A highway long-period settlement monitoring method and system based on multi-source data fusion

Abstract

The application provides a highway long-period settlement monitoring method and system based on multi-source data fusion, and belongs to the technical field of highway deformation monitoring. The method comprises the following steps: calculating original observation signals, correcting errors of the calculation results, and obtaining deformation observation signals; performing wavelet decomposition on the deformation observation signals, adaptively calculating a noise reduction threshold by using a fatigue damage factor, and generating deformation data; obtaining road inspection data, performing noise reduction processing and abnormality identification on the road inspection data, and obtaining inspection results; taking the deformation data as a reference, performing time and space alignment on the inspection results, and generating a multi-source heterogeneous fusion data set; inputting the multi-source heterogeneous fusion data set and environmental load parameters into a deep time series prediction model, predicting a future long-period settlement trend, and generating a maintenance scheme according to the prediction result.

Description

Technical Field

[0001] This invention belongs to the field of highway deformation monitoring technology, specifically relating to a method and system for long-term highway settlement monitoring based on multi-source data fusion. Background Technology

[0002] During long-term service, highways experience continuous settlement and deformation due to the combined effects of multiple factors, including foundation consolidation, traffic loads, and changes in the geological environment. To ensure operational safety and scientifically formulate maintenance plans, it is usually necessary to comprehensively utilize various spatial sensing technologies, such as satellite positioning measurement and low-altitude drone inspections, to periodically monitor the deformation status and apparent defects of the entire road.

[0003] In existing technologies, positioning monitoring methods based on the Global Navigation Satellite System can provide three-dimensional deformation observation data with a certain accuracy, while UAV inspection methods can simultaneously acquire road surface images and point cloud information to identify surface defects such as cracks and potholes. However, due to the inherent differences between different sensing methods in terms of data acquisition mechanisms, sampling frequencies, and spatial resolution, multi-source monitoring data are often processed and analyzed independently, lacking an effective mechanism for heterogeneous data fusion and collaborative utilization.

[0004] How to effectively integrate multi-source heterogeneous sensing data during long-term monitoring, and on this basis improve the accuracy of settlement trend prediction and the scientific nature of maintenance decisions, is a technical problem that urgently needs to be solved in this field. Summary of the Invention

[0005] This invention overcomes the shortcomings of the prior art and provides a method and system for monitoring long-term settlement of highways based on multi-source data fusion.

[0006] To achieve the above objectives, the technical solution adopted by this invention is: a method for long-term settlement monitoring of highways based on multi-source data fusion, comprising the following steps: S1: Solve the original observation signal, correct the error in the solution result, and obtain the deformation observation signal; S2: Perform wavelet decomposition on the deformation observation signal, adaptively calculate the noise reduction threshold using the fatigue damage factor, and generate deformation data; S3: Acquire road inspection data, perform noise reduction and anomaly identification on the road inspection data, and obtain inspection results; S4: Based on deformation data, the inspection results are spatiotemporally aligned to generate a multi-source heterogeneous fusion dataset; S5: Input the multi-source heterogeneous fusion dataset and environmental load parameters into the deep time series prediction model to predict the future long-term settlement trend, and generate a maintenance plan based on the prediction results.

[0007] Preferably, in step S2, the adaptive calculation of the noise reduction threshold using the fatigue damage factor includes obtaining the fatigue damage degree of the pavement structure corresponding to the monitoring point. fatigue damage The calculation formula is ,in The damage coefficient is related to the load stress level. This represents the actual cumulative number of standard axle loads applied to the road surface corresponding to each monitoring point since it opened to traffic. The standard axle load cycles represent the fatigue life of the pavement structure.

[0008] Preferably, in step S2, the noise reduction threshold is adaptively calculated. In order to target the The high-frequency detail components of the layer are calculated using the following formula: ,in This is the global threshold adjustment factor. For the first Noise standard deviation estimation for high-frequency detail components of the layer. The length of the deformation observation signal, This is the damage correction factor. This refers to the fatigue damage degree of the pavement structure.

[0009] Preferably, the damage correction coefficient γ ranges from 0.2 to 0.4.

[0010] Preferably, in step S2, generating deformation data includes using a soft thresholding function to shrink wavelet coefficients in each layer of high-frequency detail components whose amplitudes are lower than the corresponding noise reduction threshold, and then reconstructing them with low-frequency approximate components.

[0011] Preferably, in step S3, the anomaly identification of road inspection data includes using a lightweight target detection model that has been pruned and quantized. The backbone network of the lightweight target detection model is MobileNetV3, and depthwise separable convolution is used to replace standard convolution for real-time inference and identification of abnormal targets on the road surface in edge computing units.

[0012] Preferably, in step S3, the noise reduction processing of the road inspection data includes performing block matching three-dimensional collaborative filtering on the multispectral image and setting a noise reduction threshold. The calculation formula is ,in For estimating the standard deviation of noise in multispectral images, is the total number of pixels within the image patch, and is a correction coefficient based on the difference in road surface structure layer thickness. For asphalt mixtures during service life The measured creep modulus is below. This is the initial elastic modulus of the asphalt mixture.

[0013] Preferably, in step S4, the spatiotemporal alignment of the inspection results based on the deformation data includes using the deformation data as the baseline state and the inspection results as the observation values, employing the Kalman filter algorithm, defining the state vector as the cumulative settlement and settlement rate of the monitoring points, and the observation vector as the plane coordinates, elevation values, and outline dimensions of the disease location in the inspection results.

[0014] Preferably, in step S5, the deep time series prediction model is a long short-term memory recurrent neural network, including two hidden layers, each with 128 hidden units, the in-layer activation function is tanh, the recurrent activation function is sigmoid, the output layer is a fully connected linear layer, and the output is the predicted monthly average settlement rate for the next 3 to 12 months.

[0015] Another technical solution provided by the present invention: a highway long-term settlement monitoring system based on multi-source data fusion, used to implement the above method, including a Beidou reference station network, a UAV inspection unit and a cloud-based collaborative analysis platform, wherein the Beidou reference station network, the UAV inspection unit and the cloud-based collaborative analysis platform are connected through a communication network; The BeiDou reference station network is used to solve the original observation signals, correct the errors in the solution results, obtain the deformation observation signals, perform wavelet decomposition on the deformation observation signals, adaptively calculate the noise reduction threshold using the fatigue damage factor, and generate deformation data. The drone inspection unit is used to acquire road inspection data, perform noise reduction and anomaly identification on the road inspection data, and obtain inspection results. The cloud-based collaborative analysis platform is used to perform spatiotemporal alignment of inspection results based on deformation data, generate a multi-source heterogeneous fusion dataset, and input the multi-source heterogeneous fusion dataset and environmental load parameters into a deep time series prediction model to predict future long-term settlement trends. Based on the prediction results, a maintenance plan is generated.

