Road monitoring method, device and equipment and readable storage medium

Abstract

This application provides a road monitoring method, apparatus, device, and readable storage medium. The method includes: acquiring vehicle parameters of a first vehicle and vehicle response data of the first vehicle during its travel on a first road; determining the dynamic state of the first vehicle, the trajectory coordinates of each wheel of the first vehicle in the road coordinate system, and the effective road surface input of each wheel of the first vehicle to the first road based on the vehicle response data and vehicle parameters; determining the dynamic wheel load applied to the road surface of the first road by each wheel of the first vehicle during its travel using a first model characterizing vehicle dynamics based on the dynamic state and effective road surface input; and inputting the vehicle speed, trajectory coordinates, and dynamic wheel load from the vehicle response data into a neural network to perform road parameter inversion to obtain the road parameters of the first road. Embodiments of this application can achieve high-precision road parameter monitoring at normal driving speeds.

Description

Technical Field

[0001] This application belongs to the field of road health monitoring technology, specifically relating to a road monitoring method, device, equipment, and readable storage medium. Background Technology

[0002] With increased safety awareness, it is necessary to monitor road parameters regularly so that timely road maintenance can be carried out.

[0003] However, in the field of road parameter assessment, the relevant technical solutions mainly face some challenges. For example, the detection equipment that can provide accurate structural stiffness parameters usually needs to perform discrete detection at fixed points or at extremely low speeds, which results in low detection efficiency and requires the closure of traffic. On the other hand, vehicle-mounted detection technology that can operate at normal traffic speeds can only obtain the surface smoothness of the road surface and cannot infer the mechanical parameters of the deep structure, resulting in an inherent contradiction between detection speed and structural detection accuracy.

[0004] It is evident that there is an urgent need for a road parameter monitoring solution that can achieve high accuracy at normal driving speeds, in order to resolve the inherent contradiction between detection speed and road parameter detection accuracy in related technologies. Summary of the Invention

[0005] This application provides a road monitoring method, apparatus, device, and readable storage medium, which can resolve the inherent contradiction between detection speed and road parameter detection accuracy in related technologies, and can achieve high-precision road parameter monitoring at normal driving speeds.

[0006] In a first aspect, embodiments of this application provide a road monitoring method, the method comprising: Obtain the vehicle parameters of the first vehicle and the vehicle response data of the first vehicle during its travel on the first road; Based on the vehicle response data and the vehicle parameters, the dynamic state of the first vehicle, the trajectory coordinates of each wheel in the first vehicle in the road coordinate system, and the effective road surface input of each wheel in the first vehicle to the first road are determined. Based on the dynamic state and the effective road surface input, the dynamic wheel loads applied to the road surface of the first road by each wheel of the first vehicle during driving are determined using a first model characterizing vehicle dynamics. The vehicle speed, trajectory coordinates, and dynamic wheel load from the vehicle response data are input into a neural network to perform road parameter inversion, thereby obtaining the road parameters of the first road. The neural network includes a first branch network, a second branch network, and a third branch network. The first branch network is used to determine the static geometric elevation of the road surface of the first road in the road parameters based on the trajectory coordinates. The second branch network is used to determine the transient dynamic deflection induced by the dynamic wheel load of each wheel of the first vehicle during driving based on the vehicle speed, trajectory coordinates, and dynamic wheel load. The third branch network is used to determine the structural parameters of the first road in the road parameters based on the trajectory coordinates. The target loss function of the neural network is constructed based on the residual of the target model. The target model includes the first model and the physical constraint model of the first model. The physical constraint model includes: a second model for characterizing the contact envelope between the tire and the road, a third model for characterizing the coupled dynamics between the vehicle and the road, and a fourth model for characterizing the spatiotemporal consistency between different tires. The second model is used to constrain the static geometric elevation and transient dynamic deflection of the road surface in the effective road surface input involved in the first model. The third model is used to constrain the transient dynamic deflection induced by structural parameters and dynamic wheel loads. The fourth model is used to constrain the static geometric elevation of the road surface.

[0007] Secondly, embodiments of this application provide a road monitoring device, the device comprising: The acquisition module is used to acquire the vehicle parameters of the first vehicle and the vehicle response data of the first vehicle during its travel on the first road. The first determining module is used to determine the dynamic state of the first vehicle, the trajectory coordinates of each wheel of the first vehicle in the road coordinate system, and the effective road surface input of each wheel of the first vehicle to the first road based on the vehicle response data and the vehicle parameters. The second determining module is used to determine the dynamic wheel load applied to the road surface of the first vehicle by each wheel during the driving process, based on the dynamic state and the effective road surface input, using a first model characterizing vehicle dynamics. The inversion module is used to input the vehicle speed, trajectory coordinates, and dynamic wheel load from the vehicle response data into a neural network to perform road parameter inversion and obtain the road parameters of the first road. The neural network includes a first branch network, a second branch network, and a third branch network. The first branch network is used to determine the static geometric elevation of the road surface of the first road in the road parameters based on the trajectory coordinates. The second branch network is used to determine the transient dynamic deflection induced by the dynamic wheel load of each wheel of the first vehicle during driving based on the vehicle speed, trajectory coordinates, and dynamic wheel load. The third branch network is used to determine the structural parameters of the first road in the road parameters based on the trajectory coordinates. The target loss function of the neural network is constructed based on the residual of the target model. The target model includes the first model and the physical constraint model of the first model. The physical constraint model includes: a second model for characterizing the contact envelope between the tire and the road, a third model for characterizing the coupled dynamics between the vehicle and the road, and a fourth model for characterizing the spatiotemporal consistency between different tires. The second model is used to constrain the static geometric elevation and transient dynamic deflection of the road surface in the effective road surface input involved in the first model. The third model is used to constrain the transient dynamic deflection induced by structural parameters and dynamic wheel loads. The fourth model is used to constrain the static geometric elevation of the road surface.

[0008] Thirdly, embodiments of this application provide an electronic device, including: a processor, a memory, and a program stored in the memory and executable on the processor, wherein when the program is executed by the processor, it implements the steps of the road monitoring method as described in the first aspect.

[0009] Fourthly, embodiments of this application provide a computer-readable storage medium storing a computer program that, when executed by a processor, implements the steps of the road monitoring method as described in the first aspect.

[0010] Fifthly, embodiments of this application provide a computer program product including computer instructions that, when executed by a processor, implement the steps of the road monitoring method as described in the first aspect.

[0011] In this embodiment, a neural network with three branches is constructed. The input data (i.e., the dynamic wheel loads applied to the road surface by each wheel of the first vehicle during driving) is determined by a first model (vehicle dynamics model). A physical constraint system consisting of a second model (tire-road envelope), a third model (vehicle-road coupling), and a fourth model (spatiotemporal consistency between different tires) serves as the basis for constructing the target loss function to train the neural network. This allows for a deep integration of data-driven and physical mechanisms, enabling the neural network to achieve high-precision road parameter monitoring at normal driving speeds using the constructed three branches. Specifically, the second model corrects the high-frequency distortion caused by the single-point contact assumption; the third model establishes the mechanical balance between dynamic wheel loads and transient dynamic deflection to separate static and dynamic components; and the fourth model utilizes the physical correlation between multiple wheel tracks to suppress inversion drift under unsteady conditions. This significantly improves the accuracy and robustness of road parameter inversion while ensuring detection efficiency. Attached Figure Description

[0012] To more clearly illustrate the technical solutions of the embodiments of this application, 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 of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0013] Figure 1 This is a structural diagram of the dynamic structure monitoring system provided in the embodiments of this application; Figure 2 This is a diagram showing the vehicle's sensor configuration; Figure 3 It is a system framework for implementing road monitoring methods; Figure 4 This is a flowchart of a specific example of a road monitoring method in the embodiments of this application; Figure 5 This is an example diagram of road parameter inversion for a single vehicle; Figure 6 This is an example diagram of road parameter inversion using multiple vehicles and multiple speeds; Figure 7 This is a structural diagram of a road monitoring device provided in an embodiment of this application; Figure 8 This is a structural diagram of an electronic device provided in an embodiment of this application. Detailed Implementation

[0014] The technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, not all embodiments. Based on the embodiments of this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0015] This application relates to the fields of intelligent sensing of road infrastructure, vehicle dynamics, tire-road contact mechanics, vehicle-road-subgrade coupling dynamics, physical information neural networks, time-frequency analysis, and road structure health monitoring.

[0016] The technical problem to be solved by the embodiments of this application is as follows: (1) How to establish a common envelope filtering model of tire contact patch on static road surface elevation and dynamic deflection in vehicle response inversion, and correct the high frequency distortion caused by single point contact; (2) How to establish a two-way closed-loop coupled dynamic relationship between vehicle-road-subgrade and accurately separate the static geometric elevation of the road surface from the transient dynamic deflection induced by vehicle load; (3) How to utilize the wheelbase delay consistency and lateral coherence among multiple wheel tracks of the whole vehicle to improve the stability of inversion under non-steady-state conditions such as speed change and lane change; (4) How to perform residual projection on the fourth-order spatial derivative partial differential equation (PDE) of road structure dynamics in the wavelet time-frequency domain to solve the training collapse problem of high-order PDE in physical information neural network; (5) How to enhance the identifiability of pavement structure parameters by sharing static pavement geometric constraints under conditions of multiple vehicles, multiple speeds, and multiple axle loads.

[0017] This application's embodiments are applied to intelligent sensing scenarios for road infrastructure. Utilizing vehicles (ordinary passenger cars or commercial vehicles) equipped with multi-channel dynamic response acquisition devices, vehicle response data is continuously collected at normal traffic flow speeds (e.g., 60-120 km / h). Through the coordinated application of three types of physical constraints—namely, constructing a three-layer physical constraint collaborative inversion framework of "tire-road envelope—vehicle-roadbed closed loop—multi-wheel track coherence"—these three types of physical constraints are embedded into a multi-branch physical information neural network for training. Based on this neural network, the following implementation is achieved: High-precision reconstruction of road surface static geometric elevation (i.e., smoothness); Separation and reconstruction of the spatiotemporal field of transient dynamic deflection of road surface induced by vehicle load; Continuous inversion of pavement structural parameters (such as bending stiffness, subgrade reaction coefficient, etc.).

[0018] Specifically, the tire surface domain envelope integral and the spatiotemporal coherence of multi-track wheelbase are used as physical boundary conditions and embedded into a vehicle-road-base closed-loop system controlled by a fourth-order PDE. Residual projection training is performed in the wavelet time-frequency domain to achieve the joint inversion of the road surface static geometric elevation, transient dynamic deflection and deep structural equivalent parameters at normal driving speed.

