A curve track route monitoring system for a rail vehicle

Abstract

The application discloses a curve track line monitoring system for a rail vehicle, and particularly relates to the technical field of rail transit safety monitoring. A coupling identification module identifies a vehicle-line-bridge coupling state and outputs a feature vector; a weight adjustment module calls a preset model, dynamically adjusts a fusion weight of a geometric measurement sensor, and outputs an optimized weight; a feedback correction module fuses and estimates track geometric parameters according to the optimized weight, evaluates a confidence level, and feeds back and corrects a low-confidence-level result, and outputs a track geometric parameter with a confidence level label; a reference alignment module performs dynamic space-time reference alignment, and outputs an aligned track geometric data set; the coupling identification module identifies the vehicle-line-bridge coupling state, the weight adjustment module dynamically optimizes the fusion weight, the feedback correction module evaluates the confidence level and corrects the low-confidence-level result, and the reference alignment module realizes space-time alignment, so that a high-confidence-level track geometric data set is finally output, and the problem of measurement distortion under a dynamic coupling condition is solved.

Description

Technical Field

[0001] This invention relates to the field of rail transit safety monitoring technology, and more specifically, to a curved track route monitoring system for a track machine. Background Technology

[0002] In the field of rail transit, the geometric condition detection of curved track sections is a crucial step in ensuring operational safety and passenger comfort. Current technologies typically employ detection systems based on inertial navigation and multi-sensor fusion to acquire track geometric parameters. These systems utilize inertial measurement units, laser scanners, and image acquisition sensors mounted on track inspection vehicles (such as track inspection cars or integrated inspection trains) to collect real-time data on vehicle attitude, track cross-sectional contours, and track appearance during vehicle operation. After data fusion processing, the system outputs geometric parameters such as track gauge, alignment, elevation, and curvature. This method has been widely applied in the periodic inspection of existing lines and can meet the inspection requirements under normal operating conditions.

[0003] However, in actual railway lines, curved track sections are often located above long-span bridges or viaducts. Under dynamic conditions such as high-speed train passage, strong winds, or temperature changes, complex coupled vibrations occur between the train body, rails, and bridge structure. Existing detection systems typically use data fusion models based on quasi-static assumptions, which assume that the track geometry remains stable during the measurement period, and that changes in train body attitude primarily reflect the track's own irregularities. When the bridge undergoes elastic deformation (such as vertical deflection, lateral swaying, or torsion) at frequencies similar to the train body's vibration frequency, the sensor data simultaneously contains both the actual track geometry and the structural dynamic response information, making effective separation difficult. Therefore, this invention proposes a curved track monitoring system for track maintenance machines to address these issues. Summary of the Invention

[0004] To achieve the above objectives, the present invention provides the following technical solution: A curved track monitoring system for a track machine, comprising: The coupling identification module is used to acquire dynamic environmental parameters and vehicle attitude parameters during the operation of the track machine, and to identify the coupling state of the vehicle-track-bridge system based on the dynamic environmental parameters and vehicle attitude parameters, and output the coupling state feature vector. The weight adjustment module is used to dynamically adjust the fusion weight allocation of the geometric measurement sensor based on the coupling state feature vector and call the preset fusion weight allocation model, and output the optimized fusion weight parameters. The feedback correction module is used to perform fusion estimation of orbital geometric parameters based on the optimized fusion weight parameters, evaluate the confidence of the fusion result in real time, perform feedback correction on measurement results with confidence scores lower than the preset confidence threshold, and output orbital geometric parameters with confidence scores labels. The benchmark alignment module is used to perform dynamic spatiotemporal benchmark alignment on orbital geometric parameters with confidence labels, and output the aligned orbital geometric dataset.

[0005] Dynamic environmental parameters include real-time wind speed parameters at the location of the track machine, real-time operating speed parameters of the track machine, and real-time vibration parameters of the bridge structure; vehicle attitude parameters are collected by the inertial measurement unit, including the three-axis angular velocity parameters and three-axis acceleration parameters of the track machine.

[0006] In a preferred embodiment, the coupling state feature vector specifically refers to: Time-frequency domain analysis was performed on the real-time vibration parameters of the bridge structure to extract the deformation frequency components of the bridge structure; time-frequency domain analysis was also performed on the triaxial angular velocity parameters and triaxial acceleration parameters to extract the vehicle body vibration frequency components. Calculate the coherence coefficient between the frequency components of bridge structure deformation and vehicle vibration, and determine the coupling strength level at the current moment by calling the preset coupling strength classification model based on the coherence coefficient, real-time wind speed parameters, and real-time operating speed parameters. Based on the coupling strength level and the energy ratio of the bridge structure deformation frequency component to the vehicle vibration frequency component, a coupling state feature vector is generated to characterize the coupling state of the vehicle-track-bridge system.

[0007] The preset coupling strength classification model is a multi-input classification model trained based on historical data. The preset coupling strength classification model outputs the coupling strength level at the current moment based on the nonlinear mapping relationship between the coherence coefficient, real-time wind speed parameters and real-time operating speed parameters.

[0008] In a preferred embodiment, the preset fusion weight allocation model specifically refers to: The preset fusion weight allocation model is a multi-input multi-output mapping model trained based on historical data. The input parameter is the coupling state feature vector, and the output parameter is the fusion weight value of each geometric measurement sensor. The geometric measurement sensor includes a laser scanning sensor and an image acquisition sensor. The preset fusion weight allocation model contains multiple weight allocation sub-models corresponding to different coupling strength levels. Each weight allocation sub-model stores the optimal weight allocation scheme that minimizes the error of the fusion result under different coupling conditions.

[0009] In a preferred embodiment, a preset fusion weight allocation model is invoked to dynamically adjust the fusion weight allocation of the geometric measurement sensors, and the optimized fusion weight parameters are output, specifically including: Based on the coupling strength level in the coupling state feature vector, the corresponding weight allocation sub-model is selected from the preset fusion weight allocation model, and the weight allocation sub-model outputs the fusion weight value corresponding to each geometric measurement sensor according to the coupling state feature vector. The fusion weight values ​​are normalized to generate optimized fusion weight parameters where the sum of the weights of each geometric measurement sensor is equal to the preset total weight.

[0010] In a preferred embodiment, the orbital geometric parameters are fused and estimated, specifically including: The system acquires optimized fusion weight parameters, raw track geometric measurement data collected by the geometric measurement sensor, track cross-section contour point cloud data output by the laser scanning sensor, and track appearance image data output by the image acquisition sensor. The point cloud data of the track cross-section contour output by the laser scanning sensor is processed to extract the first track geometric parameter sequence; the track appearance image data output by the image acquisition sensor is processed to extract the second track geometric parameter sequence, wherein the image processing includes feature point extraction, stereo matching and three-dimensional coordinate transformation. The first orbital geometric parameter sequence and the second orbital geometric parameter sequence are spatiotemporally registered to place them under a unified spatial coordinate system and time reference. Based on the optimized fusion weight parameters, the first orbital geometric parameter sequence and the second orbital geometric parameter sequence after spatiotemporal registration are weighted and summed to generate preliminary fused orbital geometric parameters.

[0011] In a preferred embodiment, the confidence level of the real-time evaluation of the fusion result refers to: Obtain preliminary fused orbital geometric parameters and the first and second orbital geometric parameter sequences; Calculate the first deviation value between each data point in the first orbital geometric parameter sequence and the corresponding data point in the preliminary fused orbital geometric parameters to obtain the first deviation sequence composed of the first deviation values ​​of each data point; Calculate the second deviation value between each data point in the second orbital geometric parameter sequence and the corresponding data point in the preliminary fused orbital geometric parameters to obtain the second deviation sequence composed of the second deviation values ​​of each data point; The first deviation statistical characteristic value is calculated based on the first deviation sequence, the second deviation statistical characteristic value is calculated based on the second deviation sequence, and a confidence index is generated based on the first and second deviation statistical characteristic values. The first deviation statistical characteristic value represents the overall deviation between the first orbital geometric parameter sequence and the initially fused orbital geometric parameters, the second deviation statistical characteristic value represents the overall deviation between the second orbital geometric parameter sequence and the initially fused orbital geometric parameters, and the confidence index is negatively correlated with both the first and second deviation statistical characteristic values. The confidence index is compared with the preset confidence threshold. When the confidence index is greater than or equal to the preset confidence threshold, the preliminary fused orbital geometry parameters are directly output as orbital geometry parameters with confidence labels. When the confidence index is less than the preset confidence threshold, the feedback correction mechanism is triggered. Abnormal data points are identified based on the first deviation sequence and the second deviation sequence. The abnormal data points in the first track geometric parameter sequence and the second track geometric parameter sequence are corrected based on the preset compensation model. The weighted summation is recalculated according to the optimized fusion weight parameters to generate the corrected track geometric parameters as track geometric parameters with confidence labels as output.