[0016] This invention addresses the deficiencies in the prior art and has the following beneficial effects: By performing multi-scale discrete wavelet decomposition on the deformation observation signal and using the fatigue damage degree of the concrete pavement corresponding to the monitoring point as the core control variable, the noise reduction threshold of the high-frequency detail components of each layer is adaptively calculated. When the fatigue damage degree of the pavement structure increases under repeated axle loads, and the high-frequency micro-deformation caused by the propagation of internal micro-cracks is highly mixed with random noise in the frequency domain, the adaptive threshold increases synchronously with the increase of damage degree, reducing the shrinkage intensity of the high-frequency detail coefficients. Thus, while suppressing random noise, the low-amplitude high-frequency effective coefficients characterizing the structural state change are completely preserved. Through the deformation data obtained in step S2, the signal-to-noise ratio and the response sensitivity to the structural damage state are directly improved, which can more realistically reflect the deformation characteristics associated with the fatigue damage of the pavement structure and avoid the irreversible loss of effective information caused by the inability of fixed threshold filtering to distinguish between discrete noise and damage-related signals.

[0017] For multispectral images, block-matching 3D collaborative filtering is performed. The noise reduction threshold is dynamically modulated by introducing a creep modulus correction term, which characterizes the viscoelastic state of the asphalt mixture, making the noise reduction threshold an increasing function of the creep modulus. When the creep modulus of the asphalt mixture increases and the material becomes more brittle due to long-term aging, and the high-frequency details of the edges generated by crack initiation highly overlap with the spatial frequency of noise in the image, the image noise reduction threshold is adaptively increased, effectively preserving the detailed information of the crack edges. Simultaneously, for the 3D point cloud of the road surface, structure-aware Gaussian smoothing filtering is performed. The kernel function weight is dynamically adjusted by introducing a porosity modulation term, which characterizes the compaction state of the subgrade soil, making the smoothing weight an increasing function of the porosity. When the subgrade porosity is high in soft soil sections, and the micro-deformation features of uneven settlement caused by consolidation compression highly overlap with the spatial scale of noise in the point cloud, the point cloud smoothing intensity is adaptively reduced, effectively preserving the continuous elevation change features of the corresponding settlement micro-deformation. The two-dimensional adaptive noise reduction process in step S3 enables the inspection results to achieve noise smoothing while significantly improving the positioning accuracy and identification accuracy of road crack edges, and fully preserving the continuous micro-deformation features related to roadbed settlement.

[0018] Using high-precision deformation data acquired from a reference station network as the baseline state and inspection results obtained from UAV inspections as noisy observations, a state-space model is constructed and a recursive state estimation algorithm is run to optimally correct the spatial positioning deviation of the inspection results. This allows the prior high confidence represented by the extremely low error covariance of the deformation data to be recursively fused with the measurement information of the inspection results under the minimum mean square error criterion. Each iteration minimizes the posterior error covariance, suppressing the transmission and amplification of measurement noise. The corrected inspection results and deformation data achieve a one-to-one spatial correspondence and temporal synchronization under a unified spatiotemporal reference, eliminating the systematic correspondence deviation between settlement time series and the spatial location of defects caused by positioning drift.

[0019] A spatiotemporally aligned multi-source heterogeneous fusion dataset, along with environmental load parameters such as temperature, traffic flow, and groundwater level, is used as input features and imported into a Long Short-Term Memory (LSTM) recurrent neural network (RNN) for long-term settlement trend prediction. The LSM RNN utilizes the nonlinear mapping capability of a gating structure to the long-term temporal dependence of multiple factors. It selectively discards short-term random fluctuations through a forgetting gate and captures the lagged effects of loads and environmental factors on settlement through an input gate, storing these effects in the cell state for long-term retention. This allows it to learn the conditional probability distribution of the settlement process under the coupling of multiple factors. Based on the prediction results, settlement risk classification early warning signals and maintenance plans are generated, significantly extending the effective prediction period. Furthermore, the selection of maintenance plans relies on settlement evolution simulation and cost-benefit analysis completed using a digital twin model, enabling precise on-demand allocation of maintenance resources and quantitative support for maintenance decisions. Attached Figure Description

[0020] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments recorded in the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort. Figure 1 This is a flowchart of a long-term highway settlement monitoring method based on multi-source data fusion; Figure 2 This is an architecture diagram of a highway long-term settlement monitoring system based on multi-source data fusion. Detailed Implementation

[0021] Many specific details are set forth in the following description in order to provide a full understanding of the invention. However, the invention may also be practiced in other ways different from those described herein. Therefore, the scope of protection of the invention is not limited to the specific embodiments disclosed below.

[0022] Application Overview: In deformation observation signals, high-frequency minute deformations and random noise are highly mixed in the frequency domain. Fixed threshold denoising methods cannot distinguish between effective deformation information and noise, resulting in the irreversible filtering out of weak signals related to structural fatigue damage. At the same time, the processing of images and point cloud data in road inspection does not take into account the viscoelastic evolution of asphalt materials and the spatial heterogeneity of the subgrade soil structure. General denoising algorithms tend to treat high-frequency details at crack edges and settlement micro-deformation features as noise smoothing, resulting in missed detection of defects and loss of deformation information.

[0023] Based on the above, the present invention proposes that after obtaining deformation observation signals, wavelet decomposition is performed, a fatigue damage factor is used to adaptively calculate a noise reduction threshold to generate deformation data, material-perceptive noise reduction processing and anomaly identification are performed on road inspection data to obtain inspection results, spatiotemporal alignment is performed on the inspection results based on the deformation data to generate a multi-source heterogeneous fusion data set, then the multi-source heterogeneous fusion data set and environmental load parameters are input into a deep time series prediction model to predict the long-term settlement trend and generate a maintenance scheme.

[0024] Exemplary method: As Figure 1 shown, a long-period settlement monitoring method for highways based on multi-source data fusion comprises the following steps: S1: solving an original observation signal, performing error correction on a solving result, and acquiring a deformation observation signal; S2: performing wavelet decomposition on the deformation observation signal, adaptively calculating a noise reduction threshold by using a fatigue damage factor, and generating deformation data; S3: acquiring pavement inspection data, performing noise reduction processing and anomaly identification on the pavement inspection data, and obtaining inspection results; S4: performing spatiotemporal alignment on the inspection results based on the deformation data, and generating a multi-source heterogeneous fusion data set; S5: inputting the multi-source heterogeneous fusion data set and environmental load parameters into a deep time series prediction model, predicting a future long-period settlement trend, and generating a maintenance scheme according to a prediction result.

[0025] Hereinafter, each step will be described in detail.

[0026] In step S1, original observation signals are collected through a reference station network, and real-time dynamic carrier phase difference technology is adopted to complete solving and error correction, so as to output deformation observation signals. Wherein, the real-time dynamic carrier phase difference technology performs differential processing on synchronous observation values of a reference station and a rover station to obtain a three-dimensional positioning result in real time.

[0027] Specifically, along the trend of the highway to be monitored, continuous monitoring sections are divided at fixed intervals, positions with unobstructed vision and stable geological conditions in each monitoring section are selected to deploy dual-frequency global navigation satellite system receiving terminals, and all the dual-frequency global navigation satellite system receiving terminals jointly form a reference station network.

[0028] Illustratively, the division interval of the monitoring sections is 6km to 7km, each dual-frequency global navigation satellite system receiving terminal supports carrier phase observation of at least two frequencies, the sampling frequency is set to 1Hz to 10Hz, an observation pier matched with the dual-frequency global navigation satellite system receiving terminal is arranged in undisturbed soil, the buried depth is not less than 1.2m, and the skywards visible elevation angle of the matched measuring antenna is not less than 10°.

