Rail state real-time detection method based on multi-sensor fusion and dynamic compensation

Abstract

This invention relates to the field of mobile track geometry state detection technology, and discloses a real-time track state detection method based on multi-sensor fusion and dynamic compensation. It collects multi-source synchronous data from a global navigation satellite system receiver, inertial measurement unit, and laser profilometer mounted on the vehicle body, as well as accelerometers mounted on the bogie, and performs time synchronization and coordinate system unification. Through dynamic error correction and extended Kalman filtering fusion, it achieves high-precision estimation of vehicle attitude, velocity, and position. First-level point cloud motion compensation eliminates spatial distortion caused by the rigid body motion of the vehicle itself on the point cloud collected by the laser profilometer. Second-level point cloud motion compensation eliminates residual geometric errors caused by high-frequency vibration of the bogie. Finally, track geometric parameters are extracted and smoothness analysis is performed. This invention achieves high-precision reconstruction of track point clouds, providing technical support for real-time, high-precision track state monitoring on operating trains.

Description

Technical Field

[0001] This invention relates to the field of mobile track geometry state detection technology, specifically a real-time track state detection method based on multi-sensor fusion and dynamic compensation. Background Technology

[0002] High-speed railways have become the backbone of intercity transportation in modern society, and their operational safety and stability are a focus of public attention. During high-speed train operation, even minor changes in track geometry (such as gauge, superelevation, levelness, and alignment) can lead to increased vehicle vibration, decreased passenger comfort, and even threats to operational safety. These changes stem from a variety of complex factors, including track material fatigue, foundation settlement, wheel-rail interaction, and changes in environmental temperature and humidity. Therefore, achieving real-time, high-precision monitoring and intelligent diagnosis of track geometry has become a core technological challenge for ensuring the safe operation of high-speed railways.

[0003] With the development of detection technology, modern track inspection systems are equipped with a variety of advanced sensors to monitor the spatial topography of the track and the motion attitude of the train, including laser profilometers, inertial measurement units (IMUs), global navigation satellite systems (GNSS), and various accelerometers. Among these, track geometry is the core monitoring object because it directly determines the quality of wheel-rail interaction and the safety boundary of train operation. However, due to the complexity of the strongly coupled dynamic system consisting of train, track, and environment, the raw point cloud and pose data collected by sensors are easily affected by the combined interference of rigid body motion of the car body and high-frequency vibration of the bogie, making geometric feature extraction difficult and limiting diagnostic accuracy. While multi-source heterogeneous monitoring data provides rich information for track condition assessment, it also highlights the challenge of high-precision, real-time data fusion and compensation under strong vibration environments.

[0004] Traditional track geometry detection methods typically rely on periodic static measurements or dedicated track inspection vehicles. While these methods offer high accuracy, they are inherently offline and non-real-time. They suffer from long detection cycles, high costs, and require dedicated track space for operation, failing to capture instantaneous changes in track condition and sudden defects. In scenarios with extremely high safety requirements, such as high-speed railways, significant blind spots exist in safety monitoring. In recent years, GNSS / IMU-based integrated navigation methods have been introduced into this field, attempting to achieve motion detection. However, this approach also faces serious challenges. Since IMUs are usually installed on the vehicle body along with GNSS, they primarily sense the low-frequency, large-scale rigid body motion of the vehicle body. However, the vehicle-mounted IMU's perception of high-frequency vibrations from wheel-rail impacts (such as passing over potholes or switches) is insensitive and incomplete. Therefore, the motion compensation models of existing methods are incomplete, resulting in high-frequency jitter and microscopic distortions still existing in the compensated point cloud. The fundamental reason is that existing solutions rely on the vehicle-integrated IMU as the sole motion sensing source, failing to directly capture and separate the high-frequency vibration components from the bogie that are not transmitted to the vehicle body. Moreover, as high-speed mobile units, trains are highly susceptible to GNSS signal blockage and failure in tunnels, canyons, and other areas, leading to a sharp decline in positioning accuracy. On the other hand, IMUs based on low-cost MEMS (Micro-electromechanical Systems) exhibit significant zero-bias drift when operating alone. Even with the assistance of fusion algorithms, the attitude and position calculated by these IMUs will still accumulate considerable errors after long-term operation, making it difficult to meet the requirements for centimeter-level track detection.

[0005] Furthermore, the complex vibrations of the car body and bogie cause severe spatial distortion in the laser point cloud, and traditional simple compensation algorithms cannot effectively separate and eliminate this multi-source motion interference. The final accuracy of the point cloud is determined by both the car body motion compensation and the uncompensated high-frequency vibration error. Due to the presence of high-frequency vibration error, even if the car body motion compensation is perfect, the overall point cloud accuracy is difficult to improve further, failing to meet the centimeter-level or even higher precision detection requirements of high-speed railways. Under these circumstances, track detection systems that rely entirely on GNSS or a single IMU are unlikely to operate stably and reliably under complex track conditions.

[0006] Therefore, how to achieve high-precision fusion of multi-sensor data and real-time, accurate motion compensation of track point clouds under conditions of unstable GNSS signals, IMU drift, and complex vehicle body and bogie vibration, so as to obtain stable and reliable centimeter-level track geometry reconstruction results, has become an important technical challenge that urgently needs to be overcome in the field of intelligent track operation and maintenance. Summary of the Invention

[0007] To address the aforementioned problems, the present invention aims to provide a real-time track condition detection method based on multi-sensor fusion and dynamic compensation. This method solves the problem of severe geometric distortion of laser scanning point clouds caused by the combined effects of rigid body motion of the vehicle and high-frequency vibration of the bogie in mobile track geometric condition detection. It achieves centimeter-level accuracy reconstruction of the track point cloud, providing technical support for real-time, high-precision track condition monitoring on operating trains. The technical solution is as follows:

[0008] A real-time orbital state detection method based on multi-sensor fusion and dynamic compensation includes the following steps:

[0009] Step 1: Synchronous acquisition and preprocessing of multi-source data;

[0010] It collects multi-source synchronous data from the Global Navigation Satellite System (GNSS) receiver, Inertial Measurement Unit (IMU), and Laser Profilometer (LiDAR) installed on the vehicle body, as well as the accelerometer installed on the bogie, and performs time synchronization and coordinate system unification.

