A visually guided bogie six degree of freedom assembly system

Abstract

The present application belongs to the field of mechanical engineering, and particularly relates to a visual guidance bogie six-degree-of-freedom assembly system, aiming to solve the problems of low positioning accuracy, high artificial dependence, weak multi-degree-of-freedom coordination and poor dynamic adaptability in bogie assembly. The system comprises three-dimensional visual perception, pose solution planning, multi-axis collaborative control, actuator platform and central coordination module. Point cloud data is obtained by fusing multi-view structured light, six-degree-of-freedom deviation is calculated by combining feature matching and ICP algorithm, and a collision-free motion trajectory is generated; the wheel set is precisely moved by the six-degree-of-freedom hydraulic parallel platform, and the compliant assembly is realized by cooperating with adaptive clamping and force-position hybrid sensing; the central module uniformly schedules and supports rapid model change of multiple vehicles. The system realizes full-automatic precise positioning, the positioning accuracy reaches ±0.1 millimeter, and the assembly efficiency and flexibility level are significantly improved.

Description

Technical Field

[0001] This invention belongs to the field of mechanical engineering, and specifically relates to a vision-guided six-degree-of-freedom assembly system for bogies. Background Technology

[0002] The rail transit equipment manufacturing field encompasses multiple key technology areas, including vehicles, signaling, and traction. Among these, the bogie, as a core component of rail vehicles, directly impacts the overall vehicle's operational stability and safety through its assembly precision and efficiency. Bogie assembly involves the precise coordination of multiple parts, requiring high-precision positioning and connection under complex spatial constraints, representing a typical high-end equipment integration manufacturing process. With the development of intelligent manufacturing technology, automated assembly systems are gradually replacing traditional manual operations, becoming a key means to improve assembly quality consistency and production cycle time.

[0003] Machine vision-based guidance technology is widely used in the pose measurement and path planning stages of the assembly process. It achieves real-time acquisition of the target object's three-dimensional coordinates, attitude angles, and other degrees of freedom parameters through non-contact sensing. The basic principle of this technology is to use industrial cameras to acquire images of the assembly, combine them with calibration models and feature matching algorithms to calculate the spatial transformation relationship of the component to be assembled relative to the reference coordinate system, thereby providing precise motion commands for robots or automated guided vehicles. The application of vision-guided technology significantly improves the flexibility and environmental adaptability of assembly operations.

[0004] While existing technologies, including monocular, binocular, and structured light vision systems, have been applied in local assembly scenarios, they still face multiple technical bottlenecks when performing six-DOF full-pose assembly tasks for large and complex components such as bogies: the measurement accuracy of vision systems is easily affected by changes in illumination, surface reflection, and occlusion interference, leading to unstable pose calculation results; the multi-view data fusion strategy lacks a unified spatiotemporal reference, making it difficult to guarantee the consistency and continuity of pose information during dynamic assembly; the control model for six-DOF (three translations and three rotations) coupled motion does not fully consider the nonlinear errors caused by the elastic deformation of the mechanical structure and transmission clearance, causing the actual trajectory of the end effector to deviate from the ideal path; in addition, current systems generally adopt offline programming and fixed process control modes, which cannot dynamically adjust the assembly strategy based on real-time perception feedback, and lack online compensation capabilities for assembly stress concentration problems caused by the accumulation of small deviations. The aforementioned defects are particularly prominent in the bogie assembly process, which requires high rigidity connections. They can easily lead to quality problems such as uneven distribution of bolt preload and poor fit of mating surfaces, which seriously restricts the engineering implementation of fully automated assembly lines. Therefore, it is urgent to build a new assembly system architecture with environmental perception robustness, multi-source information collaboration, and dynamic response intelligence. Summary of the Invention

[0005] The purpose of this invention is to provide a vision-guided six-degree-of-freedom (DOF) bogie assembly system to address the technical problems in the mechanical assembly process of bogies in the field of rail transit equipment manufacturing, such as insufficient positioning accuracy, high dependence on manual intervention, weak multi-DOF collaborative adjustment capability, and poor adaptability to dynamic environments. Currently, as a core running component of rail vehicles, the assembly of the bogie frame and wheelsets requires precise alignment in six degrees of freedom. Traditional processes mainly rely on rigid tooling fixtures and operator experience for positioning, which is insufficient to cope with the assembly challenges brought about by accumulated manufacturing tolerances, structural deformation, and real-time attitude deviations. With the development of intelligent manufacturing, existing technologies have attempted to introduce laser ranging or fixed visual inspection methods for assisted positioning. However, these methods are limited by the limited dimension of single-point measurement, incomplete field of view coverage, and lack of closed-loop feedback control mechanisms. They cannot achieve real-time, full-domain perception and dynamic compensation of the three-dimensional spatial pose of the assembly, resulting in low assembly efficiency, large fluctuations in the pass rate, and a lack of rapid adaptation capability to non-standard vehicle models or flexible production lines.