[0029] Furthermore, the dual-frequency global navigation satellite system receiving terminal within the reference station network continuously and synchronously acquires carrier phase observations and pseudorange observations broadcast by the on-orbit global navigation satellite system to obtain the raw observation signal. Using the known coordinates of the reference station network as a reference, real-time dynamic carrier phase differential technology is employed to perform differential processing on the synchronous observations between the reference station network and the monitoring points, thereby completing the solution of the raw observation signal.

[0030] For example, the solution update frequency is consistent with the sampling frequency of the dual-frequency global navigation satellite system receiver terminal, the differential processing uses carrier phase smoothing pseudorange observations, the integer ambiguity solution uses the least squares downcorrelation adjustment method, and the integer ambiguity fixation success rate threshold is set to 99.9%.

[0031] Furthermore, for the solution results, an ionospheric-free linear combination method is adopted. By weighted linear combination of the phase observations of two carriers at different frequencies, the delay error of the first-order term of the ionospheric term is offset, thus completing the error correction of the solution results.

[0032] For example, the dual-frequency observations use carrier phase observations from the L1 and L2 frequency bands, and the coefficients of the linear combination are determined based on the carrier frequencies of the two frequency bands.

[0033] Furthermore, after completing the calculation of the original observation signal and the error correction of the calculation result, the deformation observation signal is output.

[0034] For example, the deformation observation signal includes a three-dimensional coordinate time series of the monitoring point, the observation residual, and the fixed state identifier of the solution. The time series resolution of the deformation observation signal is consistent with the sampling frequency of the dual-frequency global navigation satellite system receiving terminal.

[0035] In practice, monitoring sections are divided into intervals of 6km to 7km and a network of reference stations is deployed. Dual-frequency carrier phase and pseudorange observations are collected at sampling frequencies of 1Hz to 10Hz. Differential calculation and error correction are completed on the cloud platform to generate deformation observation signals containing three-dimensional coordinate time series, observation residuals and fixed state identifiers.

[0036] Step S2 achieves time-frequency domain separation of deformation observation signals through multi-scale discrete wavelet decomposition. Combined with dynamic adjustment of noise reduction thresholds for each layer based on pavement structure fatigue damage, effective deformation signals are separated from random noise to generate deformation data. Specifically, discrete wavelet decomposition decomposes the time-domain non-stationary signal into sub-components of different scales and frequency bands, enabling the separation of signal trend terms from detail terms.

[0037] Specifically, a wavelet basis that fits the characteristics of the deformation signal is selected, and the deformation observation signal output in step S1 is subjected to multi-level discrete wavelet decomposition. The deformation observation signal in the time domain is mapped to different frequency bands, and the low-frequency approximate component corresponding to the main deformation trend and the high-frequency detail component corresponding to detail fluctuations and noise are separated.

[0038] For example, the wavelet basis is the db4 wavelet basis, the number of decomposition layers is 3 to 5, and the decomposition process uses periodic extension to process the signal boundaries.

[0039] Furthermore, based on the fatigue damage evolution model, the fatigue damage degree of the concrete pavement at the current moment of the monitoring point is calculated, which is a dimensionless parameter characterizing the degree of accumulation of internal micro-defects in the pavement structure under repeated axle loads.

[0040] For example, the cumulative axle load application count is calculated according to the axle load conversion method specified in the "Specifications for Design of Cement Concrete Pavement of Highways" (JTG D40), converting the actual number of passing vehicles of different axle types into the standard axle load application count; the standard axle load application count for fatigue life is determined based on the design flexural tensile strength and load stress level of the pavement structure, according to the specifications in JTG D40; the damage coefficient related to the load stress level ranges from 0.1 to 0.5. The above-mentioned cumulative standard axle load application count is taken as the actual axle load application count. The fatigue life standard axle load cycles are taken as the fatigue life. Substitute the values ​​into the fatigue damage formula to calculate.

[0041] The formula for calculating the fatigue damage degree of concrete pavement in this step is as follows: In the formula, The fatigue damage degree of the concrete pavement; The damage coefficient is related to the load stress level; The number of times the standard axle load has been applied to the road surface corresponding to the monitoring point since it opened to traffic; This represents the standard axle load cycles for the fatigue life of the pavement structure. The calculation of fatigue damage degree of concrete pavement is based on the empirical correlation of concrete fatigue damage evolution in continuous damage mechanics, used to calculate the current fatigue damage state of the pavement structure corresponding to the monitoring point.

[0042] Fatigue damage in concrete is a gradual process of initiation, propagation, and penetration of microcracks at the interface between the internal cementitious matrix and aggregates under repeated loading. A single axle load will cause stress concentration in the interface transition zone, triggering subcritical propagation of microcracks. Each load cycle causes irreversible degradation of material stiffness, which macroscopically manifests as the accumulation of fatigue damage. As the damage level increases, the propagation of microcracks within the concrete will induce high-frequency micro-deformations in the pavement structure. The frequency band characteristics of the deformation signal highly overlap with random noise. By quantifying the damage level, a direct correlation can be established between the microscopic damage state of the material and macroscopic signal processing strategies.

[0043] Furthermore, for each high-frequency detail component obtained from the decomposition, the noise standard deviation estimate of the component is calculated. Combined with the fatigue damage degree of the concrete pavement and the preset damage correction coefficient, the noise reduction threshold of the corresponding high-frequency detail component of the decomposition layer is calculated.

[0044] For example, the global threshold adjustment factor is set to 1, and the damage correction coefficient ranges from 0.2 to 0.4. When the value of γ is below 0.2, the threshold increment caused by damage is insufficient and cannot provide enough retention gain for the high-frequency deformation signal of the corresponding pavement structure fatigue damage; when the value of γ is above 0.4, the threshold increment is too large, and a large amount of random noise will be retained, resulting in a decrease in the signal-to-noise ratio; when γ is in the range of 0.2 to 0.4, a balance between signal retention and noise suppression can be achieved.

[0045] The formula for calculating the adaptive noise reduction threshold in this step is: In the formula, For the first Noise reduction threshold for high-frequency detail components in the layer; This is the global threshold adjustment factor; For the first Noise standard deviation estimation for high-frequency detail components of the layer; The length of the deformation observation signal; This is the damage correction factor; The fatigue damage degree of the concrete pavement is considered. The adaptive noise reduction threshold is calculated based on an improved adaptive threshold calculation model using the general wavelet threshold formula. Building upon the classic Donoho threshold formula, a correction term tied to the fatigue damage state of the pavement structure is introduced to dynamically adjust the noise reduction threshold according to the structural damage state. This is used to calculate the noise reduction threshold for each layer of high-frequency detail components after wavelet decomposition, adapting it to the current pavement structural damage state, thus achieving control between noise suppression and effective preservation of minute deformation signals.

[0046] Unlike methods that set a fixed threshold to zero for low-amplitude noise coefficients, when fatigue damage to the pavement structure causes microcracks to propagate, the resulting minute deformations manifest as high-frequency, low-amplitude fluctuations in the time domain, with wavelet coefficient amplitudes highly overlapping with the noise coefficients. A fixed threshold easily sets these effective coefficients to zero, causing irreversible loss of deformation information. This formula introduces a fatigue damage degree correction term, causing the noise reduction threshold to increase synchronously with the damage degree, reducing the contraction intensity of high-frequency coefficients, thereby preserving the low-amplitude, high-frequency coefficients characterizing microcrack propagation and distinguishing damage-related signals from random noise.