[0019] See Figure 1 , Figure 1 This is a flowchart of a road monitoring method provided in an embodiment of this application. The method includes the following steps: Step 101: Obtain the vehicle parameters of the first vehicle and the vehicle response data of the first vehicle during its travel on the first road. Step 102: Based on the vehicle response data and the vehicle parameters, determine the dynamic state of the first vehicle, the trajectory coordinates of each wheel in the first vehicle in the road coordinate system, and the effective road surface input of each wheel in the first vehicle to the first road. Step 103: Based on the dynamic state and the effective road surface input, determine the dynamic wheel loads applied to the road surface of the first road by each wheel of the first vehicle during driving using a first model characterizing vehicle dynamics. Step 104: Input the vehicle speed, trajectory coordinates, and dynamic wheel load from the vehicle response data into a neural network to perform road parameter inversion and obtain the road parameters of the first road. The neural network includes a first branch network, a second branch network, and a third branch network. The first branch network is used to determine the static geometric elevation of the road surface of the first road in the road parameters based on the trajectory coordinates. The second branch network is used to determine the transient dynamic deflection induced by the dynamic wheel load of each wheel of the first vehicle during driving based on the vehicle speed, trajectory coordinates, and dynamic wheel load. The third branch network is used to determine the structural parameters of the first road in the road parameters based on the trajectory coordinates. The target loss function of the neural network is constructed based on the residual of the target model. The target model includes the first model and the physical constraint model of the first model. The physical constraint model includes: a second model for characterizing the contact envelope between the tire and the road, a third model for characterizing the coupled dynamics between the vehicle and the road, and a fourth model for characterizing the spatiotemporal consistency between different tires. The second model is used to constrain the static geometric elevation and transient dynamic deflection of the road surface in the effective road surface input involved in the first model. The third model is used to constrain the transient dynamic deflection induced by structural parameters and dynamic wheel loads. The fourth model is used to constrain the static geometric elevation of the road surface.

[0020] It should be noted that the embodiments of this application are applied to road monitoring devices, which can be deployed on vehicles or can communicate with vehicles to obtain relevant vehicle data in real time.

[0021] In step 101, the road monitoring device can acquire vehicle parameters of the first vehicle, as well as vehicle response data of the first vehicle equipped with a multi-channel dynamic response acquisition device during its travel on the first road. Among these, The number of vehicles may include, but is not limited to, vehicle body mass, unsprung mass of each wheel, front suspension stiffness, rear suspension stiffness, suspension damping coefficient, tire vertical stiffness, tire damping coefficient, pitch moment of inertia, roll moment of inertia, distance from center of gravity to front axle, distance from center of gravity to rear axle, and half track width, etc.

[0022] In some embodiments, the multi-channel dynamic response acquisition device may be a sensor.

[0023] In some embodiments, Figure 2 For the sensor configuration of the first vehicle, such as Figure 2 As shown, a vertical acceleration sensor is installed at the center of gravity of the vehicle body with a sampling rate of 1000 Hz. Each of the four wheel hubs has an unsprung vertical acceleration sensor with a sampling rate of 1000 Hz. An inertial measurement unit (IMU) is installed on the vehicle body to provide pitch and roll angular velocities with a sampling rate of 500 Hz. The vehicle speed and four-wheel speed are obtained through the Controller Area Network (CAN) bus with a sampling rate of 100 Hz. A Global Positioning System (GPS) / Real-Time Kinematic (RTK) positioning device is installed with a sampling rate of 10 Hz.

[0024] In some embodiments, vehicle response data includes the following: Vertical acceleration of vehicle body The acceleration is obtained by an acceleration sensor installed at or near the vehicle's center of gravity; the unsprung vertical acceleration of four or more wheels. ( The acceleration data is obtained from acceleration sensors installed on each wheel hub or steering knuckle; among which, j =1 indicates the left front wheel hub. j =2 indicates the right front wheel hub. j =3 indicates the left rear wheel hub. j =4 indicates the right rear wheel hub, which can be expanded to more wheels as needed; vehicle pitch rate. The vehicle roll rate is obtained from an inertial measurement unit (IMU) or a gyroscope. Vehicle longitudinal speed is obtained from an IMU or gyroscope. The wheel speed is obtained from the vehicle's CAN bus, GPS, or wheel speed calculation; The data is obtained from the wheel speed sensor.

[0025] The sampling rate of each channel should meet the following requirements: no less than 500 Hz for acceleration and angular velocity channels, and no less than 100 Hz for vehicle speed and wheel speed channels.

[0026] In some embodiments, the vehicle response data may also include one or more of the following: Suspension dynamic deflection, tire pressure, tire temperature, contact pressure, axle load, GPS / RTK positioning data, IMU yaw rate, road surface temperature, and ambient temperature.

[0027] In some embodiments, after the vehicle response data is collected, the following preprocessing steps can be performed on the data from each channel: (1) Time synchronization can unify the data of each channel to the same time base; (2) Zero bias correction can eliminate static offset of the sensor; (3) Bandpass filtering can retain effective frequency bands related to road excitation and vehicle dynamic response. For example, it can retain effective frequency bands with a lower cutoff frequency of 0.5~2Hz and an upper cutoff frequency of 200~500Hz. (4) Outlier removal can be performed by detecting and removing outlier data points based on statistical thresholds or physical constraints; (5) Coordinate system transformation can transform the response data in the vehicle coordinate system to the road coordinate system.

[0028] In step 102, the trajectory coordinates of each wheel in the road coordinate system can be reconstructed based on vehicle response data and vehicle parameters. For example, the wheel track coordinates of each wheel in the road coordinate system can be reconstructed based on vehicle speed, vehicle attitude, and vehicle geometric parameters. The formulas for calculating the wheel track coordinates of each wheel are shown in equations (1) and (2) below: (1) (2) in, Indicates the first The longitudinal and lateral coordinates of each wheel in the road coordinate system, in meters; The coordinates of the vehicle's center of gravity in the road coordinate system are expressed in meters and are obtained from GPS / RTK positioning data or by combining wheel speed integral and IMU heading integral. The vehicle heading angle (rad) is obtained from the IMU yaw rate integral or GPS heading data. Indicates the first The geometrical offset (m) of each wheel relative to the vehicle's center of gravity is a known vehicle parameter determined by the wheelbase, track width, and center of gravity position.

[0029] In some embodiments, for scenarios without high-precision positioning data, wheel track coordinate reconstruction can be achieved through the following alternative methods: the longitudinal displacement of the vehicle can be obtained by using wheel speed integral, the heading change of the vehicle can be obtained by using IMU heading integral, and constrained fusion can be performed by combining road centerline map data, thereby reconstructing the trajectory coordinates of each vehicle.

[0030] In some embodiments, the dynamic state of the first vehicle can be determined based on vehicle response data, wherein the dynamic state includes... , and , express The first derivative with respect to time, express The second derivative with respect to time. , is a generalized coordinate vector that includes the vertical displacement of the vehicle body. Pitch angle yaw angle and the vertical displacement of the unsprung mass of the four wheels .

[0031] It should be noted that in all subsequent formulas, adding a dot above the variable symbol indicates the first derivative of that variable with respect to time, adding two dots above the variable symbol indicates the second derivative of that variable with respect to time, and so on.

[0032] Correspondingly, a first model can be constructed based on the vehicle's dynamic state and vehicle parameters. The first model represents the vehicle's physical characteristics and can be a multi-degree-of-freedom dynamic model of the whole vehicle, specifically used to characterize vehicle dynamics.

[0033] In some embodiments, the first model can be the 7-DOF coupled dynamic equations of the whole vehicle, as shown in equation (3).

[0034] (3) Meaning of each parameter: Generalized coordinate vector, including the vertical displacement of the vehicle body. Pitch angle yaw angle and the vertical displacement of the unsprung mass of the four wheels In practical applications, this refers to the dynamic state of the first vehicle. : 7×7 mass matrix, composed of vehicle body mass Pitch moment of inertia Tilt moment of inertia and the unsprung mass of each wheel The matrix is ​​constructed from vehicle parameters. : 7×7 damping matrix, derived from suspension damping coefficients This matrix is ​​also constructed from vehicle parameters; : 7×7 stiffness matrix, derived from suspension stiffness and tire stiffness This matrix is ​​also constructed from vehicle parameters; Tire force vector is the effective road surface input vector. and generalized coordinates The function; : Effective road surface input vectors for the four wheels.

[0035] For half-vehicle or single-sided wheel track scenarios, the first model described above can be degenerated into a 4-DOF half-vehicle model; for basic calibration scenarios, it can be degenerated into a 2-DOF quarter-vehicle model.

[0036] In some embodiments, the effective road surface input of each wheel of the first vehicle to the first road can be determined based on vehicle response data and vehicle parameters. Two methods can be provided to determine the effective road surface input.

[0037] Method A is direct observation / semi-direct observation. When the first vehicle is equipped with a suspension dynamic deflection sensor or a tire force sensor, the effective road surface input can be obtained by inversely deducing the tire force formula, as shown in the following formula (4).

[0038] (4) in, This refers to the vertical stiffness of the tire (N / m). These are the tire damping coefficients (N·s / m), and these two parameters are vehicle parameters. Obtained from unsprung acceleration through constrained integration, Kalman filtering, or a smoother, it represents vehicle response data. It can be measured by suspension dynamic deflection sensors or tire force sensors; it is the dynamic wheel load applied by the wheel to the road surface. This is the effective road surface input for the j-th wheel pair on the first road.

[0039] Method B is a latent variable joint optimization. When the effective road surface input is not directly measured, it can be used as a latent variable and solved jointly through moving window optimization, as expressed by the following formula (5).

[0040] (5) in, The data collected includes actual acceleration, angular velocity, wheel speed, and other observational data, which are used for vehicle response. For the observation equation, and These are the weighting coefficients. This is the temporal gradient regularization term for the effective road surface input, used to constrain the temporal smoothness of the effective road surface input.

[0041] In step 103, based on the dynamic state and the effective road surface input, the dynamic wheel loads applied to the road surface by each wheel of the first vehicle during driving can be determined using a first model characterizing vehicle dynamics. These dynamic wheel loads are related to the tire force vector, and the tire force vector can be determined using the above formula (3), thereby determining the dynamic wheel loads applied to the road surface by each wheel of the first vehicle during driving.

[0042] Subsequently, in step 104, the vehicle speed, trajectory coordinates, and dynamic wheel load from the vehicle response data can be input into the neural network to perform road parameter inversion, thereby obtaining the road parameters of the first road. The vehicle speed, trajectory coordinates, and dynamic wheel load from the vehicle response data serve as the data drivers for this neural network.

[0043] It should be noted that this neural network needs to be pre-trained to integrate the physical mechanisms of road inversion. This allows for a deep fusion of data-driven approaches and physical mechanisms during application. The training of the neural network will be explained in detail below.

[0044] The construction of the first vehicle dynamics needs to be subject to multiple physical constraints. For each physical constraint, a physical constraint model can be constructed, including the second model, the third model, and the fourth model.

[0045] The second model can characterize the physical characteristics of the wheel, specifically the nonlinear contact envelope relationship of the tire path. It indicates that there is a physical relationship between the tire contact patch shape and the tire force. It is the first layer of physical constraint of the first model, and is what makes the second model... The effective road surface input for each wheel satisfies the physical constraints of the following equation (6).

