Train inspection robot cross-car navigation method and system based on polarized light and UWB

Abstract

The present application relates to train intelligent inspection and robot navigation technical field, specifically provides a kind of train inspection robot cross-carriage navigation method and system based on polarized light and UWB, method includes: robot inspection operation process simultaneously acquires wireless ranging, polarized light, angular velocity, displacement increment and car passageway geometric characteristic multi-source sensing parameter;Wireless ranging parameter is preprocessed, effective ranging set is screened and ranging quality parameter is obtained by quantization;According to ranging quality parameter, compensate polarized light sensing data, obtain polarized navigation parameter and extract navigation fluctuation feature and navigation confidence;Combining ranging quality, navigation feature, car geometric parameter divides three kinds of operation areas, car main body, local occlusion, windshield transition;According to the type and navigation confidence of region, dynamically allocate the fusion weight of each sensing parameter, output robot real-time pose by fusion solution, complete cross-carriage navigation.The present application uses UWB to combine polarized light technology, guarantees that robot continuously cross-carriage inspection.

Description

Technical Field

[0001] This invention relates to the field of intelligent train inspection and robot navigation technology, specifically to a method and system for cross-carriage navigation of a train inspection robot based on polarized light and UWB. Background Technology

[0002] With the rapid development of my country's high-speed rail network, the number of high-speed passenger trains and the pressure of daily carriage inspections are increasing. Carriage inspection is a crucial link in ensuring passenger safety and improving service quality. The high frequency of high-speed train operations, the numerous inspection points, and the complexity of service and safety checks result in a massive workload for carriage inspections. Furthermore, current high-speed rail carriage inspections in China are still primarily conducted manually by train crews, forcing major passenger transport sections to allocate significant manpower to complete routine carriage inspections.

[0003] In recent years, intelligent inspection robots have been widely used in the high-speed rail sector. These devices can not only perform maintenance and inspection of equipment and hygiene in train carriages at depots, but also conduct passenger service patrols during train operation. However, inspection robots are still limited by technical challenges such as low positioning accuracy, particularly in the absence of GPS signals inside train carriages, the large cumulative error of traditional inertial navigation, and the need for robots to move continuously between multiple carriages and accurately pass through carriage connections (aisles, windshields).

[0004] Existing technical solutions have incorporated UWB (Ultra-Wide Band) positioning and navigation technology and polarized light navigation technology to address this challenge. UWB is currently the mainstream technology for high-precision indoor wireless positioning. By pre-installing UWB base stations (anchor points) in each train carriage and at carriage connections, and attaching UWB tags to the robot, the distance between the base stations and tags is calculated using signal time-of-flight, thus determining the robot's real-time absolute coordinates. Polarized light navigation technology uses polarized light sensors on the robot to capture polarized light signals in the environment. By analyzing the direction and angle of the polarized light, it outputs the robot's heading angle, attitude, and direction of travel in real time. By using UWB to locate the robot's real-time position and polarized light to guide its direction, the challenge of moving the robot across train carriages is solved.

[0005] However, the combination of UWB and polarized light technologies still has certain technical shortcomings, specifically: 1. Main positioning area of ​​the carriage (inner passage of the carriage): The carriage is equipped with a complete UWB base station and the radio environment is good, so UWB can measure distance normally; however, the coating of the train windows and the refraction of the glass will distort the ambient natural light, causing the output heading of the airborne polarization sensor to have a fixed system deviation. Using only polarization heading will result in a large error; relying solely on UWB will fail when crossing the windshield, and continuous positioning cannot be achieved; traditional polarization navigation technology is only statically calibrated once at the factory, and the error will continue to increase after the windows age and the glass is replaced.

[0006] 2. Partial occlusion in non-transition areas (still inside the same carriage): Seats and onboard equipment form fixed occlusions, and some UWB base stations are blocked. In passenger scenarios, dynamic occlusions such as passengers and personal belongings will also be superimposed, resulting in a decrease in UWB reliability. However, the robot did not reach the windshield at the carriage connection. Conventional solutions lack an accurate area determination mechanism, which can easily lead to misjudgment of entering cross-carriage areas and incorrect switching of navigation strategies.

[0007] 3. Windshield-crossing carriage transition area (the connecting aisle between two carriages, a necessary passage across carriages): The all-metal windshield shell causes severe multipath reflection of UWB radio waves and large-scale signal blockage and failure, making UWB positioning basically ineffective; relying solely on the onboard IMU inertial measurement unit for navigation will result in accumulated yaw drift as you walk, and the onboard wheeled odometer is easily affected by wheel slippage, causing measurement errors. Conventional solutions lose positioning as soon as you reach the windshield and cannot continuously cross carriages.

[0008] To address these shortcomings, this invention provides a method and system for cross-carriage navigation of a train inspection robot based on polarized light and UWB. Summary of the Invention

[0009] To overcome the shortcomings of the prior art, the present invention provides a method and system for cross-carriage navigation of a train inspection robot based on polarized light and UWB, so as to solve the problems in the prior art.

[0010] One embodiment of the present invention provides a method for cross-car navigation of a train inspection robot based on polarized light and UWB, comprising the following steps: When the inspection robot performs its inspection task, it collects multi-source sensor parameters in real time and synchronously. The multi-source sensor parameters include at least wireless ranging parameters, polarization light sensing parameters, angular velocity parameters, displacement increment parameters, and geometric feature parameters of the carriage passage. The wireless ranging parameters are preprocessed to obtain an effective ranging set, and the ranging quality parameters are quantized based on the effective ranging set. Based on the ranging quality parameters, the polarization light sensing parameters are subjected to deviation compensation processing to obtain the compensated polarization heading parameters, and the heading fluctuation characteristics and heading parameter confidence of the polarization heading parameters are extracted. Based on the ranging quality parameters, the directional fluctuation characteristics of the polarization heading parameters, and the geometric characteristic parameters of the carriage passage, the current operating area type of the inspection robot is determined. The operating area type includes the main carriage positioning area, the partially obscured non-transition area, and the windshield-cross-carriage transition area. Based on the confidence level of the heading parameters according to the operating area type and polarization heading parameters, navigation fusion calculation is performed using the effective ranging set, polarization heading parameters, angular velocity parameters and displacement increment parameters as input items. The fusion weight of each parameter in the input item set is dynamically adjusted according to the operating area type where the inspection robot is currently located, and the real-time pose of the inspection robot is output to complete cross-carriage navigation.

[0011] This application also relates to a cross-car navigation system for a train inspection robot based on polarized light and UWB, including: The data acquisition module is used to collect multi-source sensor parameters in real time and synchronously when the inspection robot performs inspection tasks. The multi-source sensor parameters include at least wireless ranging parameters, polarization light sensing parameters, angular velocity parameters, displacement increment parameters, and geometric feature parameters of the carriage passage. The data processing module is used to preprocess the wireless ranging parameters, filter them to obtain an effective ranging set, and quantize the ranging quality parameters based on the effective ranging set. The data extraction module is used to perform deviation compensation processing on the polarization light sensing parameters based on the ranging quality parameters to obtain the compensated polarization heading parameters, and to extract the heading fluctuation characteristics and heading parameter confidence of the polarization heading parameters. The area determination module is used to determine the type of operating area where the inspection robot is currently located based on the directional fluctuation characteristics of the ranging quality parameters, polarization heading parameters, and geometric characteristic parameters of the carriage passage. The type of operating area includes the main body positioning area of ​​the carriage, the partially obscured non-transition area, and the windshield-cross-carriage transition area. The pose output module is used to perform navigation fusion calculation based on the heading parameter confidence of the operating area type and polarization heading parameters, with the effective ranging set, polarization heading parameters, angular velocity parameters and displacement increment parameters as input items. It also dynamically adjusts the fusion weight of each parameter in the input item set according to the operating area type where the inspection robot is currently located, and outputs the real-time pose of the inspection robot to complete cross-carriage navigation.

[0012] This application also relates to a computer device, including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the computer program to implement the steps of the above-described method for cross-car navigation of a train inspection robot based on polarized light and UWB.

[0013] This application also relates to a computer-readable storage medium storing a computer program that, when executed by a processor, implements the steps of the above-described method for cross-carriage navigation of a train inspection robot based on polarized light and UWB.

[0014] The above embodiments provide a method and system for cross-carriage navigation of a train inspection robot based on polarized light and UWB, which has the following beneficial effects: 1. By dynamically correcting polarization heading errors, the limitations of traditional static calibration are overcome. This invention abandons the model of relying on one-time factory calibration for polarization-based navigation. Based on UWB ranging quality parameters, it dynamically calculates heading compensation deviations, which can offset the fixed system deviations caused by window coatings and glass refraction in real time. At the same time, it adapts to environmental changes after window aging and replacement, effectively reducing polarization heading errors in the main positioning area of ​​the vehicle body and ensuring the long-term stability of directional guidance.

[0015] 2. It enables accurate differentiation of operating areas, reducing the possibility of incorrect navigation strategy switching. This invention integrates ranging quality parameters, polarization heading fluctuation characteristics, and carriage aisle geometric features to construct a multi-dimensional area determination logic. It can clearly distinguish the main carriage positioning area, the partially obscured non-transition area, and the windshield-cross-carriage transition area. This solves the problem of misjudging cross-carriage areas and incorrectly switching navigation strategies when partially obscured due to ambiguous area determination in traditional solutions, thus improving the adaptability of navigation strategies to actual operating conditions.

[0016] 3. Improves windshield area positioning performance and ensures continuity of navigation across carriages. This invention addresses scenarios where UWB signals fail in the windshield-to-carriage transition zone. By dynamically adjusting the fusion weights of various parameters, it effectively leverages the complementary roles of angular velocity and displacement increment parameters. Combined with a multi-branch navigation fusion solution model, it achieves collaborative fusion of multi-source features, replacing the traditional single inertial navigation scheme. This alleviates positioning fluctuations caused by accumulated yaw drift and wheel slippage errors in the windshield area, achieving continuous and stable full-link navigation across carriages. Attached Figure Description

[0017] Figure 1 A flowchart of a train inspection robot cross-carriage navigation method based on polarized light and UWB provided for an embodiment of the present invention; Figure 2 This is a schematic block diagram of a computer device provided in an embodiment of the present invention. Detailed Implementation

[0018] The technical solutions in the embodiments of the present invention will now be clearly and completely described in conjunction with the accompanying drawings.

[0019] Reference Figure 1One embodiment of the present invention provides a method for cross-car navigation of a train inspection robot based on polarized light and UWB, comprising the following steps: S10. When the inspection robot performs the inspection task, it collects multi-source sensor parameters in real time and synchronously. The multi-source sensor parameters include at least wireless ranging parameters, polarization light sensing parameters, angular velocity parameters, displacement increment parameters, and geometric feature parameters of the carriage passage. S20. The wireless ranging parameters are preprocessed to obtain an effective ranging set, and the ranging quality parameters are quantized based on the effective ranging set. S30. Based on the ranging quality parameters, the polarization light sensing parameters are subjected to deviation compensation processing to obtain the compensated polarization heading parameters, and the heading fluctuation characteristics and heading parameter confidence of the polarization heading parameters are extracted. S40. Based on the ranging quality parameters, the directional fluctuation characteristics of the polarization heading parameters, and the geometric characteristic parameters of the carriage passage, determine the current operating area type of the inspection robot. The operating area type includes the main carriage positioning area, the partially obscured non-transition area, and the windshield-cross-carriage transition area. S50. Based on the confidence level of the heading parameters of the operating area type and polarization heading parameters, navigation fusion calculation is performed using the effective ranging set, polarization heading parameters, angular velocity parameters and displacement increment parameters as input items. The fusion weight of each parameter in the input item set is dynamically adjusted according to the operating area type where the inspection robot is currently located, and the real-time pose of the inspection robot is output to complete cross-carriage navigation.

[0020] It should be noted that the inspection robot is an inspection robot equipped with a wireless ranging unit, a polarization light sensing unit, an inertial measurement unit, an odometer unit, and a lidar unit. The wireless ranging unit can be a UWB tag module, which works in conjunction with pre-deployed UWB anchor base stations on the train to complete wireless ranging operations. The polarization light sensing unit can be a miniature integrated polarization light sensor, the inertial measurement unit can be a MEMS-type IMU module, the odometer unit can be a wheeled incremental encoder, and the lidar unit can be a two-dimensional or three-dimensional lidar. All of the above hardware units are conventional sensing and measurement devices in this field, and specific models and specifications are not limited. Those skilled in the art can flexibly select and configure them based on the inspection accuracy requirements of the actual scenario, equipment installation space, cost budget, etc., as long as the acquisition accuracy and real-time requirements of various sensing parameters are met. Various hardware units are responsible for collecting corresponding sensing parameters: the wireless ranging unit collects wireless ranging parameters, the polarization light sensing unit collects polarization light sensing parameters, the inertial measurement unit collects angular velocity parameters, the odometer unit collects displacement increment parameters, and the lidar unit collects geometric feature parameters of the carriage passageway. UWB anchor base stations are pre-distributed evenly on the top or side walls of the passageways inside each carriage, and on both sides of the windshield transition points between carriages. All UWB base stations complete spatial position calibration in a unified coordinate system and store the corresponding coordinate data, which serves as a fixed data reference for wireless ranging. The aforementioned hardware deployment structure forms the hardware foundation for the synchronous acquisition of multi-source parameters in this embodiment.