[0012] In a preferred embodiment, abnormal data in the first and second orbital geometric parameter sequences are corrected based on a preset compensation model, specifically including: The first deviation value of each data point in the first deviation sequence is compared with the preset first anomaly threshold, and the second deviation value of each data point in the second deviation sequence is compared with the preset second anomaly threshold. When the first deviation value of any data point is greater than or equal to the first abnormal threshold, the data point in the first orbital geometric parameter sequence corresponding to that data point is marked as the first abnormal data point. When the second deviation value of any data point is greater than or equal to the second abnormal threshold, the data point in the second orbital geometric parameter sequence corresponding to that data point is marked as the second abnormal data point. The preset compensation model contains multiple compensation sub-models corresponding to different coupling strength levels. Each compensation sub-model stores abnormal data correction coefficients trained based on historical data. Based on the coupling strength level in the coupling state feature vector, a corresponding compensation sub-model is selected from the preset compensation models. The first abnormal data point and the second abnormal data point are respectively input into the selected compensation sub-model, and the compensation sub-model outputs the corrected first track geometric parameter data point and the corrected second track geometric parameter data point. Replace the corresponding outlier data points in the first orbital geometric parameter sequence with the corrected first orbital geometric parameter data points to generate the corrected first orbital geometric parameter sequence; replace the corresponding outlier data points in the second orbital geometric parameter sequence with the corrected second orbital geometric parameter data points to generate the corrected second orbital geometric parameter sequence.

[0013] In a preferred embodiment, dynamic spatiotemporal reference alignment of orbital geometric parameters with confidence labels refers to: The track geometry parameters with confidence labels are obtained from the feedback correction module. The track geometry parameters with confidence labels include the track geometry parameter values, confidence index and corresponding original acquisition timestamp and original mileage value for each data point. Based on the coupling strength level in the coupling state feature vector, query the corresponding preset spatiotemporal reference correction strategy under the current dynamic conditions; Based on the original acquisition timestamp, original mileage value, and spatiotemporal reference correction strategy, the track geometry parameters with confidence labels are aligned on the time axis to eliminate time reference deviations caused by fluctuations in track machine operating speed and data acquisition delays. Based on the bridge structure deformation frequency component in the coupled state feature vector, the time axis aligned track geometry parameters are spatially aligned to compensate for the spatial position offset caused by bridge elastic deformation and vehicle vibration, so that the track geometry parameter values ​​of each data point match the actual track physical position. The orbital geometry parameters aligned with the spatiotemporal reference are directly output as the aligned orbital geometry dataset.

[0014] The technical effects and advantages of this invention are as follows: This invention acquires dynamic environmental parameters and vehicle attitude parameters during the operation of the track machine through a coupling identification module, and identifies the coupling state of the vehicle-track-bridge system based on these parameters, outputting a coupling state feature vector. This module collects real-time wind speed parameters, real-time operating speed parameters, real-time vibration parameters of the bridge structure, and triaxial angular velocity and triaxial acceleration parameters from the inertial measurement unit. Through time-frequency domain analysis, it extracts the deformation frequency components of the bridge structure and the vibration frequency components of the vehicle body, calculates the coherence coefficient, and finally generates a coupling state feature vector containing coupling strength level, energy proportion, and characteristic frequencies. This mechanism enables the system to accurately distinguish the dynamic disturbance components generated by the actual track geometry and the vehicle-track-bridge coupled vibration under multiple dynamic loads, such as strong winds and high-speed passage on long-span bridge curves. It solves the signal aliasing problem of existing quasi-static measurement models under dynamic conditions, significantly improving the environmental adaptability and anti-interference capability of the measurement system.

[0015] This invention utilizes a weight adjustment module to dynamically adjust the fusion weight allocation of the geometric measurement sensors based on the coupling state feature vector and a preset fusion weight allocation model, outputting optimized fusion weight parameters. The preset fusion weight allocation model is a multi-input multi-output mapping model trained on historical data, containing multiple weight allocation sub-models corresponding to different coupling strength levels. Each sub-model stores the optimal weight allocation scheme that minimizes the fusion error. When the coupling strength level changes, the weight adjustment module automatically selects the corresponding sub-model, outputting the original weight values ​​of the laser scanning sensor and the image acquisition sensor. After normalization, the optimized fusion weight parameters are obtained. This dynamic adjustment mechanism balances the weights of the two sensors under weak coupling conditions and significantly improves the weight of the laser scanning sensor under strong coupling conditions. It overcomes the error amplification defect of traditional fixed-weight fusion under dynamic conditions, improving the accuracy and robustness of track geometric parameter fusion estimation.

[0016] This invention utilizes a feedback correction module to perform fusion estimation of track geometric parameters based on optimized fusion weight parameters and to evaluate the confidence level of the fusion results in real time. Measurement results with confidence levels below a preset confidence threshold are subject to feedback correction, outputting track geometric parameters with confidence labels. Simultaneously, a benchmark alignment module performs dynamic spatiotemporal benchmark alignment on the track geometric parameters with confidence labels, outputting an aligned track geometric dataset. The feedback correction module calculates the first deviation sequence between the first track geometric parameter sequence and the fused value, and the second deviation sequence between the second track geometric parameter sequence and the fused value, generating a confidence index based on the deviation statistical characteristic value. When the confidence level falls below the threshold, the system triggers feedback correction, using a preset compensation model to correct abnormal data points and re-fuse. The benchmark alignment module queries a preset spatiotemporal benchmark correction strategy based on the coupling strength level, performing time axis alignment and spatial axis alignment to compensate for deviations caused by speed fluctuations, bridge deformation, and vehicle vibration. Finally, it outputs a high-confidence, high-consistency aligned track geometric dataset, providing a reliable basis for track condition assessment and maintenance decisions. Attached Figure Description

[0017] To facilitate understanding by those skilled in the art, the present invention will be further described below with reference to the accompanying drawings; Figure 1 This is a schematic diagram of a curved track monitoring system for a track machine according to the present invention. Detailed Implementation

[0018] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those of ordinary skill in the art without creative effort are within the scope of protection of the present invention.

[0019] Reference Figure 1 The following examples were obtained: Example 1: A curved track monitoring system for a track machine, comprising: The coupling identification module is used to acquire dynamic environmental parameters and vehicle attitude parameters during the operation of the track machine, and to identify the coupling state of the vehicle-track-bridge system based on the dynamic environmental parameters and vehicle attitude parameters, and output the coupling state feature vector. The weight adjustment module is used to dynamically adjust the fusion weight allocation of the geometric measurement sensor based on the coupling state feature vector and call the preset fusion weight allocation model, and output the optimized fusion weight parameters. The feedback correction module is used to perform fusion estimation of orbital geometric parameters based on the optimized fusion weight parameters, evaluate the confidence of the fusion result in real time, perform feedback correction on measurement results with confidence scores lower than the preset confidence threshold, and output orbital geometric parameters with confidence scores labels. The benchmark alignment module is used to perform dynamic spatiotemporal benchmark alignment on orbital geometric parameters with confidence labels, and output the aligned orbital geometric dataset.

[0020] Dynamic environmental parameters include real-time wind speed at the location of the track-mounted vehicle, real-time operating speed of the track-mounted vehicle, and real-time vibration parameters of the bridge structure. Vehicle attitude parameters are collected by the inertial measurement unit, including the track-mounted vehicle's three-axis angular velocity and three-axis acceleration parameters. When the track-mounted vehicle operates on a curved section of a long-span bridge, lateral winds can cause lateral displacement and torsional deformation of the bridge, while also affecting the track-mounted vehicle's wind-receiving area and wheel-rail lateral forces. Real-time wind speed parameters are used to quantify the intensity and direction of wind loads, serving as a basis for determining whether the bridge's elastic deformation is dominated by wind-induced vibration.

[0021] For example, when a track-mounted machine passes through a curved section of a high-speed railway bridge, an anemometer installed at the mid-span of the bridge and on the roof of the track-mounted machine collects data at a frequency of 10Hz, recording an instantaneous wind speed of 18m / s and a wind direction at an angle of 75 degrees to the track. This parameter is input into the coupling identification module and analyzed together with bridge vibration parameters. If the wind speed exceeds 15m / s and the wind direction is close to the lateral direction, the system will increase the weight given to the frequency components of bridge deformation.

[0022] The operating speed of the track machine directly affects the frequency and amplitude of wheel-rail excitation, thereby altering the coupling strength between vehicle vibration and bridge response. At different speeds, the centrifugal force, suspension system dynamics, and measurement error characteristics of track geometry parameters differ for the same curve segment. Real-time operating speed parameters are used to calibrate the transformation relationship between the time axis and the spatial axis and serve as one of the key inputs to the coupling strength classification model.

[0023] For example, a track-mounted vehicle travels at 250 km / h across a bridge section with a curve radius of 5000 m. The speed sensor (such as radar speed measurement or encoder) outputs a real-time speed value that fluctuates between 248 km / h and 252 km / h. When this parameter is compared with the bridge's vibration frequency components, the system identifies that the low-frequency swaying of the vehicle body caused by the speed fluctuation is coupled with the bridge's first-order lateral bending frequency (approximately 0.8 Hz), thus determining that the vehicle is currently in a strongly coupled state.

[0024] Bridge structures undergo elastic deformation and vibration under train loads, wind loads, and temperature changes. The frequency components, amplitude, and phase information of these vibrations directly reflect the dynamic response state of the bridge. Real-time vibration parameters are typically collected by accelerometers, displacement gauges, or strain sensors installed at key sections of the bridge (such as mid-span or above the supports) to extract the frequency components of the bridge structure's deformation.