[0046] (6) Meaning of each parameter: : No. The effective road surface input for each wheel is the comprehensive road surface excitation signal after being filtered by the tire path envelope; The half-length (m) of the tire-road contact patch is determined by tire pressure and axle load. Envelope weighting function: describes the contribution of different locations within the contact patch to the effective road input, and is affected by vehicle speed. Contact pressure with the tire circuit modulation; : The integral variable within the contact patch, representing the spatial offset (m) of each point on the contact patch relative to the contact center. : No. The travel speed of each wheel (m / s) is used to modulate the spatial broadening properties of the envelope kernel function; : No. The tire contact pressure (Pa) of each wheel is used to modulate the shape and amplitude of the envelope kernel function; : Static geometric elevation of road surface (m), representing the inherent uneven profile of the road surface, which is independent of time and only changes with spatial coordinates; Transient dynamic deflection induced by vehicle load (m) represents the transient vertical deformation of the road surface under the action of vehicle dynamic wheel load; : No. The longitudinal and lateral coordinates (m) of each wheel in the road coordinate system.

[0047] In some embodiments, the envelope weight function The normalized velocity-pressure modulated Gaussian kernel function is used, as shown in equation (7) below.

[0048] (7) Among them, the envelope width parameter Modulated by vehicle speed and tire pressure, it is expressed by the following formula (8).

[0049] (8) The half-length of the contact patch is modulated by the tire pressure, and is expressed by the following formula (9).

[0050] (9) Meaning of each parameter: The initial contact patch equivalent width parameter (m) is determined by the tire specifications; Speed ​​modulation coefficient (s) describes the linear rate of change of the envelope width with vehicle speed; Pressure modulation coefficient (m / kPa) describes the linear rate of change of the envelope width with tire pressure; Nominal tire pressure (kPa) is the tire inflation pressure under reference operating conditions. : Half length of contact patch under nominal tire pressure (m).

[0051] This is a contact pressure shape function used to describe the shape characteristics of the pressure distribution within the contact patch. It can be selected from one of the following three forms: Uniformly distributed type: Assuming uniform contact pressure distribution; parabolic: Assuming the contact pressure follows a parabolic distribution; Hertz type: Hemispherical pressure distribution based on Hertz contact theory.

[0052] The above envelope parameters , , , The parameters can be determined by one or a combination of the following methods: tire bench test calibration; joint calibration of known road profile and vehicle response; joint estimation when the vehicle repeatedly passes through the same road segment at multiple speeds; using parameters provided by the tire manufacturer as initial values ​​and adding prior constraints in the neural network training.

[0053] In this embodiment, during tire-road contact modeling, the static geometric elevation of the road surface can be established simultaneously. and transient dynamic bending A nonlinear integral envelope model is constructed, with the envelope kernel modulated by vehicle speed and contact pressure. This envelope integral forms the basis for subsequent static-dynamic decoupling.

[0054] The third model can characterize the physical characteristics of the road, specifically the coupled dynamics between the vehicle and the road. It is the physical constraint of transient dynamic deflection and can establish the vehicle-road-base coupled dynamic equation satisfied by transient dynamic deflection. It is the second layer of physical constraint of the first model.

[0055] In some embodiments, the third model may be a one-dimensional strip beam model, which is a PDE equation, represented by the following equation (10).

[0056] (10) Meaning of each parameter: Equivalent bending stiffness of the road surface (N·m²) reflects the comprehensive ability of the road surface to resist bending deformation. It is a structural parameter of the road and one of the target parameters for inversion. Transient dynamic deflection Spatial coordinates The fourth-order partial derivative describes the rate of change of the road surface curvature; Equivalent mass per unit length of road surface (kg / m), reflecting the inertial characteristics of the road surface structure; Transient dynamic deflection Regarding time The second-order partial derivative, i.e. the acceleration term of transient dynamic deflection; The equivalent damping coefficient of the pavement-subsoil structure (N·s / m²) reflects the energy dissipation capacity of the pavement structure during vibration. It is a structural parameter of the road and one of the target parameters for inversion. Transient dynamic deflection Regarding time The first-order partial derivative, i.e. the velocity term of transient dynamic deflection; : Subgrade reaction coefficient (N / m²), also known as Winkler subgrade modulus, reflects the elastic support capacity of the subgrade to the road slab. It is a structural parameter of the road and one of the target parameters for inversion. : Transient dynamic deflection (m) induced by vehicle load is an unknown solution quantity for PDE; Pasternak shear layer parameter (N) reflects the shear coupling effect between foundation soil layers. It is a structural parameter of the road and one of the target parameters for inversion. Transient dynamic deflection Spatial coordinates The second-order partial derivative describes the curvature of the road surface; : No. The dynamic wheel load (N) applied to the road surface by each wheel is calculated based on the vehicle dynamics state and the effective road surface input; Distributed load kernel function, which distributes the centralized load in space according to a certain width (determined by parameters). (Control) Distribute the load distribution to avoid numerical singularities caused by concentrated loads; : No. The position (m) of each wheel in the longitudinal coordinate of the road changes over time and as the vehicle travels.

[0057] In some embodiments, the distributed wheel-loaded kernel function The following Gaussian form is used, expressed by equation (11).

[0058] (11) in, The wheel load distribution width parameter (m) can be within a certain range. That is, half to twice the length of the contact patch.

[0059] For scenarios with three-dimensional surfaces and horizontal expansion, the above one-dimensional strip beam model can be extended into a two-dimensional equivalent plate model, represented by the following equation (12).

[0060] (12) in, Let be the bending stiffness of the plate. The elastic modulus of the road surface material. For the thickness of the road slab, Poisson's ratio; and These are the biharmonic operator and the Laplace operator, respectively.

[0061] In some embodiments, initial conditions for the road under transient dynamic deflection and boundary conditions for the road ends under transient dynamic deflection can be constructed as the initial and boundary conditions of the PDE.

[0062] Initial conditions: For a length of The section of road, The initial condition is set to zero deflection. , Alternatively, the output of the previous time window can be used as the initial value for the current time window.

[0063] Boundary conditions: at both ends of the road segment and Approximate absorbing boundary conditions are set at the point to suppress boundary reflection, as shown in equations (13), (14) and (15) respectively.

[0064] (13) (14) (15) in, The boundary damping coefficient is given by formula (13), which indicates that there is no bending deformation of the road surface at both ends. Formulas (14) and (15) represent the dynamic parameters of the road surface at both ends without bending deformation and without bending deformation, respectively. In engineering implementation, a buffer zone can be extended at both ends of the detection section, and only the inversion results of the middle effective interval can be output to reduce boundary errors.

[0065] Degradation model: when When the above PDE can degenerate into a Winkler foundation beam model; when the pavement deflection is extremely small (e.g., When the vehicle load is small or the roadbed stiffness is high, the above PDE can degenerate into a rigid road envelope inversion model. In this case, the effective road surface input is approximately an envelope integral that only includes the static geometric elevation of the road surface.

[0066] The fourth model can be a multi-track spatiotemporal consistency constraint, which is the third layer of physical constraints of the first model.

[0067] In some embodiments, the fourth model may include a first sub-model, which is used to characterize the time delay consistency of the front and rear wheel axle distances on the same side of the second vehicle. That is, the first sub-model is a time delay consistency constraint of the front and rear wheel axle distances on the same side, and is represented by the following formula (16).

[0068] (16) Meaning of each parameter: Static geometric elevation of the road surface (m) represents the inherent uneven profile of the road surface and is independent of time. The rear wheel is in the... The longitudinal coordinate (m) of the side. This is the moment when the rear wheel passes through the same road surface position as the front wheel; The time interval (in seconds) between the front and rear wheels passing the same road surface position, based on the wheelbase. and vehicle speed Determined, that is ; The rear wheel is in the... The lateral coordinate (m) of the side corresponds to the lateral position of the rear wheel's track. The front wheel is in the... The vertical coordinate (m) of the side, at the current time The position of the front wheels; The front wheel is in the... The lateral coordinate (m) of the side corresponds to the lateral position of the wheel track of the front wheel.

[0069] This constraint states that the rear wheel at time... The observed static road surface geometric elevation should be consistent with the front wheel's position at time [time]. The elevations at the same observed locations are consistent. The time delay residual can be defined as the axis distance time delay loss. During the neural network training process, it is the residual of the first sub-model, expressed by the following formula (17).

[0070] (17) Among them, time delay The precise calculation formula is the integral along the wheel track path, expressed by the following formula (18).

[0071] (18) When the velocity changes slowly, it can be approximated as .

[0072] In some embodiments, the fourth model may also involve lateral coherence spectrum constraints of the left and right wheel tracks, representing the static geometric elevation of the road surface at spatial frequencies of the left and right wheel tracks. The following equation (19) defines the coherence function for estimating the left and right wheel tracks, which has statistical coherence.

[0073] (19) in, The cross-power spectral density is the static geometric elevation of the left and right wheel tracks on the road surface. and These are the self-power spectral densities of the left and right wheel tracks, respectively.

[0074] In some embodiments, the coherence function may take the form of exponential decay, as expressed by the following equation (20).

[0075] (20) in, Let be the target coherence attenuation coefficient, which describes the rate of attenuation of the coherence of the left and right wheel tracks with frequency; Wheelbase (m); To avoid small constants that divide by zero, the corresponding coherent residual is: .

[0076] In addition, the fourth model can also involve multiple vehicles sharing static road surface constraints. For scenarios where multiple vehicles (or the same vehicle multiple times) pass through the same road segment, the static road surface geometric elevation observed by all vehicles should be the same, as expressed by the following formula (21).

[0077] (twenty one) It should be noted that the transient dynamic deflection of each vehicle and dynamic wheel load They are independent of each other and vary depending on vehicle speed, axle load, tire parameters, and suspension parameters.

[0078] In this embodiment, wheelbase time delay consistency constraints and lateral coherence spectrum constraints can be established. The wheelbase time delay consistency constraints represent the static consistency of the front and rear wheels passing through the same road surface position, while the lateral coherence spectrum constraints represent the statistical coherence of the left and right wheel tracks. This can enhance the stability and identifiability of road structure parameter inversion under conditions of speed change, lane change, and multiple vehicles.

[0079] It should be noted that the three layers of physical constraints do not act independently, but rather form a synergistic mechanism, specifically as follows: The first layer (envelope) provides accurate boundary conditions for the second layer (PDE). Only when the spatial filtering effect of the tire contact patch is correctly modeled can the static geometric elevation component and transient dynamic deflection component of the effective pavement input be distinguished, and only then can the wheel load excitation in the PDE be accurate. The second layer (PDE) provides transient dynamic deflection estimation for the first layer (envelope), and the transient dynamic deflection solved by the PDE... Feedback is fed back to the envelope integral to correct the effective road surface input, forming a closed loop. Without this closed loop, the transient dynamic deflection component would be mistakenly attributed to the road surface static geometric elevation. The third layer (multi-wheel tracks) provides global consistency constraints. Even if the combination of the envelope of a single wheel track and the PDE has multiple solutions locally, the physical consistency of the multi-wheel tracks (the front and rear wheels must see the same road surface, and the left and right wheels must be statistically coherent) can eliminate unreasonable solutions and enhance the uniqueness of the inversion.