[0021] In this embodiment, as described in steps S10~S50 above, specifically: Step S10 involves the acquisition and basic preprocessing of raw sensor data during the inspection robot's carriage inspection task. Specifically, the wireless ranging parameter refers to the raw ranging data between the base station and the tag on the inspection robot, acquired using a UWB positioning architecture; this data forms the basis for subsequently calculating the robot's spatial coordinates. The polarization light sensing parameter is the raw sampling data of ambient polarized light acquired by the polarization light sensor, used to extract information related to the robot's heading. The angular velocity parameter is the rotational angular velocity data of the inspection robot acquired by the inertial sensing component, reflecting the robot's attitude changes. The displacement increment parameter is the wheel displacement change data acquired by the odometer, characterizing the actual mileage traveled by the inspection robot. The carriage aisle geometric feature parameters are the contour data of the carriage's inner wall, windshield, and aisle acquired by the environmental scanning equipment, recording the inherent structural information of the carriage space. This step involves the real-time and synchronous acquisition of multiple types of sensor parameters. During the acquisition process, basic processing such as parameter anomaly monitoring, key parameter calibration, and data association and binding are performed simultaneously. Real-time means that the inspection robot continuously collects data throughout its inspection of the carriage, while synchronous means that various sensor parameters complete data acquisition at the same sampling moment (time sequence alignment is achieved through hardware clock synchronization or software timestamp calibration), ensuring that different types of data match each other in the time dimension and that data quality is initially controllable. The core purpose of this step is to synchronously collect multi-dimensional raw information on position, heading, motion status, and carriage environment. Basic preprocessing ensures data usability and builds a high-quality multi-source data foundation. This compensates for the shortcomings of existing navigation solutions that only use single or a small amount of sensor data and lack basic processing during the acquisition stage, making them susceptible to positioning failures caused by carriage optical interference, radio obstruction, and equipment errors.

[0022] Step S20 involves the preprocessing, effective data filtering, and quality quantification of UWB wireless ranging parameters. The data processing sequentially performs three basic steps: anomaly identification and repair, partition threshold filtering, and multi-index quantification. The core of the preprocessing for wireless ranging parameters is the removal and completion of clutter from the collected raw UWB ranging data. Specifically, this includes removing persistent abnormal data caused by signal obstruction and multipath interference, interpolating missing data during transient abnormal periods using historical valid data, and matching corresponding filtering criteria to the subsequent area determination requirements. The goal is to ensure that the data meets the reliability requirements for navigation calculation. The effective ranging set, which is the collection of qualified UWB ranging data obtained after the above preprocessing and screening, removing invalid data and supplementing missing data, is a reliable data source for subsequent heading compensation and coordinate calculation. The ranging quality parameter is a comprehensive index obtained by quantifying the effective ranging set according to preset rules. It is used to intuitively reflect the reliability and availability of the current UWB ranging. The quantification process can select one or more of the following indicators for weighted calculation: number of effective base stations, ranging residual statistics, multipath disturbance intensity, and base station geometry. The core purpose of this step is to solve the technical defects of existing technologies that directly use raw UWB ranging data without screening, resulting in large calculation errors and lack of dedicated ranging quality quantification indicators to judge the reliability of UWB data.

[0023] Step S30 involves processing polarization sensing parameter compensation and feature extraction based on the ranging quality parameters. The deviation compensation process uses the ranging quality parameters output in S20 as the basis for determining the compensation strategy. It combines the dynamic adaptation quality judgment standard of the carriage section markings with the actual quality of the UWB data to select the corresponding correction scheme. When the UWB ranging quality meets the standard, the heading compensation deviation is solved in real time by analyzing the inspection robot's reference heading and optimizing the offset. When the UWB ranging quality is insufficient, the historical compensation deviations stored in the corresponding carriage section are retrieved to correct the system offset error inherent in the original polarization sensing parameters. The compensated polarization heading parameters are heading data after deviation correction and elimination of inherent deviations caused by window optical interference. The heading fluctuation characteristics and heading parameter confidence are both obtained by specifically analyzing the compensated polarization heading parameters of the corresponding section, combined with the carriage section markings. The heading fluctuation characteristics reflect the fluctuation patterns of the heading data, while the heading parameter confidence is quantified based on the fluctuation characteristics to characterize the reliability of the current polarization heading data. The core purpose of this step is to address the technical shortcomings of traditional polarized light navigation, which only performs a static calibration once at the factory stage. The continuous course deviation caused by train window coating and glass aging refraction cannot be corrected online.

[0024] Step S40 involves determining the type of operating area where the inspection robot is located by combining ranging quality parameters, heading fluctuation characteristics, and carriage aisle geometric feature parameters. The carriage aisle geometric feature parameters are used to initially define the spatial range of the inspection robot by referring to a preset contour standard library. The heading fluctuation characteristics corresponding to the polarization heading parameters, after quantification and grading, reflect the stability of the current heading data. The ranging quality parameters, after grading and delineation, reflect the actual reliability level of the UWB ranging data. The preliminary division results corresponding to the geometric parameters, the quantification and grading results of the heading fluctuation, and the ranging quality grading results are combined to distinguish three types of operating areas according to preset judgment rules: the main positioning area of ​​the carriage, the partially obscured non-transition area, and the windshield-cross-carriage transition area. The core purpose of this step is to address the technical shortcomings of existing navigation schemes, which lack multi-dimensional joint judgment logic and easily confuse partially obscured scenarios within the same carriage with cross-carriage windshield scenarios, leading to incorrect switching of navigation control strategies.

[0025] Step S50 is the core navigation step that combines the operating area type and heading parameter confidence level to complete the dynamic weighted fusion calculation of multi-source parameters. The input set for the navigation fusion calculation is explicitly defined as the effective ranging set obtained in S20, the compensated polarization heading parameters obtained in S30, and the angular velocity parameters and displacement increment parameters collected in S10. Before constructing the input set, spatiotemporal synchronization calibration is performed on various sensor parameters, and a weight adjustment rule base is pre-built based on the heading confidence level and operating area type. The core of dynamically adjusting the fusion weights is based on the current operating area type of the inspection robot (the main body of the vehicle). The system considers the positioning area, the partially obscured non-transition area, and the windshield-cross-car transition area. Combining the confidence level of the heading parameters with the polarization heading parameters, a preset weight adjustment rule library is retrieved to adjust the contribution ratio of various parameters in the input item set according to preset rules, adapting to the reliability differences of various sensor parameters in different areas. Through the aforementioned dynamically weighted multi-source parameter input fusion solution model, which is a multi-branch structure model trained offline using historical sensor data under all operating conditions, data fusion and pose calculation are completed. Finally, the real-time position and attitude information of the inspection robot is output, enabling continuous navigation across car carriages. The core purpose of this step is to address the technical shortcomings of existing technologies that use fixed weights to fuse multi-source sensor data, which cannot adapt to the sensor performance changes in different train operating areas, leading to positioning loss and navigation interruption after UWB failure in the windshield-cross-car transition area. This ensures that the inspection robot can obtain stable and reliable navigation results under all operating conditions throughout the entire journey.

[0026] In one embodiment, step S10 specifically includes the following steps: S11. When the inspection robot receives the inspection task instruction and performs the inspection task along the train, it performs time synchronization and coordinate registration of each airborne unit, establishes a unified spatiotemporal reference, and synchronously triggers each airborne unit to perform real-time and continuous data acquisition based on the spatiotemporal reference, and obtains wireless ranging parameters, polarization light sensing parameters, angular velocity parameters, displacement increment parameters and carriage passage geometric feature parameters with corresponding spatiotemporal stamps. S12. Monitor the wireless ranging parameters and polarization light sensing parameters collected during the inspection robot's movement in real time. When the wireless ranging parameters and / or polarization light sensing parameters are abnormal, mark the time period of the abnormality and cache the wireless ranging parameters and polarization light sensing parameters before the time period of the abnormality. S13. Extract the corresponding zero-bias parameters from the real-time and continuously acquired angular velocity parameters and perform zero-bias stability verification. Remove invalid angular velocity parameters whose zero bias exceeds the preset threshold. Solve the zero-bias compensation parameters using the retained valid angular velocity parameters. Update and correct the zero-bias parameters of the subsequent real-time acquired angular velocity parameters based on the zero-bias compensation parameters. S14. Based on the real-time and continuously collected displacement increment parameters, the wheel slippage state is determined. When the wheel slippage condition is determined, the initial weight of the displacement increment parameter corresponding to the time stamp is pre-adjusted in the fusion solution. S15. Based on the real-time and continuously collected geometric feature parameters of the carriage passage, analyze the cross-sectional contour features of the carriage passage, mark the corresponding carriage section identifier according to the contour features, and associate and bind the carriage section identifier with the wireless ranging parameters, polarization light sensing parameters, angular velocity parameters and displacement increment parameters collected synchronously with the same spatiotemporal reference.

[0027] In this embodiment, as described in steps S11-S15 above, specifically: Step S11 is the initial stage for acquiring raw sensor data and establishing a reference. Its core is ensuring the temporal and spatial consistency of multi-source data through a unified spatiotemporal reference. When the inspection robot receives the inspection task instruction and moves along the train track to perform the inspection, it first performs time synchronization and coordinate registration on the wireless ranging unit, polarization sensing unit, inertial measurement unit, odometer unit, and lidar unit mounted on the robot. Time synchronization refers to ensuring that the acquisition clocks of each onboard unit are consistent through hardware clock calibration or software timestamp synchronization protocols. Coordinate registration involves aligning the acquisition coordinate systems of each unit to a preset train global coordinate system or robot body coordinate system, thereby establishing a unified spatiotemporal reference. Based on this unified spatiotemporal reference, all onboard units are synchronously triggered to start data acquisition, ensuring that various sensor parameters are sampled at the same time node and in a unified spatial coordinate system. This ultimately yields wireless ranging parameters, polarization sensing parameters, angular velocity parameters, displacement increment parameters, and carriage passage geometric feature parameters with corresponding spatiotemporal stamps (timestamps record the acquisition time, and spatial stamps are associated with the unified coordinate system position information). The core objective of this step is to resolve the spatiotemporal misalignment of data caused by independent acquisition by multiple airborne units, providing spatiotemporally aligned raw data support for subsequent data association processing and fusion calculations, and ensuring the accuracy of navigation calculations.

[0028] Step S12 involves anomaly monitoring, labeling, and data caching for multi-source sensor parameters. Its core function is to identify abnormal states in wireless ranging and polarization sensing data in advance, retaining valid historical data for subsequent processing. During the inspection robot's operation along the train, the collected wireless ranging and polarization sensing parameters are monitored in real time. Parameter anomalies include signal interruptions, sudden numerical changes exceeding preset reasonable ranges, and excessive data fluctuations. When anomalies are detected in a single wireless ranging parameter, a single polarization sensing parameter, or both types of parameters simultaneously, an anomaly occurrence period label is immediately assigned to the time interval where the anomaly occurred. The label information includes the anomaly start time, anomaly type, and corresponding parameter category. Simultaneously, the continuous valid wireless ranging and polarization sensing parameters before the anomaly occurrence period label is generated are cached to ensure that valid historical data is not lost. The core purpose of this step is to provide anomaly tracing and valid data support for wireless ranging parameter preprocessing and polarization sensing parameter deviation compensation, preventing abnormal data from directly flowing into subsequent calculation stages and causing errors. It also addresses the problem of having no valid historical data available when abnormal period data is missing.

[0029] Step S13 involves the zero-bias extraction, stability verification, and real-time correction of angular velocity parameters. Its core function is to eliminate the systematic errors caused by the zero-bias drift of the inertial measurement unit (IMU) and ensure the accuracy of the angular velocity data. During the real-time acquisition of angular velocity parameters by the inspection robot, the corresponding zero-bias parameters (i.e., the inherent offset output by the IMU in a stationary or uniform motion state) are first extracted from the continuous angular velocity parameters. Then, a zero-bias stability verification is performed on the extracted zero-bias parameters. By monitoring the fluctuation amplitude and continuous stability of the zero-bias parameters, zero-bias parameters exceeding a preset zero-bias threshold are determined to be invalid. The preset zero-bias threshold is dynamically set based on the factory zero-bias stability calibration value of the IMU, combined with the vibration interference intensity of the train carriage inspection scenario, the inspection robot's travel speed, and other actual working conditions. Specifically, it can be determined by statistically analyzing the inspection robot's idle and stationary state. The sum of the mean zero-bias fluctuation and three times the standard deviation is used as the initial threshold. Those skilled in the art can flexibly adjust the appropriate threshold range based on the selected inertial measurement unit (IMU) model and the actual inspection environment. For zero-bias parameters deemed invalid, the entire segment of angular velocity data corresponding to these invalid parameters is removed. Based on the retained valid angular velocity parameters, zero-bias compensation parameters are obtained through statistical analysis or fitting calculations. These parameters can dynamically adapt to the zero-bias drift characteristics of the IMU. Finally, based on the solved zero-bias compensation parameters, the zero-bias components in the subsequently acquired real-time angular velocity parameters are updated and corrected frame by frame to continuously offset the errors caused by zero-bias drift. The core purpose of this step is to address the technical shortcomings of easy zero-bias drift during long-term operation of the IMU and the accumulation of attitude calculation errors caused by directly using raw angular velocity data.