[0011] Step 2: Fusion of dynamic error correction and extended Kalman filtering;

[0012] Establish state models and inertial measurement unit prediction equations; construct observation equations for the global navigation satellite system as well as nonholonomic constraints and zero-velocity update constraint models; achieve tight coupling fusion and dynamic correction of the global navigation satellite system and inertial measurement unit through extended Kalman filtering, and combine nonlinear feedback correction strategies to achieve high-precision estimation of vehicle attitude, velocity and position;

[0013] Step 3: Motion compensation for first-level point clouds;

[0014] For the vehicle attitude, velocity, and position data estimated in step 2, all points within the same frame are uniformly transformed to a common reference time vehicle coordinate system based on their acquisition timestamps. This eliminates the spatial distortion caused by the rigid body motion of the vehicle itself on the point cloud acquired by the laser profilometer, resulting in a first-order point cloud with motion compensation. ;

[0015] Step 4: Second-level point cloud motion compensation;

[0016] By calculating the vibration displacement measured by the accelerometers mounted on the bogie and performing reverse displacement compensation on the point cloud, a two-stage fine compensation is achieved, eliminating the point cloud distortion caused by the first-stage point cloud motion compensation. The residual geometric errors caused by the high-frequency vibration of the bogie still exist, and the final point cloud after second-level point cloud motion compensation is obtained. ;

[0017] Step 5: Track geometry parameter extraction and smoothness analysis;

[0018] Based on the final point cloud after two-stage point cloud motion compensation It automatically extracts key track geometry parameters and analyzes and evaluates their smoothness, thereby achieving intelligent diagnosis of track status.

[0019] The beneficial effects of this invention are:

[0020] This invention employs a separate sensing architecture for the vehicle body and bogie, which can separate the two different types of interference from the physical source: vehicle body motion and bogie vibration, laying the foundation for accurate compensation.

[0021] This invention employs a two-stage serial process that first compensates for vehicle body motion and then compensates for bogie vibration, which can systematically peel off complex motion distortion layer by layer, ultimately achieving centimeter-level high-precision reconstruction of point clouds.

[0022] This invention ultimately obtained high-precision point clouds, and the calculation process is suitable for real-time processing, making it possible to deploy a high-precision track detection system on daily operating EMUs. This achieves a leap from "periodic detection" to "real-time online monitoring," greatly improving detection efficiency and driving safety. Attached Figure Description

[0023] Figure 1 This is a flowchart of the dynamic error correction and extended Kalman filter fusion process.

[0024] Figure 2 This is a flowchart of the motion compensation process for first-level point clouds.

[0025] Figure 3 This is a flowchart of the motion compensation process for secondary point clouds. Detailed Implementation

[0026] The present invention will now be described in further detail with reference to the accompanying drawings and specific embodiments.

[0027] This invention addresses the severe geometric distortion of laser-scanned point clouds caused by the combined effects of rigid body motion of the vehicle and high-frequency vibration of the bogie in the geometric state detection of mobile tracks. It achieves high-precision, real-time compensation and real-time status monitoring of track point clouds on operating trains, as follows:

[0028] (1) System physical architecture: Protect the specific sensor spatial layout scheme of "mounting the laser profilometer and IMU together on the vehicle body, while mounting the high-frequency accelerometer independently on the bogie".

[0029] (2) The core method of two-level compensation: protect the specific serial compensation process of “first performing first-level point cloud motion compensation based on vehicle body IMU / GNSS data, and then performing second-level point cloud motion compensation based on bogie accelerometer data”.

[0030] (3) Specific technical means involved in the above methods:

[0031] In the motion compensation of first-level point clouds, the mathematical model of "obtaining the precise pose of each point at the time of acquisition through timestamp interpolation and uniformly transforming it to the same reference time vehicle coordinate system" is illustrated in the flowchart below. Figure 2 As shown.

[0032] In the secondary point cloud motion compensation, the specific technical means of "bandpass filtering the bogie acceleration signal, removing gravity and frequency domain integration to calculate vibration displacement, and performing reverse displacement compensation on the point cloud" are illustrated in the flowchart below. Figure 3 As shown.

[0033] In data fusion, an enhanced EKF implementation scheme is proposed, which combines NHC, ZUPT, and NFC to form a closed-loop correction system.

[0034] This protects the complete technical solution that combines the above architecture, processes, and models to achieve real-time, high-precision compensation of orbital point clouds.

[0035] Step S1: Synchronous acquisition and preprocessing of multi-source data.

[0036] 1. Data Acquisition:

[0037] The GNSS receiver mounted on the vehicle body outputs position and speed information.

[0038] The inertial measurement unit (IMU) mounted on the vehicle body outputs triaxial angular velocity and triaxial acceleration.

[0039] Raw orbital point cloud acquired by a laser profilometer (LiDAR) mounted on the vehicle body.

[0040] The vibration acceleration signal output by the high-frequency triaxial accelerometer mounted on the bogie.

[0041] 2. Time synchronization:

[0042] Since the sampling frequencies and time bases of each sensor are different, a unified time axis t needs to be established. Let the timestamps of GNSS, IMU, and LiDAR be respectively... , and Then, through interpolation transformation, it is unified to the system clock. :

[0043] ;

[0044] in, This represents the observation of sensor s at time t; and Sensors Output two adjacent valid timestamps; and These represent the time intervals of sensor s, respectively. and The sampled values; These represent different types of sensors; GNSS stands for Global Navigation Satellite System, used to collect the absolute position and speed information of vehicle operation to achieve global positioning constraints for the system; IMU stands for Inertial Measurement Unit, used to collect three-axis angular velocity and three-axis acceleration to achieve high-frequency dynamic observation of the vehicle's short-term motion state; LiDAR stands for Laser Profilometer, used to collect laser point cloud data of the track cross-section for track geometry reconstruction and analysis.

