.e., time-series features); on the other hand, it reveals the force coordination pattern of different monitoring points at the same moment (i.e., spatial features). This processing method can transform low-level raw sensor values ​​into highly abstract deep semantic features, effectively filter out environmental noise interference, and extract key feature maps that can essentially reflect the health status of the bridge structure, laying a solid data foundation for subsequent accurate condition assessment.

[0110] Specifically, to maximize the extraction of spatiotemporal correlation features, the convolutional neural network model deployed on the cloud analytics platform employs an optimized network architecture design: the model uses a dual-stream spatiotemporal residual network architecture. The main body of the network contains 3-5 stacked spatiotemporal convolutional blocks. Within each convolutional block, a decoupled design is adopted. First, a 1×3 (or similar scale) spatial convolutional layer aggregates the features of neighboring nodes, followed by a 3×1 temporal convolutional layer to extract temporal features. This sequential processing is more efficient than 3D convolutional computation. Each convolutional block is equipped with residual connections, directly adding the input to the output. This design solves the gradient vanishing problem in deep networks, allowing deep feature maps to retain weak but crucial low-frequency strain trend information from the original input. After the convolutional layers, the system preferably uses Leaky ReLU or Swish activation functions instead of the traditional ReLU. This is because bridge strain data contains a large number of negative values ​​(representing compression), and traditional ReLU would set negative values ​​to zero, resulting in information loss. Leaky ReLU allows for non-zero gradients in the negative interval, preventing neurons from entering a permanent inactive state and thus preserving the structural stress mechanics information more completely.

[0111] In one possible implementation, a deep feature map is generated by performing convolution operations on the spatiotemporal strain matrix using multiple convolutional kernels in a convolutional neural network model. Specifically, this includes: performing time-dimensional feature extraction on the spatiotemporal strain matrix using time-dimensional convolutional kernels to obtain time feature components; performing spatial-dimensional feature extraction on the spatiotemporal strain matrix using spatial-dimensional convolutional kernels to obtain spatial feature components; performing feature fusion and adaptive calibration on the time and spatial feature components to obtain a calibrated feature tensor; and performing dimensionality reduction on the calibrated feature tensor using a max-pooling layer in the convolutional neural network model to generate a deep feature map.

[0112] In this embodiment, a multi-layer convolutional kernel refers to a pre-constructed set of filter matrices with different receptive field sizes, different scanning directions, and parameters that can be trained and optimized within a convolutional neural network. These matrices are used to perform sliding window operations on specific dimensions of the input data to extract local abstract features. For example, the multi-layer convolutional kernel may include a one-dimensional convolutional kernel specifically designed for the time axis (e.g., with a size of T×1) to capture temporal fluctuations; and a one-dimensional convolutional kernel designed for the spatial node axis (e.g., with a size of 1×S) to capture structural spatial relationships.

[0113] Specifically, to decouple and extract temporal evolution patterns and spatial structural patterns from complex monitoring data, the system controls the temporal convolution kernel to perform sliding convolution operations along the time axis of the spatiotemporal strain matrix, capturing the dynamic changes in strain data over time (such as mutation rates or periodic oscillations), thereby extracting temporal feature components representing the time dimension. In parallel or serial operation, the system uses spatial convolution kernels to perform convolution operations along the spatial distribution dimension of monitoring points, aggregating the force distribution patterns between adjacent monitoring points or physically related points, thereby extracting spatial feature components representing the spatial dimension. To achieve effective complementarity of multidimensional information, the extracted temporal and spatial feature components are concatenated or superimposed along the feature channel dimension to complete feature fusion. Furthermore, an attention mechanism (such as a channel attention module) is introduced to redistribute the weights of the fused features, assigning higher weights to feature channels containing important information and suppressing irrelevant noise, thereby achieving adaptive calibration and obtaining a calibrated feature tensor with more accurate information representation. The tensor is input into the max pooling layer of the convolutional neural network model to perform a downsampling operation, which reduces the data dimensionality and computational complexity while preserving the most salient feature responses in the local region, and finally generates a deep feature map.