[0030] Step S14 involves the determination of wheel slippage status and pre-adjustment of fusion weights for displacement increment parameters. Its core is to identify displacement data distortion caused by wheel slippage in advance and reduce the impact of distorted data on navigation calculations through weight adjustment. During the real-time acquisition of displacement increment parameters by the inspection robot, slippage status is determined based on continuously sampled displacement increment data. The determination logic is centered on the changing pattern of displacement increments, combined with the correlation between the inspection robot's travel speed and angular velocity parameters for comprehensive judgment. Specifically, it includes monitoring the abrupt change in displacement increments per unit time (such as a sudden drop or rise in displacement increments within a short period) and the deviation between the displacement increment and the theoretical displacement derived from the angular velocity parameters. When the above indicators exceed the preset slippage determination threshold, it is determined that the current condition is wheel slippage. The preset slippage determination threshold is set based on parameters such as the inspection robot's wheel diameter, designed travel speed range, and the friction coefficient of the train carriage ground. Specifically, it can be determined by statistically analyzing the normal displacement increment fluctuation range under different ground conditions, using the mean + 2 to 4 times the standard deviation as the threshold range. Those skilled in the art can flexibly adjust this threshold according to the actual ground conditions of the inspection scenario. When a slippage condition is detected, the initial weight of the displacement increment parameter corresponding to the spatiotemporal stamp of the slippage period in the subsequent S50 fusion calculation is pre-adjusted. The weight reduction can be based on a preset ratio (e.g., reduced to 30%~50% of the normal weight) or dynamically adapted according to the severity of the slippage. The core purpose of this step is to address the technical defect that the odometer unit is easily affected by the smooth surface of the vehicle and the interference of debris, which can cause wheel slippage and lead to distortion of the displacement increment data. Direct participation in the fusion calculation would amplify navigation errors. By suppressing the weight of distorted data in advance, the reliability of multi-source data fusion is ensured.

[0031] Step S15 involves analyzing the geometric features of the carriage, marking and labeling sections, and associating and binding multi-source parameters. Its core is to divide the space into sections based on the geometric features of the carriage passage, so as to achieve a one-to-one correspondence between all types of sensing parameters and carriage section information. During the continuous collection of geometric feature parameters of the carriage passage by the inspection robot, the real-time transmitted contour data is first analyzed and processed to extract key cross-sectional contour features such as the cross-sectional width, side wall shape, and contour abrupt changes at the windshield transition. A standard cross-sectional contour database for each carriage is pre-stored. This standard database is established based on the factory size parameters of the train model and is used as the contour comparison benchmark. The real-time cross-sectional contour features obtained by analysis are matched one by one with the preset standard cross-sectional contours. According to the matching results, the carriage section identifier corresponding to the current position is marked. The carriage section identifier includes classification information such as carriage number, main carriage section, windshield transition section across carriage, and carriage end section. The comparison tolerance threshold used for contour matching is set by combining the laser radar ranging error and the small movement sway error of the inspection robot. It can be selected as 1.5 to 3 times the nominal ranging error of the equipment as the tolerance benchmark. Those skilled in the art can flexibly adjust the value according to the selected laser radar model and the actual installation tolerance of the carriage. After generating the carriage section identifiers, based on the unified spatiotemporal reference established in S11, the generated carriage section identifiers are associated and bound one-to-one with the wireless ranging parameters, polarization sensing parameters, angular velocity parameters, and displacement increment parameters collected synchronously with the same spatiotemporal stamp. This ensures that each set of sensor data includes the attribute information of its respective carriage section. The core purpose of this step is to achieve information coupling between multi-source sensor data and the spatial location of the carriage, solving the technical defects of various sensor data being unable to distinguish the carriage section to which they were collected and the difficulty in regionally targeted data correction.

[0032] In one embodiment, step S20 specifically includes the following steps: S21. Determine the label of the abnormal occurrence period based on the preset abnormal duration judgment threshold; when it is determined to be a continuous abnormality, remove the wireless ranging parameters corresponding to the continuous abnormality; when it is determined to be a momentary abnormality, use the wireless ranging parameters cached before the abnormal occurrence period label to interpolate and fill in the missing data of the abnormal period. S22. Within the wireless ranging parameters that have completed anomaly removal and interpolation completion, match the partition filtering threshold according to the operating area type. S23. Based on the partition filtering threshold and the geometric feature parameters of the carriage passage, filter the wireless ranging parameters corresponding to each available base station under the current spatiotemporal reference, and summarize them to obtain the effective ranging set; S24. Based on the preset index weighting coefficient corresponding to the current operating area type of the inspection robot, select at least one of the following: effective base station number, ranging residual statistics, multipath disturbance intensity, and base station geometric configuration index to perform quantification calculation on the effective ranging set to obtain ranging quality parameters.

[0033] In this embodiment, as described in steps S21-S24 above, specifically: Step S21 is the abnormal classification and data completion step for wireless ranging parameters. Its core is to differentiate the abnormal data marked in S12 based on the duration of the abnormality, remove invalid data and fill in the missing data to ensure the continuity and availability of ranging data. In specific operation, firstly, a preset threshold for judging the duration of anomalies is invoked. This threshold is set based on the occlusion characteristics of UWB signals in the train carriage environment and the duration of multipath interference. An exemplary value can be set to 50ms to 200ms (those skilled in the art can flexibly adjust it according to the carriage length, base station deployment density, and environmental interference intensity to determine a reasonable range). Then, the duration of anomalies corresponding to the anomaly occurrence period labeled by S12 is compared with the threshold. When the duration of anomalies is greater than or equal to the preset threshold, it is judged as a continuous anomaly (mostly caused by severe occlusion or temporary equipment failure, and the data has no recovery value), and all wireless ranging parameters corresponding to the continuous anomaly period are directly removed. When the duration of anomalies is less than the preset threshold, it is judged as a transient anomaly (mostly caused by transient interference or signal fluctuations). The continuous valid wireless ranging parameters before the anomaly occurred are invoked from the cache of S12, and the missing data of the anomaly period is filled in using conventional interpolation methods in the art, such as linear interpolation and nearest neighbor interpolation. The core purpose of this step is to address the technical flaw in S12, which only marks anomalies without classifying them, causing continuous invalid data and momentarily missing data to flow directly into subsequent stages and cause calculation errors. By using a differentiated strategy of classification and elimination and interpolation completion, invalid data can be filtered out completely while ensuring the continuity of the data sequence.

[0034] Step S22 is the partition threshold matching step for the preprocessed ranging data. Its core is to configure differentiated screening standards according to the different operating areas of the inspection robot, adapting to the different levels of UWB signal interference in each area. Based on the complete wireless ranging parameter dataset after anomaly removal and interpolation completion in S21, the real-time operating area type of the inspection robot is retrieved. The operating area type is divided into three categories: the main carriage positioning area, the partially obstructed non-transition area, and the windshield-cross-carriage transition area. The partition screening threshold is set according to the spatial enclosure degree, multipath interference strength, and UWB base station obstruction probability of different areas. Among them, the main carriage positioning area has less obstruction and better signal quality, so the corresponding screening threshold constraint is the most lenient. The threshold constraint for the partially obstructed non-transition area is moderate. The windshield-cross-carriage transition area has frequent obstruction and severe signal distortion, so the corresponding screening threshold constraint is the most stringent. The baseline value of the partition screening threshold can be pre-calibrated in combination with the on-site UWB base station deployment spacing and train carriage structural dimensions. Those skilled in the art can flexibly adjust the threshold size based on actual on-site test data. The core purpose of this step is to abandon the processing method of using a single filtering threshold across the entire scenario, and to solve the defects of a unified threshold that cannot adapt to the interference differences in different sections of the carriage, either by mistakenly deleting valid data or retaining poor-quality ranging data.

[0035] Step S23 involves filtering and integrating effective ranging data based on partition thresholds and geometric features. Its core is to filter compliant ranging data using both spatial geometric constraints and partition screening thresholds. After matching the partition screening thresholds in S22, using a unified spatiotemporal reference as a constraint, the geometric feature parameters of the carriage passage at the same time are retrieved. These parameters determine the current spatial range of the inspection robot, the actual deployment location of surrounding UWB base stations, and spatial occlusion relationships. Using the partition screening thresholds as numerical filtering conditions, the wireless ranging parameters fed back by each UWB base station are compared one by one. Abnormal ranging data exceeding the corresponding partition screening threshold range are eliminated, and ranging data from each base station that simultaneously meets both spatial geometric constraints and numerical threshold requirements are retained. All filtered and retained compliant wireless ranging parameters are then uniformly summarized to form a structured effective ranging set. The spatial geometric constraint rules are pre-set based on the cross-sectional profile of the carriage passage and the coordinates of the base station deployment. These rules are used to exclude invalid ranging data corresponding to physical barriers between the base station and the inspection robot caused by the carriage structure. The constraint standards can be fine-tuned by combining the measured profile error of the LiDAR. Those skilled in the art can determine the specific constraint details based on the on-site carriage structure and base station layout. The core purpose of this step is to achieve dual screening using both geometric space and numerical indicators, retaining only spatially visible and numerically valid ranging data, thus avoiding the problem that a single threshold screening cannot eliminate false ranging data caused by physical obstructions.

[0036] Step S24 is the computational step for weighted quantization of multiple indicators in the effective ranging set and generation of ranging quality parameters. Its core is to combine preset weighting coefficients for different regions and realize the numerical representation of UWB ranging quality through multi-dimensional indicator quantification. After determining the type of the inspection robot's real-time operating area, the pre-configured indicator weighting coefficients for that area are retrieved. The weighting coefficients corresponding to each indicator are set according to the differences in UWB signal environment in the main positioning area of ​​the carriage, the partial obstruction non-transition area, and the windshield-cross-carriage transition area. The initial value of the weighting coefficient can be statistically calibrated based on multiple sets of sample data measured on site. Those skilled in the art can flexibly adjust the weight ratio of each indicator in combination with the on-site base station deployment scheme and the carriage interference environment. Example weight configuration (the total weight is fixed at 1, for illustration only and does not limit the protection scope): For stable signal in the main carriage positioning area, the weights are: effective base station quantity 0.4, ranging residual statistics 0.3, multipath disturbance intensity 0.15, and base station geometry 0.15; for increased multipath interference in the non-transition area with partial obstruction, the weights are adjusted to: effective base station quantity 0.3, ranging residual statistics 0.3, multipath disturbance intensity 0.25, and base station geometry 0.15; for severe obstruction in the windshield transition area across the carriage, the weights are adjusted to: effective base station quantity 0.25, ranging residual statistics 0.25, multipath disturbance intensity 0.35, and base station geometry 0.15. During the calculation, one or more of the following indicators are selected from the four indicators: number of effective base stations, ranging residual statistics, multipath disturbance intensity, and base station geometry. The measured values ​​of each selected indicator are calculated one by one using the effective ranging set generated by S23. The number of effective base stations is the total number of UWB anchor points that can output compliant ranging data in the current time and space. The ranging residual statistics is the statistical result of the deviation between the measured ranging value and the theoretical geometric distance (such as the root mean square of the residual). The multipath disturbance intensity is obtained by converting the short-term fluctuation amplitude of the ranging data (such as the normalized variance value within the sliding window). The base station geometry is used to characterize the spatial distribution quality of available base stations relative to the inspection robot (such as the PDOP value, the smaller the value, the better the configuration). The measured results of each indicator are first normalized (mapped to the 0~1 range), then multiplied by the corresponding weighting coefficients and summed. For example, in the main positioning area of ​​the vehicle, the inspection robot is positioned with four indicators selected for calculation. After normalization, the effective base station count is 0.92, the ranging residual statistic is 0.88, the multipath disturbance intensity is 0.75, and the base station geometry is 0.80. Substituting these into the weights, we get: 0.92×0.4+0.88×0.3+0.75×0.15+0.80×0.15≈0.865. This value is the current ranging quality parameter, ultimately quantifying and generating a ranging quality parameter representing the current UWB ranging reliability. The core purpose of this step is to achieve quantitative evaluation of ranging quality, addressing the technical shortcomings of relying solely on human experience to judge ranging quality, which cannot provide quantitative judgment criteria for S30 polarization compensation. This provides quantitative input for subsequent graded compensation of polarization sensing parameters.