[0080] In some embodiments, step 103 specifically includes: Based on the dynamic state of the first vehicle at the first moment and the effective road surface input, the dynamic wheel load of each wheel of the first vehicle on the road surface at the first moment is determined by the first model, and the static geometric elevation and transient dynamic deflection of the road surface are determined by the second model to determine the effective road surface input. Based on the transient dynamic deflection induced by the dynamic wheel load of each wheel in the first vehicle at the first moment, the second model is used to correct the effective road surface input of each wheel in the first vehicle to the first road at the second moment, and the transient dynamic deflection is constrained by the third model. Based on the corrected effective road surface input and the dynamic state of the first vehicle at the second moment, the dynamic wheel load of each wheel of the first vehicle on the road surface at the second moment is determined using the first model.

[0081] Since transient dynamic deflection is constrained by PDE, when the neural network is trained based on the residual of the third model, data-driven and physical mechanisms can be integrated, so that dynamic wheel load determines transient dynamic deflection through PDE.

[0082] In other words, a two-way closed-loop coupling can be established between the PDE and the Ordinary Differential Equation (ODE) of vehicle dynamics. Specifically, the static geometric elevation and transient dynamic deflection of the road surface determine the effective road input through the envelope integral. The effective road input determines the dynamic wheel load through vehicle dynamics, and the dynamic wheel load determines the transient dynamic deflection through the PDE. The transient dynamic deflection then feeds back to the envelope integral, thus forming a two-way closed loop: effective road input → dynamic wheel load → transient dynamic deflection → effective road input. Unlike the unidirectional road input model or purely numerical simulation methods in related technologies, the constructed first, second, and third models can achieve the physical separation of the static geometric elevation and transient dynamic deflection of the road surface.

[0083] A multi-branch neural network can be constructed, including a first branch network, a second branch network, and a third branch network. The first branch network can be called the static geometry branch, which is used to predict the static geometric elevation of the road surface. The second branch network can be called the dynamic deflection branch, which is used to predict transient dynamic deflection. The third branch network can be called the structural parameter branch, which is used to predict the structural parameters of the road.

[0084] Input / output specifications for each branch: Static geometry branch The input is the trajectory coordinates of each wheel of the first vehicle in the road coordinate system. The output is the static geometric elevation of the road surface. When using a one-dimensional strip model, the input is the vertical coordinate. The output is the static geometric elevation of the road surface based on the left and right wheel tracks. Dynamic deflection branch The input is the trajectory coordinates of each wheel of the first vehicle in the road coordinate system. ,time Vehicle speed and dynamic wheel load The output is transient dynamic deflection. When using a one-dimensional strip model, the input is... The output is ; Structural parameter branches The input is the trajectory coordinates of each wheel of the first vehicle in the road coordinate system. or The output is structural parameters, including equivalent bending stiffness. Foundation reaction coefficient Pasternak shearing parameters and equivalent damping One or more of them.

[0085] In this embodiment, a three-branch network structure is used to simultaneously output the static geometric elevation of the road surface, transient dynamic deflection, and structural equivalent parameters during a single training process, achieving "one trip, three outputs".

[0086] In some embodiments, to ensure the physical rationality (positive value and boundedness) of the structural parameters, each structural parameter adopts a bounded mapping, represented by the following equations (22) and (23).

[0087] (twenty two) (twenty three) In some embodiments, softplus mapping can also be used. .in, For the Sigmoid function, , This is the original output from the network. , , , The preset parameter value range, A small constant to ensure positiveness.

[0088] In some embodiments, the input coordinates of each branch of the neural network can be Fourier feature encoded to enhance the ability to represent high-frequency road surface features and dynamic deflection, as expressed by the following equation (24).

[0089] (twenty four) in, The Fourier encoding order controls the frequency range of the input coordinates. The larger the value, the stronger the network's ability to represent high-frequency components.

[0090] In some embodiments, the target loss function of the neural network can be directly constructed based on the residuals of the first model, the residuals of the second model, the residuals of the third model, and the residuals of the fourth model.

[0091] In some embodiments, the loss constructed based on the residuals of the fourth model may include a distance delay loss and a transverse coherence spectrum loss.

[0092] In some embodiments, the target loss function further includes at least one of the following: Based on the loss of boundary conditions at both ends of the second road for transient dynamic deflection, the boundary conditions are used to indicate the dynamic parameters of the road surface without bending deformation and without bending deformation at both ends of the road, and the boundary conditions are used to constrain the transient dynamic deflection induced by dynamic wheel load. The regularization loss of the road parameters of the second road; The prior constraint loss of the target parameters, which include the vehicle parameters of the second vehicle, the parameters involved in the contact envelope between the tires of the second vehicle and the second road, and the structural parameters of the second road.

[0093] In some embodiments, the target loss function can be represented by the following equation (25).

[0094] (25) The definitions of each loss are as follows: (1) Vehicle dynamics consistency loss, based on the residuals of the first model, is expressed by the following equation (26).

[0095] (26) (2) The tire path envelope consistency loss can be constructed based on the residuals of the second model if there is a valid input of observation or estimation, and is expressed by the following equation (27).

[0096] (27) (3) PDE residual loss, based on the residual construction of the third model, is expressed by the following equation (28).

[0097] (28) (4) Wheelbase delay loss, based on the residual of the first sub-model, is expressed by the following equation (29).

[0098] (29) (5) The transverse coherence spectrum loss is constructed based on the residuals constrained by the transverse coherence spectrum of the left and right wheel tracks, and is expressed by the following formula (30).

[0099] (30) (6) Boundary condition loss, expressed by the following formula (31).

[0100] (31) (7) Regularization loss, expressed by the following formula (32).

[0101] (32) in, is the regularization coefficient, which controls the spatial smoothness constraint strength of each output quantity.

[0102] (8) The prior constraint loss is expressed by the following formula (33).

[0103] (33) in, This includes the current values ​​of vehicle parameters, tire path envelope parameters (i.e., parameters involved in the contact envelope), and structural parameters. Let be the prior mean of each parameter. Let be the prior standard deviation of each parameter.

[0104] In some embodiments, at least a portion of the residuals of the target model can be projected onto the wavelet time-frequency or spatiotemporal scale domain to construct the target loss function. Prior to step 101, the method further includes: Based on the road parameters of the second road obtained by the neural network inversion, the residual of the target model is determined; The target residual in the residual of the target model is projected onto the wavelet time frequency to obtain the wavelet frequency residual norm. The target residual includes the residual of the third model, which involves the higher-order partial derivatives of transient dynamic deflection with respect to spatial coordinates. Based on the wavelet frequency residual norm and the residuals of the target model other than the target residual, the target loss function of the neural network is constructed. The neural network is trained based on the target loss function.

[0105] In some embodiments, the target residual may include the residual of a third model.

[0106] Time signals can be used or spatial signal Perform continuous wavelet transform, expressed by the following equation (34).

[0107] (34) in, For the mother wavelet function, For scale parameters, The translation parameter is used. The mother wavelet can be the Morlet wavelet, the Mexican Hat wavelet, or other analytic wavelets with sufficient derivative order.

[0108] The fourth spatial derivative term in the PDE can be calculated using wavelet derivative kernels instead of relying directly on higher-order automatic differentiation, and is expressed as Equation (35).

[0109] (35) in, Using the fourth derivative of the mother wavelet, this method avoids the high-frequency gradient explosion problem in fourth-order automatic differentiation. In other words, wavelet domain projection can solve the training bottleneck problem of fourth-order PDEs, since the fourth spatial derivative leads to gradient explosion in direct automatic differentiation. Calculations are performed in the wavelet domain, which can form a regularization offset with the low-frequency delay residuals of the multi-round trace, thus resolving the training collapse of high-order PDEs in PINN.

[0110] For any residual The wavelet domain residual norm is defined as follows (36).

[0111] (36) in, This is a scale-location weighting function used to control the contribution of residuals at different scales and locations.

[0112] Based on the aforementioned wavelet domain residual norm, a target loss function can be constructed. In other words, the aforementioned PDE residual loss can be replaced with the wavelet domain residual norm to solve the training bottleneck problem of fourth-order PDE.

[0113] A neural network can be trained based on a target loss function, and an adaptive weighting strategy can be used to balance the various losses. In some embodiments, the method further includes: During the training of the neural network, an adaptive update strategy based on neural tangent kernel traces is used to update the weights of each loss term in the target loss function.

[0114] That is, the weights of each loss term can be assigned. An adaptive update strategy based on the Neural Tangent Kernel (NTK) trace is adopted, which is expressed by the following equation (37).

[0115] (37) in, For the first The approximate matrix of the neural tangent kernel corresponding to each loss term For matrix trace operations, The total number of loss items, It is a stable term. Every [period] The weights are updated once every training epoch to balance the contributions of different physical constraints during training.

[0116] In some embodiments, training can employ a two-stage strategy, firstly pre-training using the Adam optimizer with a learning rate set to a certain value. This reduces the magnitude of the total loss quickly; then switch to the L-BFGS optimizer for fine-tuning, where the learning rate can be set to [value missing]. This allows for a finer search in the parameter space, improving convergence accuracy.

[0117] In some embodiments, the neural network is jointly trained based on M second vehicles. The target loss function is the sum of a first loss and a second loss. The first loss is superimposed with the independent loss of each second vehicle. The independent loss includes a loss constructed based on the residuals of the first model, a loss constructed based on the residuals of the second model, a loss constructed based on the residuals of the third model, and a wheelbase delay loss. The wheelbase delay loss is constructed based on the residuals of the first sub-model. The fourth model includes the first sub-model, which is used to characterize the consistency of wheelbase delay between the front and rear wheels on the same side of the second vehicle. The second loss is superimposed with a shared loss. The shared loss includes the regularization loss of the road parameters of the second road, the prior constraint loss of the target parameters, and the lateral coherence spectrum loss. The lateral coherence spectrum loss is used to indicate the statistical coherence of the static road surface geometric elevation of the left and right wheels of the second vehicle at spatial frequencies. M is an integer greater than 1.

[0118] For multi-vehicle joint training scenarios, the objective loss function The sum of the independent losses of each second vehicle plus the shared losses is expressed by the following formula (38).

[0119] (38) in, This indicates the number of the second vehicle.

[0120] Because multiple vehicles share the same static geometric elevation of the road surface and structural parameters , , , However, each vehicle has independent transient dynamic deflection. Dynamic wheel load and tire envelope parameters Therefore, the independent losses for each vehicle include vehicle dynamics consistency loss, tire path envelope consistency loss, PDE residual loss, and wheelbase delay loss, while the shared losses include regularization loss, prior constraint loss, and lateral coherence spectrum loss.

[0121] In this embodiment, multiple vehicles share the same static road geometry, but each has its own independent transient dynamic deflection and dynamic wheel load. The diversity of different speeds, axle loads, and tire parameters can enhance the stability and uniqueness of structural parameter inversion.

[0122] After the neural network is trained, in practical applications, it can take trajectory coordinates, vehicle speed, and dynamic wheel load as input in real time, and output: Geometric state: static geometric elevation of the road surface, such as... , or three-dimensional surface layer And the roughness indices calculated from them, including the International Roughness Index (IRI), the number of rides (RN), and the longitudinal profile elevation; structural response: the spacetime field of transient dynamic deflection. Including dynamic deflection peak value, deflection basin shape characteristics, deflection recovery time, and wheel load influence length; structural parameters: an equivalent parameter of the pavement structure, including equivalent bending stiffness. Foundation reaction coefficient Pasternak shearing parameters and equivalent damping It can output the above four structural parameters under the condition of multiple vehicles working together.