[0025] For example, a triaxial accelerometer is installed at the bottom of the mid-span of a bridge, with a sampling frequency of 200Hz. When a track-mounted vehicle passes over it, the collected vibration signal, after Fourier transform, shows a significant peak at 0.6Hz with an amplitude of 0.05g, corresponding to the bridge's first-order vertical bending mode; simultaneously, a secondary peak appears at 1.2Hz, corresponding to the lateral torsional mode. After performing coherence analysis between this parameter and the vehicle's vibration parameters, the vehicle-bridge energy transfer efficiency can be quantified.

[0026] The three-axis angular velocity parameters are acquired by gyroscopes in the inertial measurement unit, measuring the angular velocity changes of the track vehicle around the X-axis (tilt), Y-axis (pitch), and Z-axis (yaw). These parameters describe the rotational motion of the track vehicle under dynamic conditions and are a key data source for separating changes in vehicle attitude from track geometry. On curved sections, the yaw angular velocity directly reflects changes in track curvature, while the roll angular velocity is related to superelevation and lateral acceleration of the vehicle body.

[0027] For example, the inertial measurement unit outputs data at 100Hz and records the triaxial angular velocities at a certain moment as follows: 0.02 rad / s around the X-axis (tilting inwards towards the curve), 0.005 rad / s around the Y-axis (tilting forward), and 0.12 rad / s around the Z-axis (yawing to the right). Combined with the real-time operating speed parameters (200 km / h), the theoretical curvature of this section can be calculated to be approximately 0.0006 rad / m, matching the designed curve radius. If the yaw angular velocity shows an abnormal jump, it indicates a possible sudden change in track orientation or sensor vibration interference.

[0028] The three-axis acceleration parameters are collected by accelerometers in the inertial measurement unit, measuring the linear acceleration of the track machine along the X-axis (lateral), Y-axis (longitudinal), and Z-axis (vertical). These parameters are used to analyze changes in the inertial force of the vehicle body, the vertical and lateral interactions between the wheel and rail, and to identify abnormal impacts caused by track irregularities or bridge deformation. On curved sections, lateral acceleration is closely related to centrifugal force and superelevation settings, while vertical acceleration reflects the excitation of the vehicle body by the vertical deflection of the bridge.

[0029] The coupling state feature vector specifically refers to: Step 1: Perform time-frequency domain analysis on the real-time vibration parameters of the bridge structure to extract the deformation frequency components of the bridge structure. For the real-time vibration parameters of the bridge collected by accelerometers or displacement sensors installed on key sections of the bridge (such as mid-span and above the supports), perform time-frequency domain analysis using short-time Fourier transform or wavelet transform methods to convert the vibration waveform in the time domain into a power spectral density distribution in the frequency domain. Extract the characteristic frequency components with concentrated energy from them. These characteristic frequency components correspond to the elastic deformation modes of the bridge structure under dynamic loads, including the first-order vertical bending frequency, the transverse bending frequency, and the torsional frequency, collectively referred to as the deformation frequency components of the bridge structure.

[0030] For example, an accelerometer installed at the bottom of the mid-span of a bridge collects vibration data at a sampling rate of 200Hz. When a track machine passes by, the system captures a 20-second vibration waveform, performs Hanning windowing, and calculates a short-time Fourier transform to obtain a frequency-time spectrum. In this spectrum, a continuous and stable energy peak appears at 0.6Hz, corresponding to the bridge's first-order vertical bending frequency; a secondary peak appears at 1.8Hz, corresponding to the second-order bending frequency. The system extracts these frequency values ​​and their corresponding amplitude and phase information as the frequency components of the bridge structure's deformation.

[0031] Step 2: Perform time-frequency domain analysis on the three-axis angular velocity and acceleration parameters to extract the vehicle body vibration frequency components. Using the same time-frequency domain analysis method as in Step 1, perform spectral transformation on the data for each axis to obtain the vibration energy distribution of the vehicle body in six degrees of freedom. From these distributions, extract significant frequency components caused by wheel-rail excitation, suspension system response, and track irregularities. These components reflect the vehicle body's pitch, roll, yaw, and heave vibration modes during operation, collectively referred to as the vehicle body vibration frequency components.

[0032] For example, the inertial measurement unit outputs triaxial angular velocity and triaxial acceleration at 100Hz. The system performs a Fast Fourier Transform on the vertical acceleration data and finds multiple energy peaks in the range of 8Hz to 12Hz, corresponding to high-frequency excitation of the wheel-rail vertical force. Simultaneously, spectral analysis of the roll velocity data reveals a significant peak at 2Hz, corresponding to the vehicle roll mode. These frequency values ​​and their corresponding amplitudes constitute the frequency components of the vehicle vibration.

[0033] Step 3: Calculate the coherence coefficient between the bridge structural deformation frequency components and the vehicle vibration frequency components. Perform frequency-point coherence calculations on the spectral sequences of the bridge structural deformation frequency components extracted in Step 1 and the vehicle vibration frequency components extracted in Step 2. The coherence coefficient characterizes the degree of linear correlation between two signals at a certain frequency, with a value between 0 and 1. The specific calculation method is as follows: For the bridge vibration signal and the vehicle vibration signal within the same time window, calculate their respective self-power spectra and the cross-power spectrum between them. Divide the square of the modulus of the cross-power spectrum by the product of the two self-power spectra to obtain the coherence function as a function of frequency. In practical engineering applications, the system focuses on the coherence coefficient values ​​in the frequency bands near the bridge's fundamental frequency and the vehicle's main frequency. The closer this value is to 1, the tighter the energy transfer between the vehicle and the bridge at that frequency, and the stronger the coupling.

[0034] For example, at the first-order vertical bending frequency of the bridge (0.6 Hz), the coherence coefficient between the vehicle's vertical acceleration and the bridge's vertical vibration was calculated to be 0.92, indicating that 92% of the energy in the vehicle's vertical vibration is synchronized with the bridge's deformation at 0.6 Hz, demonstrating a high degree of vehicle-bridge coupling at this frequency. Meanwhile, at 1.8 Hz, the coherence coefficient was calculated to be 0.31, indicating weak correlation between the vehicle and bridge responses at this frequency, primarily due to localized wheel-rail excitation.

[0035] Step 4: Based on the coherence coefficient, real-time wind speed parameters, and real-time operating speed parameters, the pre-set coupling strength classification model is invoked to determine the coupling strength level at the current moment. The coherence coefficient calculated in Step 3 (focusing on the maximum coherence value near the bridge's fundamental frequency), along with the real-time wind speed parameters (in meters per second) and the track machine's real-time operating speed parameters (in kilometers per hour) obtained in Step 2, are used as input feature vectors and input into the pre-trained coupling strength classification model. This model is a multi-input single-output classifier, internally storing nonlinear mapping relationships learned from a large amount of historical data. Based on the combination of input features, the model outputs a discrete coupling strength level label, such as Level 1 (weak coupling), Level 2 (moderate coupling), and Level 3 (strong coupling). The coupling strength level characterizes the intensity of the interaction between the vehicle-track-bridge system under dynamic conditions.

[0036] For example, the current measured coherence coefficient is 0.92, the real-time wind speed is 18 meters per second (strong crosswind), and the track machine's operating speed is 250 kilometers per hour. Inputting these three values ​​into a preset coupling strength classification model (trained using a support vector machine algorithm on historical data), the model outputs a coupling strength level of "Level 3 (Strong Coherence)". If the wind speed drops to 5 meters per second and the vehicle speed drops to 80 kilometers per hour, while the coherence coefficient remains at 0.92, the model might output "Level 2 (Medium Coherence)", indicating that wind speed and vehicle speed have a significant moderating effect on coupling strength.

[0037] Step 5: Based on the coupling strength level and the energy ratio of the bridge structure deformation frequency component and the vehicle vibration frequency component, a coupling state feature vector is generated to characterize the coupling state of the vehicle-track-bridge system. The coupling strength level obtained in Step 4 is used as the main label. Simultaneously, the percentage of the bridge's dominant frequency energy to the total vibration energy (i.e., energy ratio) is calculated from the bridge structure deformation frequency component extracted in Step 1, and the percentage of the vehicle's dominant frequency energy to the total vibration energy is calculated from the vehicle vibration frequency component extracted in Step 2. The coupling strength level, bridge energy ratio, vehicle energy ratio, and characteristic frequency values ​​of the bridge and vehicle (such as the bridge fundamental frequency and vehicle roll frequency) are combined into a multi-dimensional numerical vector, which is the coupling state feature vector. This vector comprehensively describes the coupling characteristics of the vehicle-track-bridge system in the frequency domain at the current moment, providing a quantitative basis for the subsequent weight adjustment module.

[0038] For example, with a coupling strength level of three, the energy proportion of the bridge structure deformation frequency component is 78% (i.e., the energy at 0.6Hz accounts for 78% of the total vibration energy of the bridge), and the energy proportion of the vehicle body vibration frequency component is 65% (i.e., the energy at the 2Hz roll frequency accounts for 65% of the total vibration energy of the vehicle body). The bridge's fundamental frequency is 0.6Hz, and the vehicle's roll frequency is 2.0Hz. These values ​​are arranged in a preset order to generate a coupling state feature vector: {Level 3, 0.78, 0.65, 0.6Hz, 2.0Hz}. This vector is passed to the weight adjustment module to determine the fusion weight allocation between the laser scanning sensor and the image acquisition sensor.