[0006] The technical solution of this invention includes a three-dimensional vision perception subsystem, a pose calculation and planning subsystem, a multi-axis cooperative motion control subsystem, an actuator integration platform, and a central coordination and management module. The three-dimensional vision perception subsystem consists of multiple high-resolution industrial cameras and active coded structured light projection devices, distributed around the assembly station. It continuously collects dense point cloud data of the bogie frame and wheelset component surfaces during assembly and generates a dynamic three-dimensional model with millimeter-level spatial resolution through a multi-view fusion algorithm. The pose calculation and planning subsystem receives the point cloud sequence from the three-dimensional vision perception subsystem and employs a hybrid pose estimation algorithm combining feature matching and iterative nearest point algorithms to accurately calculate the six-degree-of-freedom deviation vector of the wheelset component relative to the target frame installation reference, namely three translational components and three rotational components. Furthermore, this subsystem incorporates an assembly path optimization engine that, based on the current deviation state, mechanical constraint boundaries, and dynamic safety thresholds, generates a smooth, collision-free spatial motion trajectory from the initial pose to the target pose and decomposes it into a time-discreteized pose command sequence. The multi-axis cooperative motion control subsystem consists of six independently driven servo hydraulic cylinders and their matching proportional valve groups, pressure sensors, and displacement encoders, arranged in a spatially orthogonal layout to form a reconfigurable six-degree-of-freedom parallel motion platform. This subsystem receives a sequence of pose commands output from the pose calculation and planning subsystem. Through a feedforward-feedback composite control strategy, it adjusts the extension and retraction lengths and torques of each hydraulic cylinder in real time, ensuring that the execution platform drives the assembled wheelset to move precisely along a predetermined trajectory. The actuator integration platform is fixedly mounted on the upper part of the six-degree-of-freedom parallel motion platform and includes an adaptive clamping mechanism and a force-position hybrid interface module. The adaptive clamping mechanism is equipped with flexible contact pads and a miniature pressure array sensor, which can automatically adjust the clamping force distribution according to the shape of the wheelset axle box to avoid localized stress concentration. The force-position hybrid interface module synchronously monitors the contact force and micro-displacement response during assembly. When abnormal resistance or jamming is detected, it triggers an emergency deceleration or retraction mechanism. The central coordination and management module serves as the operational hub of the entire system, responsible for the unified scheduling of the work sequence of each subsystem, establishing a high-speed real-time communication network, and ensuring that the synchronization cycle between visual data acquisition, pose calculation, motion command issuance, and execution status feedback is less than 50 milliseconds. At the same time, this module integrates a fault diagnosis logic unit to identify and issue graded alarms in real time for working conditions such as abnormal sensor signals, delayed actuator response, or excessive path deviation.

[0007] Furthermore, the structured light projection device in the 3D vision perception subsystem employs a programmable digital micromirror device, which can dynamically adjust the frequency and phase step of the encoded pattern according to the geometric complexity of the area to be measured. In areas with drastic curvature changes, a high-density stripe pattern is activated to improve reconstruction accuracy, while in flat areas, a low-frequency pattern is switched to accelerate projection speed. As one embodiment of the invention, the pose calculation and planning subsystem, when performing feature matching, prioritizes extracting key feature point sets with high geometric stability, such as the edge of the bogie side beam weld, the center of the axle box positioning seat hole, and the brake disc mounting boss, and establishes a feature topology map corresponding to the standard CAD model. This serves as the input for initial coarse registration, significantly shortening the time required for iterative convergence. Furthermore, the assembly path optimization engine employs an improved fast expanding random tree algorithm. When constructing the search space, physical constraints such as hydraulic cylinder stroke limits, pipeline bending radii, and cable chain tension are embedded as hard boundary conditions in the node sampling process, thereby ensuring that each generated candidate path meets the actual executability requirements. As one embodiment of the present invention, the feedforward-feedback composite control strategy in the multi-axis cooperative motion control subsystem includes two parallel channels: the feedforward channel pre-calculates the target flow rate and pressure curves of each hydraulic cylinder based on the ideal motion trajectory to offset system inertia and gravity load; the feedback channel uses an adaptive gain-adjusted proportional-integral-derivative controller to correct errors based on the residual between the actual readings and set values ​​of each cylinder displacement encoder. Its proportional coefficient is dynamically updated based on the online identification results of the overall platform stiffness matrix to maintain closed-loop stability. Furthermore, the miniature pressure array sensors in the force-position hybrid interface module are distributed at a 128×128 array density on the clamping interface, with a sampling frequency set to 1 kHz. This allows them to capture the transient impact waveform at the moment of assembly contact and extract the concentrated energy frequency band through wavelet transform to determine whether metal scratching or foreign object interference exists. As another embodiment of the present invention, the central coordination and management module is equipped with edge computing nodes and deploys a lightweight neural network model to process high-frequency sensor stream data. This model, trained on a large amount of historical assembly logs, can predict potential assembly risk trends, such as progressive jamming or elastic deformation accumulation, before substantial damage occurs, and intervene in advance to adjust process parameters.

[0008] Furthermore, the entire system supports multi-model mixed-line production. Before starting a new model assembly task, the central coordination and management module automatically loads the corresponding model's process parameter package, including target pose definition, feature map template, initial preload value of hydraulic cylinder, upper and lower limit thresholds of clamping force, and weight coefficients for dedicated path planning. As one embodiment of the invention, the adaptive clamping mechanism is equipped with a replaceable chuck assembly whose shape and size match the external contour of a specific model's axle box. The replacement process is automatically completed by the robotic arm, requiring no manual intervention. Furthermore, the system features both manual teaching and automatic calibration modes. In automatic calibration mode, by driving the six-degree-of-freedom platform to execute a set of preset standard pose action sequences, combined with the reverse observation results from the vision subsystem, the platform's kinematic parameters are calibrated online to compensate for geometric mismatch errors caused by temperature drift or mechanical wear. As one embodiment of the invention, all critical communication links employ time-sensitive networking protocols, ensuring deterministic transmission delay of control commands is less than 10 microseconds and ensuring multi-axis synchronization accuracy is better than 0.05 degrees. Furthermore, the system is equipped with a remote monitoring interface, allowing engineers to access real-time 3D assembly scene reconstruction images, sensor data streams, and system health status reports through an encrypted channel, enabling cross-regional technical support and operation and maintenance decisions.