[0047] Furthermore, a soft thresholding method is employed to shrink wavelet coefficients with amplitudes below the corresponding noise reduction threshold in each layer of high-frequency detail components, while retaining coefficients with amplitudes above the threshold, thus obtaining the processed high-frequency detail components. The processed high-frequency detail components are then subjected to a discrete wavelet inverse transform with the original low-frequency approximation components to reconstruct the time-domain deformed data. The soft thresholding method is a thresholding technique that continuously shrinks wavelet coefficients to avoid local abrupt changes in the reconstructed signal.

[0048] For example, the inverse reconstruction uses the db4 wavelet basis and periodic extension method consistent with the decomposition process.

[0049] In practice, the deformation observation signal is decomposed into a 3-layer discrete wavelet based on db4 wavelet in the cloud data processing server to obtain the low-frequency approximate component and the high-frequency detail component of each layer. The standard axle load number of fatigue life of the pavement structure is calculated according to the JTGD40 standard. The current fatigue damage degree is calculated by combining the cumulative standard axle load number and the fatigue damage evolution formula. Then, the noise standard deviation estimate is calculated for each layer of high-frequency detail component. The noise reduction threshold is calculated by combining the preset parameters with the adaptive threshold formula. After processing by the soft threshold function, the low-frequency approximate component is reconstructed by inverse transformation to generate deformation data and store it.

[0050] Step S3 involves collecting multi-dimensional road surface inspection data, identifying abnormal targets at the edge, performing noise reduction processing on the inspection data to adapt to the characteristics of road surface materials and subgrade structure, preserving the characteristic information related to road surface defects and deformation, and obtaining structured inspection results.

[0051] The method for obtaining inspection results includes, in sequence, planning an inspection path along the monitored road, dispatching inspection equipment equipped with sensors and edge computing units to collect multispectral images and 3D point cloud data of the road surface to obtain raw road inspection data; performing lightweight target detection inference on the multispectral images in the raw road inspection data in the edge computing unit to complete the real-time identification of abnormal targets on the road surface and the generation of abnormal reports; performing material-aware block matching 3D collaborative filtering on the collected multispectral images to complete image noise suppression and preservation of defect edge features; performing a combined filtering process on the collected 3D point cloud of the road surface to complete outlier noise point removal and structure-aware Gaussian smoothing processing in sequence; and integrating the processed multispectral images, 3D point clouds and abnormal reports to obtain inspection results.

[0052] Specifically, drones equipped with multispectral cameras, lidar, and edge computing units are dispatched to perform periodic inspections along a preset path with fixed flight parameters, simultaneously collecting multispectral images and 3D point cloud data of the road surface to obtain raw road inspection data.

[0053] For example, the drone inspection flight altitude is 30m to 50m, the flight speed is 5m / s to 10m / s, and the inspection cycle is 1 month to 3 months; the spatial resolution of the multispectral camera is 5cm / pixel to 10cm / pixel, the number of spectral bands is 4 to 6, and the sampling frame rate is matched with the flight speed; the point cloud density of the lidar is 100 points / m². 2 Up to 500 points / m 2 The ranging accuracy is ±5mm, and the scanning frequency is 10Hz to 30Hz; the forward overlap rate of the inspection path is 60% to 80%, and the lateral overlap rate is 40% to 60%.

[0054] Furthermore, multispectral images acquired in real-time during the drone's flight are synchronously input into the edge computing unit mounted on the drone. A lightweight target detection model, after pruning and quantization processing, is run to perform real-time inference and identification of abnormal road surfaces in the multispectral images. The identified abnormal target categories, contour information, corresponding timestamps, and spatial coordinates are packaged into a lightweight anomaly message and uploaded to the cloud platform. The abnormal road surfaces include four types of road damage: cracks, potholes, loose surfaces, and subsidence.

[0055] The lightweight object detection model used in this step is an improved version of the YOLOv7 lightweight object detection model. The backbone network of the model is replaced with a MobileNetV3-Small lightweight backbone network with a width multiplier of 1.0, and depthwise separable convolutions are used instead of standard convolutions. The input layer nodes correspond to 4-6 band multispectral images acquired by a multispectral camera, with image resolution adapted to the original resolution of the camera acquisition. The neck network adopts a simplified feature pyramid network and path aggregation network structure, corresponding to the feature layers output by the MobileNetV3-Small backbone network. The head network outputs the classification results and bounding box coordinates of four types of road surface defects, and the activation function uses the Hard-Swish function to meet the acceleration requirements of edge inference. After the above lightweight modifications, the computational complexity and number of parameters of the model are significantly reduced compared to the original YOLOv7, and the frame rate requirements for real-time inference on the edge computing units can be met.

[0056] The lightweight object detection model is trained using transfer learning. The MobileNetV3-Small backbone network is initialized with weights pre-trained on the ImageNet dataset, while the neck and head networks are randomly initialized. Fine-tuning is performed on a multispectral pavement image annotation dataset containing no fewer than 10,000 multispectral images collected under different pavement types and lighting conditions. The annotation categories, according to the "Highway Technical Condition Assessment Standard" JTG 5210, include four types of pavement distress: cracks, potholes, loosening, and subsidence. The intersection-union ratio (IU) between the annotation boxes and the distress contours is no less than 0.8. The training loss function is CIoU, and the optimizer is SGD with momentum. The initial learning rate is 0.01, momentum is 0.937, weight decay is 5e-4, batch size is 16, and the training runs for 300 epochs. The learning rate decays to 1 / 10 of the previous stage in epochs 200 and 260, respectively. The training terminates when the average accuracy on the validation set does not improve by more than 0.5% in the last 30 epochs.

[0057] For example, the inference frame rate of the lightweight object detection model is no less than 10fps, the model quantization accuracy is INT8, the cross-union threshold for anomaly identification is 0.5, and the confidence threshold is 0.6 to 0.8; the data volume of the anomaly message does not exceed 5% of the original image data volume.

[0058] Furthermore, material-aware block-matching 3D collaborative filtering is applied to the acquired multispectral images to achieve image noise suppression and preservation of lesion edge features. Block-matching 3D collaborative filtering is an image denoising method that aggregates and collaboratively transforms similar image blocks, used to suppress noise while preserving image detail features.

[0059] For example, the image block size is 8×8 pixels to 16×16 pixels, the search window size for block matching is 32×32 pixels to 64×64 pixels, and the texture similarity matching threshold is 0.75 to 0.85; the correction coefficient based on the difference in road structure layer thickness ranges from 0.15 to 0.3. When the total thickness of the surface layer and the base layer is ≥20cm, the correction coefficient is 0.25 to 0.3, and when the total thickness of the surface layer and the base layer is <20cm, the correction coefficient is 0.15 to 0.25.