[0114] In one possible implementation, the temporal and spatial feature components are fused and adaptively calibrated to obtain a calibrated feature tensor. Specifically, this includes: concatenating the temporal and spatial feature components into channels to construct a fused feature tensor; performing global average pooling on the fused feature tensor to obtain a one-dimensional weight vector describing the statistical information between feature channels; inputting the one-dimensional weight vector into a preset weight generation subnetwork to calculate a sequence of weight coefficients corresponding to each feature channel, wherein the weight generation subnetwork is constructed from multiple fully connected layers; and multiplying each weight coefficient in the weight coefficient sequence by the corresponding feature channel in the fused feature tensor to obtain the calibrated feature tensor.

[0115] In this embodiment, the weight generation subnetwork refers to a neural network submodule integrated within a deep learning model, consisting of several fully connected layers and nonlinear activation functions connected in series. Its main function is to automatically learn and infer the importance of each feature channel to the current task based on the input feature statistics. For example, this subnetwork can adopt the gating mechanism in the "compression-activation" architecture, using two fully connected layers to first reduce the dimensionality and then increase it, fitting a set of normalized weight parameters for feature recalibration.

[0116] Specifically, to comprehensively utilize feature information from different dimensions and improve the model's sensitivity to key features, the system concatenates the extracted temporal and spatial feature components along the feature channel dimension, thereby constructing a fusion feature tensor containing all spatiotemporal information. Specifically, the channel attention module uses a "compression-activation" topology to recalibrate the features: In the compression stage, global average pooling is first performed on the fusion feature tensor of dimension H×W×C, compressing the two-dimensional feature map of each channel into a single real value, generating a channel statistical vector of dimension 1×1×C. In the activation stage, this vector is sequentially input into the first fully connected layer (dimensionality reduction layer, number of nodes C / r, activation function ReLU) and the second fully connected layer (dimensionality increase layer, number of nodes C, activation function Sigmoid). The nonlinear transformation of the two fully connected layers is used to fit the correlation between channels, outputting a normalized weight vector of dimension 1×1×C, where each element value is between 0 and 1. Recalibration involves performing a channel-wise dot product between the weight vector and the original fused feature tensor, thereby numerically performing adaptive weighting of the feature channels, suppressing invalid feature responses and enhancing valid feature responses, ultimately obtaining the calibrated feature tensor after feature recalibration.

[0117] S206: Input the deep feature map into the fully connected layer in the convolutional neural network model for numerical mapping to obtain the settlement state probability vector of the road and bridge at the current moment.

[0118] For example, this step primarily undertakes the crucial task of bridging the gap between feature extraction and state decision-making. Leveraging the powerful feature integration and numerical mapping capabilities of the fully connected layer, the model can project and map the high-dimensional, abstract spatiotemporal feature maps extracted by the preceding convolutional layers onto a predefined settlement state category space. This process is essentially a quantitative assessment of the bridge's current operational status. By transforming complex feature patterns into confidence values ​​belonging to different health levels (such as normal, slight settlement, severe settlement, etc.), a logical closed loop from underlying data features to upper-level physical state determination is achieved, thus providing a standardized probabilistic basis for subsequent anomaly detection and early warning.

[0119] In one possible implementation, the deep feature map is input into the fully connected layer of the convolutional neural network model for numerical mapping to obtain the settlement state probability vector of the road and bridge at the current moment. Specifically, this includes: using a flattening function to reshape the multidimensional deep feature map into a one-dimensional feature vector; performing matrix multiplication on the one-dimensional feature vector with a preset weight matrix in the fully connected layer, and superimposing a preset bias vector to obtain a logical numerical sequence; inputting the logical numerical sequence into a preset probability normalization function to calculate the probability distribution value of each settlement state category, thereby generating a settlement state probability vector.