[0009] Compared with the prior art, the advantages and positive effects of the present invention are as follows: This invention, by constructing a closed-loop assembly system composed of 3D visual perception, pose calculation and planning, multi-axis collaborative control, and intelligent actuators, achieves for the first time fully automated six-degree-of-freedom precise positioning and compliant assembly of bogie wheelset components under complex spatial constraints. This completely eliminates reliance on manual experience, improving assembly positioning accuracy from ±2 mm in traditional processes to within ±0.1 mm, with a repeatability accuracy of 0.05 mm. The invention employs a hybrid pose estimation algorithm guided by multi-view structured light fusion imaging and feature topology mapping, effectively overcoming the limitations of single-sensor field of view and feature loss. Even under strong ambient light interference or partial occlusion conditions, it maintains stable measurement performance, significantly enhancing system robustness. The feedforward-feedback composite control strategy designed in this invention... The invention incorporates an online hydraulic cylinder parameter identification mechanism, enabling the six-degree-of-freedom parallel platform to maintain excellent dynamic response characteristics and trajectory tracking accuracy even when carrying ton-level loads. This solves the bottleneck of insufficient thrust and severe overheating in traditional electric drive systems under heavy-load conditions. Furthermore, the invention introduces a force-position hybrid perception and risk prediction model, endowing the assembly process with human-like "tactile" and "predictive" capabilities. This allows for proactive avoidance of hard collisions during the contact phase, achieving "trial-and-error" compliant assembly under micron-level force control, significantly reducing the risk of damage to precision components. Finally, the invention supports rapid model changeover and parameterized process configuration for multiple vehicle types. Combined with automatic chuck replacement and online calibration functions, it reduces the entire line changeover time from several hours to less than 30 minutes, fully meeting the needs of flexible and intelligent production in modern rail transit equipment. Attached Figure Description

[0010] Figure 1 This is a schematic diagram of the overall technical solution architecture proposed in this invention; Figure 2 This is a schematic diagram of the core principle framework of the hybrid pose estimation algorithm based on multi-view structured light fusion and feature topology map guidance in this invention. Detailed Implementation

[0011] Please refer to Figure 1 and Figure 2 This invention provides a vision-guided six-degree-of-freedom bogie assembly system for fully automated, high-precision, and compliant assembly of bogie frames and wheelset components in a three-dimensional spatial environment for rail transit vehicles. The system integrates three-dimensional vision perception, pose calculation and path planning, multi-axis cooperative motion control, an actuator integration platform, and a central coordination and management module to construct an intelligent assembly system with real-time closed-loop feedback capabilities. This addresses the technical bottlenecks of traditional assembly processes, such as reliance on manual experience, low positioning accuracy, poor dynamic adaptability, difficulties in multi-degree-of-freedom linkage, and insufficient flexibility in model changeover.

[0012] The system's overall operation begins with the receipt of the assembly task initiation command. The central coordination and management module then initializes the system state, loads the corresponding vehicle model's process parameter package, and activates the 3D vision perception subsystem to enter continuous data acquisition mode. The 3D vision perception subsystem consists of six high-resolution industrial cameras distributed around the assembly station and three sets of active-encoding structured light projection devices. These devices form a spatially distributed sensing network after precise calibration. Each industrial camera uses a global shutter CMOS image sensor with a resolution of 2048×2048 pixels and a frame rate of 60 frames per second to ensure effective capture of high-speed dynamic scenes. The structured light projection device is based on a programmable digital micromirror device (DMD) chip, supporting dynamic pattern generation and automatically switching the fringe frequency and phase step size according to the geometric characteristics of the area to be measured. In flat surface areas, the system uses low-frequency sinusoidal fringe projection, completing one depth map reconstruction with a single three-phase phase shift, and the projection period is 16 milliseconds. In areas with drastic curvature changes or deep hole structures, such as around the axle box mounting base, the system automatically switches to high-frequency dense fringe mode and increases the phase step to 5 steps to improve local point cloud density and reconstruction accuracy. In this case, the single-frame acquisition time is extended to 28 milliseconds. All cameras and projectors are equipped with narrowband filters and synchronous trigger interfaces, and a hardware-level clock synchronization mechanism ensures strict alignment of multi-view data in the time dimension.

[0013] The raw image sequences acquired by the 3D vision perception subsystem are decoded and reconstructed via a dedicated image processing pipeline. Each camera-projector pair first performs a phase unrolling operation to convert the wrapped phase map into absolute phase values, and then calculates the corresponding spatial point coordinates using triangulation principles. To eliminate blind spots in single-view measurements, the system employs a multi-view point cloud registration algorithm to achieve full-domain 3D model fusion. In the initial coarse registration stage, highly stable geometric features commonly visible from various views are extracted, including the edges of weld seams on the frame side beams, the center point of the brake disc mounting boss, and the intersection of the traction rod seat hole axes, identifying no fewer than 12 feature point groups. The positional information of these feature points is fitted to their theoretical coordinates in the standard CAD model using the least squares method, establishing an affine transformation matrix as initial input. Subsequently, the Iterative Closest Point (ICP) algorithm is executed for fine registration. During the registration process, a point-to-plane distance metric is introduced, and a robust kernel function is used to suppress outlier interference. The calculation terminates when the root mean square error between two adjacent iterations decreases by less than 0.01 mm or the total number of iterations reaches 50. The final generated fused point cloud data has a spatial resolution better than 0.1 mm, fully covers more than 98% of the outer surface area of ​​the frame and wheelset components, and is continuously updated at a rate of 20 frames per second, and transmitted to the pose calculation and planning subsystem.