[0045] Time synchronization of multi-source data can be achieved through linear or spline interpolation.

[0046] 3. Coordinate System One:

[0047] From the sensor coordinate system (L) to the vehicle coordinate system (B):

[0048] ;

[0049] The purpose of this step is to transform the local point seen in the sensor coordinate system to the view of the entire vehicle body. The formula represents the position of point P in the vehicle coordinate system B. It is equal to its position in the sensor coordinate system L. After a rotation matrix Translation vector It was obtained later.

[0050] From vehicle coordinate system (B) to world coordinate system (W):

[0051] ;

[0052] The purpose of this step is to transform points from the vehicle's perspective to a fixed global world coordinate system. In the formula... This indicates the absolute position of point P in the world coordinate system W, which is equal to its position in the vehicle coordinate system B. The vehicle rotates and translates together to its current position in the world coordinate system.

[0053] Step S2: Fusion of dynamic error correction and extended Kalman filtering.

[0054] Dynamic error correction and state fusion estimation are performed on the multi-source synchronized data acquired in step 1. Tightly coupled fusion of GNSS and IMU is achieved through Extended Kalman Filter (EKF). Combined with Nonholonomic Constraint (NHC), Zero-speed Update (ZUPT), and Nonlinear Feedback Correction (NFC) strategies, high-precision estimation of vehicle attitude, velocity, and position is achieved, suppressing drift accumulation and providing a highly stable trajectory for subsequent point cloud motion compensation. The flowchart of the dynamic error correction and EKF fusion process is shown below. Figure 1 As shown.

[0055] 1. Establish the state model and IMU prediction equations.

[0056] 1) Establish a state model:

[0057] A data fusion state model based on GNSS and IMU is established to describe the dynamic attitude, velocity, and position changes of the track inspection vehicle in the world coordinate system. State vectors are defined as follows:

[0058] ;

[0059] in, This represents the absolute position in the world coordinate system W. The velocity in world coordinates W; The attitude of the vehicle coordinate system B relative to the world coordinate system W; These are the zero biases of the gyroscope and accelerometer, respectively. This is the root cause of drift when the IMU operates independently, so it must be estimated and corrected as a state.

[0060] 2) IMU prediction equation:

[0061] ;

[0062] The equation is the rate of change equation for position. It shows that the position vector... The rate of change over time is equal to the current velocity. .

[0063] ;

[0064] This equation is the rate of change equation for velocity, and it is the core of inertial navigation. The acceleration measured by the IMU (after deducting zero bias) Rotate from vehicle coordinate system B to world coordinate system W. Since the IMU measures "specific force" which includes the influence of gravity, the gravity vector in the world coordinate system must be added. Only then can the true acceleration of the vehicle body in space be obtained.

[0065] ;

[0066] This equation is the rate of change equation for attitude quaternions. It describes how the vehicle's attitude changes with the angular velocity measured by the gyroscope (after deducting zero bias). And change. Quaternion operation matrix It realizes the conversion from angular velocity vector to quaternion differential operation.

[0067] ;

[0068] 3) State prediction and covariance prediction:

[0069] ;

[0070] This formula is a discrete-time prediction of the state. Function f (·) represents the numerical integration process of the above continuous-time differential equation. It utilizes the prior estimate of the optimal state at the previous time step. and current IMU measurements By time step The integral is used to predict the prior state estimate at the current moment. .

[0071] ;

[0072] This formula is a discrete-time prediction of the error covariance matrix. It predicts how the uncertainty of the state estimate propagates over time. It is a function f The Jacobian matrix (·) represents the covariance of the previous time step. Perform linearized propagation; It is the process noise covariance matrix, representing the new uncertainties introduced by IMU measurement noise and zero-bias random walk noise.

[0073] 2. Construct GNSS observations and NHC / ZUPT constraints.

[0074] GNSS provides absolute position and velocity observations to correct for IMU accumulated errors. When other external observation data becomes available, they are used to correct the IMU's predictions. These observations are considered to be true at a given moment. An extended Kalman filter compares the differences between the "predicted values" and the "observed values" to optimally adjust the state vector. Installation deviations between the IMU and GNSS are considered, and vehicle characteristic constraints (NHC, ZUPT) are introduced to enhance system observability.

[0075] 1) GNSS observation equation:

[0076] ;

[0077] This formula uses the observed position of the GNSS receiver antenna phase center. It is associated with the system state. Among them, It is the predicted origin position of the vehicle body in the state vector; The GNSS antenna is installed relative to the vehicle's origin. The lever arm compensation term is due to the fact that GNSS observations are based on antenna positions, while state estimates are based on the IMU vehicle center. This represents the noise in GNSS position measurements. The filter will compare the predicted noise... Compared with actual GNSS observations To correct the state vector.

[0078] ;

[0079] This formula will use the velocity vector measured by GNSS. It is associated with the system state. It is the predicted velocity of the vehicle's origin in the state vector; It is the lever arm speed compensation term, which takes into account the angular velocity of the vehicle body rotation. The resulting linear velocity at the GNSS antenna mounting point; This is GNSS velocity measurement noise. This observation is used to directly correct the system's velocity state.