[0120] In this embodiment, the settlement state probability vector refers to a set of normalized values ​​output by a deep learning model after comprehensive reasoning and analysis of the input spatiotemporal features. This set of values ​​is used to quantify the probability that the bridge structure is in various preset settlement levels (such as healthy, first-level warning, second-level warning, etc.) at the current moment. For example, if the system presets three health state categories, the vector can be represented as [0.05, 0.92, 0.03], corresponding to the confidence levels of the three states, respectively.

[0121] Specifically, to transform the high-dimensional abstract features extracted by the convolutional layer into intuitive classification results, the system calls a flattening function to perform dimensionality reduction and rearrangement on the multi-dimensional deep feature map (usually a three-dimensional tensor) output by the previous layer. This stretches and unfolds the data elements in both spatial and channel dimensions, reshaping them into a one-dimensional feature vector. This vector is then input into the model's classification layer, where the one-dimensional feature vector is multiplied by a preset weight matrix in the fully connected layer. This integrates global features through weighted combination, and a preset bias vector is superimposed to adjust the activation threshold, thereby calculating a set of unnormalized raw scores, i.e., a logistic numerical sequence. To ensure the model output has clear statistical significance, the logistic numerical sequence is input into a preset probability normalization function (usually the Softmax function). Through exponential operations and normalization, the raw scores are mapped to a probability distribution with a sum of 1, thereby calculating the probability distribution values ​​for each settlement state category, and finally generating a settlement state probability vector.

[0122] S207: Calculate the numerical deviation between the settlement state probability vector and the preset baseline health state vector. If the numerical deviation exceeds the preset safety threshold, generate a settlement early warning command.

[0123] In this embodiment, the baseline health state vector refers to a standardized probability distribution reference that is predefined and stored in the system to characterize whether the bridge structure is in an ideal and intact state or a state with zero settlement risk. For example, if the system divides the settlement monitoring status into three categories: "healthy", "Level 1 warning" and "Level 2 warning", then the vector is usually set as a one-hot encoded vector corresponding to a "healthy" state with a probability of 100% and other states with a probability of 0, such as [1, 0, 0].

[0124] Specifically, to quantitatively assess the degree to which the current bridge structure deviates from its ideal healthy state and to promptly identify potential risks, the system obtains the settlement state probability vector output from previous steps and simultaneously reads a preset baseline healthy state vector from memory. A vector space distance calculation algorithm (such as the Euclidean distance formula or the Kullback-Leibler divergence algorithm) is used to calculate the distance or difference between these two vectors in the multidimensional feature space, thereby calculating the numerical deviation between the settlement state probability vector and the preset baseline healthy state vector. This calculated numerical deviation is then compared with a preset safety threshold (usually based on the statistical distribution of historical monitoring data or expert experience). If the numerical deviation exceeds the preset safety threshold, the settlement state of the current monitoring point is determined to have exceeded the normal fluctuation range, triggering an alarm logic. This involves constructing a data packet containing information such as the time of the anomaly, the anomaly monitoring point number, and the degree of deviation, ultimately generating a settlement warning command.

[0125] Furthermore, considering the changes in the physical characteristics of the bridge structure throughout its entire life cycle, the preset baseline health state vector possesses dynamic adaptability. The system includes the following baseline update mechanism: Considering that factors such as concrete shrinkage and creep, and bearing rubber aging can cause irreversible, slow changes in the bridge's baseline position (long-term trend), a static baseline vector may lead to an increase in false alarm rate over time. Therefore, the system has a baseline dynamic drift calibration function. The system extracts historical monitoring data from periods deemed healthy on a monthly or quarterly basis for statistical analysis, calculates a new statistical distribution center, and fine-tunes the baseline vector parameters accordingly. Specifically, the system's judgment logic for performing baseline dynamic drift calibration is as follows: The system maintains a historical deviation dataset of length N days (e.g., N=30). The deviation value D between the settlement state probability vector of the current day and the current baseline vector is calculated daily. t Calculate the set D of deviation values ​​for the most recent N days. t-N , ..., D t variance S 2 If S 2 Less than the preset stability threshold ε stable (This indicates that the data fluctuations are minimal and there are no sudden anomalies), and all D t All values ​​are less than the alarm threshold; calculate the arithmetic mean D of this set. If D is greater than the preset drift trigger threshold ε... drift (This indicates the existence of a non-zero constant offset); when the above conditions are met, the system does not directly replace the benchmark, but instead updates the benchmark vector using the exponential moving average (EMA) algorithm: V new =α·V currentavg +(1-α)·V old , where α is the update coefficient (e.g., 0.1), to achieve a smooth transition of the reference vector.