[0039] The pre-defined fusion weight allocation model is a multi-input multi-output mapping model trained on historical data. The model's input parameters are the coupling state feature vectors, and its output parameters are the fusion weight values ​​for each of the geometric measurement sensors. These sensors include laser scanning sensors and image acquisition sensors. The model's core function is to assign appropriate fusion weights to the two sensors under different vehicle-track-bridge coupling states, minimizing the error in the weighted fusion of track geometry parameters.

[0040] For example, when the coupling state feature vector is {Level 3 coupling strength, bridge energy percentage 0.78, vehicle energy percentage 0.65, bridge fundamental frequency 0.6Hz, vehicle roll frequency 2.0Hz}, the model output laser scanning sensor weight is 0.7 and the image acquisition sensor weight is 0.3. When the coupling state feature vector is {Level 1 coupling strength, bridge energy percentage 0.20, vehicle energy percentage 0.15, bridge fundamental frequency 0.6Hz, vehicle roll frequency 1.5Hz}, the model output laser scanning sensor weight is 0.5 and the image acquisition sensor weight is 0.5.

[0041] Training this model requires constructing a historical dataset containing a large number of samples. Each sample consists of a set of input features (coupled state feature vectors) and a set of output labels (optimal fusion weight values). The output labels are obtained as follows: Under various dynamic conditions (different wind speeds, vehicle speeds, bridge types, curve radii), the geometric parameters (such as track gauge, curvature, and alignment) of the same track section are independently measured using both laser scanning sensors and image acquisition sensors. Simultaneously, the true value of this section is obtained using a high-precision reference measuring device (such as a total station or static track inspection instrument). For each condition, an enumeration or optimization algorithm (such as grid search or particle swarm optimization) is used to find a set of fusion weights that minimizes the root mean square error between the weighted fusion result and the reference true value. This set of weights that minimizes the error is the output label for that sample.

[0042] The process involves collecting samples covering various coupling strength levels (weak, medium, strong), wind speed ranges (0 to 30 m / s), vehicle speed ranges (50 to 350 km / h), and different bridge structure types, typically requiring thousands of sets. These samples are divided into training and validation sets and trained using machine learning regression algorithms (such as neural networks, random forests, or support vector regression). During training, the model learns a non-linear mapping from coupling state feature vectors to optimal fusion weights. After training, the model can directly output the optimal fusion weight allocation for the current operating condition based on the real-time input coupling state feature vectors during actual operation.

[0043] For example, during the training phase, on a curved section of a high-speed railway bridge, under conditions of wind speed of 12 m / s, vehicle speed of 200 km / h, and coupling strength of level two, the fusion error was minimized when the weights of the laser scanning sensor (0.65) and the image acquisition sensor (0.35) were determined using an enumeration method. The coupling state feature vector under this condition and this set of weights were used as a training sample. After collecting 5000 similar samples, a three-layer neural network was used for training. The network input layer had a dimension of 5 (corresponding to the 5 components of the feature vector), the hidden layer had 10 nodes, and the output layer had 2 nodes (corresponding to the weights of the two sensors). Training continued until the error on the validation set no longer decreased, resulting in a well-trained model.

[0044] The pre-defined fusion weight allocation model contains multiple weight allocation sub-models corresponding to different coupling strength levels. The coupling strength levels are typically divided into three levels (weak, medium, and strong), with each level corresponding to an independent sub-model. This divide-and-conquer structure reduces the complexity of each sub-model and improves mapping accuracy. Each sub-model is trained and predicts only within its corresponding coupling strength level, avoiding interference between different coupling characteristics.

[0045] For example: The weakly coupled sub-model is suitable for load conditions with a coupling strength level of one, where bridge deformation is small and vehicle vibration mainly comes from wheel-rail excitation; the weights of the laser scanning sensor and the image acquisition sensor are nearly equal. The medium-coupling sub-model is suitable for load conditions two, where energy transfer begins between the vehicle and the bridge; the weight of the laser scanning sensor is slightly higher. The strongly coupled sub-model is suitable for load conditions three, where the bridge deforms significantly and the vehicle follows the bridge's movement; the weight of the laser scanning sensor is significantly higher than that of the image acquisition sensor because point cloud data is more sensitive to structural deformation.

[0046] Each weight allocation sub-model stores the optimal weight allocation scheme that minimizes the error of the fusion result under different coupling conditions. These schemes exist in the form of mapping functions or lookup tables. For continuously input coupling state feature vectors, the sub-model directly calculates the corresponding weight values ​​through its internal regression function, rather than only for discrete points. The optimal scheme ensures that the fused orbital geometry parameters are as close as possible to the true values ​​under various dynamic conditions, thereby overcoming the measurement distortion defects of traditional fixed weights or simple rule weights under strong coupling conditions.

[0047] For example, in a strongly coupled sub-model, when the bridge energy proportion in the input feature vector is 0.78 and the vehicle energy proportion is 0.65, the model outputs a laser weight of 0.70 and an image weight of 0.30. When the bridge energy proportion increases to 0.90 and the vehicle energy proportion to 0.80, the model outputs a laser weight of 0.80 and an image weight of 0.20. This trend reflects that the more severe the bridge deformation, the higher the weight of the laser scanning sensor, because the laser directly measures the structural contour and is less affected by dynamic coupling contamination.

[0048] The system dynamically adjusts the fusion weight allocation of the geometric measurement sensors by calling a preset fusion weight allocation model and outputs the optimized fusion weight parameters. Specifically, it selects the corresponding weight allocation sub-model from the preset fusion weight allocation model based on the coupling strength level in the coupling state feature vector, and the weight allocation sub-model outputs the corresponding fusion weight value of each geometric measurement sensor based on the coupling state feature vector. The fusion weight values ​​are normalized to generate optimized fusion weight parameters where the sum of the weights of each geometric measurement sensor equals the preset total weight. The original weight values ​​output by the weight allocation sub-model may have dimensional differences or inconsistent numerical ranges, and their sum may not necessarily equal the system's preset total weight (usually the preset total weight is 1, indicating that the sum of the weights of all sensors is 100%). To meet the basic requirement of weight normalization in weighted fusion calculations, the system normalizes the original weight values. The normalization method is as follows: divide the original weight value of each sensor by the sum of the original weight values ​​of all sensors, and then multiply by the preset total weight. After normalization, the sum of the fusion weight values ​​of all sensors is strictly equal to the preset total weight, and each weight value lies between 0 and the preset total weight. The normalized weight values ​​are the optimized fusion weight parameters, which are output to the feedback correction module.

[0049] For example: The original weights are 0.82 for laser and 0.34 for image, with a sum of 1.16. The default total weight is 1. Normalization is then performed: laser normalized weight = 0.82 / 1.16 × 1 ≈ 0.707; image normalized weight = 0.34 / 1.16 × 1 ≈ 0.293. The optimized fusion weight parameters are {laser sensor weight 0.707, image sensor weight 0.293}. This parameter satisfies the condition that the sum of the two weights is 1, and can be directly used for subsequent weighted fusion calculations. If, under a certain operating condition, the original weights for laser and image are both 0.50, and after normalization, both are 0.50, it indicates that both sensors are equally reliable.

[0050] The fusion estimation of orbital geometric parameters specifically includes: The system acquires optimized fusion weight parameters, raw track geometric measurement data collected by the geometric measurement sensor, track cross-section contour point cloud data output by the laser scanning sensor, and track appearance image data output by the image acquisition sensor. It also receives optimized fusion weight parameters (e.g., laser sensor weight 0.707, image sensor weight 0.293) from the weight adjustment module. Simultaneously, it acquires track cross-section contour point cloud data from the laser scanning sensor, which records the surface geometry information of the rail and track slab in three-dimensional point coordinates; and acquires track appearance image data from the image acquisition sensor, which is a two-dimensional pixel matrix recording the optical image of the track area.

[0051] Point cloud processing is performed on the track cross-section contour point cloud data output by the laser scanning sensor to extract the first track geometric parameter sequence. The point cloud processing includes filtering, segmentation, and feature fitting. First, outlier filtering and noise reduction are performed on the original point cloud to remove isolated noise points caused by rain, fog, or vibration. Then, based on intensity information or spatial distribution, the point cloud is segmented into rail region, fastener region, and track bed region. For the rail region, the least squares method is used to fit the spatial straight line or curve at the top surface of the rail and the gauge point, from which geometric parameters such as gauge, alignment, elevation, and curvature are calculated for each mileage section. These parameters are arranged in mileage order to form the first track geometric parameter sequence. Each data point in this sequence corresponds to a track cross-section location and includes multiple geometric indices for that location.

[0052] For example, a 200-meter-long point cloud dataset of track cross-sections is processed, and a cross-section is extracted every 0.125 meters. At mileage K100+000, the fitted values ​​are: track gauge 1435.2 mm, track orientation deviation +1.2 mm, elevation deviation -0.8 mm, and curvature 0.0006 radians per meter. These values, along with the mileage labels, are stored in the first track geometry parameter sequence.