[0080] 2) NHC / ZUPT constraint model:

[0081] NHC (Nonholonomic Constraint) is a key virtual observation technology introduced in the multi-sensor fusion navigation algorithm of this system. Based on the mechanical properties of a train's rigid body constrained motion on the rails: neglecting extremely small wheel-rail clearances and bounce, the train cannot slip in the direction perpendicular to the rails (lateral), nor can it jump off the track in the vertical direction. The train's lateral and vertical velocities are theoretically close to zero. This constraint can be expressed as:

[0082] ;

[0083] in, This refers to the train's forward speed. The core function of the NHC is to indirectly improve attitude estimation by constraining the speed, thereby suppressing heading drift caused by gyroscope integration.

[0084] The core value of introducing NHC lies in its ability to provide the system with a high-frequency, independent auxiliary observation source, regardless of whether GNSS signals are available or unavailable. This technology effectively suppresses IMU heading angle drift and significantly improves the accuracy of horizontal velocity estimation, thereby greatly enhancing the overall stability and reliability of train attitude calculation. Especially under conditions of prolonged satellite signal blockage, it can maintain system navigation accuracy even when GNSS is unavailable.

[0085] ZUPT (Zero-speed Update) is a core signal processing technique used in this invention to suppress the divergence of inertial navigation errors. Its basic principle is based on the inevitable brief periods of stillness that occur during train operation, such as stopping at a station or waiting for a signal. During these periods of stillness, the train's true three-dimensional velocity vector is an absolute physical truth value—zero.

[0086] Based on this absolute truth value, the system automatically identifies the train's stationary state using a zero-speed detector (typically achieved by analyzing the variance between the acceleration output from the IMU and the gyroscope data). Once a zero-speed state is determined, the system introduces a strongly constrained observation equation into the Kalman filter update process:

[0087] ;

[0088] The core function of ZUPT is to directly estimate and suppress the zero bias of the accelerometer.

[0089] The core value of introducing ZUPT lies in its role as a "software sensor" requiring no additional hardware, providing the system with a periodic absolute reference. This enables low-cost IMU-based navigation systems to fundamentally overcome the divergent trend of position errors accumulating over time, significantly improving the system's autonomous navigation accuracy during GNSS signal interruptions and its stability during long-term operation. It is a crucial guarantee for achieving high-precision, low-cost track detection.

[0090] 3. Extended Kalman filter fusion and dynamic correction.

[0091] This step aims to perform data fusion and state estimation of the models established in steps 1 and 2 using the EKF (Extended Kalman Filter) algorithm. EKF combines short-term, high-precision dynamic predictions from the IMU with absolutely constrained observations from GNSS, NHC, and ZUPT through a "prediction-update" loop, achieving high-precision, drift-free estimation of vehicle attitude, velocity, and position, and correcting IMU sensor bias in real time. The specific process is as follows:

[0092] 1) Prediction steps (based on IMU):

[0093] When a new IMU measurement is received Perform the state prediction and covariance prediction in step 1.

[0094] 2) Update steps (based on observations and constraints):

[0095] When GNSS observation data is available, or the system determines that the NHC / ZUPT conditions are met, an update step is performed. This step corrects the predicted values ​​by fusing the observations.

[0096] a) Calculate the Kalman gain:

[0097] According to step 2, the observation equation is... State vector The obtained partial derivatives are the observation matrix. Then calculate the Kalman gain. The observation matrix is ​​obtained by linearizing the observation equation. And combined with the predicted covariance at the current time Covariance matrix of observation noise The Kalman gain is calculated to weigh the reliability of predictions against observations.

[0098] ;

[0099] in, The observation noise covariance matrix is ​​used, and its magnitude represents the degree of confidence in the observed value.

[0100] b) Status update:

[0101] State updates include joint corrections to position, velocity, attitude, and bias. Attitude is updated using error quaternions and normalized after each update. The observation residuals, i.e., the actual observed values, are calculated. Compared with state-based prediction of observed values The difference. Utilizing Kalman gain. The prior state estimate is corrected to obtain the posterior state estimate. :

[0102] ;

[0103] When the GNSS signal is valid The absolute position and velocity provided for the GNSS receiver.

[0104] When the vehicle is in motion and the NHC conditions are met virtual observations And call the corresponding NHC observation model.

[0105] When the zero-speed detector determines that the vehicle is stationary virtual observations And call the ZUPT observation model.

[0106] c) Covariance update:

[0107] Update the error covariance matrix of the state estimate to reflect the reduction in uncertainty after this observation update.

[0108] ;

[0109] 4. NFC (Nonlinear Feedback Correction) mechanism.

[0110] Introducing an NFC (Nonlinear Feedback Correction) mechanism after each EKF update step can improve the stability and accuracy of the system.

[0111] The sensor zero bias estimated after EKF update The data is fed back to the IMU preprocessing stage in real time.

[0112] Before the next IMU data enters the prediction step, it is first corrected using the estimated zero bias:

[0113] ;

[0114] ;

[0115] This mechanism constitutes a closed-loop correction system that can suppress integral drift caused by zero bias at the source and significantly improve the system's stability under long-term GNSS signal interruption.

[0116] 5. Output results.

[0117] After each EKF filtering cycle, the final posterior state estimate is output. This refers to the high-precision navigation solution after dynamic correction and multi-source fusion, which includes:

[0118] High-precision absolute position and speed ;

[0119] High-stability attitude quaternions It can be converted into a rotation matrix. Or Euler angles;

[0120] Estimated IMU sensor zero bias .

[0121] This output provides stable and accurate pose information for the point cloud motion compensation in the subsequent step S3.

[0122] Step S3: First-level point cloud motion compensation (eliminating vehicle body motion).

[0123] The purpose of this step is to eliminate the spatial distortion caused by the rigid body motion of the detection vehicle itself on the point cloud acquired by the laser profilometer. In this invention, the laser profilometer and the IMU are rigidly mounted together on the vehicle body, forming an integrated vehicle sensing unit. Although the high-precision trajectory and attitude of this unit are obtained through step S2, the single-frame point cloud of the laser profilometer is formed by scanning within a non-zero time window. During this scanning period, the vehicle body and the integrated laser profilometer move as a whole, causing different points in the same frame of point cloud to be acquired under different sensor poses, resulting in point cloud artifacts or distortion.