[0126] Figure 3 This is a schematic diagram of a bridge monitoring system based on deep learning, as described in an embodiment of this application. This system can be implemented through software, hardware, or a combination of both, forming all or part of the overall system. Figure 3 As shown, the system includes:

[0127] The data acquisition module 301 is configured to collect the original strain monitoring data sequence of each monitoring point on the road and bridge, and to collect the environmental parameter data sequence of the area where each monitoring point is located.

[0128] The data processing module 302 is configured to calculate the theoretical thermal strain numerical sequence of each monitoring point at different sampling times based on the environmental parameter data sequence and the preset material thermal expansion coefficient, and to align the original strain monitoring data sequence with the theoretical thermal strain numerical sequence in time and subtract them one by one to obtain the corrected structural strain sequence.

[0129] The matrix construction module 303 is configured to arrange the corrected structural strain sequence according to the monitoring point location and acquisition time order to construct a spatiotemporal strain matrix within a preset time window;

[0130] The data transmission module 304 is configured to transmit the spatiotemporal strain matrix to the cloud analysis platform and input the spatiotemporal strain matrix into a convolutional neural network model deployed on the cloud analysis platform.

[0131] The feature extraction module 305 is configured to perform convolution operations on the spatiotemporal strain matrix through multiple convolutional kernels in the convolutional neural network model to generate a deep feature map. The multiple convolutional kernels include temporal convolutional kernels and spatial convolutional kernels.

[0132] The state analysis module 306 is configured to input deep feature maps into the fully connected layer of the convolutional neural network model for numerical mapping to obtain the settlement state probability vector of the road and bridge at the current moment.

[0133] The settlement early warning module 307 is configured to calculate the numerical deviation between the settlement state probability vector and the preset benchmark health state vector. If the numerical deviation exceeds the preset safety threshold, a settlement early warning command is generated.

[0134] Based on the above embodiments, as an optional embodiment, the data processing module 302 is specifically used for: traversing the environmental parameter data sequence through a sliding time window of a preset length, and determining the sampling time located at the center of the sliding time window as the target sampling time; based on a preset weight decay function, calculating the time distance weight corresponding to each environmental temperature value according to the time interval between the sampling time of each environmental temperature value contained in the sliding time window and the target sampling time, wherein the time distance weight is negatively correlated with the time interval; using the time distance weight to perform a weighted average calculation on each environmental temperature value contained in the sliding time window to obtain the effective temperature value corresponding to the target sampling time, and determining the sequence composed of multiple effective temperature values ​​as the 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 a preset reference temperature value from each temperature value in the aligned temperature sequence to obtain a temperature difference sequence; and multiplying each value in the temperature difference sequence by the material's thermal expansion coefficient to obtain a theoretical thermal strain value sequence.

[0135] Based on the above embodiments, as an optional embodiment, the matrix construction module 303 is specifically used for: performing a sliding slice operation on the corrected structural strain sequence according to a preset time step and a preset window length, and calculating the difference sequence of the corrected structural strain sequence on the time axis after the slice operation to obtain a strain rate subsequence; calculating the correlation coefficient values ​​between the target monitoring point and each adjacent monitoring point based on a preset correlation calculation model, and comparing the correlation coefficient values ​​with a preset spatial correlation threshold to determine the set of effective monitoring points with correlation coefficient values ​​greater than the spatial correlation threshold; using the strain rate subsequence corresponding to the set of effective monitoring points as the first feature channel, using the corrected structural strain sequence corresponding to the set of effective monitoring points as the second feature channel, and combining the first feature channel and the second feature channel to obtain a multi-channel matrix; and normalizing the multi-channel matrix to generate a spatiotemporal strain matrix.