[0037] In one embodiment, step S30 specifically includes the following steps: S31. Retrieve the preset quality pass threshold, and dynamically correct the quality pass threshold according to the carriage section identification to obtain the corrected quality pass threshold. S32. Compare the ranging quality parameter with the corrected quality qualification threshold. When the ranging quality parameter is greater than or equal to the corrected quality qualification threshold, analyze the reference heading of the inspection robot based on the effective ranging set, and calculate the heading compensation deviation in real time based on the reference heading of the inspection robot and the original polarization heading parameter obtained by analyzing the original polarization light sensing parameters collected from the same spatiotemporal reference. When the ranging quality parameter is less than the corrected quality qualification threshold, retrieve the historical heading compensation deviation pre-stored in the corresponding carriage section. S33. The original polarization light sensing parameters collected under the same spatiotemporal reference are compensated for by the heading compensation deviation to obtain the compensated polarization heading parameters. S34. Extract the heading fluctuation characteristics and heading parameter confidence levels corresponding to the polarization heading parameters after compensation for different sections based on the carriage section identification.

[0038] In this embodiment, as described in steps S31-S34 above, specifically: Step S31 is the dynamic correction step for the quality qualification threshold. Its core is to make targeted corrections to the initial preset quality qualification threshold based on the differences in environmental interference in the corresponding sections of the carriage, so as to adapt to the inherent interference characteristics of UWB signals in different sections. In the execution phase, the basic quality qualification threshold pre-saved by the system is first retrieved. This basic quality qualification threshold is based on the statistical results of several measured distance quality samples under the condition that there is no obvious obstruction in the main positioning area of ​​the carriage. Since the spatial sealing degree, obstacle obstruction probability, and multipath interference intensity of the main carriage section, the partially obstructed section, and the windshield-cross-carriage transition section are different, directly using a uniform basic threshold can easily cause the quality judgment to be distorted. The system pre-assigns a unique threshold correction coefficient to each type of carriage section identifier. The correction coefficient is statistically calibrated based on historical full-volume ranging quality data for each section. The main carriage section experiences minimal interference, with a correction coefficient approaching 1; partially obstructed non-transition sections experience moderate interference, resulting in a moderately lower correction coefficient; and the windshield-crossing carriage transition section experiences the most severe interference, leading to the lowest correction coefficient value. Those skilled in the art can flexibly optimize the correction coefficient values ​​for each section based on on-site retesting data. The system calculates the corrected quality pass threshold as follows: Corrected quality pass threshold = Basic quality pass threshold × Correction coefficient corresponding to the current carriage section identifier, generating a corrected quality pass threshold adapted to the current section. The core purpose of this step is to abandon the model of using a single judgment threshold across the entire section, eliminating the adverse effects of section environmental differences on ranging quality judgment results. Step S32 is the dual-path determination and solution step for heading compensation deviation. Its core is to adapt two compensation deviation acquisition schemes—real-time calculation or historical retrieval—based on the UWB ranging quality compliance status, ensuring reliable heading compensation data under different signal quality conditions. Specifically, the ranging quality parameters generated in S24 are first compared one-to-one with the corrected quality compliance threshold obtained in S31. Based on the comparison results, two scenarios are processed: When the ranging quality parameter is greater than or equal to the corrected quality qualification threshold, the current UWB signal quality is determined to be reliable, and the conditions for real-time solution of heading compensation deviation are met: First, based on the effective ranging set generated by S23, the reference heading of the inspection robot is obtained by analyzing the pose data of the inspection robot at continuous sampling time and fitting the motion trajectory (this reference heading is not affected by polarized light interference, and its accuracy is guaranteed by the UWB ranging quality); Second, the original polarization heading parameters corresponding to the original polarized light sensing parameters are extracted from the original sensing data collected from the same spatiotemporal reference using conventional analytical methods in this field; Third, the difference between the reference heading and the original polarization heading parameters is calculated, and the difference is dynamically corrected by combining iterative optimization algorithms to obtain the real-time updated heading compensation deviation, ensuring that the deviation can offset the current polarized light sensing system offset.

[0039] When the ranging quality parameter is less than the corrected quality qualification threshold, it is determined that the current UWB signal is significantly affected by obstruction and multipath interference, resulting in insufficient real-time accuracy of the reference heading. In this case, the system directly retrieves the pre-stored historical heading compensation deviation for the corresponding carriage section. The historical heading compensation deviation is a set of effective heading compensation deviations calculated and stored in real time during past inspections of this carriage section when the UWB signal quality met the standards. The system selects the optimal value based on the principle of prioritizing the most recent acquisition and optimizing the average of multiple acquisitions, ensuring that the retrieved historical deviation can adapt to the polarization interference characteristics of the current section. The storage of historical deviations is categorized and managed using the carriage section identifier as an index, and is periodically updated and iterated with new effective real-time deviations to avoid compensation failure due to outdated historical data. The core purpose of this step is to solve the defect that a single compensation deviation acquisition scheme cannot adapt to the fluctuations in UWB signal quality. By adopting a dual-path strategy of real-time calculation when the quality meets the standards and historical data retrieval when the quality does not meet the standards, the system can ensure the compensation accuracy when the signal is good and avoid the problem of having no reliable deviations available when the signal is poor.

[0040] Step S33 is the deviation correction step for the original polarization sensing parameters. Its core is to use the heading compensation deviation output from S32 to perform orientation correction on the original data, eliminating the inherent system offset error of the polarization sensing. After obtaining the heading compensation deviation output from S32, the original polarization sensing parameters corresponding to the same spatiotemporal stamp are retrieved, and the heading compensation deviation is used as a correction amount to perform compensation calculations on the original polarization heading values. The compensation calculation adopts the conventional numerical addition and subtraction correction method in this field, superimposing or subtracting the corresponding heading compensation deviation from the original polarization heading parameters to offset the systematic heading offset caused by changes in window light transmission, stray light interference from the carriage, and temperature drift of the sensor devices. After unified spatiotemporal reference matching and numerical compensation processing, the system fixed offset component is removed, generating optimized compensated polarization heading parameters. The core purpose of this step is to correct the systematic heading deviation of the polarization sensor caused by interference from the complex operating conditions of the train carriage, improving the accuracy of the polarization heading data.

[0041] Step S34 is the segmented heading feature extraction and confidence conversion step. Its core is to achieve data partitioning and integration based on the carriage segment identifiers, extract heading fluctuation features by region, and convert the heading parameter confidence. Combining the carriage segment identifiers pre-bound to the parameters in S15, the polarization heading parameters after the compensation processing in S33 are classified and collected according to different carriage segments, and split into independent parameter datasets for the main carriage positioning area, the partially obscured non-transition area, and the windshield cross-carriage transition area. For the polarization heading data within each zone, heading fluctuation features are extracted within a fixed-length sliding statistical window. These features include quantitative indicators such as the mean, short-time variance, and maximum / minimum fluctuation difference of the heading data, used to intuitively characterize the stability of the polarization heading of the corresponding section. A mapping relationship between fluctuation features and confidence levels is pre-established based on historical calibration data for each section. The smaller the heading fluctuation, the higher the confidence level of the converted heading parameter; conversely, severe fluctuations result in a lower confidence level. This mapping conversion standard can be flexibly fine-tuned and the conversion rules reasonably determined based on the actual measured conditions of each carriage section. Based on the aforementioned mapping relationship, the confidence levels of the corresponding heading parameters are generated one by one from the extracted heading fluctuation features of each section. The core objective of this step is to achieve segment-based statistical analysis of polarization heading data and to quantify the reliability of data by region.

[0042] In one embodiment, step S32, when the ranging quality parameter is greater than or equal to the corrected quality qualification threshold, involves resolving the reference heading of the inspection robot based on the effective ranging set, and calculating the heading compensation deviation in real time based on the reference heading of the inspection robot and the original polarization heading parameter obtained by resolving the original polarization light sensing parameters collected from the same spatiotemporal reference. This step specifically includes: S321. When the ranging quality parameter is greater than or equal to the corrected quality qualification threshold, the robot pose data corresponding to each continuous sampling time is obtained by parsing the effective ranging set. S322. Taking the pose data of the inspection robot corresponding to the continuous sampling time as the processing object, the sliding window is used to segment the motion trajectory interval of the inspection robot along the time axis. Abnormal intervals are filtered out within a single window, and the effective trajectory points corresponding to uniform and stable driving are retained. The effective trajectory points are fitted with lines by the least squares method, and the azimuth angle corresponding to the fitted line is used as the reference heading of the inspection robot. S323. Obtain the original polarization heading parameters generated by the original polarization light sensing parameters under the same spatiotemporal reference, calculate the difference between the original polarization heading parameters and the reference heading of the inspection robot as the slow-varying offset, and iteratively optimize and update the slow-varying offset frame by frame through a preset iterative algorithm. When the slow-varying offset converges, the real-time heading compensation deviation is obtained.

[0043] In this embodiment, as described in steps S321-S323 above, specifically: Step S321 is the pre-step for solving the original pose of the reference heading. Its core is to calculate the temporally continuous pose data of the inspection robot based on the selected effective ranging set, provided that the UWB ranging quality meets the standards. After the threshold comparison confirms that the ranging quality parameters meet the judgment conditions, the effective ranging set with a unified spatiotemporal stamp generated in S23 is retrieved. The effective ranging set contains the measured ranging values ​​of multiple compliant UWB base stations relative to the inspection robot at the same acquisition time, and the global coordinates of each base station are pre-calibrated and stored in the system in the previous coordinate registration stage of S11. Using a common solution algorithm in this field, such as polygonal positioning, the planar position coordinates of the inspection robot in the train's global coordinate system are calculated using the multi-base station ranging data at a single moment. Then, combined with the position change relationship between adjacent sampling times and the synchronously acquired angular velocity parameters, the fuselage heading information at the corresponding moment is calculated, and integrated to form complete pose data for a single sampling moment. All effective sampling points are traversed sequentially according to the acquisition time sequence to generate multiple consecutive pose sequences arranged in chronological order, and each set of pose data is bound to a corresponding original spatiotemporal stamp. Those skilled in the art can flexibly select conventional algorithms such as polygonal positioning or fusion filtering positioning based on the density of base station deployment on site. The core purpose of this step is to output a continuous pose dataset with regular time sequence and unified coordinates, providing complete raw data support for the S322 sliding window to segment and extract trajectories, select effective points, and fit a reference heading.

[0044] Step S322 involves trajectory segmentation and filtering, and baseline heading fitting. Its core is to process the temporal pose segment by segment using a sliding window, eliminate abnormal trajectory points, and obtain a high-precision baseline heading using least-squares fitting. This step uses the continuous sampled pose sequence output in S321 as the data source. A sliding window with a fixed number of sampling points or a fixed time length is pre-set, and slides frame by frame along the time axis, segmenting and extracting the motion trajectory interval of the inspection robot. The size of the sliding window is determined comprehensively based on the inspection robot's normal travel speed and the sensor unit's sampling frequency. For example, a single window can contain 20 to 50 sets of continuous pose data. Those skilled in the art can flexibly modify the window parameters according to the actual sampling conditions on site to determine a reasonable range. For each trajectory point within the window, validity is judged based on the displacement change rate of adjacent points and the fluctuation amplitude of travel speed at adjacent times. Abnormal points caused by sudden bumps, brief starts and stops, or instantaneous wheel slippage are deemed invalid and eliminated, retaining only valid trajectory points corresponding to a uniform and stable travel state. After filtering within a single window, the remaining valid trajectory points are fitted with a spatial straight line using the general least squares method. The equation of the fitted straight line is obtained, and the azimuth angle corresponding to this fitted straight line in the train's global coordinate system serves as the reference heading for the inspection robot in the current window. The core purpose of this step is to filter out trajectory distortion data caused by travel disturbances, avoid interference from abnormal points in the heading calculation accuracy, and obtain a reference heading unaffected by polarization sensor errors through steady-state trajectory fitting, providing a reliable reference for S323 to calculate the heading offset.