[0124] First-level point cloud motion compensation involves uniformly transforming all points within the same frame to a common reference time (such as the end time of scanning) in the vehicle coordinate system based on their acquisition timestamps, in order to eliminate intra-frame distortion caused by the joint motion of the vehicle and the laser profilometer.

[0125] 1. Determine the point cloud timestamp and pose interpolation:

[0126] For each frame of laser point cloud, let its scanning start time be... The end time is Each point in this frame of the point cloud Each corresponds to a precise collection timestamp. .

[0127] Through the extended Kalman filter fusion output in step S2, a series of discrete time points have been obtained. High-precision vehicle body pose, including position vectors and attitude rotation matrix To compensate for the points The positional deviation caused by vehicle movement needs to be obtained by interpolation to get its data collection timestamp. precise pose Spherical linear interpolation is used for attitude quaternion interpolation to obtain... And linear interpolation is used for the position vector to obtain .

[0128] 2. Coordinate transformation and point cloud relocation:

[0129] The point cloud acquired by the laser profilometer is initially located in the sensor coordinate system L. Using a pre-calibrated extrinsic parameter matrix, the points can be... Transform from the sensor coordinate system to the vehicle body coordinate system B.

[0130] ;

[0131] Then, using the timestamp obtained by interpolation The corresponding pose will be the point Points transformed from vehicle coordinate system B to a unified world coordinate system W :

[0132] .

[0133] 3. Unify to the reference coordinate system:

[0134] To completely eliminate intra-frame distortion, all points within the same frame need to be uniformly transformed to a common reference pose. This invention chooses to uniformly transform the point cloud of this frame to the vehicle coordinate system at the end of the scan.

[0135] set up for The vehicle's position and orientation at any given moment. So, point... After exercise compensation, The correct position of the vehicle in the vehicle coordinate system at any given time is:

[0136] ;

[0137] The core of this step lies in calculating the timestamp collected from the point. to reference time The relative pose transformation, and the point This transformation is used for correction. After the above processing, the points originally collected in different poses are unified to... In the vehicle coordinate system at a given time, the intra-frame distortion caused by the rigid body motion of the vehicle is effectively eliminated, resulting in a point cloud after first-order point cloud motion compensation. .

[0138] Step S4: Second-level point cloud motion compensation (removal of high-frequency vibration of bogie).

[0139] The purpose of this step is to calculate the vibration displacement measured by the accelerometers mounted on the bogie and perform reverse displacement compensation on the point cloud to achieve secondary fine compensation. This eliminates the point cloud after primary motion compensation. Residual geometric errors caused by high-frequency bogie vibrations still exist. Although the first-level point cloud motion compensation in step S3 effectively eliminates the influence of rigid body motion, the laser profilometer is still modulated by the high-frequency vibrations of the bogie through the transmission path of the car body-bogie-rail. This vibration mainly originates from the direct impact between the wheels and rails during train operation (such as passing through uneven sections like welds and potholes). Its high frequency and small amplitude make it impossible for the IMU / GNSS fusion system installed on the car body to fully capture it. This leads to high-frequency jitter and blurring of the point cloud at the microscale, directly affecting the extraction accuracy of track geometric parameters (especially gauge and ride comfort).

[0140] To address this problem, this invention proposes a two-stage fine motion compensation method based on acceleration signals directly measured on the bogie.

[0141] 1. Vibration acceleration acquisition and preprocessing.

[0142] At least one high-sampling-rate (typically ≥1000Hz) triaxial accelerometer is rigidly mounted on the track inspection beam, with its coordinate system fixed to the beam. This sensor directly measures the raw vibration acceleration signals of the track inspection beam in three orthogonal directions. The acquired raw vibration acceleration signal Preprocessing:

[0143] a) Noise reduction filtering:

[0144] A bandpass filter is used to retain the vibration frequency band related to the first and main higher-order natural frequencies of the track inspection beam, while removing high-frequency measurement noise and low-frequency trend terms.

[0145] b) Remove the gravitational component:

[0146] Using the vehicle body attitude output in step S2 Calculate and subtract the component of gravitational acceleration in the accelerometer coordinate system.

[0147] 2. Vibration displacement calculation.

[0148] By analyzing the pre-processed pure vibration acceleration By performing a second integration, the vibration displacement of the track inspection beam relative to its static equilibrium position can be solved. .

[0149] In practical digital signal processing, this integration process is performed in the frequency domain, and a high-pass filter is introduced to suppress the inherent trend divergence problem during integration. The final result is the triaxial vibration displacement time series at the laser profilometer mounting point in the sensor coordinate system. .

[0150] 3. Point cloud reverse displacement compensation.

[0151] For point clouds after motion compensation of primary point clouds Each point in Its collection timestamp is From the vibration displacement time series Interpolation yields the vibration displacement at that precise moment. .

[0152] Subsequently, point A reverse displacement compensation is applied to counteract the effects of track inspection beam vibration, thereby obtaining the final high-precision point cloud after two-stage point cloud motion compensation. .

[0153] This step allows for the reverse correction of point cloud offsets caused by high-frequency vibrations of the bogie, thereby restoring the true geometric position.

[0154] The output of this step is the final point cloud after secondary point cloud motion compensation. This point cloud simultaneously eliminates distortions caused by rigid body motion of the vehicle and elastic vibration of the bogie, restoring the true geometric shape of the track surface to the greatest extent possible, thus laying a solid foundation for the high-precision extraction of subsequent track geometric parameters.

[0155] Step S5: Track geometry parameter extraction and smoothness analysis.

[0156] The purpose of this step is to automatically extract key track geometric parameters based on the final high-precision point cloud after two levels of motion compensation, and to analyze and evaluate its smoothness. This step is the final stage in realizing intelligent track condition diagnosis and maintenance decision-making.