[0136] Based on the above embodiments, as an optional embodiment, the matrix construction module 303 is specifically used for: extracting a first numerical sequence of the target monitoring point within a preset time window from the corrected structural strain sequence, and extracting a second numerical sequence of adjacent monitoring points within the preset time window; setting multiple discrete time lags within a preset phase offset search range, and performing position offset processing on the data elements in the second numerical sequence according to each time lag, generating multiple sets of candidate test sequences corresponding to different time lags; calculating the standard cross-correlation values ​​between the first numerical sequence and each candidate test sequence, obtaining a cross-correlation value set; selecting 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.

[0137] Based on the above embodiments, as an optional embodiment, the feature extraction module 305 is specifically used to: perform time dimension feature extraction on the spatiotemporal strain matrix using a time-dimensional convolution kernel to obtain time feature components; perform spatial dimension feature extraction on the spatiotemporal strain matrix using a spatial-dimensional convolution kernel to obtain spatial feature components; perform feature fusion and adaptive calibration on the time feature components and spatial feature components to obtain a calibrated feature tensor; and perform dimensionality reduction processing on the calibrated feature tensor through the max pooling layer in the convolutional neural network model to generate a deep feature map.

[0138] Based on the above embodiments, as an optional embodiment, the feature extraction module 305 is specifically used for: concatenating the temporal feature components and spatial feature components into channels to construct a fused feature tensor; performing global average pooling on the fused feature tensor to obtain a one-dimensional weight vector describing the statistical information between feature channels; inputting the one-dimensional weight vector into a preset weight generation sub-network to calculate the weight coefficient sequence corresponding to each feature channel, wherein the weight generation sub-network is constructed by multiple fully connected layers; and multiplying each weight coefficient in the weight coefficient sequence with the corresponding feature channel in the fused feature tensor to obtain the calibrated feature tensor.

[0139] Based on the above embodiments, as an optional embodiment, the state analysis module 306 is specifically used to: reshape the multidimensional deep feature map into a one-dimensional feature vector using a flattening function; perform matrix multiplication operation between the one-dimensional feature vector and the preset weight matrix in the fully connected layer, and superimpose a preset bias vector to obtain a logical value sequence; input the logical value sequence into a preset probability normalization function to calculate the probability distribution value of each settlement state category and generate a settlement state probability vector.

[0140] It should be noted that the system provided in the above embodiments is only illustrated by the division of the above functional modules. In actual applications, the above functions can be assigned to different functional modules as needed, that is, the internal structure of the device can be divided into different functional modules to complete all or part of the functions described above. In addition, the system and method embodiments provided in the above embodiments belong to the same concept, and the specific implementation process can be found in the method embodiments, which will not be repeated here.

[0141] This embodiment also discloses an electronic device, as shown in the reference. Figure 4 The electronic device may include: at least one processor 401, at least one communication bus 402, user interface 403, network interface 404, and at least one memory 405.

[0142] The communication bus 402 is used to enable communication between these components.

[0143] The user interface 403 may include a display screen and a camera. Optionally, the user interface 403 may also include a standard wired interface and a wireless interface.

[0144] The network interface 404 may optionally include a standard wired interface or a wireless interface (such as a Wi-Fi interface).

[0145] The processor 401 may include one or more processing cores. The processor 401 connects to various parts of the server using various interfaces and lines, and performs various server functions and processes data by running or executing instructions, programs, code sets, or instruction sets stored in memory 405, and by calling data stored in memory 405. Optionally, the processor 401 may be implemented using at least one hardware form of Digital Signal Processing (DSP), Field-Programmable Gate Array (FPGA), or Programmable Logic Array (PLA). The processor 401 may integrate one or a combination of several of the following: Central Processing Unit (CPU), Graphics Processing Unit (GPU), and modem. The CPU primarily handles the operating system, user interface, and applications; the GPU is responsible for rendering and drawing the content required for display; and the modem handles wireless communication. It is understood that the modem may also be implemented as a separate chip without being integrated into the processor 401.