[0060] The formula for calculating the noise reduction threshold of block-matched 3D collaborative filtering in this step is as follows: In the formula, The noise reduction threshold for the three-dimensional transform domain coefficients; This is an estimate of the standard deviation of noise in a multispectral image; The total number of pixels within a single image block; This is a correction factor based on the difference in pavement structure layer thickness; For asphalt mixtures during service life The measured creep modulus is given in MPa. The initial elastic modulus of the asphalt mixture is expressed in MPa. The noise reduction threshold calculation formula is a material-aware adaptive threshold calculation model based on an improved version of the classic block-matching three-dimensional filtering threshold formula. On the basis of the general threshold formula, a creep modulus correction term characterizing the viscoelastic state of the asphalt mixture is introduced to achieve dynamic adjustment of the noise reduction threshold according to the service state of the asphalt material. This is used to calculate the dynamic noise reduction threshold in the multispectral image collaborative filtering process, achieving control between noise suppression and preservation of pavement crack edge features, avoiding the erroneous filtering of crack edge details by general fixed threshold filtering.

[0061] The creep modulus of asphalt mixtures after long-term service and aging As the temperature rises and material brittleness increases, the image edge features formed by crack initiation highly overlap with noise in spatial frequency. General fixed-threshold filtering, based on the Gaussian noise assumption, sets low-amplitude transform domain coefficients to zero, easily filtering out high-frequency coefficients at crack edges, leading to irreversible loss of edge details. The formula introduces a creep modulus correction term, making the threshold variable... Increase and enhance, dynamically reduce coefficient of shrinkage strength, and preserve high-frequency details at crack edges.

[0062] Furthermore, a combined filtering process is performed on the acquired 3D point cloud of the road surface. First, statistical filtering is used to remove outlier noise points, and then Gaussian smoothing filtering, which integrates the porosity modulation of the subgrade soil, is performed to smooth the point cloud noise and preserve the settlement micro-deformation features. Statistical filtering is a filtering method that removes outliers based on the statistical features of the point cloud's neighborhood distance.

[0063] For example, the number of neighborhood search points for statistical filtering is 15 to 20, and the distance threshold is 1.5 to 2 times the average neighborhood distance of the point cloud; Gaussian kernel standard deviation. Based on the average pavement texture depth calibration, the value range corresponds to an average pavement texture depth of 0.8mm to 1.2mm; subgrade soil void ratio The value ranges from 0.6 to 0.9, and the porosity sensitivity coefficient is... The value ranges from 0.08 to 0.12; the higher the roadbed compaction degree, the better. The smaller the value, the higher the porosity sensitivity coefficient. To verify the control parameters of the corresponding experimental variables, when When the value is below 0.08, the kernel function weight increment caused by the porosity is insufficient, failing to provide enough retention gain for the uneven settlement and micro-deformation characteristics of the soft soil subgrade section, and the filtering process will smooth out the micro-deformation characteristics; when When the value is higher than 0.12, the kernel function weight increment is too large, which will misjudge the normal texture structure of the road surface as deformation features and retain them, resulting in burrs in the point cloud data and reducing the smoothing accuracy of the point cloud. Within the range of 0.08 to 0.12, a balance can be achieved between noise smoothing and preservation of micro-deformation features.

[0064] The formula for the structure-aware Gaussian smoothing filter kernel function in this step is: In the formula, The two-dimensional Gaussian smoothing kernel function in coordinates The weight value at the location; The standard deviation is a Gaussian kernel, calibrated by the average surface texture depth. The relative coordinates of the pixels within the kernel function; Pore ​​sensitivity coefficient; The measured void ratio of the subgrade soil in the current road section; This represents the maximum void ratio of this type of subgrade soil. The Gaussian smoothing filter kernel function is an improved two-dimensional Gaussian smoothing kernel function. Based on the standard Gaussian kernel function, a void ratio modulation term characterizing the structural state of the subgrade soil is introduced to achieve dynamic adjustment of the Gaussian smoothing weights according to the subgrade structural state. This is used to calculate the dynamic kernel function weights for smoothing the three-dimensional point cloud of the pavement, thereby achieving smoothing suppression of random noise in the point cloud and preservation of micro-deformation features caused by uneven settlement of the subgrade, avoiding the erroneous smoothing of micro-deformation features by the general Gaussian filter.

[0065] The higher the porosity of soft soil subgrades, the more significant the uneven settlement caused by consolidation. Macroscopically, this manifests as continuous, low-amplitude micro-deformations of the pavement, which highly overlap with the spatial scale of noise in point cloud elevation data. The fixed weights of general Gaussian filters are only related to spatial distance and independent of the subgrade structural state. In soft soil sections, this can easily treat elevation changes in micro-deformations as noise smoothing, resulting in irreversible loss of settlement characteristics. By introducing a porosity modulation term into the kernel function, the smoothing weight decreases as the porosity increases, reducing the Gaussian smoothing intensity and thus preserving continuous micro-deformation characteristics, enabling the identification of random noise and settlement signals.

[0066] Furthermore, the processed multispectral images, 3D point clouds, and anomaly messages are associated and bound by timestamps and spatial coordinates to form inspection results, with the spatial coordinate system consistent with step S1.

[0067] In practice, the drone performs periodic inspections along a preset route. The lightweight target detection model at the edge identifies abnormal targets in real time and uploads messages. The cloud performs material perception BM3D noise reduction on the multispectral images and statistical filtering and structure perception Gaussian smoothing on the 3D point cloud, and integrates the results to generate inspection results.

[0068] Step S4 uses deformation data as a benchmark to perform optimal correction of the positioning deviation of the inspection results through Kalman filtering, unifies the spatiotemporal resolution, and generates a multi-source heterogeneous fusion dataset.

[0069] Specifically, deformation data and inspection results are fully collected. After invalid data frames are removed through validity verification, a standardized spatiotemporal index containing timestamps, spatial coordinates, and data type identifiers is constructed for each set of valid data, which is used for rapid association and matching of multi-source data.

[0070] For example, the validity verification rules include: the fixed state identifier of the deformation data solution is a fixed solution; the residual of the observed value is less than 3 times the standard error; the image frame integrity of the inspection results is 100%; and the positioning status of the point cloud data frame is a valid solution. The timestamp adopts the Coordinated Universal Time format, the spatial coordinates are uniformly adopted using the 2000 National Geodetic Coordinate System, and the elevation datum adopts the 1985 National Elevation Datum, which is consistent with the coordinate system adopted in step S1.

[0071] Furthermore, the deformation data is defined as the baseline state, the inspection results are defined as noisy observations, and state equations and observation equations adapted to the settlement monitoring scenario are established, and corresponding state vectors and observation vectors are defined.

[0072] For example, the state vector is a two-dimensional column vector containing the cumulative settlement and settlement rate of the monitoring points, with units of mm and mm / day, respectively; the observation vector is a column vector, with elements containing the plane coordinates, elevation values ​​and contour dimensions of the disease location in the inspection results, with units of m, mm and mm, respectively; the discrete sampling period is consistent with the UAV inspection period, which is 1 to 3 months.