[0014] The pose estimation and planning subsystem receives dynamic point cloud sequences from the 3D vision perception subsystem and executes a six-degree-of-freedom relative pose estimation algorithm. This subsystem incorporates a dual-channel parallel processing architecture: the first channel performs coarse pose estimation based on feature topology maps, and the second channel performs fine pose optimization based on dense point cloud matching. The results of both are fused to output the final deviation vector. In the coarse pose estimation channel, the system calls a target vehicle model feature topology map template pre-stored in the database. This template records the design coordinates of no fewer than 20 key assembly reference points and their topological connections, such as the spatial vector between the center of the axle box positioning cone and the reference hole of the frame beam. After downsampling, the actual acquired point cloud is used for feature extraction using the Fast Point Feature Histogram (FPFH) descriptor to search for the most similar set of feature points to the template. The matching process uses the Random Consistent Sampling (RANSAC) algorithm to eliminate mismatches. A successful match is considered achieved when the proportion of interior points exceeds 70% and the matching residual is less than 0.5 mm. The rigid body transformation matrix is ​​then solved to obtain the target initial pose estimate of the wheelset assembly relative to the frame, which includes three translational components ΔX, ΔY, and ΔZ and three rotational components Δα, Δβ, and Δγ, where the rotational order follows the ZYX Euler angle convention.

[0015] In the fine pose optimization channel, the system uses the coarse estimation result as the initial value to drive the improved iterative nearest-point algorithm for global optimization. The algorithm adopts a multi-scale strategy. In the initial stage, the point cloud is downsampled by voxelization, with the voxel size set to 2 mm to accelerate the nearest-point search. As the iteration progresses, the voxel size is gradually reduced to 0.2 mm to improve the matching accuracy. To enhance the convergence ability of the algorithm in weak texture regions, a normal vector consistency constraint term is introduced, defining that the validity of point pair matching depends not only on the Euclidean distance but also on the fact that the average normal angle between the two points within a 5 mm radius neighborhood is less than 15 degrees. The optimization objective function is as follows: in, This represents the rigid body transformation matrix to be solved; and The first point in the source cloud The coordinates of each point and its unit normal vector; and This corresponds to its point and normal vector in the target point cloud; and These are the weighting coefficients for the distance and normal terms, which in this embodiment are 1.0 and 0.8, respectively. To effectively match the total number of point pairs, this composite objective function significantly improves the algorithm's tolerance to non-ideal contact surfaces (such as oil stains and rust) while reducing registration errors, enabling pose calculation to maintain stable output even under harsh conditions.

[0016] The pose calculation and planning subsystem further integrates an assembly path optimization engine to generate safe, smooth, and executable motion trajectories from the current pose to the target pose. The engine employs an improved Fast Expanding Random Tree (RRT*) algorithm to construct the search space. Its core innovation lies in explicitly embedding physical constraints into the node sampling and path pruning processes. The search space is defined as a six-dimensional pose space, where each node represents a possible intermediate pose. During the node expansion phase, the system starts from the current configuration and randomly generates a directional increment Δq, but imposes multiple boundary constraints on this increment: the maximum stroke limit of the hydraulic cylinder is converted into the maximum displacement range of the platform end in the X, Y, and Z directions of ±150 mm, ±120 mm, and ±100 mm, respectively; the rotational degree of freedom is limited by the hinge joint activity angle, with a rotation range of ±8 degrees around the X and Y axes and ±12 degrees around the Z axis; in addition, the maximum allowable bending radius of the cable carrier is 300 mm, and the minimum bending radius of the pipeline is 250 mm. These geometric constraints are modeled as virtual obstacles that cannot be crossed, and any candidate node that may cause excessive stretching of the cable carrier or bending of the pipeline is directly eliminated.

[0017] The path cost function comprehensively considers three indicators: motion smoothness, energy efficiency, and obstacle avoidance margin. in, and These are the generalized velocity and acceleration vectors, respectively; This represents the reciprocal of the distance between a waypoint and the nearest obstacle, used to penalize trajectories that approach dangerous areas. , , The weighting coefficients are set to 0.6, 0.3, and 0.1 respectively in this embodiment to prioritize motion stability. The algorithm performs a progressive rewiring operation after each new node is added, checking for a better path to connect the node, until no cost reduction occurs after 10 consecutive iterations. The final output trajectory is a set of time-discrete pose command sequences with a sampling interval of 10 milliseconds, containing no fewer than 500 discrete path points to ensure the trajectory is continuous and differentiable.

[0018] The multi-axis cooperative motion control subsystem receives the pose command sequence output by the pose calculation and planning subsystem, driving a six-DOF parallel motion platform to accurately track a predetermined trajectory. This subsystem consists of six servo hydraulic cylinders arranged in a Stewart platform structure, with the upper and lower platforms connected by Hooke's joints and ball joints, forming a spatial six-DOF controllable mechanism. Each hydraulic cylinder has a rated thrust of 200 kN and a maximum stroke of 300 mm, equipped with a magnetostrictive displacement sensor with a resolution better than 1 μm and a repeatability of 0.5 μm; it also features a piezoresistive pressure sensor with a range of 0 to 40 MPa and an accuracy class of 0.2%FS, used for real-time monitoring of the cylinder's oil pressure. The proportional directional valve assembly uses a pilot-operated electro-hydraulic proportional control valve with a response time of less than 20 milliseconds and a flow gain linearity better than 95%, working in conjunction with a high-performance servo amplifier to achieve closed-loop flow regulation.