[0157] 1. Establishment of orbital reference coordinate system and point cloud slicing.

[0158] To achieve accurate calculation of geometric parameters, a reference coordinate system extending along the track is first established:

[0159] a) Fitting the centerline of the track;

[0160] Based on the final clean point cloud, the rail head centerlines of the left and right rails are generated through feature extraction and model fitting. The centerline of the two rail head centerlines is the track centerline.

[0161] b) Definition of the reference coordinate system;

[0162] Establish a local coordinate system that varies with track mileage, using the track centerline as a reference. ;in, This represents the longitudinal mileage along the centerline of the track. The horizontal coordinates are perpendicular to the tangent plane of the track centerline. These are the vertical coordinates.

[0163] A series of cross-sectional slices perpendicular to the track centerline are generated at fixed intervals (e.g., 0.25 meters). All 3D point cloud data are projected onto these cross-sections to form a 2D profile point set for subsequent parameter calculations.

[0164] 2. Extraction of orbital geometric parameters.

[0165] Within each cross-sectional slice, the two-dimensional point cloud projected thereon is processed and calculated as follows:

[0166] a) Rail profile recognition;

[0167] Using a pre-established standard rail profile library, the complete profiles of the left and right rails are accurately identified from the cross-sectional point cloud through the Iterative Closest Point (ICP) algorithm or model matching.

[0168] b) Track gauge calculate;

[0169] At a specific distance (usually 16mm or 14mm) below the top surface of the rail, locate points on the working edges of the left and right rails respectively. The horizontal distance between these two points is the track gauge.

[0170] ;

[0171] c) Extremely high calculate;

[0172] Calculate the vertical height difference between the highest points of the tops of the left and right rails. If the right rail is higher than the left rail, the superelevation is positive; otherwise, it is negative.

[0173] ;

[0174] in, , These are the heights of the highest points of the tops of the left and right rails, respectively.

[0175] d) Horizontal (elevation and track orientation) extraction;

[0176] Elevation / Lower Slope: The longitudinal smoothness of the track is assessed by extracting the vertical coordinates of the centerline of the top of a single rail and analyzing their variation relative to the design longitudinal slope or smoothing reference.

[0177] Track alignment: The smoothness of the track in direction is evaluated by extracting the lateral coordinates of the centerline of the rail head of a single rail and analyzing its lateral offset relative to the centerline of the track.

[0178] e) Distortion calculate;

[0179] Twist is defined as the superelevation algebraic difference between two adjacent cross sections within a certain base length (e.g., 3 meters).

[0180] ;

[0181] in, and This represents two cross sections that are one base length apart.

[0182] 3. Track smoothness analysis and evaluation.

[0183] The extracted continuous sequence of geometric parameters (gauge, superelevation, elevation, and alignment) is used as input for track smoothness analysis.

[0184] a) Wavelength analysis;

[0185] Performing Fourier transforms or wavelet analyses on the sequence of geometric parameters to identify specific wavelength components present in the track geometry is crucial for analyzing vehicle-track dynamic responses.

[0186] b) Deviation statistics;

[0187] Calculate the deviation of each geometric parameter from its design value or management standard, such as standard deviation and extreme values.

[0188] c) TQI (track quality index) calculation;

[0189] Within a 200-meter or specified unit interval, the standard deviations of the measured geometric parameters—left and right elevation, left and right alignment, track gauge, levelness, and torsion—are calculated and summed to form the comprehensive quality evaluation index for that section. A smaller TQI value indicates better track smoothness.

Claims

1. A method for real-time detection of track state based on multi-sensor fusion and dynamic compensation, characterized in that, Includes the following steps: Step 1: Synchronous acquisition and preprocessing of multi-source data; Collect multi-source synchronized data from GNSS receivers, IMUs, and LiDARs mounted on the vehicle body, as well as accelerometers mounted on the bogies, and perform time synchronization and coordinate system unification; Step 2: Fusion of dynamic error correction and extended Kalman filtering; Establish state models and IMU prediction equations; construct GNSS observation equations as well as nonholonomic constraint and zero-velocity update constraint models; achieve tight coupling fusion and dynamic correction of GNSS and IMU through extended Kalman filtering, and combine nonlinear feedback correction strategies to achieve high-precision estimation of vehicle attitude, speed and position; Step 3: Motion compensation for first-level point clouds; For the vehicle attitude, velocity, and position data estimated in step 2, all points within the same frame are uniformly transformed to a common reference time vehicle coordinate system based on their acquisition timestamps. This eliminates the spatial distortion caused by the rigid body motion of the vehicle itself on the LiDAR-acquired point cloud, resulting in a motion-compensated point cloud. ; Step 4: Second-level point cloud motion compensation; By calculating the vibration displacement measured by the accelerometers mounted on the bogie and performing reverse displacement compensation on the point cloud, a two-stage fine compensation is achieved, eliminating the point cloud distortion caused by the first-stage point cloud motion compensation. The residual geometric errors caused by the high-frequency vibration of the bogie still exist, and the final point cloud after second-level point cloud motion compensation is obtained. ; Step 5: Track geometry parameter extraction and smoothness analysis; Based on the final point cloud after two-stage point cloud motion compensation It automatically extracts key track geometry parameters and analyzes and evaluates their smoothness, thereby achieving intelligent diagnosis of track status.