[0146] The memory 405 may include random access memory (RAM) or read-only memory. Optionally, the memory 405 may include a non-transitory computer-readable storage medium. The memory 405 may be used to store instructions, programs, code, code sets, or instruction sets. The memory 405 may include a program storage area and a data storage area, wherein the program storage area may store instructions for implementing an operating system, instructions for at least one function (such as touch function, sound playback function, image playback function, etc.), instructions for implementing the above-described method embodiments, etc.; the data storage area may store data involved in the above-described method embodiments, etc. Optionally, the memory 405 may also be at least one storage device located remotely from the aforementioned processor 401. Figure 4 As shown, the memory 405, which serves as a computer storage medium, may include an operating system, a network communication module, a user interface module, and an application program for a bridge monitoring method based on deep learning.

[0147] exist Figure 4 In the electronic device shown, the user interface 403 is mainly used to provide an input interface for the user and to obtain the user input data; while the processor 401 can be used to call an application program stored in the memory 405 that is a bridge monitoring method based on deep learning. When executed by one or more processors 401, the electronic device executes one or more methods as described in the above embodiments.

[0148] It should be noted that, for the sake of simplicity, the foregoing method embodiments are all described as a series of actions. However, those skilled in the art should understand that this application is not limited to the described order of actions, as some steps may be performed in other orders or simultaneously according to this application. Furthermore, those skilled in the art should also understand that the embodiments described in the specification are preferred embodiments, and the actions and modules involved are not necessarily essential to this application.

[0149] In the above embodiments, the descriptions of each embodiment have different focuses. For parts not described in detail in a certain embodiment, please refer to the relevant descriptions in other embodiments.

[0150] In the several embodiments provided in this application, it should be understood that the disclosed apparatus can be implemented in other ways. For example, the apparatus embodiments described above are merely illustrative; for instance, the division of units is only a logical functional division, and in actual implementation, there may be other division methods. For example, multiple units or components may be combined or integrated into another system, or some features may be ignored or not executed. Furthermore, the shown or discussed mutual couplings or direct couplings or communication connections may be through some service interfaces; indirect couplings or communication connections between apparatuses or units may be electrical or other forms.

[0151] The units described as separate components may or may not be physically separate. The components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the units can be selected to achieve the purpose of this embodiment according to actual needs.

[0152] Furthermore, the functional units in the various embodiments of this application can be integrated into one processing unit, or each unit can exist physically separately, or two or more units can be integrated into one unit. The integrated unit can be implemented in hardware or as a software functional unit.

[0153] If the integrated unit is implemented as a software functional unit and sold or used as an independent product, it can be stored in a computer-readable storage device (CMD). Based on this understanding, the technical solution of this application, in essence, or the part that contributes to the prior art, or all or part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a memory 405 and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods of the various embodiments of this application. The aforementioned memory 405 includes various media capable of storing program code, such as a USB flash drive, external hard drive, magnetic disk, or optical disk.

[0154] The foregoing description is merely an exemplary embodiment of this disclosure and should not be construed as limiting the scope of this disclosure. Any equivalent changes and modifications made in accordance with the teachings of this disclosure shall still fall within the scope of this disclosure. Those skilled in the art will readily conceive of other embodiments of this disclosure upon considering the disclosure in this specification. This application is intended to cover any variations, uses, or adaptations of this disclosure that follow the general principles of this disclosure and include common knowledge or customary techniques in the art not described in this disclosure. The specification and embodiments are considered exemplary only, and the scope of this application is defined by the claims.