[0019] The control system employs a feedforward-feedback composite control strategy, comprising two parallel operating channels. The feedforward channel pre-calculates the target flow and pressure curves required for each hydraulic cylinder based on the ideal trajectory. This is given by the desired motion trajectory at the end of the platform. The desired length of each cylinder is calculated using an inverse kinematics model. Then, taking the derivative gives the target speed. With acceleration Based on the system dynamics model, the feedforward controller outputs a compensation signal: in, For feedforward compensation signal, The second derivative of the desired trajectory. The first derivative of the desired trajectory. The equivalent mass matrix of the hydraulic cylinder. For the Coriolis and centrifugal force terms, The gravity load term, along with all three factors, was identified through an offline system. This feedforward signal is used to counteract the effects of inertial force, gravity, and coupling force, reducing the burden on the feedback loop.

[0020] The feedback channel is based on the actual readings of the displacement sensors for each cylinder. With set value residual The correction signal is generated using a proportional-integral-derivative (PID) controller with adaptive gain adjustment. in, For feedback correction signal, For positional error, This is the integral gain coefficient. The differential gain coefficient and the proportional coefficient are given. Instead of being a fixed constant, the stiffness parameters are dynamically updated based on the online identification results of the overall platform stiffness matrix. The system injects a small excitation signal every 50 milliseconds, measures the platform end response without affecting the main trajectory tracking, and uses the recursive least squares (RLS) method to identify the current stiffness parameters online. and with nominal stiffness Compare, adjust This is to maintain a constant damping ratio in the closed-loop system and prevent oscillation instability caused by load changes or structural loosening. The total control output is... After D / A conversion, it drives the proportional valve to achieve high-precision trajectory tracking.

[0021] The actuator integration platform is mounted on top of a six-degree-of-freedom parallel motion platform, directly bearing and controlling the mounted wheelset assembly. The platform body is a high-strength aluminum alloy casting structure, weighing 85 kg, with a static load capacity of no less than 10 tons and dynamic impact overload resistance up to 3 times the rated load. The platform's front end integrates an adaptive clamping mechanism. Its clamp body is made of titanium alloy, with an embedded flexible silicone contact pad, 12 mm thick, with a Shore hardness of 60A, capable of elastic deformation upon contact, uniformly distributing clamping stress. Symmetrically arranged miniature pressure array sensors are positioned on both sides of the clamp, each side having a 128×128 element array, a total area of ​​64 square centimeters, a spatial resolution of 0.5 mm, and a sampling frequency of 1 kHz, enabling real-time capture of the pressure distribution field on the clamping interface. Sensor data, after compression and encoding, is uploaded to the central coordination and management module via gigabit Ethernet for evaluating clamping safety and contact consistency.

[0022] The adaptive clamping mechanism also features a force-position hybrid interface, enabling simultaneous monitoring of contact force and micro-displacement response during assembly. When the wheelset journal approaches the frame mounting hole, the system switches to compliant control mode, setting the maximum allowable contact force threshold to 5 kN. If the normal force measured in any direction exceeds this limit, an emergency deceleration mechanism is immediately triggered, reducing the propulsion speed from the original 10 mm / s to 1 mm / s. If the force persists for more than 3 seconds without release, a retraction action is executed, retreating 5 mm and readjusting the posture before reassembly. In the initial contact phase, the system activates transient force waveform analysis, performing wavelet transform processing on the impact signal captured by the pressure array sensor. The Daubechies 4 wavelet basis function is used to decompose the signal to the 5th level, extracting the amplitude characteristics of the energy concentration frequency band. If a spike pulse with a frequency higher than 8 kHz and a duration of less than 0.5 milliseconds is detected, it is determined to be a metal scratching event. The system automatically records the location of the event and notifies maintenance personnel to check the wear of the guide structure. If a vibration signal with a frequency band concentrated in the range of 2 kHz to 4 kHz and exhibiting periodic fluctuations is found, it indicates that there may be foreign object interference. The assembly process is suspended and a cleaning procedure is initiated.

[0023] The central coordination and management module, serving as the operational hub of the entire system, is deployed within an industrial-grade server rack. It features dual Intel Xeon processors with a 3.2 GHz clock speed, 128 gigabytes of memory, and runs a real-time operating system with a kernel real-time clock interrupt cycle of 1 millisecond. The module constructs a high-speed communication backbone network via a Time-Sensitive Network (TSN) switch. All subsystems are connected to the same TSN domain, employing the IEEE 802.1Qbv time-aware shaping mechanism to allocate dedicated time windows for control command streams, ensuring end-to-end transmission latency is consistently below 10 microseconds and jitter is less than 1 microsecond. The overall system control cycle is set to 50 milliseconds, with 20 milliseconds for visual data acquisition, 15 milliseconds for pose calculation, and 15 milliseconds for control command issuance and execution feedback, meeting stringent real-time requirements.

[0024] The central coordination and management module incorporates a fault diagnosis logic unit to continuously monitor the operational status of each subsystem. The diagnostic logic employs a tiered alarm mechanism, defining three levels of anomaly response strategies. Level 1 warning addresses minor sensor signal drift, such as camera temperature drift causing an overall point cloud offset between 0.2 mm and 0.5 mm; the system automatically initiates an online calibration program to compensate. Level 2 alarms apply to actuator response lag or path deviation exceeding limits, such as a hydraulic cylinder response delay exceeding 5 milliseconds or a cumulative trajectory tracking error reaching 1 mm. In this case, the system pauses the current action, switches to a safety hold mode, and awaits manual confirmation before deciding whether to continue. Level 3 emergency shutdown handles severe faults, such as hydraulic leaks, communication interruptions, or collision detection triggers; the system immediately cuts off the power source, locks all actuators, and activates audible and visual alarms. All fault events generate structured logs, including timestamps, fault codes, snapshots of associated parameters, and suggested handling measures, stored in a solid-state drive array, with a retention period of no less than 5 years.