[0073] The discrete-time state equation in this step is: In the formula, for The state vector at time step is a two-dimensional column vector, with the first element being... The cumulative settlement at each monitoring point, in mm; the second element is... Settlement rate at each monitoring point, in mm / day; The state transition matrix is ​​a 2×2 matrix, constructed based on a uniform settlement model, specifically in the form of... ,in for Time's up The time interval between moments, in days; for The process noise vector at time t follows a zero-mean multivariate normal distribution, and its covariance matrix is... It is used to characterize unmodeled random disturbances during the settlement process; This represents the discrete time step, corresponding to the sequence number of the inspection cycle. The discrete-time state equation is a linear discrete state evolution equation, belonging to a linear dynamic model of deterministic evolution and random perturbation superposition, used to describe the change of state over time, and is used in this step to achieve prior prediction.

[0074] The discrete-time observation equation in this step is: In the formula, for The observation vector at any given time is a column vector derived from the measured values ​​of the inspection results. Its elements include the elevation values ​​of the monitoring points obtained during the inspection and the plane coordinate offset of the location of the disease, with units of mm and m, respectively. This is the observation matrix, used to map the state vector to the observation vector space, with dimensions matching the length of the observation vector; for The measurement noise vector at time t follows a zero-mean multivariate normal distribution, with a covariance matrix of... This is used to characterize random disturbances such as positioning drift and sensor measurement errors in UAV inspections. The discrete-time observation equation is a linear observation model that describes the linear mapping relationship between the internal state and the external observable values. In this step, it is used to establish the mathematical correlation between the measured values ​​of the inspection results and the baseline state.

[0075] Furthermore, Kalman filtering is employed as a recursive state estimation algorithm. Using deformation data from the baseline state as the initial posterior state value, the algorithm completes the time update of the prior state estimation and the prior error covariance matrix based on the state equation. Substituting the measured values ​​from the inspection results into the observation equation, the algorithm completes the measurement update steps of Kalman gain calculation, posterior state estimation update, and posterior error covariance matrix update. Using each inspection cycle as the iteration step size, the algorithm iterates through all monitoring points along the entire road segment to obtain the optimal posterior spatial location estimate of the inspection results under a unified spatiotemporal reference. Kalman filtering is a linear optimal state estimation algorithm based on the minimum mean square error criterion, used to recursively solve for the optimal estimate from noisy observations.

[0076] For example, the range of the process noise covariance matrix is ​​1e-6mm. 2 / sky 2 Up to 1e-4mm 2 / sky 2 The planar positioning variance of the measurement noise covariance matrix is ​​0.01 mm. 2 up to 0.05mm 2 The variance of elevation measurement is 1 mm. 2 Up to 5mm 2 The convergence condition is that the trace of the posterior error covariance matrix is ​​less than 1e-3mm. 2 .

[0077] The Kalman filter time update formula is: , In the formula, for The prior state estimation vector at time t; for The posterior state estimation vector at time step S2 is initially derived from the deformation data from step S2. for The prior error covariance matrix at time t; for The posterior error covariance matrix at time t; State transition matrix The transpose of the matrix; This is the process noise covariance matrix. The prediction step recursive formula of the Kalman filter time update equation, based on the optimal state estimate of the previous time step, predicts the prior state and corresponding error covariance at the current time step. In this step, it is used to predict the prior distribution of the current inspection cycle based on the baseline deformation data.

[0078] The Kalman filter measurement update formula is as follows: , , In the formula, for Kalman gain matrix at time step; It is an identity matrix, with dimensions matching the dimensions of the state vector; for The optimal posterior state estimation vector at time t is the corrected spatial position of the inspection results under a unified spatiotemporal reference. for The posterior error covariance matrix at time t; This is the measurement noise covariance matrix. The correction step recursive formula of the Kalman filter measurement update formula corrects the prior state estimate based on the current observation value, and solves for the optimal state estimate under the minimum mean square error criterion. In this step, it is used to correct the positioning drift and measurement noise in the inspection results using high-precision deformation data, and achieve optimal alignment of the spatial reference of the two types of data.

[0079] In the recursive iteration, the minimum error covariance of the deformation data is used as a high-confidence prior, and the measurement value of the inspection results is used as a noisy observation. The Kalman gain dynamically allocates weights between the prior and the observation. Each iteration minimizes the posterior error covariance, suppresses the transmission and amplification of measurement noise, and thus achieves spatiotemporal alignment of heterogeneous data under the statistical optimality criterion, solving the error accumulation problem of direct coordinate registration.

[0080] Furthermore, the inspection results after the optimal posterior estimation correction are compared with the deformation data used as the benchmark. Interpolation resampling in the time dimension and rasterization in the spatial dimension are unified to make the two types of data have the same time resolution and spatial resolution. The deformation data, road surface defect information and three-dimensional point cloud elevation data after unified resolution are structurally integrated to construct a multi-source heterogeneous fusion dataset.

[0081] For example, the time resolution is uniformly set to 1 day, and the discrete inspection data is resampled using a linear interpolation method; the spatial resolution is uniformly set to 0.5m×0.5m grid cells, and the spatial discrete data is rasterized using an inverse distance weighted interpolation method; the dataset is stored in a geographic information database format, and each grid cell contains a unique spatial index, which is associated with the full set of feature data of the corresponding location.

[0082] In practice, after validating the deformation data and inspection results, a standardized spatiotemporal index is constructed. Kalman filtering is run to complete the optimal posterior estimation and spatial registration. After time-dimensional interpolation resampling and spatial-dimensional rasterization unification, a multi-source heterogeneous fusion dataset is constructed and stored.

[0083] Step S5 extracts and standardizes features from the multi-source heterogeneous fusion dataset, inputs them into a long short-term memory recurrent neural network to predict the settlement trend in the next 3 to 12 months, compares them with preset grading thresholds to generate early warning signals, and uses a digital twin model to simulate the settlement evolution and cost-benefit analysis of the maintenance plan to generate a maintenance plan.

[0084] Among them, the Long Short-Term Memory Recurrent Neural Network is used to handle the long-period dependency problem of time-series data; the digital twin model is a digital model built based on geological survey data, structural design parameters and historical monitoring data, used to map the real state of roadbed and pavement structure and the law of settlement evolution.

[0085] Specifically, from the multi-source heterogeneous fusion dataset generated in step S4, the daily-scale historical settlement time series corresponding to each spatial grid cell is extracted, and the environmental load parameters of the corresponding time period and road section are extracted simultaneously. Standardized preprocessing is performed on all extracted features to construct an input feature vector that matches the input dimension of the time series prediction model.

[0086] Environmental load parameters include the daily average road surface temperature, the daily cumulative traffic flow at the cross-section, the daily average groundwater level, and the daily cumulative number of standard axle load applications. Specifically, the daily average road surface temperature is obtained from temperature sensors deployed every 5 to 10 km along the monitored road section, and the arithmetic mean of the hourly temperature readings over the 24 hours is taken. The daily cumulative traffic flow at the cross-section is obtained from the traffic volume detection equipment at the highway toll management system or traffic dispatch station, and the cumulative equivalent axle load applications for each vehicle type are taken after conversion using the standard axle load conversion factor. The daily average groundwater level is obtained from groundwater level monitoring wells deployed every 5 to 10 km along the monitored road section, and the arithmetic mean of the hourly water level readings over the 24 hours is taken. The daily cumulative number of standard axle load applications is accumulated from vehicle type classification data obtained by the dynamic weighing system or highway toll management system after conversion using the axle load spectrum.