[0045] Step S323 is the slow-varying offset calculation and iterative convergence optimization stage. Its core is to construct the offset based on the difference between the reference heading and the original polarization heading, and then obtain the final usable real-time heading compensation deviation through iterative optimization and convergence. First, the original polarization light sensor parameters with the same time and space stamp as the reference heading are retrieved, and the original polarization heading parameters are generated by parsing using the conventional polarization azimuth angle calculation method. The difference between the original polarization heading parameters and the reference heading output in S322 is calculated, and the result is defined as the slow-varying offset. This offset is mainly caused by long-term systematic errors such as polarization sensor temperature drift and slow changes in stray light in the carriage environment, which is different from the random abrupt error caused by instantaneous interference. The system is equipped with a preset iterative optimization algorithm, with options including mature iterative methods in the field such as recursive weighted averaging and Kalman filtering iteration. It continuously iterates and corrects the slowly varying offset by acquiring new time-series data frame by frame. Simultaneously, a preset offset convergence threshold is set, which is determined based on the nominal angular accuracy of the selected polarization sensor, and can be exemplarily set to 0.1°–0.3°. When the offset change after a preset number of consecutive frame iterations (e.g., 8–15 consecutive frames) is less than this convergence threshold, the slowly varying offset is considered to have converged. Those skilled in the art can flexibly adjust the convergence threshold and the number of consecutive frames based on the actual sensor accuracy and the strength of interference in the field environment. The slowly varying offset after convergence is the real-time heading compensation deviation, which is passed down to S33 for correction and compensation of the original polarization heading. The core purpose of this step is to filter out the interference of random instantaneous noise on the offset, eliminate the influence of accidental errors in single sampling, obtain a stable systematic compensation deviation, and further improve the accuracy of subsequent polarization heading compensation.

[0046] In one embodiment, step S40 specifically includes the following steps: S41. Based on the geometric feature parameters of the carriage passage, analyze the cross-sectional dimensions and contour structure of the carriage. Use the cross-sectional dimensions and contour structure of the carriage to match the preset contour standard library to obtain the area division result corresponding to the location of the inspection robot. S42. Extract the heading fluctuation amplitude index of the polarization heading parameters, and use a preset classification threshold to quantify and classify the heading fluctuation index to obtain the fluctuation amplitude classification result. S43. Based on the preset grading threshold, the ranging quality parameters are compared and the corresponding quality levels are defined to obtain the ranging quality grading results. S44. By integrating the results of the region division, the fluctuation amplitude classification, and the ranging quality classification, the inspection robot is determined to be in the current operating area type according to the preset judgment rules.

[0047] In this embodiment, as described in steps S41-S44 above, specifically: Step S41 is the preliminary determination of region division. Its core is to achieve the preliminary positioning of the spatial area where the inspection robot is located by analyzing the geometric features of the carriage and matching it with a standard contour library, providing a basic spatial basis for subsequent multi-dimensional fusion determination. This step uses the geometric feature parameters of the carriage passage collected in real time as the data source. First, the parameters are analyzed and processed: the key external dimensions of the carriage cross-section (such as cross-section width, upper and lower height, and side wall spacing) and core contour structures (such as side wall straightness, whether there are local protrusions / depressions, contour abrupt change points unique to the windshield transition area, and the closing angle of the carriage end) are extracted to form a structured real-time contour feature dataset. The analysis dimensions are consistent with the contour analysis logic of S15 to ensure data continuity. The system pre-stores a preset contour standard library, which is established based on the factory design drawings of the target train model and the actual measured contour data of the whole train. It includes standard contour templates for three types of areas: the main body positioning area of ​​the carriage, the partially obscured non-transition area, and the windshield transition area across the carriage. The contour template of the main body positioning area of ​​the carriage is characterized by stable dimensions and straight side walls without obvious abrupt changes. The template of the partially obscured non-transition area includes the contour recess / protrusion features corresponding to local obstacles such as seat protrusions and equipment boxes. The template of the windshield transition area across the carriage is marked with typical structures such as glass mounting grooves, thickness abrupt changes, and dimensional narrowing corresponding to the transition frame. During the matching process, the real-time parsed contour features are compared one by one with three types of templates in the standard library. The comparison dimensions include deviations in external dimensions (such as width deviation and height deviation) and similarity of contour structure (such as the overlap of abrupt change points and the consistency of curve curvature). The tolerance threshold used for matching is set comprehensively based on the nominal ranging error of the LiDAR and the small sway error of the inspection robot. A basic tolerance of 1.5 to 3 times the equipment error can be selected. Those skilled in the art can flexibly fine-tune it according to the contour matching accuracy measured on site. When the matching degree between the real-time contour features and a certain type of region template exceeds the preset similarity threshold (such as above 85%), the region label corresponding to that template is output, which is the region division result corresponding to the location of the inspection robot. The core purpose of this step is to use the inherent differences in the geometry of the carriage to achieve preliminary region screening, avoid the problem of easy interference from relying solely on sensor data, and provide reliable spatial feature support for the S44 to fuse the fluctuation classification and quality classification results to complete the final region determination.

[0048] Step S42 is the polarization heading fluctuation quantification and grading process. Its core is to extract the heading fluctuation amplitude index from the compensated polarization heading parameters, complete the numerical quantification and grading using a preset grading threshold, and output the grading label. This step uses the polarization heading parameters after compensation and correction in S33 as the raw data, and completes the extraction of the heading fluctuation amplitude index within a fixed-length time-series sliding statistical window. The standard deviation of the heading values ​​and the maximum and minimum range of the heading values ​​within the window can be selected as the fluctuation amplitude quantification index. The preset grading threshold is statistically calibrated based on historical measured heading data for each carriage section, combined with inherent interference from the main carriage positioning area, partially obstructed non-transition areas, and windshield-cross-carriage transition areas. The system sets multiple grading boundaries to differentiate between strong and weak variability. For example, it divides the threshold into three levels: low variability, medium variability, and high variability. The threshold boundaries corresponding to the same level in different regions can be fine-tuned individually according to the environmental conditions of the section. Those skilled in the art can optimize the threshold values ​​of each level by combining multiple batches of inspection samples collected on-site to determine a reasonable threshold range. The real-time calculated heading variability amplitude index is compared with the threshold values ​​of each level one by one. The corresponding level is determined according to the threshold range in which the index is located, and finally a standardized variability amplitude grading result is generated. The core purpose of this step is to convert the continuously changing heading variability values ​​into discrete grading indicators to achieve standardized quantification of polarization heading stability.

[0049] Step S43 is the ranging quality parameter grading and calibration step. Its core is to classify the ranging quality parameters according to pre-set grading thresholds, achieving discrete grading of UWB ranging quality. This step uses the ranging quality parameters generated in S24 as the processing object. The preset grading thresholds are based on long-term accumulated ranging quality statistical samples from different carriage sections, dividing them into multiple boundary values. The grading standards are differentiated based on signal interference differences in the main carriage positioning area, partial obstruction non-transition area, and windshield-cross-carriage transition area, also classifying them into three quality levels: excellent, medium, and poor. The boundary thresholds for each level can be referenced from the on-site base station deployment. By conducting on-site measurements and calibrations under environmental and carriage obstruction conditions, those skilled in the art can flexibly adjust the threshold size based on historical inspection data. The real-time acquired ranging quality parameters are compared with the threshold values ​​of each level in sequence. Based on the threshold range in which the parameter is located, the corresponding quality level label is matched, and a standardized ranging quality grading result is formed. The core purpose of this step is to convert continuously measured ranging quality values ​​into fixed level labels, and form a unified format of discrete judgment data with the S41 area division results and the S42 heading fluctuation grading results, so as to provide standardized grading input data for the S44 multi-feature fusion determination of the operating area type.

[0050] Step S44 is the multi-source classification result fusion and final determination of the operating area. Its core is to integrate the region division results obtained from geometric contour matching in S41, the polarization heading fluctuation amplitude classification results in S42, and the ranging quality classification results in S43. Based on pre-set judgment rules, a comprehensive verification and determination of the operating area type is completed. The system pre-forms standardized preset judgment rules based on a large number of historical inspection samples from each carriage section of the entire line. The rules use the initial judgment zone obtained from the carriage cross-section contour as the primary judgment basis, and the heading fluctuation classification and ranging quality classification as auxiliary correction and verification conditions. An exemplary judgment logic can be set as follows: when the initial contour judgment result is the main positioning area of ​​the carriage, and the heading fluctuation is at a low fluctuation level and the ranging quality is excellent... At the first level, the final operating area is directly determined as the main positioning area of ​​the carriage. If the initial contour judgment result is unclear or contradicts the two sensor classification features, the judgment result is corrected to the windshield-to-carriage transition area based on the features of high fluctuation and poor ranging quality. If the fluctuation is moderate and the ranging quality is moderate, it is corrected to the partial occlusion non-transition area. Each judgment detail can be flexibly optimized and adjusted in combination with the working condition data of repeated on-site measurements. The cross-verification of the three types of results is completed by comparing each item with the preset judgment rules, and the final operating area type after multi-dimensional verification is output. The core purpose of this step is to overcome the drawback of single contour recognition being easily misjudged by local bumps and temporary debris obstruction. By using polarization heading and UWB ranging as two sensor indicators for cross-verification, the accuracy of operating area recognition is improved.

[0051] In one embodiment, step S50 specifically includes the following steps: S51. Calculate the confidence level of the heading parameter confidence using the fluctuation amplitude classification results, and construct a weight adjustment rule base based on the confidence level and the current operating area type of the inspection robot. S52. Perform spatiotemporal synchronization calibration on the effective ranging set, polarization heading parameters, angular velocity parameters and displacement increment parameters to construct an input item set; S53. Based on the type of the current operating area of ​​the inspection robot and the confidence level of the heading parameter, retrieve the weight adjustment rule base to assign dynamic fusion weights to each corresponding parameter in the input item set. S54. Input the set of input items after configuring dynamic fusion weights into the trained multi-branch navigation fusion solution model for fusion solution, and output the real-time pose of the inspection robot to complete cross-carriage navigation.

[0052] In this embodiment, as described in steps S51-S54 above, specifically: Step S51 is the construction of the weight adjustment rule base. Its core is to establish differentiated parameter weight allocation rules by combining the confidence level of the heading parameters with the current operating area type, providing a standardized basis for subsequent dynamic weight adaptation. This step first uses the fluctuation amplitude classification results output in S42 as a foundation, and associates them with the preset fluctuation characteristics and confidence mapping relationship in S34 to complete the confidence level conversion. Specifically, a low fluctuation amplitude classification corresponds to a high confidence level; medium fluctuation corresponds to medium confidence; and high fluctuation corresponds to low confidence. The conversion logic is consistent with S34 to ensure data consistency. Subsequently, based on the type of the inspection robot's current operating area determined by S44 (main carriage positioning area, partial obstruction non-transition area, windshield cross-carriage transition area), a weight adjustment rule base with area type and confidence level as dual indexes is constructed. The rule base is constructed based on the statistical analysis results of historical navigation fusion data under multiple working conditions across the entire line. The core rule logic is: to differentiate the basic weight range of effective ranging set (UWB), polarization heading parameters, angular velocity parameters, and displacement increment parameters according to the differences in the reliability of sensor parameters in different areas and the stability of parameters corresponding to the confidence level. An example rule configuration is as follows: When the operating area is the main positioning area of ​​the carriage and the confidence level is high, the effective ranging set weight is allocated as follows: 0.4-0.5, polarization heading parameter weight is allocated as follows: 0.3-0.35, angular velocity parameter weight is allocated as follows: 0.1-0.15, and displacement increment parameter weight is allocated as follows: When the operating area is a partially obstructed non-transition area and the confidence level is medium, the effective ranging set weight is adjusted to 0.3-0.35, polarization heading parameter weight is allocated as follows: 0.25-0.3, angular velocity parameter weight is allocated as follows: 0.2-0.25, and displacement increment parameter weight is allocated as follows: When the operating area is a windshield-crossing carriage transition area and the confidence level is low, the effective ranging set weight and polarization heading parameter weight are further reduced to 0.2-0.25 and 0.15-0.2, respectively, while the angular velocity parameter weight and displacement increment parameter weight are simultaneously increased to 0.3-0.35 and 0.2-0.25, respectively. The weight ranges of each parameter in the rule base can be flexibly optimized by those skilled in the art based on navigation error data measured in the field. By statistically analyzing the fusion solution accuracy under different weight ratios, the boundaries of the weight intervals can be iteratively corrected. The core objective of this step is to establish a weight allocation rule system with two dimensions: regional characteristics and parameter reliability, thereby addressing the technical shortcomings of fixed weights being unable to adapt to different regions and different levels of parameter stability.

[0053] Step S52 is the spatiotemporal calibration and input set construction step for multi-source sensor data. Its core is to unify the time reference and spatial coordinate system of the four types of sensor data, eliminating data misalignment issues caused by sampling time differences and coordinate deviations, and then integrating them to generate a standardized input item set. This step uses the effective ranging set generated in S23, the polarization heading parameters after compensation in S33, and the angular velocity and displacement increment parameters synchronously collected during the inspection process as the original data sources. The global spatiotemporal stamp uniformly calibrated in stage S11 is used as the alignment reference for time synchronization processing. To address the data sampling time misalignment caused by inconsistent sampling frequencies of various sensor hardware, industry-standard processing methods such as nearest neighbor matching or linear interpolation are used to fill in missing sampling points, ensuring that all parameters are strictly aligned. At the same acquisition time, spatial dimension calibration is performed to convert all four types of parameters into the train's global coordinate system, correcting spatial coordinate errors caused by the offset of different sensor installation positions. After all calibration work is completed, the effective ranging data, polarization heading data, angular velocity data, and displacement increment data corresponding to the same time and space stamp are grouped and packaged one by one to form a well-structured set of input items. Those skilled in the art can adjust the synchronization implementation scheme based on the actual sampling rate and installation calibration information of each sensor device on site. The core purpose of this step is to solve the data mismatch defects caused by asynchronous acquisition of multiple sensors and inconsistent coordinate systems, and to ensure that all parameters correspond one by one in the subsequent dynamic weight allocation and model fusion calculation process, thereby improving the computational reliability of multi-source data fusion.