[0123] In some embodiments, after step 104, the method further includes: Based on the road parameters of the first road, a health assessment is performed on the first road to obtain health evaluation index information of the first road.

[0124] In some embodiments, a health assessment of the first road can be performed based on the road parameters of the first road using a deep learning model or other models.

[0125] The health assessment indicators can include road sections with abnormal structural stiffness, locations of weak roadbed layers, locations of excessive dynamic deflection, locations of abnormal smoothness, and maintenance priorities. The results of these health assessment indicators can be visualized to support decision-making.

[0126] The following is a detailed explanation of the road monitoring method provided in the embodiments of this application using a specific example.

[0127] The core of this application's embodiments is the construction of a three-layer physical constraint collaborative inversion framework: "tire track envelope—vehicle-roadbed closed loop—multi-wheel track coherence." This framework embeds these three types of physical constraints into a multi-branch neural network and performs residual projection training in the wavelet time-frequency domain. The system framework for implementing the road monitoring method is as follows: Figure 3 As shown, its implementation flowchart is as follows: Figure 4 As shown.

[0128] The system includes a data acquisition layer, a data preprocessing layer, a three-layer physical constraint layer, a wavelet domain projection layer, a neural network inversion layer, and an output layer.

[0129] The multi-channel vehicle dynamic response acquisition module is located in the data acquisition layer. It is used to collect multi-channel dynamic response data of vehicles during road driving, including at least: vehicle vertical acceleration, unsprung vertical acceleration of four or more wheels, vehicle pitch rate, vehicle roll rate, vehicle longitudinal speed, and wheel speed. It can also collect suspension dynamic deflection, tire pressure, tire temperature, contact pressure, axle load, GPS / RTK positioning data, IMU yaw rate, road surface temperature, and ambient temperature. The data from each channel can undergo preprocessing such as time synchronization, zero-bias correction, bandpass filtering, outlier removal, and coordinate system transformation.

[0130] The wheel track coordinate reconstruction module is located in the data preprocessing layer. It can calculate the trajectory coordinates of each wheel in the road coordinate system based on vehicle speed, wheel speed, IMU, and positioning data. For scenarios without high-precision positioning data, constraint fusion is performed using wheel speed integral, IMU heading integral, and road centerline map.

[0131] The vehicle multi-DOF dynamics modeling and state observation module is also located in the data preprocessing layer. It can establish a 7-DOF coupled dynamics model of the vehicle, i.e., the first model, with the 7 degrees of freedom being vertical, pitch, roll, and four unsprung masses. For effective road surface input, two implementation methods can be provided: Method A is to directly observe and back-calculate through suspension dynamic deflection or tire force sensors; Method B is to treat the effective road surface input as a latent variable and solve it jointly through a state-space observer or moving window optimization.

[0132] The tire track nonlinear contact envelope modeling module is located in the first layer of the three-layer physical constraint layer. It can establish a nonlinear integral model, i.e., the second model, that simultaneously applies envelope filtering to the tire contact patch on both static geometric elevation and transient dynamic deflection. The envelope kernel function is modulated by vehicle speed and tire track contact pressure, and includes a Gaussian attenuation term and a contact pressure shape function term. This module corrects the high-frequency distortion caused by the single-point contact assumption and is a prerequisite for decoupling static elevation and dynamic deflection.

[0133] The vehicle-road-subgrade closed-loop coupled dynamic constraint module is located in the second layer of the three-layer physical constraint layer. It can establish the Pasternak viscoelastic foundation beam partial differential equation satisfied by transient dynamic deflection, i.e., the third model, which includes the fourth-order spatial derivative, inertial term, damping term, elastic foundation term and shear term.

[0134] The first and second layers of physical constraints form a closed-loop coupling relationship. Specifically, the static geometric elevation and transient dynamic deflection of the road surface determine the effective road surface input through envelope integral. The effective road surface input determines the dynamic wheel load through vehicle dynamics. The dynamic wheel load determines the transient dynamic deflection through PDE. The transient dynamic deflection is then fed back to the envelope integral to correct the effective road surface input, forming a two-way closed loop. This module can achieve the physical separation of the static geometric elevation and transient dynamic deflection of the road surface.

[0135] The multi-track spatiotemporal coherence constraint module is located in the third layer of the three-layer physical constraint layer. It can establish two types of spatiotemporal consistency constraints: first, the same-side front and rear wheel axle distance time delay consistency constraint, indicating that the observations of the front and rear wheels at the same static road surface geometry should be consistent, used to suppress drift under variable speed conditions; second, the left and right wheel track lateral coherence spectrum constraint, indicating that the road surface elevation of the left and right wheel tracks has statistical coherence within a certain frequency band, used to enhance spatial consistency. It can also be extended to allow multiple vehicles to share static road surface geometry constraints, enhancing the stability of structural parameter inversion.

[0136] The wavelet time-spectrum domain physical operator projection module is located in the wavelet domain projection layer. It can perform wavelet transform multi-scale projection on the residuals of the above three types of physical constraints. Its key role is to calculate the fourth-order spatial derivative term in the PDE using the wavelet derivative kernel (instead of automatic differentiation) to avoid high-frequency gradient explosion. At the same time, it forms a regularization mechanism with the multi-round trace time delay residual (constraining low-frequency consistency) to solve the training instability problem of the fourth-order PDE in PINN.

[0137] The multi-branch physical constraint neural network inversion module is located in the neural network inversion layer. The neural network contains three branches: a static geometry branch (input trajectory coordinates, output static geometric elevation of the road surface), a dynamic deflection branch (input trajectory coordinates, time, velocity, and dynamic wheel load, output transient dynamic deflection), and a structural parameter branch (input trajectory coordinates, output an equivalent parameter of the road surface structure, including equivalent bending stiffness). Foundation reaction coefficient Pasternak shearing parameters and equivalent damping The structural parameters are guaranteed to be positive and bounded through Sigmoid or softplus mapping. Input coordinates can be Fourier feature encoded to enhance high-frequency representation capabilities.

[0138] The structural parameter prior constraint and adaptive weight training module is also located in the neural network inversion layer. It can use an adaptive update strategy based on the NTK trace for the weights of each loss term to balance the contributions of different physical constraints during training. The structural parameters are constrained by the prior range (vehicle parameters, tire path envelope parameters, and prior values ​​of structural parameters), and the parameter constraints are achieved through the prior loss term.

[0139] The pavement condition, structural parameters and health index output module is located in the output layer. It can output the pavement static geometric elevation (including smoothness indexes such as IRI), transient dynamic deflection spatiotemporal field (including deflection peak value, deflection basin characteristics, and deflection recovery time), a structural equivalent parameter, and health evaluation indexes (road sections with abnormal structural stiffness, location of weak subgrade, maintenance priority, etc.).

[0140] The modules are interconnected, specifically: the data acquisition layer output provides input to the data preprocessing layer, the data preprocessing layer provides observation data to the three-layer physical constraint layer, the three-layer physical constraint layer transmits multi-scale residuals to the neural network through the wavelet domain projection module, the target loss function of the neural network inversion layer is composed of the residuals of the physical constraints, there is a bidirectional feedback loop between the vehicle-road-subgrade closed-loop coupled dynamic constraint module and the tire-road nonlinear contact envelope modeling module, and finally the output module summarizes the inversion results.

[0141] The embodiments of this application have the following technical effects.

[0142] (1) Break through the inherent bottlenecks in speed and accuracy in the industry and achieve an order-of-magnitude improvement in detection efficiency.

[0143] Traditional high-precision structural load-bearing capacity testing (such as FWD) has an efficiency of only a few kilometers per hour and requires traffic closure. The embodiments of this application support joint inversion of road parameters for vehicles at normal traffic flow speeds of 60~120 km / h, realizing a leap from "closed traffic, fixed-point low-speed testing" to "open traffic, full-network full-speed continuous monitoring", improving testing efficiency by more than an order of magnitude.

[0144] (2) Achieve precise decoupling between static geometric elevation and transient dynamic deflection of the road surface.

[0145] Traditional vehicle-mounted vibration inversion methods often result in road surface unevenness detection errors exceeding 15% on heavily loaded road sections due to the inability to separate transient dynamic deflection. This application's embodiment, through the collaborative constraint of "tire-road envelope—vehicle-roadbed closed loop," achieves precise separation between the road surface's static geometric elevation and the transient dynamic deflection induced by vehicle loads under high-speed conditions. This reduces the road surface's static geometric elevation reconstruction error to below 3%, while simultaneously making it possible to continuously invert deep equivalent parameters of the road surface solely through vehicle response.

[0146] (3) It can realize the continuous inversion of the deep structural parameters of the road surface under high-speed vehicle conditions.

[0147] This application's embodiments not only output road surface smoothness indicators, but also continuously invert the equivalent bending stiffness of the road surface at normal driving speeds. Foundation reaction coefficient The parameters of the same depth structure fill the technical gap in the state assessment of high-speed continuous structural layers.

[0148] (4) Multi-round track coherence eliminates inversion drift and significantly improves robustness under complex working conditions.

[0149] Traditional single-wheel or quarter-car models are prone to divergence under speed change and lane change conditions. The embodiments of this application introduce the wheelbase delay residual and the lateral coherence spectrum residual with physical constraints. Under non-steady driving conditions such as vehicle acceleration, deceleration and lane change, the spatial consistency of the inversion results remains stable, which greatly improves the generalization ability and engineering practical value of the system.

[0150] (5) Solve the training crash problem of fourth-order PDE in PINN.

[0151] By leveraging the mutual constraints between the multi-round track delay residuals (i.e., the wheelbase delay loss, which constrains low-frequency static consistency) and the fourth-order derivative projection in the wavelet domain (which suppresses high-frequency noise in higher-order derivatives), a computational path for the joint inversion of higher-order dynamic partial differential equations under on-board stochastic excitation can be established.

[0152] The following example illustrates the implementation process of road monitoring methods.

[0153] Example 1: Joint Inversion of Static Geometric Elevation and Transient Dynamic Deflection of Single-Vehicle Asphalt Pavement like Figure 5 As shown in the example, this embodiment illustrates how to jointly invert the static geometric elevation and transient dynamic deflection of a single passenger vehicle traveling at a constant speed on an asphalt road surface.

[0154] 1. Implementation conditions The vehicle parameters are shown in Table 1 below.

[0155] Table 1 Vehicle Parameter Table

[0156] Sensor configuration: A vertical acceleration sensor (sampling rate 1000Hz) is installed at the center of gravity of the vehicle body; an unsprung vertical acceleration sensor (sampling rate 1000Hz) is installed at each of the four wheel hubs; an IMU is installed on the vehicle body to provide pitch and roll angular velocities (sampling rate 500Hz); vehicle speed and four-wheel wheel speed are obtained via CAN bus (sampling rate 100Hz); and a GPS / RTK positioning device (sampling rate 10Hz) is installed.