[0053] Image processing is performed on the track appearance image data output by the image acquisition sensor to extract the second track geometric parameter sequence. Image processing includes feature point extraction, stereo matching, and 3D coordinate transformation. First, feature point extraction is performed using scale-invariant feature transformation or existing accelerated robust feature algorithms to identify significant feature points such as rail edges, fastener centers, and sleeper corners in the image. Then, stereo matching is performed using disparity information obtained from a binocular camera (or a monocular camera with structured light) to associate corresponding feature points in the left and right images, calculating the depth value of each feature point in the camera coordinate system. Finally, 3D coordinate transformation is performed using pre-calibrated camera extrinsic parameters (rotation matrix and translation vector) to transform the feature point coordinates from the camera coordinate system to the same world coordinate system as the laser scanning sensor. In the world coordinate system, geometric parameters such as track gauge, relative rail position, and track slab flatness are calculated based on the 3D coordinates of the feature points. These parameters are arranged in mileage order to form the second track geometric parameter sequence.

[0054] For example, in the image at mileage K100+000, feature points on the inner edge of the rail are extracted. After stereo matching, the three-dimensional coordinates of the left and right rail edge points are obtained, and the track gauge is calculated to be 1435.5 mm. Multiple cross-sections are processed consecutively to obtain a second sequence of track geometric parameters, where the track gauge value at K100+000 is 1435.5 mm, slightly different from the 1435.2 mm in the first sequence.

[0055] Due to differences in sampling frequency, initial acquisition time, and installation location between the laser scanning sensor and the image acquisition sensor, the two sequences may have offsets in mileage and time. Spatiotemporal registration first performs time alignment, using the synchronization clock or encoder pulses on the track machine to associate the data points of the two sequences with the same acquisition time. Then, spatial alignment is performed, employing an iterative nearest-point algorithm or a registration method based on common features (such as fastener positions) to transform the coordinate values ​​in the second track geometry parameter sequence to the coordinate system used by the first track geometry parameter sequence. After registration, the two sequences have a one-to-one correspondence of geometry parameter values ​​at every mileage section location.

[0056] For example, in the laser sequence, the location K100+000 corresponds to the acquisition time t=120.5 seconds. In the image sequence, interpolation using the encoder yields a mileage of K100+000.3 corresponding to the same time t=120.5 seconds, resulting in a mileage discrepancy of 0.3 meters. After registration, the mileage label in the image sequence is corrected to K100+000 to align it with the laser sequence.

[0057] For each mileage section after spatiotemporal registration, the geometric parameter values ​​from the first track geometric parameter sequence (e.g., gauge value of 1435.2 mm) and the same geometric parameter value from the second track geometric parameter sequence (e.g., gauge value of 1435.5 mm) are taken and weighted and summed according to the optimized fusion weight parameters (laser weight 0.707, image weight 0.293). The calculation formula is: Fusion value = Laser weight × First sequence value + Image weight × Second sequence value. Weighted summations are performed for all sections and all geometric parameter types (gauge, alignment, elevation, curvature, etc.) to obtain the fusion estimate of each geometric parameter on each section. These fusion estimates are arranged in mileage order to generate preliminary fused track geometric parameters. This parameter combines the advantages of both sensors and is closer to the true geometric state of the track than the measurement results of a single sensor.

[0058] For example, at section K100+000, the fused gauge value = 0.707 × 1435.2 mm + 0.293 × 1435.5 mm = 1014.5 + 420.6 ≈ 1435.3 mm (decimals are rounded in the calculation). The fused alignment value is calculated similarly. The fused results of all sections constitute the preliminary fused track geometry parameters, which are output to the feedback correction module for confidence assessment.

[0059] Real-time assessment of the confidence level of the fusion results refers to: acquiring the preliminary fused orbital geometric parameters and the first and second orbital geometric parameter sequences; calculating the first deviation value between each data point in the first orbital geometric parameter sequence and the corresponding data point in the preliminary fused orbital geometric parameters, resulting in a first deviation sequence composed of the first deviation values ​​of each data point: For each mileage section, subtract the geometric parameter value (e.g., gauge value) in the first orbital geometric parameter sequence from the corresponding parameter value in the preliminary fused orbital geometric parameters for the same section, and take the absolute value (or retain the signed difference, depending on specific design requirements). The absolute value of this difference is the first deviation value for that data point. This operation is repeated for all sections to obtain a set of first deviation values ​​arranged in mileage order, called the first deviation sequence. This sequence reflects the degree of local deviation between the laser measurement results and the fusion results at each location.

[0060] For example: At section K100+000, the first sequence gauge is 1435.2 mm, the initial merged gauge is 1435.3 mm, and the first deviation is 0.1 mm. At section K100+001, the first sequence gauge is 1435.1 mm, the initial merged gauge is 1435.3 mm, and the first deviation is 0.2 mm. Calculating this for all sections sequentially yields the first deviation sequence {0.1, 0.2, ...}.

[0061] The second deviation value between each data point in the second orbit geometric parameter sequence and the corresponding data point in the preliminary fused orbit geometric parameters is calculated, resulting in a second deviation sequence composed of the second deviation values ​​of each data point. Using the exact same calculation method, the geometric parameter values ​​in the second orbit geometric parameter sequence are subtracted from the parameter values ​​of the corresponding sections in the preliminary fused orbit geometric parameters, and the absolute value is taken to obtain the second deviation value for each section. The second deviation values ​​of all sections are arranged in mileage order, forming the second deviation sequence. This sequence reflects the degree of local deviation between the image measurement results and the fusion results at each location.

[0062] For example: At section K100+000, the second sequence gauge is 1435.5 mm, the initial fused gauge is 1435.3 mm, and the second deviation value is 0.2 mm. At section K100+001, the second sequence gauge is 1435.4 mm, the initial fused gauge is 1435.3 mm, and the second deviation value is 0.1 mm. This yields the second deviation sequence {0.2, 0.1, ...}.

[0063] The first deviation statistical feature value is calculated based on the first deviation sequence, and the second deviation statistical feature value is calculated based on the second deviation sequence. A confidence index is then generated based on the first and second deviation statistical feature values. To comprehensively assess the reliability of the fusion result, the local deviation sequences need to be compressed into global statistical feature values. Commonly used statistical feature values ​​include: Root mean square error: Calculated by taking the square root of the sum of the squares of all deviations divided by the number of samples. It is sensitive to large deviations.

[0064] Mean absolute deviation: Calculates the arithmetic mean of all deviation values, reflecting the average degree of deviation.

[0065] Maximum absolute deviation: The maximum value in the deviation sequence is taken, reflecting the degree of deviation under the worst case.

[0066] Percentiles (e.g., 95th percentile): Sort the deviation values ​​from smallest to largest and take the value at the 95th position, taking into account both the overall value and extreme values.

[0067] The first and second statistical characteristic values ​​of deviation are calculated using the same statistical method based on their respective deviation sequences. Then, a single confidence index is generated based on these two statistical characteristic values. The design principle of the confidence index is: the larger the statistical characteristic value (i.e., the more severe the deviation), the smaller the confidence index; the two are negatively correlated. Several existing technologies are available for the specific generation method, including: Method 1: Based on the exponential decay function, the statistical characteristic values ​​of the first and second deviations are denoted as E1 and E2, respectively, and a reference deviation value E0 is set (e.g., typical deviation under historical good operating conditions). Confidence index = exp(-(E1+E2) / (2×E0)). The confidence level is 1 when the sum of deviations is zero, and the larger the deviation, the closer the confidence level is to 0.

[0068] Method 2: Based on the inverse mapping of the Sigmoid function, first calculate the overall deviation S = (E1 + E2) / 2, then calculate the confidence index = 1 - 1 / (1 + exp(-k × (S0 - S))), where S0 is the deviation threshold and k is the steepness coefficient. This function has a confidence level close to 1 when S is much smaller than S0, and close to 0 when S is much larger than S0.

[0069] Method 3: Based on tiered linear interpolation, pre-divide the deviation intervals: 0 to 0.1 mm corresponds to a confidence level of 1.0, 0.1 to 0.3 mm corresponds to a confidence level of 0.8, 0.3 to 0.5 mm corresponds to a confidence level of 0.5, and greater than 0.5 mm corresponds to a confidence level of 0.2. Obtain the two confidence level components E1 and E2 from the table respectively, and then take the minimum value or weighted average as the final confidence level index.

[0070] Method 4: Based on the Gaussian distribution assumption, assuming the deviation follows a zero-mean Gaussian distribution, the standard deviation σ is estimated using historical data. The current comprehensive deviation S = (E1 + E2) / 2, and the confidence index = 2 × (1 - Φ(S / σ)), where Φ is the standard normal cumulative distribution function. This index represents the probability that the deviation is less than the current value.

[0071] Method 5: Based on fuzzy logic, E1 and E2 are used as fuzzy inputs. Membership functions such as "small deviation", "medium deviation" and "large deviation" are defined. The confidence index is obtained by reasoning through preset fuzzy rules (such as "if E1 is small and E2 is small, then the confidence is high"), and finally the output is defuzzified.

[0072] For example, using root mean square error (RMSE) as a statistical characteristic value, the RMSE of the first deviation sequence is calculated to be 0.15 mm, and the RMSE of the second deviation sequence is 0.18 mm. Using Method 1, setting E0 = 0.10 mm, the combined deviation S = (0.15 + 0.18) / 2 = 0.165 mm, and the confidence index = exp(-0.165 / 0.10) = exp(-1.65) ≈ 0.192. This confidence index is low, indicating that the fusion result deviates significantly from the original data of the two sensors, possibly due to measurement distortion caused by strong dynamic coupling. If Method 3 is used, 0.15 mm corresponds to a confidence index of 0.8, and 0.18 mm corresponds to a confidence index of 0.8; taking the minimum value yields 0.8.