[0087] The daily average road surface temperature and groundwater level are hourly sampled meteorological and hydrological monitoring data. The daily cumulative traffic flow and daily cumulative standard axle load application times are daily summarized traffic volume data. The historical settlement time series generated in step S4 is daily sampled deformation data. Before constructing the LSTM model input, the following preprocessing alignment is performed on each environmental load parameter: using the calendar day as the standard alignment benchmark, data is bound in segments according to the mileage markers of the road sections, and the original values ​​of each environmental load parameter are aggregated by date and road section; for data sampled at a higher frequency, the arithmetic mean is first aggregated into daily values; for dates with missing values, the average of 3 adjacent time periods is used to fill in the missing values; all parameters aligned on a daily scale and the historical settlement data of the corresponding road sections and dates are merged into a unified time series record by timestamp, and then standardized preprocessing is performed.

[0088] For example, the standardization preprocessing adopts the Z-score standardization method, so that the mean of all features is 0 and the variance is 1; the time sliding window length is set to 90 days, that is, the input feature vector contains all feature data of the 90 consecutive days before the current time; the scope of feature standardization is the full historical data of a single monitoring segment, and the standardization parameters are updated synchronously with the dataset update of each inspection cycle; the time resolution of the environmental load parameters is consistent with the historical settlement time series, which is 1 day.

[0089] Furthermore, a long short-term memory recurrent neural network adapted to highway settlement scenarios is constructed, and supervised learning is used to complete the model training. The input feature vector is then imported into the trained model to perform inference and output the long-term settlement trend prediction results for the next 3 to 12 months.

[0090] For example, the input layer nodes of the Long Short-Term Memory Recurrent Neural Network correspond one-to-one with the dimensions of the input feature vectors, and each input node corresponds to a standardized feature data. The hidden layer is set as two layers of Long Short-Term Memory Recurrent Neural Network, with 128 hidden units in each layer. The activation function within the layer is the tanh function, and the recurrent activation function is the sigmoid function. A dropout layer with a dropout rate of 0.2 to 0.3 is set between the two hidden layers. The output layer is a fully connected layer with a linear activation function, and the output node corresponds to the predicted monthly average settlement rate for the next 3 to 12 months.

[0091] During model training, the dataset is divided into training and validation sets in an 8:2 ratio. The loss function is mean squared error, the optimizer is Adam, the initial learning rate is 1e-3, and it decays by 10% every 10 rounds. The termination condition is that the validation set loss does not decrease for 15 consecutive rounds and the average absolute error of prediction is less than 0.5mm / month. The model is trained offline and inferred online, and fine-tuned once every inspection cycle.

[0092] Settlement is influenced by the lag coupling of factors such as accumulated axle load, soil consolidation, temperature changes, and groundwater level fluctuations, with a time-series dependency spanning several months to a year. Linear trend extrapolation based on the stationary assumption cannot cover its nonlinear time-delay characteristics, which is the root cause of the short prediction period and large errors of traditional methods. Long Short-Term Memory (LSTM) recurrent neural networks achieve selective memory of long-range time-series dependencies through a gating structure: the forget gate discards short-term random fluctuations, the input gate captures the lag effects of load and environmental factors and stores them in the cell state, and the output gate outputs long-term predictions based on the accumulated state. Through multi-dimensional nonlinear mapping, the model learns the conditional probability distribution of the settlement process, achieving effective control of long-term prediction bias.

[0093] The mean squared error loss function used in this step of model training is formulated as follows: In the formula, This represents the mean squared error loss value; This refers to the number of samples within a training batch. For the first The measured settlement rate for each sample is labeled in mm / month. For the first The model predicts the settlement rate for each sample, in mm / month. The mean squared error loss function is the standard loss function for regression tasks, belonging to the empirical risk minimization model. It is used to quantify the deviation between the model's predicted values ​​and the measured labeled values. In this step, it serves as the optimization objective for model training, guiding the model weights to be updated in the direction of minimizing the prediction deviation.

[0094] Furthermore, the monthly average settlement rate for the next 12 months output by the model is converted into an annualized settlement rate. The annualized settlement rate is then compared with the preset three-level warning thresholds one by one, and a settlement risk warning signal of the corresponding level is generated based on the comparison results.

[0095] For example, the trigger condition for the first-level warning is an annualized settlement rate greater than 5 mm / year, the trigger condition for the second-level warning is an annualized settlement rate of 3 mm / year to 5 mm / year, and the trigger condition for the third-level warning is an annualized settlement rate of 1 mm / year to 3 mm / year; an annualized settlement rate less than 1 mm / year is determined to be a normal state and no warning signal is generated.

[0096] Furthermore, for the monitoring sections that generate Level 1 and Level 2 warnings, the digital twin models of the corresponding road sections are invoked, and preset maintenance schemes are applied to the models respectively. The roadbed settlement evolution process under different maintenance schemes over the next 10 years is simulated, and the full life cycle operation cost corresponding to each maintenance scheme is calculated simultaneously. Through cost-benefit analysis, maintenance schemes that meet the settlement control requirements are selected.

[0097] For example, the preset maintenance schemes include four types: static pressure grouting reinforcement, high-pressure jet grouting pile reinforcement, geogrid reinforcement, and road milling and repaving; the simulation time step of the digital twin model is 1 month, and the simulation period is 10 years; the evaluation index of cost-benefit analysis is the total life cycle cost corresponding to the unit settlement control volume, with the unit being yuan / mm.

[0098] The digital twin model quantifies the effects of different maintenance schemes on soil consolidation rate and structural stiffness improvement into a predictable state evolution law. The cost-benefit analysis transforms the multi-objective problem of settlement control and cost minimization into a cost-benefit ratio optimization problem. Through simulation comparison, the scheme with the lowest life-cycle cost under the constraints is selected.

[0099] In practice, features are extracted and standardized using a 90-day sliding window, then imported into a trained long short-term memory recurrent neural network to perform inference and output the monthly average settlement rate prediction. After comparison with graded early warnings, a digital twin model is called to complete the maintenance plan simulation and cost analysis, and a maintenance plan is generated.

[0100] Exemplary system: like Figure 2 As shown, a highway long-term settlement monitoring system based on multi-source data fusion includes a Beidou reference station network, a UAV inspection unit, and a cloud-based collaborative analysis platform. The Beidou reference station network, the UAV inspection unit, and the cloud-based collaborative analysis platform complete data transmission between each other through a communication network.

[0101] The BeiDou reference station network is used to solve the original observation signals, correct the errors in the solution results, obtain the deformation observation signals, perform wavelet decomposition on the deformation observation signals, adaptively calculate the noise reduction threshold using the fatigue damage factor, and generate deformation data.

[0102] For example, the BeiDou reference station network includes dual-frequency global navigation satellite system receiving terminals, measurement antennas, and observation piers deployed along the monitored roads. It completes the calculation and error correction of the original observation signals through real-time dynamic carrier phase differential and ionospheric linear combination, generates deformation data by adaptive threshold wavelet denoising, and transmits it to the cloud collaborative analysis platform through the communication network.

[0103] The drone inspection unit is used to acquire road inspection data, perform noise reduction and anomaly identification on the road inspection data, obtain inspection results, and transmit the inspection results to the cloud collaborative analysis platform through the communication network.