[0157] Road conditions: 500m asphalt pavement, pavement structure from top to bottom is 4cm SMA-13 ​​modified asphalt surface layer, 6cm AC-20 intermediate layer, 8cm AC-25 base layer, cement stabilized crushed stone base course, and subgrade.

[0158] Driving speed: 80km / h (approximately 22.2m / s), constant speed.

[0159] 2. Data Acquisition and Preprocessing The vehicle travels at a constant speed of 80 km / h through a 500m section of road, with a data collection time of approximately 22.5 seconds. Data from each channel undergoes the following preprocessing: Time synchronization: Based on the PPS second pulse of GPS, all sensor data are unified to the same time base, with a synchronization accuracy better than 0.5ms; Zero bias correction: The average value of each channel when the vehicle is stationary is taken as the zero bias and subtracted from the dynamic data; Bandpass filtering: Apply Butterworth bandpass filtering of 0.5~400Hz to acceleration and angular velocity signals to remove static drift and noise above the Nyquist frequency; Outlier removal: Outlier data points are detected and removed based on the 3σ criterion, and replaced by linear interpolation of adjacent sampling points; Coordinate transformation: Transforming the response data in the vehicle coordinate system using the heading angle. Convert to the road coordinate system.

[0160] 3. Wheel track coordinate reconstruction Based on the vehicle's center of gravity trajectory and heading angle Calculate the wheel track coordinates of the four wheels. Taking the left wheel as an example, the calculation formula is expressed by the following formulas (39) and (40).

[0161] (39) (40) The other three wheels are similar, only and The values ​​are different. This embodiment involves uniform linear travel, therefore... A constant, the lateral distance between the left and right wheel tracks is approximately equal to the wheelbase. m.

[0162] 4. Vehicle dynamics modeling and effective road surface input construction Establish a 7-DOF vehicle dynamics model, with generalized coordinate vectors. A 7×7 mass matrix is ​​constructed based on the above vehicle parameters. Damping matrix and stiffness matrix .

[0163] This embodiment uses method B (latent variable joint optimization) to construct an effective road surface input: As latent variables, they are jointly solved using a moving window optimization method. In the optimization objective function, the observed data... Including vehicle body acceleration (by (Integral verification), unsprung acceleration of the four wheels, pitch rate, and roll rate; weighting coefficients are taken as follows: , .

[0164] 5. Nonlinear contact envelope modeling of tire tracks The Gauss-Hertz kernel function is adopted, which is a composite form of the Gaussian envelope term and the Hertz contact pressure distribution, and is expressed by the following equation (41).

[0165] (41) The envelope parameter values ​​are shown in Table 2 below.

[0166] Table 2 Envelope Parameter Table

[0167] exist m / s Under kPa conditions, m, m.

[0168] 6. Vehicle-Road-Subgrade Closed-Loop Coupled PDE Modeling Establish a Pasternak viscoelastic foundation beam model. Initial conditions are set to zero deflection: , Absorbing boundary conditions are set at both ends, as expressed by the following equations (42) to (44).

[0169] (42) (43) (44) The distributed round-robin kernel function adopts a Gaussian form. m (equal to) Extend the buffer zone by 50 m at both ends of the road segment. The actual calculation interval is 600 m. Output the results for the middle 500 m effective interval.

[0170] Among them, effective road surface input Includes static geometric elevation and transient dynamic deflection of the road surface → Determines dynamic wheel load through vehicle dynamics → Determining transient dynamic deflection using PDE → Transient dynamic deflection is fed back to the envelope integral correction of the effective road surface input, thus forming a closed loop.

[0171] 7. Spatiotemporal consistency constraints for multi-round tracks Establish two types of constraints: (1) Consistency constraint on the time delay of the front and rear wheel track on the same side, wherein the wheel track calculation is expressed as: m, the time delay calculation is expressed as s. For the left front wheel and left rear wheel, the constraint is: The right front wheel is similar to the right rear wheel.

[0172] (2) Lateral coherence spectrum constraint of left and right wheel tracks: wheel spacing m, target coherence attenuation coefficient Target coherence function In the low-frequency range (spatial wavelength > 3m), the left and right wheel tracks are almost completely coherent; in the high-frequency range (spatial wavelength < 0.5m), the coherence decays to near zero.

[0173] 8. Construction and Training of Multi-Branch PINN The neural network structure is shown in Table 3 below.

[0174] Table 3 Neural Network Structure Table

[0175] The input coordinates of each branch are Fourier feature encoded, and the encoding order is... Mapping 1D input to 3D feature vector. The output of the structural parameter branch is mapped to a preset range by the Sigmoid function, and is represented by the following equation (45).

[0176] (45) in, kN·m², kN·m². , , Similar processing.

[0177] Wavelet domain residual projection: using Morlet mother wavelet The fourth spatial derivative term in the PDE residual is obtained by wavelet derivative kernel. Calculation, scale range (7 scale levels).

[0178] The target loss function is expressed by the following equation (46).

[0179] (46) Initial weights The regularization coefficient is taken as... .

[0180] Training strategy: Use Adam pre-training for 5000 rounds (learning rate...) ) Combined with L-BFGS fine-tuning for 1000 rounds (learning rate) A two-stage strategy. NTK trace adaptive weights every... The data is updated once per round. The training data is organized in time windows, with each window containing 4096 sampling points and a 50% overlap between time windows.

[0181] Training hardware: Completed on a single GPU (such as NVIDIA RTX 3090), with a single training session taking approximately 15 minutes.

[0182] 9. Output Results After training is complete, the following will be output: Static geometric elevation of road surface: left wheel track and right wheel tracks The spatial resolution is approximately 0.5m. It can be further used to calculate the International Roughness Index (IRI) and longitudinal profile elevation. Transient dynamic bending spacetime field: With a spatial resolution of approximately 0.5m and a temporal resolution of 1ms, it can extract the distribution curve of dynamic deflection peak along mileage, the shape characteristics of the deflection basin, and the deflection recovery time. Equivalent parameters of pavement structure: , , , Under single-vehicle conditions, and Its inversion stability is better than and .

[0183] 10. Ablation Experiments and Technical Effects To verify the contribution of each technical feature, an ablation comparison experiment was conducted. Using known static geometric elevation of the road surface and FWD deflection basin data as the true values, six schemes were compared, and the comparison table is shown in Table 4 below.

[0184] Table 4 Comparison of Schemes

[0185] Analysis shows that, in scheme A→B (tire track envelope), the distortion of high-frequency road surface components is corrected after introducing the tire track nonlinear contact envelope, and the static elevation reconstruction error is reduced from >15% to ~8%. This verifies the necessity of modeling the spatial filtering effect of tire contact patches.

[0186] Scheme B→C (PDE dynamic deflection): After introducing the Pasternak foundation beam PDE, the transient dynamic deflection component is separated, and the static geometric elevation error of the pavement is further reduced to ~5%. However, the fourth spatial derivative in the PDE exhibits gradient oscillation in the early stage of training, resulting in a relatively high dynamic deflection separation error (~12%).

[0187] Scheme C→D (Multi-wheel track constraint): After introducing the front and rear wheel track time delay constraint and the left and right wheel track lateral coherence constraint, the drift of static elevation under variable speed conditions is effectively suppressed, and the two errors are reduced to ~4% and ~8%, respectively.

[0188] Scheme D→E (wavelet domain projection): After projecting the residual to the wavelet domain, the fourth derivative is calculated through the wavelet derivative kernel, which significantly improves the training stability and reduces the dynamic deflection separation error to ~5%.

[0189] Scheme E→F (NTK Adaptive Weights): After introducing NTK trace adaptive weight updates, the training progress of each loss term tends to be balanced, the convergence speed is improved by about 30%, and the final road surface static geometric elevation reconstruction error is <3%, and the transient dynamic deflection separation error is <5%.

[0190] Example 2: Joint Inversion of Structural Parameters for Multiple Vehicles and Multiple Speeds like Figure 6 As shown, this embodiment illustrates how to jointly invert the equivalent parameters of the road structure when multiple vehicles of different models and speeds pass through the same road segment, and verifies the enhancement effect of multi-vehicle constraints on the identifiability of road parameters.

[0191] 1. Implementation conditions Vehicle platoon: 10 vehicles of different types, covering SUVs, light trucks, medium trucks, and heavy trucks, with different weights, axle loads, suspension parameters, and tire parameters.

[0192] The vehicle parameter range is shown in Table 5 below.

[0193] Table 5 Vehicle Parameter Range Table

[0194] Speed ​​range: Vehicles pass through the same road section at different speeds, with a speed range of 40~100km / h (11.1~27.8m / s), covering the full operating conditions from low speed to high speed.

[0195] Road conditions: 500m asphalt pavement, same as in Example 1. All vehicles share the same static geometric elevation of the road surface. However, each vehicle has independent transient dynamic deflection. and dynamic wheel load .

[0196] Sensor configuration: Each vehicle is equipped with the same type of multi-channel sensors as in Example 1 (vehicle acceleration, unsprung acceleration, IMU, vehicle speed, wheel speed, GPS), and the sampling rate is required to be consistent.

[0197] 2. Independent data processing for each vehicle Ten vehicles each travel through a 500m section of road at their own speed, and each vehicle independently performs the following steps: Data acquisition and preprocessing: Each vehicle independently acquires response data, and performs time synchronization, zero bias correction, bandpass filtering (cutoff frequency range is consistent with Example 1), outlier removal, and coordinate transformation. Wheel track coordinate reconstruction: Each vehicle reconstructs its wheel track coordinates based on its own speed, heading angle, and geometric parameters. Because the wheelbase and track width of each vehicle are different, the position of the wheel track coordinates in the road coordinate system varies slightly; Vehicle dynamics modeling: Each vehicle establishes its own vehicle dynamics model based on its own parameters. The mass matrix, damping matrix, and stiffness matrix of the vehicle are respectively , , All vehicles use Method B to construct an effective road surface input.

[0198] 3. Modeling shared physical constraints Tire path envelope: Each vehicle establishes an independent tire path envelope model based on its own tire specifications and tire pressure. The envelope parameters of the vehicle are , , , , Its envelope weight function It varies depending on the tire model.

[0199] Vehicle-Road-Subgrade Closed-Loop PDE: All vehicles share the same Pasternak viscoelastic subgrade beam PDE, i.e., pavement structure parameters. , , , The same applies to all vehicles. However, the dynamic wheel load varies from vehicle to vehicle. The transient dynamic deflection induced by the loads of each vehicle is different. They are mutually independent. The excitation term on the right side of PD is either the superposition of the wheel loads of all vehicles (when multiple vehicles are traveling at the same time) or independent (when each vehicle passes by in turn).

[0200] This embodiment uses a scenario where each vehicle passes through in sequence, and each vehicle independently calculates its own PDE residual.

[0201] Multi-round track spatiotemporal consistency constraints: Each vehicle independently establishes front and rear wheelbase delay constraints (wheelbase of each vehicle). and speed Different, therefore time delay different); Each vehicle independently establishes lateral coherence spectrum constraints for its left and right wheel tracks; Shared static road surface constraints for multiple vehicles: The static geometric elevation of the road surface observed when all 10 vehicles pass through the same road segment should be exactly the same, i.e. .