2. The real-time track state detection method based on multi-sensor fusion and dynamic compensation according to claim 1, characterized in that, In step 1, the GNSS receiver outputs position and velocity information, the IMU outputs triaxial angular velocity and triaxial acceleration, the LiDAR outputs the original orbit point cloud, and the accelerometer outputs vibration acceleration signals; the time synchronization specifically involves: Establish a unified time axis t, and let the timestamps of GNSS, IMU, and LiDAR be respectively... , and Then, after interpolation transformation, it is unified to the system clock. , is represented as: ； in, This represents the observation of sensor s at time t; and These are two adjacent valid timestamps output by sensor s; and These represent the time intervals of sensor s, respectively. and The sampled values; Indicates different types of sensors; The coordinate system one is specifically as follows: Transforming a local point seen in the sensor coordinate system to the view of the entire vehicle body, it is represented as follows: ； in, Let P be the position of point P in the vehicle coordinate system B; Let P be the position of point P in the sensor coordinate system L; Let L be the rotation matrix from the sensor coordinate system L to the vehicle coordinate system B. Let L be the translation vector from the sensor coordinate system L to the vehicle coordinate system B; Points viewed from the vehicle's perspective are transformed to a fixed global world coordinate system, as shown below: ； in, Let P be the absolute position of point P in the world coordinate system W. Let W be the rotation matrix from the vehicle coordinate system B to the world coordinate system W. Let be the translation vector from the vehicle coordinate system B to the world coordinate system W.

3. The real-time track state detection method based on multi-sensor fusion and dynamic compensation according to claim 2, characterized in that, In step 2, the state model and IMU prediction equations are established as follows: Step 2.1.1: Establish a data fusion state model centered on a GNSS receiver and IMU to describe the dynamic attitude, velocity, and position changes of the track inspection vehicle in the world coordinate system; define the state vector. : ； in, This represents the absolute position in the world coordinate system W. The velocity in world coordinates W; The attitude of the vehicle coordinate system B relative to the world coordinate system W; These are the zero bias values ​​for the gyroscope and accelerometer, respectively. T It is the transpose symbol; Step 2.1.2: Construct the IMU prediction equation; Includes: Equation of the rate of change of position: ； Equation for the rate of change of velocity: ； Equation of the rate of change of attitude quaternions: ； in, It is a position vector; It is a velocity vector; These are triaxial acceleration observations obtained from the IMU; This is the gravity vector in the world coordinate system; This is the attitude vector; This is a quaternion operation matrix; These are the triaxial angular velocity observations obtained from the IMU; Zero bias vector of a gyroscope and the zero bias vector of the accelerometer for: ； in, This represents the random walk noise at zero bias of the gyroscope. This refers to the random walk noise at the zero bias of the accelerometer. Step 2.1.3: State prediction and covariance prediction; Using the optimal state prior estimate from the previous time step and current IMU measurements By time step The integral is used to perform discrete-time prediction of the state, and the predicted prior state estimate at the current time is obtained. for: ； Among them, the function f (·) represents the numerical integration process of a continuous-time differential equation; Using the covariance of the previous time step sum function f Jacobian matrix of (·) Perform discrete-time prediction of the error covariance matrix, and predict the covariance at the current time. for: ； in, It is the process noise covariance matrix, representing the new uncertainties introduced by IMU measurement noise and zero-bias random walk noise.

4. The real-time track state detection method based on multi-sensor fusion and dynamic compensation according to claim 3, characterized in that, In step 2, the GNSS observation equations, as well as the nonholonomic constraints and zero-rate update constraint models, are constructed as follows: Step 2.2.1: Observe the position of the GNSS receiver antenna phase center. In relation to the system state, the details are as follows: ； in, It is the predicted origin position of the vehicle body in the state vector; The GNSS antenna is installed relative to the vehicle's origin. Lever compensation item; Noise for GNSS position measurement; This refers to the measured position of the GNSS antenna phase center in the world coordinate system W. To transfer from the world coordinate system W to the vehicle coordinate system The rotation matrix; velocity vector measured by GNSS In relation to the system state, the details are as follows: ； in, It is the predicted velocity of the vehicle's origin in the state vector; It is the lever arm speed compensation term. The angular velocity of the vehicle body; It is GNSS velocity measurement noise; The velocity observation of the vehicle in the world coordinate system W, as measured by GNSS; Step 2.2.2: Establish nonholonomic constraints By constraining the velocity, attitude estimation can be indirectly improved, thereby suppressing heading drift caused by gyroscope integration, as shown below: ； in, It is the forward speed of the train body; and These are the lateral speed and vertical speed of the train body, respectively. Zero-rate update constraint The zero bias of the accelerometer is directly estimated and suppressed, as follows: 。 5. The real-time track state detection method based on multi-sensor fusion and dynamic compensation according to claim 4, characterized in that, In step 2, the tight coupling fusion and dynamic correction of GNSS and IMU are achieved through extended Kalman filtering as follows: Step 2.3.1: Prediction based on IMU; When a new IMU measurement is received Then perform the state prediction and covariance prediction in step 2.1.3; Step 2.3.2: Update based on observations and constraints; When GNSS observation data is available, or the system determines that the nonholonomic constraints and zero-rate update constraints are met, the following update steps are executed: a) Calculate the Kalman gain; According to step 2.2.1, the observation equation is in State vector The obtained partial derivatives are the observation matrix. Then calculate the Kalman gain. The observation matrix is ​​obtained by linearizing the observation equation. And combined with the predicted covariance at the current time Covariance matrix of observation noise Calculate Kalman gain To weigh the credibility of predictions against observations; ； b) Status update; Calculate the observation residuals, i.e., the actual observed values. Compared with state-based prediction of observed values The difference; utilizing Kalman gain The prior state estimate is corrected to obtain the posterior state estimate. : ； When GNSS signal is valid Absolute position and velocity provided to the GNSS receiver; When the vehicle is in motion and the IMU conditions are met virtual observations And call the corresponding IMU observation model; When the zero-speed detector determines that the vehicle is stationary virtual observations And call the zero-rate update constraint model; c) Covariance update; Update the error covariance matrix of the state estimate to reflect the reduction in uncertainty after this observation update; ； in, The updated covariance; It is an identity matrix.