[0054] Step S53 is the dynamic fusion weight matching and assignment step. Its core is to combine the current operating area type and the heading parameter confidence level, and match the corresponding weight interval from the weight adjustment rule base established in S51. It then assigns specific dynamic fusion weights to each of the four types of sensor parameters within the input item set generated in S52. This step first reads the operating area type finally determined in S44 and the heading parameter confidence level converted in S51, using these two indicators as a joint retrieval index. It then locates matching entries in the weight adjustment rule base indexed by area type and confidence level, retrieving the valid ranging set, polarization heading parameter, and angle within each entry. The weight ranges for velocity parameters and displacement increment parameters are determined by combining the real-time performance of the current single-frame data and selecting specific weight values ​​within the corresponding ranges. This completes the weight binding of each parameter in the input set. The final values ​​of various parameters can be slightly adaptively adjusted within the range given by the rules. Those skilled in the art can optimize the value logic within the range by combining the actual error performance of on-site navigation. The core purpose of this step is to get rid of the limitation of fixed weights across the entire scene and dynamically adjust the fusion ratio of each sensor data according to the real-time operating conditions and the reliability of the parameters, so that the sensor parameters with higher reliability receive a larger weight ratio, while the weights of parameters with lower reliability are appropriately reduced.

[0055] Step S54 is the multi-branch model weighted fusion solution and real-time pose output stage. Its core is to input the input item set, configured with dynamic fusion weights in S53, into the pre-trained multi-branch navigation fusion solution model. Through branch feature extraction, weighted fusion, and pose calculation, the real-time pose of the inspection robot is output, ensuring the continuity of navigation across carriages. This step calls the input item set constructed in S52 and dynamically weighted in S53 as the model input. Four types of parameters in this input item set correspond one-to-one with model branches: the effective ranging set corresponds to the effective ranging feature branch of the model; the compensated polarization heading parameters correspond to the polarization heading feature branch; the angular velocity parameters correspond to the angular velocity feature branch; and the displacement increment parameters correspond to the displacement increment feature branch. During the solution process, each branch processes its corresponding input parameters synchronously and in parallel. After receiving the effective ranging feature branch, it first filters out abnormal jumping ranging values ​​caused by occlusion and multipath interference by combining the preset spatial coordinates of the base station. Then, based on the ranging constraint relationship of multiple base stations at the same time, it extracts the spatial position association features of the robot relative to each base station, the signal-to-noise ratio features of the ranging data, and the discrete features of the spatial distribution of multiple points. After standardized coding processing, it generates a regular effective ranging feature vector. The polarization heading feature branch extracts heading stability and orientation features from polarization heading data. First, it removes single-point pulse heading anomalies caused by instantaneous stray light in the environment. Then, it calculates heading fluctuation statistics based on a fixed-length time window and extracts heading reference system offset features, short-term heading stability features, and segmented azimuth gradual change features. These features are then integrated and spliced ​​together to form a complete polarization heading feature vector. The angular velocity feature branch analyzes the robot rotation motion features contained in the angular velocity parameters. First, the inherent zero bias of the sensing hardware is compensated and corrected. The original angular velocity sequence is separated into two types of data components: the angular velocity of travel and the angular velocity of stationary spin. The instantaneous angular velocity magnitude feature, the cumulative angular velocity feature over a fixed period, and the angular fluctuation stability feature are extracted and combined to obtain the motion feature vector corresponding to the angular velocity. The displacement increment feature branch mines the mileage accumulation features corresponding to the displacement increment data. First, it corrects the abnormal displacement increment data caused by wheel spin and ground bumps. Then, it accumulates the mileage change according to time sequence segments, extracts the single-step displacement amplitude features, continuous mileage increase and decrease trend features, and displacement-to-speed correlation features, and finally generates the mileage feature vector corresponding to the displacement increment. The feature vectors output by each branch are synchronously fed into the model's fusion solution layer. The fusion solution layer first performs dimension unification preprocessing on the feature vectors of the four types of branches; it then converts the heterogeneous feature vectors output by different branches into the same dimension (exemplarily set to 256 dimensions) through a feature dimension mapping matrix to ensure the feasibility of weighted operations; subsequently, it strictly follows the dynamic fusion weights allocated by S53 to perform element-wise weighted operations of "branch feature × corresponding weight" on the normalized feature vectors, and then performs element-wise summation and fusion on all weighted feature vectors. At the same time, it introduces an L2 regularization term to constrain feature redundancy in the fusion process and avoid overfitting. For example, if the weight of the effective ranging feature branch is 0.4, the weight of the polarization heading feature branch is 0.3, the weight of the angular velocity feature branch is 0.2, and the weight of the displacement increment feature branch is 0.1, and the i-th element of the 256-dimensional feature vector output by the four branches at a certain moment is a, b, c, d respectively, then the i-th element of the fused feature vector is (a×0.4+b×0.3+c×0.2+d×0.1), and all elements are normalized to the [0,1] interval to ensure that the parameter features with high reliability play a dominant role in the fusion result, while suppressing the interference of low reliability features.

[0056] The 256-dimensional global fused feature vector generated after fusion is input into the pose output layer. The pose output layer consists of three fully connected network layers. The first layer compresses the fused feature vector from 256 dimensions to 64 dimensions and extracts the core nonlinear correlation features through the ReLU activation function. The second layer further compresses it to 16 dimensions to strengthen the mapping correlation between features and pose parameters. The third layer is the output layer, which maps the 16-dimensional features into three continuous values ​​through a linear activation function, corresponding to the X-axis coordinate, Y-axis coordinate, and heading angle in the train's global coordinate system, respectively. This mapping relationship is learned through massive labeled samples during the model training phase. For example, when the core feature representing "stable spatial distance + smooth heading + uniform motion" in the fused feature vector is activated, it will output the pose result "X=12.5m, Y=4.8m, heading angle=89.2°". After the output, it will also be checked through a preset range (such as the X / Y coordinates need to be within the constraints of the train car length / width, and the heading angle is limited to 0°~360°) to ensure the physical rationality of the pose result. Finally, it is combined to form a complete real-time pose including planar coordinates and heading angle, providing accurate navigation and positioning basis for the inspection robot.

[0057] The core objective of this step is to utilize the multi-branch structure to adapt to the feature differences of multi-source heterogeneous sensor data, combine dynamic fusion weights to match the current operating area and parameter confidence conditions, and offset the limitations of single sensor data through model fusion calculation to continuously output stable real-time pose, ultimately achieving seamless cross-carriage autonomous navigation of the robot in various complex sections such as the main body of the carriage, partial occlusion, and windshield transitions between carriages.

[0058] It should be noted that the construction and training of the multi-branch navigation fusion solution model specifically includes the following steps: S501. Collect historical sensor data under different operating conditions in different areas of the train carriage. Based on the historical sensor data, construct a model training sample set including several sets of samples. Each set of samples includes effective ranging data, polarization heading data, angular velocity data and displacement increment data. Each set of samples is also labeled with the corresponding actual pose label of the inspection robot. S502. Construct a multi-branch neural network model, which includes: The effective ranging feature branch is used to receive effective ranging data and extract effective ranging features. The polarization heading feature branch is used to receive polarization heading data and extract polarization heading features. The angular velocity feature branch is used to receive angular velocity data and extract the motion features corresponding to the angular velocity. The displacement increment feature branch is used to receive displacement increment data and extract the mileage features corresponding to the displacement increment. The fusion solution layer is used to receive effective ranging features, polarization heading features, motion features and odometer features, and perform feature weighted fusion according to dynamic fusion weights to output a fused feature vector; The pose output layer is used to receive fused feature vectors, perform feature calculations, and output the real-time pose of the inspection robot. S503. Preprocess the model training sample set and divide it into training set and validation set according to a preset ratio. Input the training set into the multi-branch neural network model in batches for feature extraction, weighted fusion and feature calculation, and output the predicted pose of the inspection robot. S504. Configure weight constraint rules for the multi-branch neural network model. The weight constraint rules are set based on the corresponding relationship between different operating area types of the carriage and the confidence level of the heading parameter, and are used to guide the model to learn the contribution rules of each parameter feature under different working conditions. S505. With the goal of minimizing the error between the model-predicted pose of the inspection robot and the actual pose label of the inspection robot, a loss function is constructed, and the gradient descent algorithm is used to iteratively optimize the model parameters until the pose prediction error of the validation set is lower than the preset error threshold. Then, training is stopped, and the trained multi-branch navigation fusion solution model is obtained.

[0059] In this embodiment, as described in steps S501~S505 above, specifically: Step S501 is the model training sample set construction stage. Its core is to collect historical sensor data under all working conditions and complete pose annotation, providing high-quality training materials for multi-branch models and ensuring that the models have cross-regional navigation adaptation capabilities. This step first targets three types of operating areas in the train carriage (main carriage positioning area, partially obstructed non-transition area, and windshield cross-carriage transition area), covering different typical working conditions to collect historical data, including scenarios such as robot uniform speed movement, start-stop switching, slight bumps, and temporary obstruction by local debris. Four types of core sensor data are collected simultaneously: effective ranging data output by the UWB device, polarization heading data collected by the polarization sensor and compensated, angular velocity data synchronously acquired by the inertial measurement unit, and displacement increment data output by the odometer. The collection frequency is consistent with the sensor sampling frequency during actual on-site inspection (exemplarily set to 100Hz) to ensure that the data timing characteristics match the actual application scenario. Subsequently, a sample set was constructed based on the collected time-series sensor data. Each sample set was extracted using a fixed-length time-series window (an example of a 100ms window containing 10 sets of continuous sampling data). Each sample set completely contained the four types of sensor data sequences within the corresponding window. To ensure the effectiveness of training supervision, mature industry benchmark positioning solutions such as laser positioning systems and visual positioning markers were used to obtain the actual pose of the inspection robot in the train's global coordinate system at the time of sample collection (including X-axis coordinates, Y-axis coordinates, and heading angle). This actual pose was used as a label and bound to the corresponding sensor data sequence, forming a complete sample structure of four types of sensor data plus the actual pose label. During the construction of the sample set, it is necessary to ensure data diversity, covering sensor data from different vehicle models, different lighting conditions, and different equipment wear states. An exemplary sample set can be configured to have 100,000+ valid samples, and the sample proportions of the three types of areas and each typical working condition should match the regional distribution and probability of occurrence of the actual inspection scenario. After collection, the raw data also needs to be preprocessed to remove invalid data caused by sensor failures and extreme interference, ensuring the validity of each data set and the accuracy of the labels. Those skilled in the art can flexibly adjust the sample collection scale and window length according to the number of target vehicle models and the complexity of the working conditions. The core purpose of this step is to provide supervised samples that cover the entire scenario and are labeled accordingly for model training, so that the model can learn the mapping relationship between sensor features and the actual pose of the robot under different areas and working conditions, providing a data foundation for subsequent model training convergence.