[0202] 4. Multi-branch PINN architecture and sharing mechanism The network architecture adopts a hybrid structure of "shared branches combined with independent branches", as shown in Table 6 below.

[0203] Table 6. Hybrid Neural Network Structures

[0204] Shared branches ensure consistent estimations of road surface static geometry and structural parameters across all vehicles; independent branches allow each vehicle's transient dynamic deflection to vary independently based on its dynamic wheel load and speed. Each branch network structure consists of 4 fully connected layers, each with 256 neurons and Tanh activation. The input is encoded using Fourier features. The output of the structural parameter branch is mapped to a preset positive range via a Sigmoid function.

[0205] 5. Multi-vehicle joint loss function and training The target loss function is expressed by the following equation (47).

[0206] (47) in, : No. The vehicle dynamics consistency loss of the vehicle, using the first Vehicle parameters , , calculate; : No. The tire envelope consistency loss of the vehicle, using the first vehicle envelope kernel calculate; : No. The PDE residual loss of the vehicle, the structural parameters are shared by the branch Provided by independent branches, transient dynamic deflection Provided, wheeled By the The dynamic state of the vehicle is determined; : No. The wheelbase delay loss of the vehicle, using the first Wheelbase of the vehicle and speed Calculate latency ; The transverse coherence spectrum loss can be calculated separately for each vehicle and then summed. Regularization loss constrains the spatial smoothness of the outputs of shared and independent branches; Prior constraint loss constrains the prior range of vehicle parameters, tire path envelope parameters, and structural parameters.

[0207] All loss terms can be projected onto the wavelet domain (Morlet mother wavelet, 7 scale levels).

[0208] The training parameters are shown in Table 7 below.

[0209] Table 7 Training Parameter Table

[0210] Training hardware: Completed on a single GPU (such as NVIDIA RTX 3090), joint training takes about 2 hours.

[0211] 6. Output Results After training is complete, the following will be output: Static geometric elevation of road surface: With a spatial resolution of approximately 0.5m, constrained by 10 vehicles, its accuracy surpasses that of any single-vehicle result. Transient dynamic deflection of each vehicle: These reflect the dynamic road surface response of each vehicle under different speeds and dynamic wheel loads; Equivalent parameters of pavement structure: , , , This is the result of joint inversion of 10 vehicles, with a spatial resolution of approximately 1 m; Health evaluation indicators include road sections with abnormal structural stiffness, locations of weak roadbed layers, locations with excessive dynamic deflection, locations with abnormal smoothness, and maintenance priorities.

[0212] 7. Verification of Technical Effectiveness (1) Accuracy of road parameter inversion Using the FWD deflection basin inversion results and core sampling laboratory test results as reference true values, the relative errors of the inversion of four structural parameters under single-vehicle and multi-vehicle joint conditions are compared, as shown in Table 8 below.

[0213] Table 8 Comparison of Relative Errors in Inversion

[0214] (2) Gain analysis of multi-vehicle constraints As can be seen from Table 8 above, the multi-vehicle joint inversion has a significant gain compared to the single-vehicle inversion. The reason is as follows: The excitation frequency bands are complementary, and vehicles at different speeds elicit different time-frequency responses in the road surface PDE. Low-speed vehicles (40 km / h) provide low-frequency excitation, which is beneficial for identification. High-speed vehicles (100 km / h) provide high-frequency excitation, which is beneficial for identification. ; The difference in load amplitude and the different dynamic wheel loads (3000~35000 N) of vehicles produce transient dynamic deflection of different amplitudes, which causes the relative contribution of each parameter in PDE to change, thereby breaking the coupling ambiguity between parameters. By sharing the static geometric constraints of the road surface, 10 vehicles jointly constrain the same static geometric elevation of the road surface, which is equivalent to enhancing the observation constraints of the static geometric branch by about 10 times, significantly reducing the interference of the uncertainty of the road surface static geometric elevation estimation on the structural parameter inversion. Diversity of envelope parameters: Vehicles with different tire parameters (contact patch size, tire pressure) provide road surface information with different spatial resolutions, which is beneficial for separating the static geometric elevation of the road surface from the transient dynamic deflection.

[0215] (3) The comparison results with the single vehicle scheme are shown in Table 9 below.

[0216] Table 9 Comparison of Different Schemes

[0217] See Figure 7 , Figure 7 This is a structural diagram of a road monitoring device provided in an embodiment of this application, as shown below. Figure 7 As shown, the device includes: The acquisition module 701 is used to acquire the vehicle parameters of the first vehicle and the vehicle response data of the first vehicle during its driving process on the first road. The first determining module 702 is used to determine the dynamic state of the first vehicle, the trajectory coordinates of each wheel in the first vehicle in the road coordinate system, and the effective road surface input of each wheel in the first vehicle to the first road based on the vehicle response data and the vehicle parameters. The second determining module 703 is used to determine the dynamic wheel load applied to the road surface of the first vehicle by each wheel during the driving process based on the dynamic state and the effective road surface input, using a first model characterizing vehicle dynamics. The inversion module 704 is used to input the vehicle speed, trajectory coordinates, and dynamic wheel load from the vehicle response data into a neural network to perform road parameter inversion and obtain the road parameters of the first road. The neural network includes a first branch network, a second branch network, and a third branch network. The first branch network is used to determine the static geometric elevation of the road surface of the first road in the road parameters based on the trajectory coordinates. The second branch network is used to determine the transient dynamic deflection induced by the dynamic wheel load of each wheel of the first vehicle during driving based on the vehicle speed, trajectory coordinates, and dynamic wheel load. The third branch network is used to determine the structural parameters of the first road in the road parameters based on the trajectory coordinates. The target loss function of the neural network is constructed based on the residual of the target model. The target model includes the first model and the physical constraint model of the first model. The physical constraint model includes: a second model for characterizing the contact envelope between the tire and the road, a third model for characterizing the coupled dynamics between the vehicle and the road, and a fourth model for characterizing the spatiotemporal consistency between different tires. The second model is used to constrain the static geometric elevation and transient dynamic deflection of the road surface in the effective road surface input involved in the first model. The third model is used to constrain the transient dynamic deflection induced by structural parameters and dynamic wheel loads. The fourth model is used to constrain the static geometric elevation of the road surface.

[0218] Optionally, the second determining module 703 is specifically used for: Based on the dynamic state of the first vehicle at the first moment and the effective road surface input, the dynamic wheel load of each wheel of the first vehicle on the road surface at the first moment is determined by the first model, and the static geometric elevation and transient dynamic deflection of the road surface are determined by the second model to determine the effective road surface input. Based on the transient dynamic deflection induced by the dynamic wheel load of each wheel in the first vehicle at the first moment, the second model is used to correct the effective road surface input of each wheel in the first vehicle to the first road at the second moment, and the transient dynamic deflection is constrained by the third model. Based on the corrected effective road surface input and the dynamic state of the first vehicle at the second moment, the dynamic wheel load of each wheel of the first vehicle on the road surface at the second moment is determined using the first model.

[0219] Optionally, the device further includes: The assessment module is used to perform a health assessment on the first road based on the road parameters of the first road, and obtain the health evaluation index information of the first road.

[0220] Optionally, the device further includes: The third determining module is used to determine the residual of the target model based on the road parameters of the second road obtained by the neural network inversion. The projection module is used to project the target residual in the residual of the target model to the wavelet time frequency to obtain the wavelet frequency residual norm. The target residual includes the residual of the third model, which involves the higher-order partial derivatives of transient dynamic deflection with respect to spatial coordinates. A construction module is used to construct the target loss function of the neural network based on the wavelet frequency residual norm and other residuals in the residuals of the target model other than the target residual; The training module is used to train the neural network based on the target loss function.

[0221] Optionally, the target loss function further includes at least one of the following: Based on the loss of boundary conditions at both ends of the second road for transient dynamic deflection, the boundary conditions are used to indicate the dynamic parameters of the road surface without bending deformation and without bending deformation at both ends of the road, and the boundary conditions are used to constrain the transient dynamic deflection induced by dynamic wheel load. The regularization loss of the road parameters of the second road; The prior constraint loss of the target parameters, which include the vehicle parameters of the second vehicle, the parameters involved in the contact envelope between the tires of the second vehicle and the second road, and the structural parameters of the second road.

[0222] Optionally, the device further includes: An update module is used to update the weights of each loss term in the target loss function during the training process of the neural network using an adaptive update strategy based on neural tangent kernel traces.

[0223] Optionally, the neural network is jointly trained based on M second vehicles. The target loss function is the sum of the first loss and the second loss. The first loss is superimposed with the independent loss of each second vehicle. The independent loss includes the loss constructed based on the residuals of the first model, the loss constructed based on the residuals of the second model, the loss constructed based on the residuals of the third model, and the wheelbase delay loss. The wheelbase delay loss is constructed based on the residuals of the first sub-model. The fourth model includes the first sub-model, which is used to characterize the consistency of the wheelbase delay of the front and rear wheels on the same side of the second vehicle. The second loss is superimposed with a shared loss. The shared loss includes the regularization loss of the road parameters of the second road, the prior constraint loss of the target parameters, and the lateral coherence spectrum loss. The lateral coherence spectrum loss is used to indicate the statistical coherence of the static road surface geometric elevation of the left and right wheels of the second vehicle at spatial frequencies. M is an integer greater than 1.

[0224] The road monitoring device can realize all the processes implemented in the above road monitoring method embodiments and achieve the same technical effect. To avoid repetition, it will not be described again here.

[0225] In some embodiments, see Figure 8 The figure shows a structural diagram of an electronic device provided in an embodiment of the present invention. Figure 8 As shown, the electronic device 800 includes: a processor 801, a memory 802, a user interface 803, and a bus interface 804.

[0226] Processor 801 is used to read the program from memory 802 and execute the following procedures: Obtain the vehicle parameters of the first vehicle and the vehicle response data of the first vehicle during its travel on the first road; Based on the vehicle response data and the vehicle parameters, the dynamic state of the first vehicle, the trajectory coordinates of each wheel in the first vehicle in the road coordinate system, and the effective road surface input of each wheel in the first vehicle to the first road are determined. Based on the dynamic state and the effective road surface input, the dynamic wheel loads applied to the road surface of the first road by each wheel of the first vehicle during driving are determined using a first model characterizing vehicle dynamics. The vehicle speed, trajectory coordinates, and dynamic wheel load from the vehicle response data are input into a neural network to perform road parameter inversion, thereby obtaining the road parameters of the first road. The neural network includes a first branch network, a second branch network, and a third branch network. The first branch network is used to determine the static geometric elevation of the road surface of the first road in the road parameters based on the trajectory coordinates. The second branch network is used to determine the transient dynamic deflection induced by the dynamic wheel load of each wheel of the first vehicle during driving based on the vehicle speed, trajectory coordinates, and dynamic wheel load. The third branch network is used to determine the structural parameters of the first road in the road parameters based on the trajectory coordinates. The target loss function of the neural network is constructed based on the residual of the target model. The target model includes the first model and the physical constraint model of the first model. The physical constraint model includes: a second model for characterizing the contact envelope between the tire and the road, a third model for characterizing the coupled dynamics between the vehicle and the road, and a fourth model for characterizing the spatiotemporal consistency between different tires. The second model is used to constrain the static geometric elevation and transient dynamic deflection of the road surface in the effective road surface input involved in the first model. The third model is used to constrain the transient dynamic deflection induced by structural parameters and dynamic wheel loads. The fourth model is used to constrain the static geometric elevation of the road surface.