[0073] The confidence index is compared with a preset confidence threshold. When the confidence index is greater than or equal to the preset confidence threshold, the preliminary fused orbital geometry parameters are directly output as orbital geometry parameters with confidence labels. A confidence threshold is preset in the system, for example, 0.7 (which can be adjusted based on engineering experience). The confidence index calculated in the previous step is compared with this threshold. If the confidence index is greater than or equal to the threshold, it indicates that the fusion result is in good agreement with the original measurement results of the two sensors, and the fusion result is reliable. In this case, the preliminary fused orbital geometry parameters are directly output as orbital geometry parameters with confidence labels. Each output data point is accompanied by a confidence label (i.e., the confidence index value or confidence level of that point) for use by the downstream benchmark alignment module.

[0074] For example, if the preset confidence threshold is 0.7, and method three calculates a confidence level of 0.8, then the calculation passes, and the orbital geometry parameters with a confidence level label of 0.8 are output. When the confidence level is less than the preset confidence threshold, a feedback correction mechanism is triggered. Abnormal data points are identified based on the first and second deviation sequences. The abnormal data points in the first and second orbital geometry parameter sequences are corrected based on a preset compensation model. The weighted summation is then recalculated according to the optimized fusion weight parameters to generate corrected orbital geometry parameters as the output orbital geometry parameters with a confidence level label.

[0075] Based on a pre-defined compensation model, abnormal data in the first and second orbital geometric parameter sequences are corrected, specifically including: The system compares the first deviation value of each data point in the first deviation sequence with a preset first anomaly threshold, and the second deviation value of each data point in the second deviation sequence with a preset second anomaly threshold. The system pre-sets two anomaly thresholds: a first threshold to determine if the laser measurement data is abnormal, and a second threshold to determine if the image measurement data is abnormal. These thresholds can be fixed values ​​(e.g., the 99th percentile based on historical deviation distribution under good operating conditions) or dynamically adjusted values ​​(e.g., the threshold is appropriately relaxed as the coupling strength increases). For each data point in the first deviation sequence, its deviation value is compared with the first anomaly threshold; for each data point in the second deviation sequence, its deviation value is compared with the second anomaly threshold. The purpose of the comparison is to identify which data points have measurement results that differ significantly from the preliminary fusion results, possibly due to sensor measurement errors or dynamic interference.

[0076] For example: The first anomaly threshold is preset to 0.3 mm, and the second anomaly threshold is 0.4 mm. In the first deviation sequence, the deviation value at section K100+001 is 0.2 mm, which is less than the threshold and is considered normal; the deviation value at section K100+005 is 0.5 mm, which is greater than the threshold and is considered abnormal. In the second deviation sequence, the deviation value at section K100+005 is 0.6 mm, which is greater than the second anomaly threshold of 0.4 mm and is also considered abnormal.

[0077] When the first deviation value of any data point is greater than or equal to the first anomaly threshold, the data point in the first orbital geometric parameter sequence corresponding to that data point is marked as the first anomaly data point. For first deviation sequence data points whose deviation values ​​reach or exceed the first anomaly threshold, the system finds data points with the same mileage section in their corresponding first orbital geometric parameter sequence and marks those data points as the first anomaly data points. The marking method can be adding attribute tags or storing them in an anomaly list for subsequent specialized processing. Data points that do not reach the threshold remain in a normal state.

[0078] For example, at section K100+005, the first deviation value is 0.5 mm ≥ 0.3 mm. Therefore, the track gauge, track orientation and other data points at section K100+005 in the first track geometry parameter sequence are marked as the first abnormal data point.

[0079] When the second deviation value of any data point is greater than or equal to the second abnormal threshold, the data point in the second orbital geometric parameter sequence corresponding to that data point is marked as the second abnormal data point. Similarly, for data points in the second deviation sequence that reach or exceed the second abnormal threshold, the system marks the data points with the same cross section in the corresponding second orbital geometric parameter sequence as the second abnormal data points.

[0080] For example: at section K100+005, the second deviation value is 0.6 mm ≥ 0.4 mm. The data point at section K100+005 in the second orbital geometric parameter sequence is marked as the second abnormal data point.

[0081] The pre-defined compensation model contains multiple compensation sub-models corresponding to different coupling strength levels. Each sub-model stores outlier correction coefficients trained based on historical data. The pre-defined compensation model is a holistic framework, internally divided into multiple independent compensation sub-models according to coupling strength levels (Level 1, Level 2, Level 3). Each sub-model specifically handles the outlier correction task at its corresponding coupling strength level. Instead of storing a complete mapping function, each sub-model stores a set of outlier correction coefficients. These coefficients typically include a multiplicative factor (for scaling the bias), an additive factor (for translating the baseline), and possible high-frequency compensation coefficients (for restoring details). These coefficients are obtained through offline training. The training process is as follows: A large amount of historical data is collected, including pairs of original sensor measurements (abnormal data points) and corresponding true orbital geometric parameters (obtained through high-precision reference equipment) occurring under various coupling strength levels and anomaly types. For each sample, the difference between the true value and the abnormal measurement is calculated, and then the optimal transformation relationship from the abnormal measurement to the true value is found through regression analysis or optimization algorithms (such as least squares, ridge regression, or neural network fitting). This relationship is usually simplified to a linear or low-order polynomial form, and its parameters are the correction coefficients. For all samples under the same coupling strength level, the coefficient that minimizes the overall error after correction is selected as the stored coefficients for that sub-model. Multiple sets of coefficients can also be stored according to the anomaly type (such as excessive deviation, oscillation, offset, etc.), and selected based on the deviation characteristics during actual use.

[0082] For example, under a strong coupling level, 1000 laser measurement anomalies and their corresponding true track gauge values ​​were collected. Analysis revealed that when the measured value deviated from the true value, the deviation was approximately linearly related to the coupling strength and vibration amplitude. Correction coefficients were obtained through linear regression: multiplicative factor a = 1.02, additive factor b = -0.05 mm. That is, the correction value = 1.02 × measured value - 0.05 mm. This coefficient was stored in the strong coupling compensation sub-model.

[0083] Based on the coupling strength level in the coupling state feature vector, a corresponding compensation sub-model is selected from the preset compensation models. The first and second abnormal data points are input into the selected compensation sub-model, and the compensation sub-model outputs the corrected first and second track geometric parameter data points. The system reads the coupling strength level in the current coupling state feature vector and activates the corresponding compensation sub-model from the preset compensation models based on this level. Then, for each data point marked as abnormal (first or second abnormal data point), its geometric parameter value (e.g., track gauge, track direction, etc.) is used as input and sent to the selected compensation sub-model. The compensation sub-model uses internally stored correction coefficients to calculate the input value and outputs a corrected geometric parameter value. The specific formula for the correction calculation depends on the correction form adopted by the sub-model: for a linear model, correction value = a × input value + b; for a polynomial model, correction value = a0 + a1 × input value + a2 × input value² + ... This calculation process is performed independently for each abnormal data point.

[0084] Example: The current coupling strength level is level three (strong coupling), and the system selects a strong coupling compensation sub-model. For the first abnormal data point at section K100+005, the laser-measured track gauge is 1434.2 mm (the true value should be 1435.3 mm). The sub-model stores coefficients a=1.02, b=-0.05 mm, and calculates the correction value = 1.02×1434.2-0.05=1462.884-0.05≈1462.834 mm, which is clearly incorrect. This indicates that the coefficients in the example need to be redesigned. A more reasonable example: The true value is 1435.3 mm, but the measured value is compressed to 1434.2 mm due to bridge deformation, a deviation of -1.1 mm. If the correction coefficients a=1.0008, b=+0.5 mm, then the correction value = 1.0008×1434.2+0.5=1435.3 mm, which is correct. The correction coefficients should ensure that the corrected value approximates the true value.

[0085] The corrected first orbital geometric parameter sequence is generated by replacing the corresponding outlier data points in the first orbital geometric parameter sequence with the corrected first orbital geometric parameter data points. For each data point marked as outlier in the first orbital geometric parameter sequence, its original value is replaced with the corrected value output by the compensation sub-model. For normal data points (where the deviation does not exceed the threshold), they remain unchanged. After the replacement is completed, the entire first orbital geometric parameter sequence becomes the corrected first orbital geometric parameter sequence, where the outlier segments have been corrected and the original normal segments have been preserved.

[0086] For example, in the first track geometry parameter sequence, the original gauge value of section K100+005 was 1434.2 mm, which is corrected to 1435.3 mm; the original gauge value of section K100+006 was 1434.3 mm, which is corrected to 1435.4 mm. Other normal sections remain unchanged, generating the corrected first track geometry parameter sequence.

[0087] Replace the corresponding outlier data points in the second orbital geometry parameter sequence with the corrected second orbital geometry parameter data points to generate a corrected second orbital geometry parameter sequence. Perform the same operation on the second orbital geometry parameter sequence: replace the original value of each data point marked as an outlier with the corrected value output by the compensator sub-model, while leaving normal data points unchanged. This finally generates the corrected second orbital geometry parameter sequence. These two corrected sequences will then be reused for weighted fusion to generate corrected orbital geometry parameters, resulting in a more accurate final output.