Claims

1. A bridge monitoring method based on deep learning, characterized in that, The method includes: The original strain monitoring data sequence of each monitoring point on the road and bridge was collected, and the environmental parameter data sequence of the area where each monitoring point is located was also collected. Based on the environmental parameter data sequence and the preset material thermal expansion coefficient, the theoretical thermal strain numerical sequence of each monitoring point at different sampling times is calculated, and the original strain monitoring data sequence and the theoretical thermal strain numerical sequence are time-aligned and subtracted one by one to obtain the corrected structural strain sequence. The corrected structural strain sequence is arranged according to the location of the monitoring points and the order of acquisition time to construct a spatiotemporal strain matrix within a preset time window; The spatiotemporal strain matrix is ​​transmitted to the cloud analysis platform, and the spatiotemporal strain matrix is ​​input into the convolutional neural network model deployed on the cloud analysis platform; The spatiotemporal strain matrix is ​​convolved using multiple convolutional kernels in the convolutional neural network model to generate a deep feature map. The multiple convolutional kernels include temporal convolutional kernels and spatial convolutional kernels. The deep feature map is input into the fully connected layer of the convolutional neural network model for numerical mapping to obtain the settlement state probability vector of the road and bridge at the current time. Calculate the numerical deviation between the settlement state probability vector and the preset baseline health state vector. If the numerical deviation exceeds a preset safety threshold, generate a settlement early warning command.

2. The method according to claim 1, characterized in that, The step of calculating the theoretical thermal strain value sequence of each monitoring point at different sampling times based on the environmental parameter data sequence and the preset material thermal expansion coefficient specifically includes: The environmental parameter data sequence is traversed by a sliding time window of a preset length, and the sampling time located at the center of the sliding time window is determined as the target sampling time. Based on a preset weight decay function, the time distance weight corresponding to each ambient temperature value is calculated according to the time interval between the sampling time of each ambient temperature value included in the sliding time window and the target sampling time. The time distance weight is negatively correlated with the time interval. The time distance weight is used to calculate the weighted average of each ambient temperature value contained in the sliding time window to obtain the effective temperature value corresponding to the target sampling time, and the sequence composed of multiple effective temperature values ​​is determined as the effective temperature sequence. The effective temperature sequence is resampled using a preset interpolation algorithm to generate an aligned temperature sequence corresponding to the timestamp of the original strain monitoring data sequence. Subtract a preset reference temperature value from each temperature value in the aligned temperature sequence to obtain a temperature difference sequence. The theoretical thermal strain numerical sequence is obtained by multiplying each value in the temperature difference sequence by the thermal expansion coefficient of the material.

3. The method according to claim 1, characterized in that, The step of arranging the corrected structural strain sequence according to the monitoring point location and acquisition time sequence to construct a spatiotemporal strain matrix within a preset time window specifically includes: A sliding slice operation is performed on the modified structural strain sequence according to a preset time step and a preset window length, and the difference sequence of the modified structural strain sequence on the time axis after the slice operation is calculated to obtain the strain rate subsequence. Based on a preset correlation calculation model, the correlation coefficient between the target monitoring point and each adjacent monitoring point is calculated, and the correlation coefficient is compared with a preset spatial correlation threshold to determine the set of valid monitoring points whose correlation coefficient is greater than the spatial correlation threshold. The strain rate subsequence corresponding to the effective monitoring point set is used as the first feature channel, and the corrected structural strain sequence corresponding to the effective monitoring point set is used as the second feature channel. The first feature channel and the second feature channel are combined to obtain a multi-channel matrix. The multi-channel matrix is ​​normalized to generate the spatiotemporal strain matrix.

4. The method according to claim 3, characterized in that, The correlation coefficient between the target monitoring point and each adjacent monitoring point is calculated based on a preset correlation calculation model, specifically including: Extract the first numerical sequence of the target monitoring point within the preset time window from the corrected structural strain sequence, and extract the second numerical sequence of the adjacent monitoring point within the preset time window; Within a preset phase offset search range, multiple discrete time lags are set, and the data elements in the second numerical sequence are shifted according to each of the time lags to generate multiple sets of candidate test sequences corresponding to different time lags. Calculate the standard cross-correlation values ​​between the first numerical sequence and each of the candidate test sequences to obtain a set of cross-correlation values; The value with the largest absolute value is selected from the set of cross-correlation values, and the value with the largest absolute value is determined as the correlation coefficient value.