[0060] Step S502 involves building the multi-branch neural network model architecture. Its core is constructing a modular network structure adapted to four types of heterogeneous sensor data, enabling feature extraction and weighted fusion calculation, ensuring the model can match dynamic weight allocation logic and output high-precision pose. The overall model adopts a three-stage architecture of multi-branch feature extraction + fusion calculation + pose output. The functions and structural designs of each module are as follows: The effective ranging feature branch is designed for the spatial correlation of multi-base station UWB ranging data. It consists of 3 layers of 1D convolutional layers and 2 layers of fully connected layers. The 1D convolutional layers extract the spatial distribution features and signal-to-noise ratio correlation features of the temporal ranging data through 5×1 convolutional kernels. After convolution, the dimension is compressed by the max pooling layer. The subsequent fully connected layers map the feature dimension to 256 dimensions and enhance the non-linear expression capability through the ReLU activation function. Finally, a regular effective ranging feature vector is output. The polarization heading feature branch adapts to the temporal stationarity of heading data. It adopts a 3-layer convolutional layer and 2-layer fully connected layer structure consistent with the effective ranging feature branch. The 1D convolutional layer captures the temporal fluctuation pattern of heading, and the fully connected layer completes the unification of feature dimensions, outputting a 256-dimensional polarization heading feature vector. At the same time, the BatchNorm layer is introduced to suppress the training fluctuation caused by data distribution shift. The angular velocity feature branch focuses on extracting rotational motion features. It also adopts a 3-layer convolutional layer and 2-layer fully connected layer architecture. The convolutional layer splits key features such as instantaneous angular velocity and cumulative rotation angle, and the fully connected layer maps to 256 dimensions to output a motion feature vector, ensuring that the feature dimensions are consistent with those of other branches. The displacement increment feature branch is designed for the trend of mileage accumulation data. Structurally, it maintains the modularity of 3 convolutional layers and 2 fully connected layers. The convolutional layers extract features such as single-step displacement amplitude and mileage change trend, and the fully connected layers complete the output of 256-dimensional feature vectors, achieving complete alignment of the feature dimensions of the four types of branches. The fusion solution layer is a fully connected fusion module with built-in feature dimension verification interface and dynamic weight adaptation interface. First, it performs dimension consistency verification on the four types of 256-dimensional feature vectors. Then, it receives the dynamic fusion weights assigned by S53 through the weight mapping interface, performs element-wise weighted summation operation of "branch feature × corresponding weight", and simultaneously introduces L2 regularization term to constrain feature redundancy. Finally, it outputs a 256-dimensional global fusion feature vector. The pose output layer consists of three fully connected network layers. The first layer compresses the 256-dimensional fused features to 64 dimensions and extracts the core associated features through the ReLU activation function. The second layer further compresses the features to 16 dimensions to enhance the mapping relationship between the features and the pose. The third layer is a linear output layer that maps the 16-dimensional features to three consecutive values, corresponding to the X-axis coordinate, Y-axis coordinate, and heading angle in the train's global coordinate system, respectively. The output interface is adapted to the subsequent pose verification logic.

[0061] The structural parameters of each branch and layer (such as the size of the convolution kernel and the number of neurons in the fully connected layer) can be fine-tuned by those skilled in the art according to the sample set size and feature complexity. The modular design ensures that the model has good scalability and adaptability. The core purpose is to realize the corresponding extraction and dynamic weighted fusion of multi-source sensing features through a structured architecture, so as to provide a stable network carrier for model training and inference.

[0062] Step S503 involves sample preprocessing, dataset partitioning, and model forward training initiation. Its core is to improve training data quality, rationally split the data distribution, and initiate the batch feature learning and prediction process of the model, providing a stable training foundation for subsequent parameter optimization. This step first performs full preprocessing on the model training sample set constructed in S501: targeted cleaning operations are performed for the four types of sensor data; effective ranging data is processed using the 3σ principle to remove abnormal jump values ​​caused by multipath interference; polarization heading data is filtered using median filtering to remove impulse noise caused by instantaneous stray light; angular velocity data is used to compensate for hardware zero-bias errors; and displacement increment data is used to correct abnormal cumulative values ​​caused by wheel idling. Subsequently, the cleaned data is standardized, using the Z-score standardization formula to uniformly map the four types of data to the [-1,1] interval, eliminating the impact of dimensional differences on training convergence; simultaneously, the temporal consistency of each sample group is verified, and linear interpolation is used to fill in a small number of missing temporal data due to sampling interruptions, ensuring the integrity of the data within the sample. After preprocessing, the sample set is randomly divided into training and validation sets according to a preset ratio (e.g., 8:2 or 7:3). During the division, the sample proportions of the three operating regions and typical working conditions in the two datasets are kept consistent with the original sample set to avoid data skew that could lead to a decrease in the model's generalization ability. Then, the training batch size is set (e.g., 64 or 128 groups / batch, which can be flexibly adjusted according to the hardware memory capacity). The training set samples are input into the multi-branch neural network model built by S502 in batches. The four types of sensor data in each batch are sent to the corresponding feature branches, and feature extraction is completed synchronously. The feature vectors output by each branch are fed into the fusion solution layer and weighted fusion is performed according to the initial weights (the median value of the S51 weight rule library can be used as the initial weights in the early stage of training). After the fused feature vectors are solved by the pose output layer, the predicted pose (including X-axis coordinates, Y-axis coordinates, and heading angle) corresponding to the batch of samples is output. Those skilled in the art can adjust the batch size and division ratio according to the sample set size. The preprocessing method can be flexibly replaced according to the type of data noise (such as replacing median filtering with wavelet filtering). The core purpose of this step is to improve the quality of training data through data cleaning and standardization, ensure the objectivity of model training and validation by reasonably dividing the dataset, and start the feature learning process of the model through batch forward propagation, so as to provide forward prediction results to support weight constraints and parameter optimization.

[0063] Step S504 is the branch weight constraint configuration step in the model training process. Its core is to set the parameter constraint boundary according to the correspondence between the type of carriage operating area and the confidence level of the heading parameter. The constraint conditions are used to regulate the update range of the fusion weight of each feature branch during model backpropagation, guide the model to learn the contribution rules of each sensor feature in accordance with the real working conditions, and realize the consistency and unity of the offline training rules, the weight adjustment rule library constructed in S51, and the online dynamic weight allocation logic in S53. This step references the pre-built weight interval standard in S51, combining three regions—the main carriage positioning area, the partially occluded non-transition area, and the windshield-cross-carriage transition area—with three confidence levels: high confidence, medium confidence, and low confidence. Upper and lower constraint thresholds are defined for the fusion weights of the effective ranging feature branch, polarization heading feature branch, angular velocity feature branch, and displacement increment feature branch, respectively. When reading a single batch of samples for model training, the region type labeled during sample acquisition and the confidence level converted from heading fluctuation classification are retrieved first, and the corresponding weight constraint interval is matched. For example, when the sample label is "main carriage positioning area + high confidence," the training value of the effective ranging branch fusion weight is limited. When the weights are 0.4–0.5, the polarization heading branch is limited to 0.3–0.35, and the sample label is the windshield-to-carriage transition zone + low confidence, the effective ranging branch weights are constrained to 0.2–0.25, and the polarization heading branch weights are constrained to 0.15–0.2. After the model updates the weight parameters of the fusion solution layer during gradient backpropagation, the system will perform boundary checks on the updated weight values. Once the parameters exceed the constraint range corresponding to the current working condition, automatic truncation is performed to force the weights to fall within the specified range. This prevents extreme distorted parameters from occurring during training, such as a single feature branch weight approaching 1 and other branch weights approaching 0, and prevents the model from overfitting to a certain type of sensor data. Those skilled in the art can make minor adjustments to the upper and lower limits of the weight constraints under each combination of working conditions based on the error performance of the newly added actual inspection samples. The core purpose of this step is to use the weight constraints bound to the working conditions to guide the model to learn the actual reliability of various types of sensor data under different environments, so that the feature contribution rules learned after the model training converges match the actual dynamic weighting requirements of navigation, and ensure that the dynamic weight allocation in the S54 online inference stage is consistent with the feature preferences of the model itself.

[0064] Step S505 is the model loss function construction, iterative optimization and convergence finalization stage. Its core is to minimize the error between the model's predicted pose and the actual pose label as the optimization objective. All network parameters are iteratively updated through standard optimization algorithms, and the model training convergence is completed by combining the validation set error judgment. Finally, a multi-branch navigation fusion solution model that can be directly deployed and applied is obtained. This step first constructs a composite loss function adapted to the pose calculation task, selecting mean squared error (MSE) as the core basic loss term. It calculates the sum of squared errors of the predicted X-axis coordinates, Y-axis coordinates, and heading angles output by the model, corresponding to the dimensions of the actual pose labels in the sample annotations. Simultaneously, it overlays the L2 regularization term and weight constraint penalty term set in S502 and S504, ensuring maximum overall pose prediction accuracy while constraining network parameter complexity and strictly adhering to the interval constraint rules for each branch weight, avoiding overfitting or abnormal weight updates. Then, it selects the industry-standard Adam adaptive gradient descent algorithm as the optimizer. Based on the predicted pose obtained from forward propagation and the loss function calculation results, it solves for the loss function with respect to all trainable parameters in the model (including the convolutional kernel weights of each feature branch, the weights of fully connected layers, etc.). The gradient values ​​of the weighted coefficients of the fusion solution layer and the network parameters of the pose output layer are used to update the model parameters batch by batch and round by round according to the gradient backpropagation mechanism, so as to gradually reduce the deviation between the predicted pose and the actual pose. During the iterative training process, after completing the full learning of the training set in each round, the validation set samples are immediately input into the model for gradient-free forward inference. The overall pose prediction error of the validation set (including the average coordinate error and the average heading angle error) is statistically analyzed, and the real-time error is compared with the preset error threshold (exemplarily set as the coordinate prediction error inside the carriage ≤ 5cm and the heading angle error ≤ 1°). If the pose prediction error of the validation set is consistently lower than the preset threshold and there is no significant fluctuation for several rounds, it is determined that the model has reached the convergence state, the training process is stopped immediately, and the current optimal parameters are saved. If the error does not meet the standard, iterative optimization continues until the convergence condition is met. Those skilled in the art can flexibly adjust hyperparameters such as loss function weights, optimizer learning rate, and preset error thresholds according to the on-site navigation accuracy requirements. The core purpose of this step is to complete the full parameter optimization of the model through a standardized supervised learning process, so that the model can accurately learn the mapping rules between multi-source sensor features and robot pose. Combined with weight constraint rules, the model is ensured to adapt to the navigation requirements of cross-carriage full working conditions, and finally a stable and reliable multi-branch navigation fusion solution model that can be directly used for S54 online inference is obtained.

[0065] It should be understood that the sequence number of each step in the above embodiments 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.

[0066] In one embodiment, a train inspection robot cross-car navigation system based on polarized light and UWB is provided. This train inspection robot cross-car navigation system based on polarized light and UWB corresponds to the train inspection robot cross-car navigation method based on polarized light and UWB in the above embodiment. The train inspection robot cross-car navigation system based on polarized light and UWB includes: The data acquisition module is used to collect multi-source sensor parameters in real time and synchronously when the inspection robot performs inspection tasks. The multi-source sensor parameters include at least wireless ranging parameters, polarization light sensing parameters, angular velocity parameters, displacement increment parameters, and geometric feature parameters of the carriage passage. The data processing module is used to preprocess the wireless ranging parameters, filter them to obtain an effective ranging set, and quantize the ranging quality parameters based on the effective ranging set. The data extraction module is used to perform deviation compensation processing on the polarization light sensing parameters based on the ranging quality parameters to obtain the compensated polarization heading parameters, and to extract the heading fluctuation characteristics and heading parameter confidence of the polarization heading parameters. The area determination module is used to determine the type of operating area where the inspection robot is currently located based on the directional fluctuation characteristics of the ranging quality parameters, polarization heading parameters, and geometric characteristic parameters of the carriage passage. The type of operating area includes the main body positioning area of ​​the carriage, the partially obscured non-transition area, and the windshield-cross-carriage transition area. The pose output module is used to perform navigation fusion calculation based on the heading parameter confidence of the operating area type and polarization heading parameters, with the effective ranging set, polarization heading parameters, angular velocity parameters and displacement increment parameters as input items. It also dynamically adjusts the fusion weight of each parameter in the input item set according to the operating area type where the inspection robot is currently located, and outputs the real-time pose of the inspection robot to complete cross-carriage navigation.

[0067] Specific limitations regarding the cross-car navigation system for a train inspection robot based on polarized light and UWB can be found in the above description of the method for cross-car navigation of a train inspection robot based on polarized light and UWB, and will not be repeated here. The modules in the aforementioned cross-car navigation system for a train inspection robot based on polarized light and UWB can be implemented entirely or partially through software, hardware, or a combination thereof. These modules can be embedded in or independent of the processor in a computer device, or stored in the memory of a computer device as software, so that the processor can call and execute the corresponding operations of each module.

[0068] In one embodiment, a computer device is provided, which may be a server, and its internal structure diagram may be as follows. Figure 2As shown, the computer device includes a processor, memory, network interface, and database connected via a system bus. The processor provides computing and control capabilities. The memory includes non-volatile storage media and internal memory. The non-volatile storage media stores the operating system, computer programs, and database. The internal memory provides the environment for the operation of the operating system and computer programs stored in the non-volatile storage media. The database is used for data storage, data processing, and data analysis. The network interface is used for communication with external terminals via a network connection. When the computer program is executed by the processor, it implements a cross-car navigation method for a train inspection robot based on polarized light and UWB.

[0069] In one embodiment, a computer device is provided, including a memory, a processor, and a computer program stored in the memory and executable on the processor. When the processor executes the computer program, it implements a method for cross-car navigation of a train inspection robot based on polarized light and UWB.

[0070] In one embodiment, a computer-readable storage medium is provided having a computer program stored thereon, which, when executed by a processor, implements a method for cross-car navigation of a train inspection robot based on polarized light and UWB.