[0088] For example, in the second track geometry parameter sequence, the original track gauge value at section K100+005 was 1436.1 mm (due to excessive image jitter deviation), which was corrected to 1435.4 mm by the compensation sub-model. This replacement generated the corrected second track geometry parameter sequence. Subsequently, the system recalculated the weighted sum according to the optimized fusion weight parameters (laser 0.707, image 0.293), obtaining a corrected track gauge value of approximately 0.707 × 1435.3 + 0.293 × 1435.4 = 1435.33 mm, which is close to the true value.

[0089] Dynamic spatiotemporal reference alignment of orbital geometry parameters with confidence labels refers to: The system acquires track geometry parameters with confidence labels from the feedback correction module. These parameters include the track geometry parameter values, confidence indices, and corresponding original acquisition timestamps and original odometer values ​​for each data point. The system also receives track geometry parameters from the feedback correction module after confidence evaluation (which may include anomaly correction and re-fusion). Each data point includes not only geometric parameter values ​​such as gauge, heading, and elevation, but also a confidence index (e.g., 0.92 indicates 92% confidence), and the original timestamp (milliseconds since system startup) and original odometer value (rough odometer provided by wheel encoders or GPS) recorded by the sensors. This information forms the basis for subsequent alignment operations.

[0090] For example, a data point is recorded as: original mileage K100+000, timestamp 120500 milliseconds, track gauge 1435.3 mm, confidence level 0.92. The next data point: original mileage K100+125, timestamp 120600 milliseconds, track gauge 1435.4 mm, confidence level 0.90. Note that the mileage interval is 125 mm, the time interval is 100 milliseconds, and the calculation speed is approximately 1.25 m / s. This is significantly different from the actual operating speed of the track machine (e.g., 250 km / h ≈ 69.4 m / s), indicating a large error in the original mileage values.

[0091] The system queries the preset spatiotemporal reference correction strategy corresponding to the current dynamic condition based on the coupling strength level in the coupling state feature vector. The preset spatiotemporal reference correction strategy is a set of predefined processing rules and parameters specifically designed to guide spatiotemporal reference alignment operations under different dynamic conditions. Since the sources and severity of time and spatial reference deviations differ under different coupling strength levels, differentiated correction strategies are required. These strategies are stored in the system's internal strategy library, indexed by the coupling strength level (Level 1, Level 2, Level 3). Specific strategy details may include: the interpolation method used for time axis alignment (linear interpolation, cubic spline interpolation, polynomial fitting, etc.), whether velocity smoothing filtering is enabled and its window size, whether bridge dynamic deformation is compensated for during spatial axis alignment, the value of the compensation coefficient, and whether auxiliary sensors (such as GPS or accelerometers) are referenced. The system reads the coupling strength level in the current coupling state feature vector and retrieves the complete correction strategy corresponding to that level from the strategy library.

[0092] For example, when the coupling strength level is Level 1 (weak coupling), the preset strategy is: time axis alignment uses linear interpolation, and velocity smoothing filtering is not enabled; spatial axis alignment only compensates for known static installation deviations (such as the fixed offset between the laser sensor and the image sensor), and does not compensate for dynamic deformation of the bridge. When the coupling strength level is Level 3 (strong coupling), the preset strategy is: time axis alignment uses cubic spline interpolation, and five-point moving average velocity smoothing is enabled; spatial axis alignment requires dynamic compensation using the real-time amplitude and phase of the bridge structure deformation frequency components, with a compensation coefficient of 0.85 (i.e., deducting 85% of the bridge deformation influence), and time synchronization calibration is performed by referring to the GPS second pulse signal. By querying, the system obtains the specific strategy parameters corresponding to the current level.

[0093] Based on the original acquisition timestamps, original mileage values, and a spatiotemporal reference correction strategy, the track geometry parameters with confidence labels are time-axis aligned to eliminate time reference deviations caused by fluctuations in track machine speed and data acquisition delays. Because the track machine speed is not constant (acceleration / deceleration, natural deceleration over curves, turnout sections, etc.), and there are slight delays in the acquisition times of different sensors, the relationship between the original timestamps and mileage values ​​does not conform to actual motion patterns. The goal of time-axis alignment is to establish a uniform time-mileage correspondence that conforms to physical motion. Specifically: First, the instantaneous velocity corresponding to each data point is calculated using the original timestamps and original mileage values. Then, the velocity curve is smoothed according to the smoothing method specified in the preset spatiotemporal reference correction strategy (such as moving average or spline fitting) to eliminate noise caused by encoder slippage or electromagnetic interference. Next, the time axis is reconstructed at a fixed time interval (e.g., every 10 milliseconds), and the accurate mileage value corresponding to each new time point is obtained by integrating the smoothed velocity curve. Finally, the geometric parameters and confidence labels of the original data points are interpolated and mapped to these uniform time points, so that the time axis and the mileage axis form a monotonic, smooth correspondence that conforms to the actual vehicle speed.

[0094] For example, in the original data, timestamp 120500 milliseconds corresponds to the original mileage K100+000, and timestamp 120600 milliseconds corresponds to K100+125, with a calculated speed of only 1.25 m / s, while the actual track machine speed is 69.4 m / s. According to a strong coupling strategy, the system uses cubic spline interpolation to smooth the speed curve and, referring to the mileage change trend of adjacent sections, identifies that the encoder experienced severe slippage during this period. After time axis alignment, the system corrects the mileage at timestamp 120500 milliseconds to K100+000 and the mileage at timestamp 120505 milliseconds to K100+347 (69.4 m / s × 0.005 s = 0.347 m), making the time-mileage relationship consistent with the actual speed.

[0095] Based on the bridge structural deformation frequency components in the coupled state feature vector, the track geometry parameters after time-axis alignment are spatially aligned to compensate for spatial position offsets caused by bridge elastic deformation and vehicle vibration, ensuring that the track geometry parameter values ​​at each data point match the actual physical position of the track. After time-axis alignment, the data points have a reasonable time-mileage relationship, but the geometry parameter values ​​measured by the sensors on the track machine actually include spurious components caused by bridge elastic deformation (such as vertical deflection, lateral sway, and torsion) and vehicle vibration (such as roll and pitch). The goal of spatial alignment is to subtract these dynamic components from the measured values, restoring the inherent geometry of the track when there is no dynamic deformation. Specifically, the bridge structural deformation frequency components (e.g., amplitude and phase at 0.6 Hz) are extracted from the coupled state feature vector. Based on the compensation coefficients given in the preset spatiotemporal reference correction strategy, the displacement of the bridge deformation in the measurement direction at each moment is calculated. For lateral parameters such as track gauge and track alignment, corrections are made based on the bridge's lateral displacement and torsional angle; for vertical parameters such as elevation, the bridge's vertical deflection is subtracted. In addition, for sensor attitude changes caused by vehicle vibration, the angular velocity data recorded by the inertial measurement unit is used for reverse compensation to transform the measured values ​​in the sensor coordinate system to the geodetic coordinate system.

[0096] For example, under strong coupling conditions, the bridge's vertical deflection amplitude at 0.6Hz is 15 mm, and its lateral sway amplitude is 8 mm. After time axis alignment, the measured rail top elevation (relative to the track machine reference) at a certain moment is 1250 mm. According to the bridge vibration phase, at this time, the bridge deflects downward at mid-span by 10 mm, and the vehicle body tilt causes the laser sensor to tilt outward by 0.2 degrees. When aligning the spatial axis, first subtract the bridge deflection: 1250 mm + 10 mm = 1260 mm (converting the sensor measurement to the geodetic reference); then compensate for the lateral offset caused by the vehicle body tilt, finally obtaining the true rail elevation in the geodetic coordinate system as 1260.5 mm. Similarly, the measured track alignment deviation includes the offset caused by the bridge's lateral sway; subtracting the lateral sway component yields the true track alignment deviation.

[0097] The track geometry parameters aligned to the spatiotemporal reference are directly output as the aligned track geometry dataset. After time axis alignment and spatial axis alignment, each data point has an accurate time label, a mileage label that strictly matches the actual physical location, and track geometry parameter values ​​that eliminate the influence of bridge deformation and vehicle vibration. At this point, the system does not need to perform additional resampling or format conversion, and directly outputs the set of all aligned data points as the final track geometry dataset. This dataset can be directly used for track condition assessment, defect identification, historical trend comparison, or maintenance operation guidance.

[0098] For example, the final output aligned track geometry dataset contains data for all sections from K100+000 to K101+000. The mileage error for each section is less than 10 mm, and parameters such as gauge, orientation, and elevation accurately reflect the track's geometric state. For instance, section K100+050 has an accurate mileage of K100+050.002, a gauge of 1435.3 mm, an elevation deviation of -0.5 mm, and a confidence level of 0.95. This dataset can be imported into a track geometry management system to generate accurate maintenance decisions.

[0099] The above-mentioned models or function formulas are all dimensionless and numerical calculations. The models or function formulas are obtained by software simulation based on a large amount of collected data to obtain the most recent real situation. The preset parameters in the models or function formulas are set by those skilled in the art according to the actual situation.

[0100] It should be understood that in the various embodiments of this application, the order of the above-mentioned processes does not imply the order of execution. The execution order of each process should be determined by its function and internal logic, and should not constitute any limitation on the implementation process of the embodiments of this application.

[0101] 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 implementation should not be considered beyond the scope of this application.