5. The method according to claim 1, characterized in that, The step of performing convolution operations on the spatiotemporal strain matrix using multiple convolutional kernels in the convolutional neural network model to generate a deep feature map specifically includes: The time-dimensional convolution kernel is used to perform time-dimensional feature extraction on the spatiotemporal strain matrix to obtain time feature components; The spatial dimension features of the spatiotemporal strain matrix are extracted using the spatial dimension convolution kernel to obtain spatial feature components. The temporal feature components and the spatial feature components are fused and adaptively calibrated to obtain the calibrated feature tensor. The dimensionality of the calibrated feature tensor is reduced by the max pooling layer in the convolutional neural network model to generate the deep feature map.

6. The method according to claim 5, characterized in that, The step of fusing and adaptively calibrating the temporal and spatial feature components to obtain a calibrated feature tensor specifically includes: The temporal feature components and the spatial feature components are concatenated by channels to construct a fused feature tensor. Global average pooling is performed on the fused feature tensor to obtain a one-dimensional weight vector describing the statistical information between feature channels; The one-dimensional weight vector is input into a preset weight generation sub-network to calculate the weight coefficient sequence corresponding to each feature channel, wherein the weight generation sub-network is constructed from multiple fully connected layers. Each weight coefficient in the weight coefficient sequence is multiplied by the corresponding feature channel in the fused feature tensor to obtain the calibrated feature tensor.

7. The method according to claim 1, characterized in that, The step of inputting the deep feature map into the fully connected layer of the convolutional neural network model for numerical mapping to obtain the settlement state probability vector of the road and bridge at the current moment specifically includes: The multidimensional deep feature map is reshaped into a one-dimensional feature vector using a flattening function; The one-dimensional feature vector is multiplied by a preset weight matrix in the fully connected layer, and a preset bias vector is superimposed to obtain a logical numerical sequence. The logical numerical sequence is input into a preset probability normalization function to calculate the probability distribution value of each settlement state category, thereby generating the settlement state probability vector.

8. A bridge monitoring system based on deep learning, characterized in that the system... include: The data acquisition module is configured to collect the original strain monitoring data sequence of each monitoring point on the road and bridge, and to collect the environmental parameter data sequence of the area where each monitoring point is located; The data processing module is configured to calculate the theoretical thermal strain numerical sequence of each monitoring point at different sampling times based on the environmental parameter data sequence and the preset material thermal expansion coefficient, and to time-align the original strain monitoring data sequence with the theoretical thermal strain numerical sequence and subtract them one by one to obtain the corrected structural strain sequence. The matrix construction module is configured to arrange the corrected structural strain sequence according to the monitoring point location and acquisition time order to construct a spatiotemporal strain matrix within a preset time window; The data transmission module is configured to transmit the spatiotemporal strain matrix to a cloud analysis platform and input the spatiotemporal strain matrix into a convolutional neural network model deployed on the cloud analysis platform. The feature extraction module is configured to perform convolution operations on the spatiotemporal strain matrix through multiple convolutional kernels in the convolutional neural network model to generate a deep feature map. The multiple convolutional kernels include temporal convolutional kernels and spatial convolutional kernels. The state analysis module is configured to input the deep feature map into the fully connected layer in the convolutional neural network model for numerical mapping to obtain the settlement state probability vector of the road and bridge at the current moment. The settlement early warning module is configured to calculate the numerical deviation between the settlement state probability vector and the preset benchmark health state vector. If the numerical deviation exceeds a preset safety threshold, a settlement early warning command is generated.

9. An electronic device, characterized in that, The device includes a processor, a memory, a user interface, and a network interface. The memory is used to store instructions. The user interface and the network interface are both used to communicate with other devices. The processor is used to execute the instructions stored in the memory to cause the electronic device to perform the method as described in any one of claims 1-7.

10. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores instructions that, when executed, perform the method as described in any one of claims 1-7.

Images