[0104] For example, the UAV inspection unit includes an inspection UAV equipped with a multispectral camera, LiDAR, and an edge computing unit. The UAV inspection unit performs periodic inspections along the monitored road according to a preset flight path. It acquires multispectral images of the road surface using the multispectral camera and 3D point cloud data of the road surface using the LiDAR, obtaining raw road surface inspection data. The edge computing unit runs a lightweight target detection model that has undergone pruning and quantization processing, performing real-time inference and identification of abnormal targets on the road surface from the multispectral images, generating anomaly reports containing the anomaly target category, contour information, and spatiotemporal coordinates. The UAV inspection unit performs material-aware block matching 3D collaborative filtering on the multispectral images, with the noise reduction threshold dynamically modulated by the asphalt mixture creep modulus; it first performs statistical filtering to remove outlier noise points from the 3D point cloud of the road surface, and then performs Gaussian smoothing filtering that integrates the porosity modulation of the subgrade soil; finally, it correlates and integrates the processed multispectral images, 3D point cloud, and anomaly reports to obtain the inspection results.

[0105] The cloud-based collaborative analysis platform is used to perform spatiotemporal alignment of inspection results based on deformation data, generate a multi-source heterogeneous fusion dataset, and input the multi-source heterogeneous fusion dataset and environmental load parameters into a deep time series prediction model to predict future long-term settlement trends. Based on the prediction results, a maintenance plan is generated.

[0106] For example, the cloud-based collaborative analysis platform includes computing servers, storage servers, and geographic information data processing services deployed in the cloud. The platform aggregates deformation data transmitted from the BeiDou reference station network and inspection results transmitted from UAV inspection units. Using the deformation data as the baseline state and the inspection results as noisy observations, it runs a recursive state estimation algorithm to correct spatial positioning deviations in the inspection results, achieving spatiotemporal benchmark unification for the two types of heterogeneous data. It then performs spatiotemporal resolution unification and structured integration on the corrected inspection results and deformation data to generate a multi-source heterogeneous fusion dataset. The platform extracts input features from the multi-source heterogeneous fusion dataset and environmental load parameters, constructs and trains a long short-term memory recurrent neural network, and outputs long-term settlement trend prediction results for the next 3 to 12 months. The prediction results are compared with preset grading thresholds to generate settlement risk warning signals. For the warning sections, a digital twin model is invoked to simulate the settlement evolution and cost-benefit analysis of the maintenance plan, generating a maintenance plan.

[0107] Based on the preferred embodiments of the present invention described above, those skilled in the art can make various changes and modifications without departing from the inventive concept. The technical scope of this invention is not limited to the contents of the specification, but must be determined according to the scope of the claims.

Claims

1. A method for long-term settlement monitoring of highways based on multi-source data fusion, characterized in that, Includes the following steps: S1: Solve the original observation signal, correct the error in the solution result, and obtain the deformation observation signal; S2: Perform wavelet decomposition on the deformation observation signal, adaptively calculate the noise reduction threshold using the fatigue damage factor, and generate deformation data; S3: Acquire road inspection data, perform noise reduction processing and anomaly identification on the road inspection data, and obtain inspection results; wherein, adaptively calculate the noise reduction threshold. In order to target the The high-frequency detail components of the layer are calculated using the following formula: In the formula This is the global threshold adjustment factor. For the first Noise standard deviation estimation for high-frequency detail components of the layer. The length of the deformation observation signal, This is the damage correction factor. The fatigue damage degree of the pavement structure; the damage correction coefficient The value range is from 0.2 to 0.4; S4: Using the deformation data as the baseline state and the inspection results as the observation values, the Kalman filter algorithm is used to perform spatiotemporal alignment on the inspection results to generate a multi-source heterogeneous fusion dataset; wherein, the state vector contains the cumulative settlement and settlement rate of the monitoring points, and the observation vector contains the plane coordinates, elevation values ​​and disease outline dimensions of the disease location in the inspection results; S5: Input the multi-source heterogeneous fusion dataset and environmental load parameters into the deep time series prediction model to predict the future long-term settlement trend, and generate a maintenance plan based on the prediction results.

2. The method according to claim 1, characterized in that, In step S2, the adaptive calculation of the noise reduction threshold using the fatigue damage factor includes obtaining the fatigue damage degree of the pavement structure corresponding to the monitoring point. The fatigue damage degree The calculation formula is: ,in The damage coefficient is related to the load stress level. This represents the actual cumulative number of standard axle loads applied to the road surface corresponding to each monitoring point since it opened to traffic. The standard axle load cycles represent the fatigue life of the pavement structure.

3. The method according to claim 1, characterized in that, In step S2, generating deformation data includes using a soft thresholding function to shrink the wavelet coefficients in each layer of high-frequency detail components whose amplitudes are lower than the corresponding noise reduction threshold, to obtain the processed high-frequency detail components; and performing discrete wavelet inverse transform on the processed high-frequency detail components and the original low-frequency approximation components to reconstruct the deformation data.

4. The method according to claim 1, characterized in that, In step S3, anomaly identification of road inspection data includes using a lightweight target detection model that has been pruned and quantized. The backbone network of the lightweight target detection model is MobileNetV3, and depthwise separable convolution is used to replace standard convolution for real-time inference and identification of abnormal road targets in edge computing units.

5. The method according to claim 1, characterized in that, In step S3, the noise reduction processing of the road inspection data includes performing block matching three-dimensional collaborative filtering on the multispectral image and setting a noise reduction threshold. The calculation formula is: ,in For estimating the standard deviation of noise in multispectral images, This represents the total number of pixels within an image block. This is a correction factor based on the difference in pavement structure layer thickness. For asphalt mixtures during service life The measured creep modulus is as follows. This is the initial elastic modulus of the asphalt mixture.

6. The method according to claim 1, characterized in that, In step S5, the deep time series prediction model is a long short-term memory recurrent neural network, which includes two hidden layers, each with 128 hidden units. The activation function within the layer is tanh, the recurrent activation function is sigmoid, and the output layer is a fully connected linear layer that outputs the predicted monthly average settlement rate for the next 3 to 12 months.

7. A highway long-term settlement monitoring system based on multi-source data fusion, used to implement the method according to any one of claims 1-6, characterized in that, It includes a BeiDou reference station network, a UAV inspection unit, and a cloud-based collaborative analysis platform, wherein the BeiDou reference station network, the UAV inspection unit, and the cloud-based collaborative analysis platform are connected through a communication network; The BeiDou reference station network is used to solve the original observation signal, correct the error of the solution result, obtain the deformation observation signal, perform wavelet decomposition on the deformation observation signal, adaptively calculate the noise reduction threshold using the fatigue damage factor, and generate deformation data. The drone inspection unit is used to acquire road inspection data, perform noise reduction processing and anomaly identification on the road inspection data, and obtain inspection results. The cloud-based collaborative analysis platform is used to perform spatiotemporal alignment of the inspection results based on the deformation data, generate a multi-source heterogeneous fusion dataset, and input the multi-source heterogeneous fusion dataset and environmental load parameters into a deep time series prediction model to predict the future long-term settlement trend and generate a maintenance plan based on the prediction results.

Images