6. The real-time track state detection method based on multi-sensor fusion and dynamic compensation according to claim 5, characterized in that, In step 2, the nonlinear feedback correction strategy specifically includes: The estimated IMU sensor bias after Kalman filtering update Real-time feedback is sent to the IMU preprocessing stage; Before the next IMU data enters the prediction step, it is first corrected using the estimated zero bias: ； ； in, The angular velocity observation of the IMU after zero bias compensation; The IMU acceleration observations are after zero bias compensation.

7. The real-time track state detection method based on multi-sensor fusion and dynamic compensation according to claim 6, characterized in that, Step 3 is as follows: Step 3.1: Determine the point cloud timestamp and pose interpolation; For each frame of laser point cloud, let its scanning start time be... The end time is Each point in the point cloud of this frame Each corresponds to a precise collection timestamp. Then, by fusing the output through the extended Kalman filter in step 2, a series of discrete time points are obtained. High-precision vehicle body pose, including position vectors and attitude rotation matrix Spherical linear interpolation is used for attitude quaternion interpolation to obtain the timestamp. Corresponding attitude rotation matrix Linear interpolation is used for the position vector to obtain the timestamp. Corresponding position vector ; Step 3.2: Coordinate transformation and point cloud relocation; Using a pre-calibrated extrinsic matrix, the points in the point cloud acquired by LiDAR are... Points transformed from sensor coordinate system L to vehicle coordinate system B : ； Then use the timestamp obtained by interpolation The corresponding pose will be the point Points transformed from vehicle coordinate system B to a unified world coordinate system W : ； in, and They represent The pose rotation matrix and translation vector from the vehicle coordinate system B to the world coordinate system W at any given time; Step 3.3: Unify to the reference coordinate system; set up for The vehicle's position and posture at any given moment; then, point After exercise compensation, Correct position in the vehicle body coordinate system at any moment for: 。 8. The real-time track state detection method based on multi-sensor fusion and dynamic compensation according to claim 7, characterized in that, Step 4 is as follows: Step 4.1: Vibration acceleration acquisition and preprocessing; At least one high-sampling-rate triaxial accelerometer is rigidly mounted on the track inspection beam, with its coordinate system fixed to the beam, to collect the vibration acceleration of the track inspection beam in three orthogonal directions. And perform noise reduction filtering to remove the gravity component; Step 4.2: Vibration displacement calculation; By analyzing the pre-processed pure vibration acceleration By performing a second integration, the vibration displacement of the track inspection beam relative to its static equilibrium position can be solved. Finally, the triaxial vibration displacement time series at the LiDAR mounting point in the sensor coordinate system was obtained. ; , and They are respectively in x , y and z Displacement in direction; Step 4.3: Point cloud reverse displacement compensation; For point clouds after motion compensation of primary point clouds Each point in Its collection timestamp is From vibration displacement time series Interpolation yields the vibration displacement at that precise moment. ; Subsequently, point A reverse displacement compensation is applied to counteract the effects of track inspection beam vibration, thereby obtaining the final high-precision point cloud after two-stage point cloud motion compensation. .

9. The real-time track state detection method based on multi-sensor fusion and dynamic compensation according to claim 8, characterized in that, Step 5 is as follows: Step 5.1: Establishing the orbital reference coordinate system and slicing the point cloud; Orbit centerline fitting: based on final point cloud By extracting features and fitting the model, the center lines of the rail heads of the left and right rails are generated. Definition of reference coordinate system: A local coordinate system that varies with track mileage, with the track centerline as the reference. ;in, This represents the longitudinal mileage along the centerline of the track. The horizontal coordinates are perpendicular to the tangent plane of the track centerline. Vertical coordinates; A series of cross-sectional slices perpendicular to the centerline of the track are generated at fixed intervals. All three-dimensional point cloud data are projected onto these cross-sections to form a two-dimensional profile point set for subsequent parameter calculation. Step 5.2: Extraction of orbital geometry parameters; Within each cross-sectional slice, the two-dimensional point cloud projected thereon is processed and calculated as follows: a) Rail profile recognition; Using a pre-established standard rail profile library, the complete profiles of the left and right rails are accurately identified from the cross-sectional point cloud through iterative nearest point algorithm or model matching. b) Track gauge calculate; At a specific distance below the top surface of the rail, locate points on the working edges of the left and right rails respectively. , The horizontal distance between these two points is the track gauge. ； c) Extremely high calculate; Calculate the vertical height difference between the highest points of the tops of the left and right rails; if the right rail is higher than the left rail, the superelevation is positive, otherwise it is negative; ； in, , These are the heights of the highest points of the tops of the left and right rails, respectively. d) Horizontal extraction; Elevation / Lower Slope: The longitudinal smoothness of the track is assessed by extracting the vertical coordinates of the centerline of the top of a single rail and analyzing their variation relative to the design longitudinal slope or smoothing reference. Track alignment: The smoothness of the track in the direction is evaluated by extracting the lateral coordinates of the centerline of the rail head of a single rail and analyzing its lateral offset relative to the centerline of the track. e) Distortion calculate: Twisting Defined as the algebraic difference in superelevation between two adjacent cross sections within a certain base length; ； in, and This represents two cross sections that are one base length apart; and These are the superelevation of two adjacent cross sections; Step 5.3: Track smoothness analysis and evaluation; The extracted continuous geometric parameter sequence is used as input for track smoothness analysis: a) Wavelength analysis; Perform Fourier transform or wavelet analysis on the sequence of geometric parameters to identify specific wavelength components existing in the orbital geometry; b) Deviation statistics; Calculate the deviation of each geometric parameter from its design value or management standard; c) Calculation of the track quality index; Within the defined unit interval, the standard deviations of the measured values ​​of the following geometric parameters are calculated: left and right elevation, left and right track orientation, track gauge, level, and torsion. These standard deviations are then summed to form the comprehensive quality evaluation index for the unit interval.

Images