[0227] exist Figure 8 In this context, the bus architecture can include any number of interconnected buses and bridges, specifically linking various circuits of one or more processors represented by processor 801 and memory represented by memory 802 together. The bus architecture can also link various other circuits such as peripheral devices, voltage regulators, and power management circuits, which are well known in the art and therefore will not be described further herein. Bus interface 804 provides an interface. For different user devices, user interface 803 can also be an interface capable of connecting external or internal devices, including but not limited to keypads, displays, speakers, microphones, joysticks, etc.

[0228] The processor 801 is responsible for managing the bus architecture and general processing, while the memory 802 can store the data used by the processor 801 when performing operations.

[0229] Preferably, the present invention also provides an electronic device 800, including a processor 801, a memory 802, and a computer program stored in the memory 802 and executable on the processor 801. When the computer program is executed by the processor 801, it implements the various processes of the above-described road monitoring method embodiments and achieves the same technical effect. To avoid repetition, it will not be described again here.

[0230] This invention also provides a readable storage medium storing a computer program. When executed by a processor, this computer program implements the various processes of the road monitoring method embodiments described above and achieves the same technical effects. To avoid repetition, it will not be described again here. The readable storage medium may be a read-only memory (ROM), a random access memory (RAM), a magnetic disk, or an optical disk, etc.

[0231] This application also provides a computer program product, including computer instructions. When executed by a processor, these computer instructions implement the various processes of the above-described road monitoring method embodiments and achieve the same technical effects. To avoid repetition, they will not be described again here.

[0232] Those skilled in the art will recognize that the units and algorithm steps of the various examples described in conjunction with the embodiments disclosed herein can be implemented in electronic hardware, or a combination of computer software and electronic hardware. Whether these functions are implemented in hardware or software depends on the specific application and design constraints of the technical solution. Those skilled in the art can use different methods to implement the described functions for each specific application, but such implementations should not be considered beyond the scope of this invention.

[0233] Those skilled in the art will clearly understand that, for the sake of convenience and brevity, the specific working processes of the systems, devices, and units described above can be referred to the corresponding processes in the foregoing method embodiments, and will not be repeated here.

[0234] In the embodiments provided in this application, it should be understood that the disclosed systems and methods can be implemented in other ways. For example, the system 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 coupling or direct coupling or communication connection shown or discussed may be an indirect coupling or communication connection through some interfaces, devices, or units, and may be electrical, mechanical, or other forms.

[0235] 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 the embodiments of the present invention, depending on actual needs.

[0236] In addition, the functional units in the various embodiments of the present invention 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.

[0237] If the aforementioned functions are implemented as software functional units and sold or used as independent products, they can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of this invention, essentially, or the part that contributes to the prior art, or a portion of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium 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 described in the various embodiments of this invention. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, ROM, RAM, magnetic disks, or optical disks.

[0238] The above description is merely a specific embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the technical scope disclosed in the present invention should be included within the scope of protection of the present invention. Therefore, the scope of protection of the present invention should be determined by the scope of the claims.

Claims

1. A road monitoring method, characterized in that, The method includes: Obtain the vehicle parameters of the first vehicle and the vehicle response data of the first vehicle during its travel on the first road; Based on the vehicle response data and the vehicle parameters, the dynamic state of the first vehicle, the trajectory coordinates of each wheel in the first vehicle in the road coordinate system, and the effective road surface input of each wheel in the first vehicle to the first road are determined. Based on the dynamic state and the effective road surface input, the dynamic wheel loads applied to the road surface of the first road by each wheel of the first vehicle during driving are determined using a first model characterizing vehicle dynamics. The vehicle speed, trajectory coordinates, and dynamic wheel load from the vehicle response data are input into a neural network to perform road parameter inversion, thereby obtaining the road parameters of the first road. The neural network includes a first branch network, a second branch network, and a third branch network. The first branch network is used to determine the static geometric elevation of the road surface of the first road in the road parameters based on the trajectory coordinates. The second branch network is used to determine the transient dynamic deflection induced by the dynamic wheel load of each wheel of the first vehicle during driving based on the vehicle speed, trajectory coordinates, and dynamic wheel load. The third branch network is used to determine the structural parameters of the first road in the road parameters based on the trajectory coordinates. The target loss function of the neural network is constructed based on the residual of the target model. The target model includes the first model and the physical constraint model of the first model. The physical constraint model includes: a second model for characterizing the contact envelope between the tire and the road, a third model for characterizing the coupled dynamics between the vehicle and the road, and a fourth model for characterizing the spatiotemporal consistency between different tires. The second model is used to constrain the static geometric elevation and transient dynamic deflection of the road surface in the effective road surface input involved in the first model. The third model is used to constrain the transient dynamic deflection induced by structural parameters and dynamic wheel loads. The fourth model is used to constrain the static geometric elevation of the road surface.

2. The method according to claim 1, characterized in that, The step of determining the dynamic wheel loads applied to the road surface by each wheel of the first vehicle during its operation, based on the dynamic state and the effective road surface input, using a first model characterizing vehicle dynamics, includes: Based on the dynamic state of the first vehicle at the first moment and the effective road surface input, the dynamic wheel load of each wheel of the first vehicle on the road surface at the first moment is determined by the first model, and the static geometric elevation and transient dynamic deflection of the road surface are determined by the second model to determine the effective road surface input. Based on the transient dynamic deflection induced by the dynamic wheel load of each wheel in the first vehicle at the first moment, the second model is used to correct the effective road surface input of each wheel in the first vehicle to the first road at the second moment, and the transient dynamic deflection is constrained by the third model. Based on the corrected effective road surface input and the dynamic state of the first vehicle at the second moment, the dynamic wheel load of each wheel of the first vehicle on the road surface at the second moment is determined using the first model.

3. The method according to claim 1, characterized in that, After inputting the vehicle speed, trajectory coordinates, and dynamic wheel load from the vehicle response data into a neural network to perform road parameter inversion and obtain the road parameters of the first road, the method further includes: Based on the road parameters of the first road, a health assessment is performed on the first road to obtain health evaluation index information of the first road.

4. The method according to claim 1, characterized in that, Before acquiring the vehicle parameters of the first vehicle and the vehicle response data of the first vehicle during its travel on the first road, the method further includes: Based on the road parameters of the second road obtained by the neural network inversion, the residual of the target model is determined; The target residual in the residual of the target model is projected onto the wavelet time frequency to obtain the wavelet frequency residual norm. The target residual includes the residual of the third model, which involves the higher-order partial derivatives of transient dynamic deflection with respect to spatial coordinates. Based on the wavelet frequency residual norm and the residuals of the target model other than the target residual, the target loss function of the neural network is constructed. The neural network is trained based on the target loss function.

5. The method according to claim 4, characterized in that, The target loss function also includes at least one of the following: Based on the loss of boundary conditions at both ends of the second road for transient dynamic deflection, the boundary conditions are used to indicate the dynamic parameters of the road surface without bending deformation and without bending deformation at both ends of the road, and the boundary conditions are used to constrain the transient dynamic deflection induced by dynamic wheel load. The regularization loss of the road parameters of the second road; The prior constraint loss of the target parameters, which include the vehicle parameters of the second vehicle, the parameters involved in the contact envelope between the tires of the second vehicle and the second road, and the structural parameters of the second road.

6. The method according to any one of claims 1 to 5, characterized in that, The method further includes: During the training of the neural network, an adaptive update strategy based on neural tangent kernel traces is used to update the weights of each loss term in the target loss function.

7. The method according to any one of claims 1 to 5, characterized in that, The neural network is jointly trained based on M second vehicles. The target loss function is the sum of the first loss and the second loss. The first loss is superimposed with the independent loss of each second vehicle. The independent loss includes the loss constructed based on the residuals of the first model, the loss constructed based on the residuals of the second model, the loss constructed based on the residuals of the third model, and the wheelbase delay loss. The wheelbase delay loss is constructed based on the residuals of the first sub-model. The fourth model includes the first sub-model, which is used to characterize the consistency of the wheelbase delay of the front and rear wheels on the same side of the second vehicle. The second loss is superimposed with a shared loss. The shared loss includes the regularization loss of the road parameters of the second road, the prior constraint loss of the target parameters, and the lateral coherence spectrum loss. The lateral coherence spectrum loss is used to indicate the statistical coherence of the static road surface geometric elevation of the left and right wheels of the second vehicle at spatial frequencies. M is an integer greater than 1.

8. A road monitoring device, characterized in that, The device includes: The acquisition module is used to acquire the vehicle parameters of the first vehicle and the vehicle response data of the first vehicle during its travel on the first road. The first determining module is used to determine the dynamic state of the first vehicle, the trajectory coordinates of each wheel of the first vehicle in the road coordinate system, and the effective road surface input of each wheel of the first vehicle to the first road based on the vehicle response data and the vehicle parameters. The second determining module is used to determine the dynamic wheel load applied to the road surface of the first vehicle by each wheel during the driving process, based on the dynamic state and the effective road surface input, using a first model characterizing vehicle dynamics. The inversion module is used to input the vehicle speed, trajectory coordinates, and dynamic wheel load from the vehicle response data into a neural network to perform road parameter inversion and obtain the road parameters of the first road. The neural network includes a first branch network, a second branch network, and a third branch network. The first branch network is used to determine the static geometric elevation of the road surface of the first road in the road parameters based on the trajectory coordinates. The second branch network is used to determine the transient dynamic deflection induced by the dynamic wheel load of each wheel of the first vehicle during driving based on the vehicle speed, trajectory coordinates, and dynamic wheel load. The third branch network is used to determine the structural parameters of the first road in the road parameters based on the trajectory coordinates. The target loss function of the neural network is constructed based on the residual of the target model. The target model includes the first model and the physical constraint model of the first model. The physical constraint model includes: a second model for characterizing the contact envelope between the tire and the road, a third model for characterizing the coupled dynamics between the vehicle and the road, and a fourth model for characterizing the spatiotemporal consistency between different tires. The second model is used to constrain the static geometric elevation and transient dynamic deflection of the road surface in the effective road surface input involved in the first model. The third model is used to constrain the transient dynamic deflection induced by structural parameters and dynamic wheel loads. The fourth model is used to constrain the static geometric elevation of the road surface.

9. An electronic device, characterized in that, include: A processor, a memory, and a program stored in the memory and executable on the processor, wherein the program, when executed by the processor, implements the steps of the road monitoring method as described in any one of claims 1 to 7.

10. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores a computer program that, when executed by a processor, implements the steps of the road monitoring method as described in any one of claims 1 to 7.

Images