[0071] Those skilled in the art will understand that all or part of the processes in the methods of the above embodiments can be implemented by a computer program instructing related hardware. The computer program can be stored in a non-volatile computer-readable storage medium. When executed, the computer program can include the processes of the embodiments of the above methods. Any references to memory, storage, databases, or other media used in the embodiments provided in this application can include non-volatile and / or volatile memory. Non-volatile memory may include read-only memory (ROM), programmable ROM (PROM), electrically programmable ROM (EPROM), electrically erasable programmable ROM (EEPROM), or flash memory. Volatile memory may include random access memory (RAM) or external cache memory. By way of illustration and not limitation, RAM is available in a variety of forms, such as static RAM (SRAM), dynamic RAM (DRAM), synchronous DRAM (SDRAM), dual data rate SDRAM (DDRSDRAM), enhanced SDRAM (ESDRAM), synchronous link DRAM (SLDRAM), RAMbus direct RAM (RDRAM), direct memory bus dynamic RAM (DRDRAM), and memory bus dynamic RAM (RDRAM), etc.

[0072] Those skilled in the art will clearly understand that, for the sake of convenience and brevity, the above-described division of functional units and modules is used as an example. In practical applications, the above functions can be assigned to different functional units and modules as needed, that is, the internal structure of the device can be divided into different functional units or modules to complete all or part of the functions described above.

[0073] The above-described embodiments are only used to illustrate the technical solutions of this application, and are not intended to limit them. Although this application has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of this application, and should all be included within the protection scope of this application.

Claims

1. A method for cross-car navigation of a train inspection robot based on polarized light and UWB, characterized in that, Includes the following steps: When the inspection robot performs its inspection task, it collects multi-source sensor parameters in real time and synchronously. The multi-source sensor parameters include at least wireless ranging parameters, polarization light sensing parameters, angular velocity parameters, displacement increment parameters, and geometric feature parameters of the carriage passage. The wireless ranging parameters are preprocessed to obtain an effective ranging set, and the ranging quality parameters are quantized based on the effective ranging set. Based on the ranging quality parameters, the polarization light sensing parameters are subjected to deviation compensation processing to obtain the compensated polarization heading parameters, and the heading fluctuation characteristics and heading parameter confidence of the polarization heading parameters are extracted. Based on the ranging quality parameters, the directional fluctuation characteristics of the polarization heading parameters, and the geometric characteristic parameters of the carriage passage, the current operating area type of the inspection robot is determined. The operating area type includes the main positioning area of ​​the carriage, the partially obscured non-transition area, and the windshield-cross-carriage transition area. Based on the confidence level of the heading parameters according to the operating area type and polarization heading parameters, navigation fusion calculation is performed using the effective ranging set, polarization heading parameters, angular velocity parameters and displacement increment parameters as input items. The fusion weight of each parameter in the input item set is dynamically adjusted according to the operating area type where the inspection robot is currently located, and the real-time pose of the inspection robot is output to complete cross-carriage navigation.

2. The method for cross-car navigation of a train inspection robot based on polarized light and UWB as described in claim 1, characterized in that, The inspection robot is equipped with a wireless ranging unit, a polarization light sensing unit, an inertial measurement unit, an odometer unit, and a lidar unit. When performing an inspection task, the inspection robot performs a real-time and synchronous acquisition of multi-source sensor parameters, specifically including: When the inspection robot receives the inspection task instruction and performs the inspection task along the train, it performs time synchronization and coordinate registration of each airborne unit to establish a unified spatiotemporal reference. Based on this spatiotemporal reference, it synchronously triggers each airborne unit to perform real-time and continuous data acquisition, and obtains wireless ranging parameters, polarization light sensing parameters, angular velocity parameters, displacement increment parameters and carriage passage geometric feature parameters with corresponding spatiotemporal stamps. The system monitors the wireless ranging parameters and polarization light sensing parameters collected during the inspection robot's movement in real time. When the wireless ranging parameters and / or polarization light sensing parameters are abnormal, the system marks the time period of the abnormality and caches the wireless ranging parameters and polarization light sensing parameters before the time period of the abnormality. Extract the corresponding zero-bias parameters from the real-time and continuously acquired angular velocity parameters and perform zero-bias stability verification. Remove invalid angular velocity parameters whose zero bias exceeds a preset threshold. Use the retained valid angular velocity parameters to solve for the zero-bias compensation parameters. Update and correct the zero-bias parameters of the subsequently acquired angular velocity parameters based on the zero-bias compensation parameters. The wheel slippage state is determined based on the real-time and continuously collected displacement increment parameters. When the wheel slippage condition is determined, the initial weight of the displacement increment parameter corresponding to the time stamp is pre-adjusted in the fusion solution. Based on the real-time and continuously collected geometric feature parameters of the carriage passage, the cross-sectional contour features of the carriage passage are analyzed, and the corresponding carriage section identifiers are marked according to the contour features. The carriage section identifiers are then associated and bound with the wireless ranging parameters, polarization light sensing parameters, angular velocity parameters, and displacement increment parameters collected synchronously with the same spatiotemporal reference.

3. The method for cross-car navigation of a train inspection robot based on polarized light and UWB as described in claim 2, characterized in that, The steps of preprocessing the wireless ranging parameters, filtering to obtain an effective ranging set, and quantizing the ranging quality parameters based on the effective ranging set specifically include: The abnormal occurrence period label is determined based on the preset abnormal duration judgment threshold; when it is determined to be a continuous abnormality, the wireless ranging parameters corresponding to the continuous abnormality are removed; when it is determined to be a momentary abnormality, the wireless ranging parameters cached before the abnormal occurrence period label are used for interpolation to fill in the missing data of the abnormal period. Within the wireless ranging parameters that have undergone anomaly removal and interpolation completion, the partition filtering threshold is matched according to the operating area type; Based on the partitioning screening threshold and the geometric feature parameters of the carriage passage, the wireless ranging parameters corresponding to each available base station under the same spatiotemporal reference are screened and summarized to obtain the effective ranging set; Based on the preset index weighting coefficients corresponding to the current operating area type of the inspection robot, at least one of the following is selected: effective base station number, ranging residual statistics, multipath disturbance intensity, and base station geometric configuration index. The effective ranging set is then quantified to obtain ranging quality parameters.

4. The method for cross-car navigation of a train inspection robot based on polarized light and UWB as described in claim 2, characterized in that, The steps of performing deviation compensation processing on the polarization sensing parameters based on the ranging quality parameters to obtain the compensated polarization heading parameters, and extracting the heading fluctuation characteristics and heading parameter confidence levels of the polarization heading parameters, specifically include: A preset quality pass threshold is retrieved, and the quality pass threshold is dynamically corrected according to the carriage section markings to obtain the corrected quality pass threshold. The ranging quality parameter is compared with the corrected quality qualification threshold. When the ranging quality parameter is greater than or equal to the corrected quality qualification threshold, the reference heading of the inspection robot is analyzed based on the effective ranging set. Based on the reference heading of the inspection robot and the original polarization heading parameter obtained by analyzing the original polarization light sensing parameters collected from the same spatiotemporal reference, the heading compensation deviation is calculated in real time. When the ranging quality parameter is less than the corrected quality qualification threshold, the historical heading compensation deviation pre-stored in the corresponding carriage section is retrieved. By using heading compensation deviation, the original polarization light sensing parameters collected under the same spatiotemporal reference are offset to obtain the compensated polarization heading parameters; Based on the carriage section identification, extract the heading fluctuation characteristics and heading parameter confidence levels corresponding to the polarization heading parameters after compensation for different sections.

5. The method for cross-car navigation of a train inspection robot based on polarized light and UWB as described in claim 4, characterized in that, The step of calculating the heading compensation deviation in real time when the ranging quality parameter is greater than or equal to the corrected quality qualification threshold, based on the effective ranging set, the reference heading of the inspection robot, and the original polarization heading parameter obtained by analyzing the original polarization light sensing parameters collected from the same spatiotemporal reference, specifically includes: When the ranging quality parameter is greater than or equal to the corrected quality qualification threshold, the robot pose data corresponding to each continuous sampling time is obtained by parsing the effective ranging set. Using the pose data of the inspection robot corresponding to continuous sampling time as the processing object, the motion trajectory interval of the inspection robot is segmented along the time axis by a sliding window. Abnormal intervals are filtered out within a single window, and the effective trajectory points corresponding to uniform and stable driving are retained. The effective trajectory points are fitted with lines by the least squares method, and the azimuth angle corresponding to the fitted line is used as the reference heading of the inspection robot. The original polarization heading parameters generated by the original polarization light sensing parameters under the same spatiotemporal reference are obtained. The difference between the original polarization heading parameters and the reference heading of the inspection robot is calculated as the slow-varying offset. The slow-varying offset is iteratively optimized and updated frame by frame through a preset iterative algorithm. When the slow-varying offset converges, the real-time heading compensation deviation is obtained.

6. The method for cross-car navigation of a train inspection robot based on polarized light and UWB as described in claim 1, characterized in that, The step of determining the type of operating area currently occupied by the inspection robot based on the azimuth fluctuation characteristics of the ranging quality parameters, polarization heading parameters, and geometric characteristic parameters of the carriage passage specifically includes: Based on the geometric feature parameters of the carriage passage, the cross-sectional dimensions and contour structure of the carriage are analyzed. The cross-sectional dimensions and contour structure of the carriage are matched with a preset contour standard library to obtain the area division result corresponding to the location of the inspection robot. Extract the heading fluctuation amplitude index of the polarization heading parameters, and use a preset classification threshold to quantify and classify the heading fluctuation amplitude index to obtain the fluctuation amplitude classification result. The ranging quality parameters are compared with the preset grading threshold and the corresponding quality levels are determined to obtain the ranging quality grading results. By integrating the results of regional division, fluctuation amplitude classification, and ranging quality classification, the current operating area type of the inspection robot is determined according to the preset judgment rules.

7. The method for cross-car navigation of a train inspection robot based on polarized light and UWB as described in claim 6, characterized in that, The navigation fusion calculation, based on the confidence level of the heading parameters according to the operating area type and polarization heading parameters, uses the effective ranging set, polarization heading parameters, angular velocity parameters, and displacement increment parameters as input items. It dynamically adjusts the fusion weights of each parameter in the input item set according to the current operating area type of the inspection robot, outputting the real-time pose of the inspection robot to complete the cross-carriage navigation process. Specifically, this includes: The confidence level of the heading parameter confidence is calculated using the fluctuation amplitude classification results. Based on the confidence level and the type of the current operating area of ​​the inspection robot, a weight adjustment rule base is constructed. The effective ranging set, polarization heading parameters, angular velocity parameters, and displacement increment parameters are spatiotemporally synchronized and calibrated to construct an input item set; Based on the type of the current operating area of ​​the inspection robot and the confidence level of the heading parameter, the weight adjustment rule base is retrieved to assign dynamic fusion weights to each corresponding parameter in the input item set; The set of input items with configured dynamic fusion weights is input into the trained multi-branch navigation fusion solution model for fusion solution, and the real-time pose of the inspection robot is output to complete cross-carriage navigation.

8. A train inspection robot cross-car navigation system based on polarized light and UWB, used to implement the steps of the train inspection robot cross-car navigation method based on polarized light and UWB as described in any one of claims 1-7, characterized in that, include: The data acquisition module is used to collect multi-source sensor parameters in real time and synchronously when the inspection robot performs inspection tasks. The multi-source sensor parameters include at least wireless ranging parameters, polarization light sensing parameters, angular velocity parameters, displacement increment parameters, and geometric feature parameters of the carriage passage. The data processing module is used to preprocess the wireless ranging parameters, filter them to obtain an effective ranging set, and quantize the ranging quality parameters based on the effective ranging set. The data extraction module is used to perform deviation compensation processing on the polarization light sensing parameters based on the ranging quality parameters to obtain the compensated polarization heading parameters, and to extract the heading fluctuation characteristics and heading parameter confidence of the polarization heading parameters. The area determination module is used to determine the type of operating area where the inspection robot is currently located based on the directional fluctuation characteristics of the ranging quality parameters, polarization heading parameters, and geometric characteristic parameters of the carriage passage. The type of operating area includes the main body positioning area of ​​the carriage, the partially obscured non-transition area, and the windshield-cross-carriage transition area. The pose output module is used to perform navigation fusion calculation based on the heading parameter confidence of the operating area type and polarization heading parameters, with the effective ranging set, polarization heading parameters, angular velocity parameters and displacement increment parameters as input items. It also dynamically adjusts the fusion weight of each parameter in the input item set according to the operating area type where the inspection robot is currently located, and outputs the real-time pose of the inspection robot to complete cross-carriage navigation.

9. A computer device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that, When the processor executes the computer program, it implements the steps of the train inspection robot cross-carriage navigation method based on polarized light and UWB as described in any one of claims 1-7.

10. A computer-readable storage medium storing a computer program, characterized in that, When the computer program is executed by the processor, it implements the steps of the train inspection robot cross-carriage navigation method based on polarized light and UWB as described in any one of claims 1-7.

Images