[0025] To further enhance the system's intelligence, the central coordination and management module is configured with edge computing nodes, deploying a lightweight convolutional neural network (CNN) model for high-frequency sensor stream data analysis. This model uses a variant of MobileNetV3, with parameters compressed to 1.2 million, enabling inference speeds of 100 frames per second on embedded GPUs. The model input consists of a sequence of pressure array images acquired over the past 10 seconds, six-axis torque sensor readings, and hydraulic cylinder pressure fluctuation curves. The output is an assembly risk trend score ranging from 0 to 100. The model training data comes from 24,000 real assembly records accumulated over the past three years, covering various working conditions including normal completion, minor jamming, severe blockage, and component damage. A transfer learning strategy is employed during training, first pre-training the feature extraction layer on a publicly available mechanical fault dataset, and then fine-tuning the classification head using data from this field. Validation shows that the model can predict progressive jamming trends at least 8 seconds before substantial damage occurs, achieving an accuracy of 93.7% and a recall of 89.2%, effectively supporting preventative intervention.

[0026] The system supports multi-model mixed-line production and can switch between different bogie models without stopping the system. When a new model assembly task is issued, the central coordination and management module automatically downloads the corresponding process parameter package from the cloud database. This package includes at least 156 configuration parameters, such as: target pose definition file (containing 6 degrees of freedom design values), feature topology map template, initial preload value of hydraulic cylinders (used to eliminate mechanical backlash), upper and lower limit thresholds of clamping force (set according to axle box material), and dedicated path planning weight coefficients (for optimizing obstacle avoidance strategies for specific structures). After the parameters are loaded, the system automatically performs an integrity check, comparing key parameters such as maximum stroke and rotation angle to see if they exceed the platform's physical limits. If conflicts are found, loading is rejected and an error code is reported.

[0027] The adaptive gripping mechanism is equipped with replaceable gripper assemblies, offering six specifications to accommodate the axle box shapes of CRH series, Fuxing bullet trains, metro types A and B, trams, and light rail vehicles. Gripper replacement is automated by a six-axis industrial robot with a quick-change flange interface at the robot's end effector. The gripper body features a positioning pin and a pneumatic locking mechanism. The replacement process is as follows: the central coordination and management module sends a replacement command; the robot moves to the old gripper position, performs an unlocking action, removes it, and places it on the designated rack; then, it picks up the new gripper, aligns it with the positioning hole with visual assistance, inserts it, and locks it in place. The entire process takes no more than 90 seconds. After replacement, the system automatically triggers a no-load test cycle to verify the uniformity of the gripping force distribution and the normality of sensor communication.

[0028] The system features both manual teaching and automatic calibration modes. In automatic calibration mode, the central coordination and management module instructs the six-DOF platform to execute a set of preset standard pose sequences, encompassing 21 typical attitude points covering the entire workspace boundary and central area. During the 2-second dwell time at each attitude point, the 3D vision perception subsystem performs reverse observation of the platform's end effector from the outside, acquiring its actual spatial coordinates. The measured values ​​are compared with the calculated kinematics of the platform to construct an error mapping field. A spatial compensation lookup table is generated using the radial basis function (RBF) interpolation method for subsequent geometric error correction during operation. This online calibration process is executed automatically once per shift and can also be manually triggered by the operator at any time. It effectively compensates for geometric mismatches caused by temperature changes (the working environment temperature difference can reach ±15 degrees Celsius) or mechanical wear due to long-term operation, ensuring an absolute positioning accuracy better than 0.1 mm under long-term operation.

[0029] All critical communication links employ Time-Sensitive Networking Protocol (TSP), with the physical layer based on shielded twisted-pair Cat6A cable, supporting a transmission rate of 1 gigabit per second. TSN switches are configured with a bandwidth reservation policy, allocating dedicated virtual channels for core control flow, limiting bandwidth utilization to 70%, with the remaining bandwidth shared for non-real-time services such as video streaming and diagnostic logs. The network topology uses a dual-ring redundancy structure; if any link fails, path switching can be completed within 50 milliseconds, ensuring continuous system operation. For network security, all remote access must be encrypted via IPSec tunnels, and authentication uses certificate-based two-way authentication to prevent unauthorized access.

[0030] The system is equipped with a remote monitoring interface, allowing authorized engineers to access real-time 3D assembly scene reconstruction images, sensor data streams, and system health status reports via an encrypted channel. The monitoring client software supports simultaneous display of multiple views: the left window presents a 3D assembly animation rendered with fused point cloud data at a refresh rate of 30 frames per second; the upper right area displays six-degree-of-freedom deviation curves and time-series force control graphs; the lower right area displays real-time trend graphs of hydraulic system pressure, temperature, and flow. Users can drag and drop data snapshots at any historical moment along the timeline, and the system supports exporting raw data in CSV format for offline analysis. This functionality enables cross-regional technical support and operational decision-making, making it particularly suitable for remote assistance scenarios at overseas project sites.