[0102] Those skilled in the art will 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.

[0103] The above are merely specific embodiments of this application, but the scope of protection of this application is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in this application should be included within the scope of protection of this application. Therefore, the scope of protection of this application should be determined by the scope of the claims.

Claims

1. A curved track monitoring system for a track machine, characterized in that, include: The coupling identification module is used to acquire dynamic environmental parameters and vehicle attitude parameters during the operation of the track machine, and to identify the coupling state of the vehicle-track-bridge system based on the dynamic environmental parameters and vehicle attitude parameters, and output the coupling state feature vector. The weight adjustment module is used to dynamically adjust the fusion weight allocation of the geometric measurement sensor based on the coupling state feature vector and call the preset fusion weight allocation model, and output the optimized fusion weight parameters. The feedback correction module is used to perform fusion estimation of orbital geometric parameters based on the optimized fusion weight parameters, evaluate the confidence of the fusion result in real time, perform feedback correction on measurement results with confidence scores lower than the preset confidence threshold, and output orbital geometric parameters with confidence scores labels. The benchmark alignment module is used to perform dynamic spatiotemporal benchmark alignment on orbital geometric parameters with confidence labels, and output the aligned orbital geometric dataset.

2. The curved track monitoring system for a track machine according to claim 1, characterized in that, Dynamic environmental parameters include real-time wind speed parameters at the location of the track machine, real-time operating speed parameters of the track machine, and real-time vibration parameters of the bridge structure; vehicle attitude parameters are collected by the inertial measurement unit, including the three-axis angular velocity parameters and three-axis acceleration parameters of the track machine.

3. The curved track monitoring system for a track machine according to claim 1, characterized in that, The coupling state feature vector specifically refers to: Time-frequency domain analysis was performed on the real-time vibration parameters of the bridge structure to extract the deformation frequency components of the bridge structure; time-frequency domain analysis was also performed on the triaxial angular velocity parameters and triaxial acceleration parameters to extract the vehicle body vibration frequency components. Calculate the coherence coefficient between the frequency components of bridge structure deformation and vehicle vibration, and determine the coupling strength level at the current moment by calling the preset coupling strength classification model based on the coherence coefficient, real-time wind speed parameters, and real-time operating speed parameters. Based on the coupling strength level and the energy ratio of the bridge structure deformation frequency component to the vehicle vibration frequency component, a coupling state feature vector is generated to characterize the coupling state of the vehicle-track-bridge system.

4. The curved track monitoring system for a track machine according to claim 3, characterized in that, The preset coupling strength classification model is a multi-input classification model trained based on historical data. The preset coupling strength classification model outputs the coupling strength level at the current moment based on the nonlinear mapping relationship between the coherence coefficient, real-time wind speed parameters and real-time operating speed parameters.

5. A curved track monitoring system for a track machine according to claim 1, characterized in that, The pre-defined fusion weight allocation model specifically refers to: The preset fusion weight allocation model is a multi-input multi-output mapping model trained based on historical data. The input parameter is the coupling state feature vector, and the output parameter is the fusion weight value of each geometric measurement sensor. The geometric measurement sensor includes a laser scanning sensor and an image acquisition sensor. The preset fusion weight allocation model contains multiple weight allocation sub-models corresponding to different coupling strength levels. Each weight allocation sub-model stores the optimal weight allocation scheme that minimizes the error of the fusion result under different coupling conditions.

6. A curved track monitoring system for a track machine according to claim 1, characterized in that, The system dynamically adjusts the fusion weight allocation of the geometric measurement sensors by calling a preset fusion weight allocation model, and outputs the optimized fusion weight parameters, specifically including: Based on the coupling strength level in the coupling state feature vector, the corresponding weight allocation sub-model is selected from the preset fusion weight allocation model, and the weight allocation sub-model outputs the fusion weight value corresponding to each geometric measurement sensor according to the coupling state feature vector. The fusion weight values ​​are normalized to generate optimized fusion weight parameters where the sum of the weights of each geometric measurement sensor is equal to the preset total weight.

7. A curved track monitoring system for a track machine according to claim 6, characterized in that, The fusion estimation of orbital geometric parameters specifically includes: The system acquires optimized fusion weight parameters, raw track geometric measurement data collected by the geometric measurement sensor, track cross-section contour point cloud data output by the laser scanning sensor, and track appearance image data output by the image acquisition sensor. The point cloud data of the track cross-section contour output by the laser scanning sensor is processed to extract the first track geometric parameter sequence; the track appearance image data output by the image acquisition sensor is processed to extract the second track geometric parameter sequence, wherein the image processing includes feature point extraction, stereo matching and three-dimensional coordinate transformation. The first orbital geometric parameter sequence and the second orbital geometric parameter sequence are spatiotemporally registered to place them under a unified spatial coordinate system and time reference. Based on the optimized fusion weight parameters, the first orbital geometric parameter sequence and the second orbital geometric parameter sequence after spatiotemporal registration are weighted and summed to generate preliminary fused orbital geometric parameters.

8. A curved track monitoring system for a track machine according to claim 7, characterized in that, The confidence level of real-time evaluation of fusion results refers to: Obtain preliminary fused orbital geometric parameters and the first and second orbital geometric parameter sequences; Calculate the first deviation value between each data point in the first orbital geometric parameter sequence and the corresponding data point in the preliminary fused orbital geometric parameters to obtain the first deviation sequence composed of the first deviation values ​​of each data point; Calculate the second deviation value between each data point in the second orbital geometric parameter sequence and the corresponding data point in the preliminary fused orbital geometric parameters to obtain the second deviation sequence composed of the second deviation values ​​of each data point; The first deviation statistical characteristic value is calculated based on the first deviation sequence, the second deviation statistical characteristic value is calculated based on the second deviation sequence, and a confidence index is generated based on the first and second deviation statistical characteristic values. The first deviation statistical characteristic value represents the overall deviation between the first orbital geometric parameter sequence and the initially fused orbital geometric parameters, the second deviation statistical characteristic value represents the overall deviation between the second orbital geometric parameter sequence and the initially fused orbital geometric parameters, and the confidence index is negatively correlated with both the first and second deviation statistical characteristic values. The confidence index is compared with the preset confidence threshold. When the confidence index is greater than or equal to the preset confidence threshold, the preliminary fused orbital geometry parameters are directly output as orbital geometry parameters with confidence labels. When the confidence index is less than the preset confidence threshold, the feedback correction mechanism is triggered. Abnormal data points are identified based on the first deviation sequence and the second deviation sequence. The abnormal data points in the first track geometric parameter sequence and the second track geometric parameter sequence are corrected based on the preset compensation model. The weighted summation is recalculated according to the optimized fusion weight parameters to generate the corrected track geometric parameters as track geometric parameters with confidence labels as output.

9. A curved track monitoring system for a track machine according to claim 8, characterized in that, Based on a pre-defined compensation model, abnormal data in the first and second orbital geometric parameter sequences are corrected, specifically including: The first deviation value of each data point in the first deviation sequence is compared with the preset first anomaly threshold, and the second deviation value of each data point in the second deviation sequence is compared with the preset second anomaly threshold. When the first deviation value of any data point is greater than or equal to the first abnormal threshold, the data point in the first orbital geometric parameter sequence corresponding to that data point is marked as the first abnormal data point. When the second deviation value of any data point is greater than or equal to the second abnormal threshold, the data point in the second orbital geometric parameter sequence corresponding to that data point is marked as the second abnormal data point. The preset compensation model contains multiple compensation sub-models corresponding to different coupling strength levels. Each compensation sub-model stores abnormal data correction coefficients trained based on historical data. Based on the coupling strength level in the coupling state feature vector, a corresponding compensation sub-model is selected from the preset compensation models. The first abnormal data point and the second abnormal data point are respectively input into the selected compensation sub-model, and the compensation sub-model outputs the corrected first track geometric parameter data point and the corrected second track geometric parameter data point. Replace the corresponding outlier data points in the first orbital geometric parameter sequence with the corrected first orbital geometric parameter data points to generate the corrected first orbital geometric parameter sequence; replace the corresponding outlier data points in the second orbital geometric parameter sequence with the corrected second orbital geometric parameter data points to generate the corrected second orbital geometric parameter sequence.

10. A curved track monitoring system for a track machine according to claim 9, characterized in that, Dynamic spatiotemporal reference alignment of orbital geometry parameters with confidence labels refers to: The track geometry parameters with confidence labels are obtained from the feedback correction module. The track geometry parameters with confidence labels include the track geometry parameter values, confidence index and corresponding original acquisition timestamp and original mileage value for each data point. Based on the coupling strength level in the coupling state feature vector, query the corresponding preset spatiotemporal reference correction strategy under the current dynamic conditions; Based on the original acquisition timestamp, original mileage value, and spatiotemporal reference correction strategy, the track geometry parameters with confidence labels are aligned on the time axis to eliminate time reference deviations caused by fluctuations in track machine operating speed and data acquisition delays. Based on the bridge structure deformation frequency component in the coupled state feature vector, the time axis aligned track geometry parameters are spatially aligned to compensate for the spatial position offset caused by bridge elastic deformation and vehicle vibration, so that the track geometry parameter values ​​of each data point match the actual track physical position. The orbital geometry parameters aligned with the spatiotemporal reference are directly output as the aligned orbital geometry dataset.

Images