[0031] In practical applications, the standard procedure for completing a full bogie wheelset assembly operation using this system is as follows: First, the wheelset assembly is transported to the assembly station via a conveyor line, and a positioning fixture initially fixes it. After the central coordination and management module confirms the workpiece's arrival signal, it initiates the 3D vision perception subsystem to perform the first scan, acquiring the initial pose deviation. Based on this, the pose calculation and planning subsystem generates the first coarse adjustment trajectory, and the multi-axis collaborative motion control subsystem drives the platform to quickly approach the target area, reducing the deviation from the initial ±50 mm to within ±5 mm. Subsequently, it switches to fine adjustment mode, enabling higher frequency visual sampling and a more refined force control strategy to gradually complete the final alignment. When all six degrees of freedom deviations enter the ±0.1 mm tolerance zone, the system issues an assembly ready signal and initiates the pressing process. Throughout the process, the force-position hybrid interface module continuously monitors the contact status, and immediately intervenes to adjust once abnormal resistance is detected, avoiding damage caused by hard pressing. The entire process is 98% automated, the average assembly time per session is reduced from 45 minutes in the traditional manual method to 12 minutes, the positioning accuracy is improved by two orders of magnitude, and the first-pass yield rate of products increases from 87% to 99.6%.

[0032] This embodiment constructs an intelligent assembly system with closed-loop capabilities of perception, decision-making, execution, and feedback through the deep integration and collaborative operation of the aforementioned subsystems. The 3D vision perception subsystem provides full-domain, high-resolution, dynamic point cloud data, laying the data foundation for accurate pose calculation; the pose calculation and planning subsystem, combining a hybrid algorithm of feature topology guidance and dense point cloud optimization, significantly improves the robustness of pose estimation under complex working conditions; the multi-axis cooperative motion control subsystem, with feedforward-feedback composite control and an online parameter identification mechanism, ensures high dynamic response and trajectory tracking accuracy under heavy load conditions; the actuator integration platform endows the system with human-like compliant assembly capabilities, effectively preventing damage to precision components; the central coordination and management module, through high-speed deterministic networks and intelligent diagnostic mechanisms, ensures coordinated, consistent, safe, and reliable operation of the entire system. The entire system achieves a fundamental shift from "experience-driven" to "data-driven," fully meeting the intelligent manufacturing requirements of modern rail transit equipment for high precision, high efficiency, high flexibility, and high reliability.

[0033] In existing technologies, bogie assembly mainly relies on large rigid positioning fixtures and visual judgment by operators for adjustments. Assembly accuracy is greatly affected by human factors, repeatability is poor, and it is difficult to cope with the inevitable manufacturing tolerance fluctuations in mass production. Some advanced production lines have attempted to introduce laser trackers or monocular vision-assisted positioning, but due to the limited dimensions of single-point or multi-point measurement, complete six-degree-of-freedom state information cannot be obtained, still requiring significant manual intervention for attitude fine-tuning. Furthermore, existing equipment generally lacks closed-loop control capabilities, unable to dynamically correct motion trajectories based on real-time deviations, leading to unpredictable and high-risk assembly processes. In contrast, this invention, by constructing a multi-view structured light fusion three-dimensional imaging system, achieves full-field, continuous, and high-density data acquisition of the assembly surface, fundamentally overcoming the technical shortcomings of single-sensor limited field of view and susceptibility to occlusion. Combined with a hybrid pose estimation algorithm guided by feature topology maps, stable measurement performance can be maintained even under conditions of missing local features or changing ambient lighting, significantly improving system availability.

[0034] Furthermore, this invention incorporates physical constraints such as hydraulic cylinder stroke, pipeline bending, and cable tension as hard boundary conditions into the node sampling process of the RRT* algorithm during the path planning stage. This ensures that each generated motion trajectory naturally meets the executability requirements, avoiding the repeated iterations and failure risks associated with the "plan first, verify later" approach in traditional methods. This design enables the system to automatically generate safe, efficient, and smooth assembly paths under complex spatial constraints, greatly improving the feasibility and reliability of automated assembly.

[0035] At the control level, the feedforward-feedback composite control strategy proposed in this invention, combined with an online identification mechanism for hydraulic cylinder parameters, solves the closed-loop stability problem of heavy-duty parallel platforms under varying load and stiffness conditions. Traditional electric drive systems, limited by motor power density, are prone to insufficient thrust and severe overheating under ton-level loads. While hydraulic drives offer the advantage of high power density, they face challenges such as nonlinear response and time-varying parameters. This solution actively compensates for inertial and gravitational loads through a feedforward channel, significantly reducing the adjustment pressure on the feedback loop. Simultaneously, it dynamically adjusts the PID gain using online identification results, enabling the control system to adapt to changes in platform stiffness and maintain optimal damping characteristics, thus ensuring both high precision and dynamic response speed.

[0036] More importantly, this invention is the first to introduce a force-position hybrid sensing and risk prediction model into the bogie assembly process, endowing the machine system with "tactile" and "predictive" capabilities. Traditional assembly often adopts a "blind pressure" strategy, which is highly susceptible to component damage due to minor misalignments. This system monitors the stress distribution at the clamping interface in real time through a high-density pressure array sensor. Combined with transient force waveform analysis technology, it can identify potential risks such as metal scratches and foreign object interference at the moment of contact and immediately initiate deceleration, retraction, or attitude readjustment mechanisms, achieving truly compliant assembly. The lightweight neural network model deployed on edge computing nodes further enhances the system's foresight, enabling it to mine early anomaly patterns from massive sensor data and issue warnings before faults manifest, achieving a leap from "post-event processing" to "pre-event prevention."

[0037] At the production organization level, this invention supports rapid changeover for multiple vehicle models and parameterized process configuration. Combined with automatic chuck replacement and online calibration functions, it reduces the changeover time for the entire production line from several hours in the traditional method to less than 30 minutes. This capability is of great significance to the growing demand for customized, small-batch, and multi-variety production in the current rail transit equipment manufacturing field, significantly improving production line utilization and market responsiveness.

[0038] In summary, this invention, through systematic technological innovation, overcomes multiple bottlenecks in traditional assembly technology regarding precision, efficiency, flexibility, and reliability, constructing a complete, autonomous, and controllable intelligent assembly solution. The implementation of this system not only significantly improves product quality and production efficiency but also promotes the transformation and upgrading of rail transit equipment manufacturing towards digitalization, networking, and intelligence, demonstrating significant technological advancement and broad application prospects.

Claims

1. A vision-guided six-degree-of-freedom bogie assembly system, characterized in that, include: The 3D vision perception subsystem is used to continuously collect point cloud data of the surface of the bogie frame and wheelset components during the assembly process, and generate dynamic 3D models through multi-view fusion algorithms. The pose calculation and planning subsystem is used to receive the dynamic three-dimensional model, and adopt a hybrid pose estimation algorithm based on feature matching and iterative nearest point algorithm to calculate the six-degree-of-freedom deviation vector of the wheelset assembly relative to the target frame installation reference. Based on the current deviation state, mechanical constraint boundary and dynamic safety threshold, it generates the spatial motion trajectory from the initial pose to the target pose and decomposes it into a time-discrete pose command sequence. The multi-axis cooperative motion control subsystem consists of a six-degree-of-freedom parallel motion platform with a spatial orthogonal layout composed of six independently driven servo hydraulic cylinders. It is used to receive the posture command sequence and adjust the extension and retraction length of each hydraulic cylinder in real time through a feedforward-feedback composite control strategy to drive the execution platform to move the mounted wheelset precisely along a predetermined trajectory. An actuator integration platform is fixedly installed on the upper part of the six-degree-of-freedom parallel motion platform. It includes an adaptive clamping mechanism and a force-position hybrid interface module. The adaptive clamping mechanism is used to automatically adjust the clamping force distribution according to the shape of the wheelset axle box. The force-position hybrid interface module is used to synchronously monitor the contact force and micro-displacement response during the assembly process, and trigger an emergency deceleration or retraction mechanism when abnormal resistance or jamming is detected. The central coordination and management module serves as the operational hub of the entire system, used to uniformly schedule the working sequence of each subsystem.

2. The vision-guided six-degree-of-freedom bogie assembly system according to claim 1, characterized in that, The structured light projection device in the three-dimensional vision perception subsystem uses a programmable digital micromirror device, which can dynamically adjust the frequency and phase step of the encoded pattern according to the geometric complexity of the area to be measured.

3. The vision-guided six-degree-of-freedom bogie assembly system according to claim 2, characterized in that, When performing feature matching, the pose calculation and planning subsystem prioritizes extracting key feature point sets such as the edge of the bogie side beam weld, the center of the axle box positioning seat hole, and the brake disc mounting boss, and establishes a feature topology map corresponding to the standard CAD model, which serves as the input for the initial coarse registration.

4. The vision-guided six-degree-of-freedom bogie assembly system according to claim 3, characterized in that, The assembly path optimization engine in the pose calculation and planning subsystem adopts an improved fast expanding random tree algorithm. When constructing the search space, the physical constraints of hydraulic cylinder stroke limit, pipeline bending radius and cable drag chain tension are embedded as hard boundary conditions into the node sampling process.

5. The vision-guided six-degree-of-freedom bogie assembly system according to claim 4, characterized in that, The feedforward-feedback composite control strategy in the multi-axis cooperative motion control subsystem includes two parallel channels: the feedforward channel pre-calculates the target flow rate and pressure curves of each hydraulic cylinder based on the ideal motion trajectory to offset the system's inertia and gravity load; the feedback channel uses an adaptive gain-adjusting proportional-integral-derivative controller to correct errors based on the residual between the actual readings and set values ​​of the displacement encoders of each cylinder.

6. The vision-guided six-degree-of-freedom bogie assembly system according to claim 5, characterized in that, The proportional coefficient in the feedback channel is dynamically updated based on the online identification results of the overall platform stiffness matrix to maintain closed-loop stability.

7. The vision-guided six-degree-of-freedom bogie assembly system according to claim 6, characterized in that, The miniature pressure array sensors in the force-position hybrid interface module are distributed on the clamping interface to capture the transient impact waveform at the moment of assembly contact, and extract the concentrated energy frequency band through wavelet transform to determine whether there is metal scratching or foreign object interference.

8. The vision-guided six-degree-of-freedom bogie assembly system according to claim 7, characterized in that, The central coordination and management module is equipped with edge computing nodes and deploys a lightweight neural network model trained on historical assembly logs to process high-frequency sensor stream data to predict potential assembly risk trends and intervene in advance to adjust process parameters.

9. The vision-guided six-degree-of-freedom bogie assembly system according to claim 8, characterized in that, The adaptive clamping mechanism is equipped with a replaceable chuck assembly whose shape and size match the external contour of a specific model of axle box, and the replacement process is automatically completed by a robotic arm.

10. The vision-guided six-degree-of-freedom bogie assembly system according to claim 9, characterized in that, The system has an automatic calibration mode. In this mode, the six-degree-of-freedom platform is driven to execute a preset standard pose sequence, and the kinematic parameters of the platform itself are calibrated online in combination with the reverse observation results of the three-dimensional vision perception subsystem to compensate for geometric mismatch errors caused by temperature drift or mechanical wear.

Images