Railway vacuum decontamination operation robot intelligent control method and system

Abstract

The application discloses a railway vacuum decontamination operation robot intelligent control method and system, relates to the technical field of robot control, obtains decontamination operation plans, station environment data, robot state data, interface pose data, vacuum working condition data, energy margin data and hidden danger perception data, performs time sequence alignment, dimensionless mapping and weight fusion on different source data, obtains an operation adaptation state, determines an operation task sequence, a driving path, a docking control boundary, a safety response level and an energy return condition according to the state, and controls the robot to complete autonomous navigation, a sewage outlet identification, end compliance docking, valve linkage, vacuum decontamination, abnormal disposal and maintenance records, and the system comprises a cloud management and control layer, an edge control layer and a robot execution layer, and is suitable for unmanned operation of train vacuum decontamination in railway passenger stations, motor train depots and vehicle depots.

Description

Technical Field

[0001] This invention relates to the field of robot control technology, specifically to an intelligent control method and system for a railway vacuum sewage unloading robot. Background Technology

[0002] Railway passenger stations, EMU depots and rolling stock depots need to complete train vacuum sewage unloading operations within the preparation window. Existing fixed or mobile vacuum sewage unloading equipment usually relies on manual labor to complete hose pulling, interface docking, valve opening and closing and operation confirmation. Personnel need to be close to the bottom of the train and the sewage tank interface. Due to the narrow working space, sewage corrosion, odor leakage and differences in train formation, the labor intensity and safety risks are high.

[0003] While existing wastewater unloading robots can perform some movement or docking actions, task scheduling, station navigation, interface identification, smooth docking, vacuum condition monitoring, energy return, and maintenance decisions are mostly set up in a decentralized manner. There is a lack of unified linkage between the robot's status and train arrival and departure plans, work locations, safety zones, and equipment health information. When temporary obstacles occur, it is difficult to adjust the task sequence and handling strategies according to the overall situation on site, resulting in insufficient continuity, traceability, and safety closed-loop capabilities of unmanned operations. Summary of the Invention

[0004] The purpose of this invention is to provide an intelligent control method and system for a railway vacuum sewage unloading robot, so as to solve the problems in the background technology.

[0005] To achieve the above objectives, the present invention provides the following technical solution: an intelligent control method for a railway vacuum sewage unloading robot, comprising:

[0006] After acquiring multi-source data, mapping and weight fusion are performed to obtain the job adaptation status;

[0007] Determine the task sequence, travel path, docking control boundary, safety response level, and energy return conditions based on the operational adaptation status.

[0008] According to the work task sequence, the robot is controlled to autonomously navigate to the target sewage unloading point, identify the train's sewage outlet, and perform a smooth docking at the end.

[0009] Multimodal data is monitored during the unloading process, and the operational adaptation status is updated based on the monitoring results to adjust subsequent operations.

[0010] As a preferred implementation method, the multi-source data includes unloading operation plan, station environment data, robot status data, interface pose data, vacuum condition data, energy reserve data, and hazard perception data;

[0011] Multimodal data includes vacuum negative pressure, pipeline pressure, flow rate, docking force, valve status, and safe zone status.

[0012] As a preferred embodiment, determining the sequence of tasks includes:

[0013] Based on train formation, unloading points, operation time window, robot real-time position, remaining power, and path conflict, the unloading points to be executed are sorted, and the execution order of incomplete tasks is regenerated when temporary obstacles, equipment malfunctions, or task changes are detected.

[0014] As a preferred embodiment, determining the travel path includes:

[0015] An initial path is generated based on the target unloading point and the safe area in the site map. Feasible path points are selected by combining real-time obstacles and the robot's turning ability, and the path points are smoothed so that the robot can reach the target unloading point along the smooth path.

[0016] As a preferred implementation, the operational adaptation status includes at least one of the following state quantities: task urgency, path accessibility, interface compatibility, vacuum condition stability, energy sufficiency, component health, and hazard suppression. The state quantity is obtained by normalizing the corresponding sensor data or business data.

[0017] As a preferred implementation, identifying the train's sewage outlet and performing a compliant end-to-end docking includes:

[0018] The image location of the drain outlet or drain valve is obtained through visual recognition;

[0019] The relative distance and angle of the interface are obtained through laser positioning;

[0020] Obtain the robot's global position using ultra-wideband positioning;

[0021] The positioning results are fused into the docking pose, and the joint angles and driving force of the robotic arm are adjusted according to the docking force error.

[0022] As a preferred implementation, during end-effector compliant docking, the edge control layer determines the force-position control amount based on the expected docking force, the actual docking force, and the safety force threshold. When the actual docking force exceeds the safety force threshold or the interface pose deviation exceeds the allowable range, the robotic arm is controlled to decelerate, retract, or reposition.

[0023] As a preferred implementation, the sewage unloading process includes: controlling the robotic arm to connect with the ground vacuum tube interface, opening the train sewage outlet door, connecting the end effector with the sewage tank drain valve, opening the relevant valves in sequence, using vacuum negative pressure for suction, closing the valves after the sewage unloading conditions are met, disengaging from the interface, closing the door and retracting the hose.

[0024] As a preferred implementation, updating the job adaptation status includes the following steps:

[0025] The hazard level is determined based on at least one of the following events: collision risk, personnel intrusion, leakage characteristics, odor leakage, pipeline blockage, vacuum abnormality, motor stall, communication interruption, and battery abnormality. The response is executed according to at least one of the following strategies: warning, deceleration, suspension, valve closure, emergency stop, or return to charging.

[0026] This application also provides an intelligent control system for a railway vacuum sewage unloading robot, including a cloud management layer, an edge control layer, and a robot execution layer;

[0027] The cloud-based management layer is used for job planning, collaborative scheduling, data storage, global monitoring, maintenance decision-making, and hazard control.

[0028] The edge control layer is used for path planning and navigation, visual recognition and pose calculation, end-effector compliant control, workflow control, safety protection, energy management and status monitoring;

[0029] The robot's execution layer includes a walking mechanism, a waste unloading robotic arm, a ground docking device, a hose reel module, a sensor array, and an end effector.

[0030] The technical effects and advantages provided by the present invention in the above technical solution are as follows:

[0031] Compared to existing sewage unloading equipment that only performs movement, docking, suction or alarm control separately, making it difficult to integrate the working environment, train status, robot status and vacuum sewage unloading process into the control judgment, this application can form a linkage control effect for the entire process of railway vacuum sewage unloading.

[0032] The intelligent control method for railway vacuum sewage unloading robots proposed in this application first acquires multi-source data and then performs mapping and weight fusion to obtain the operation adaptation state. Secondly, based on the operation adaptation state, the operation task sequence, travel path, docking control boundary, safety response level, and energy return conditions are determined. Then, the robot is controlled to autonomously navigate to the target sewage unloading point according to the operation task sequence, identify the train sewage outlet, and perform end-point compliant docking. Finally, multi-modal data is monitored during the sewage unloading process, and the operation adaptation state is updated based on the multi-modal data monitoring results to adjust subsequent operations.

[0033] Because the operational adaptation status is formed by the fusion of multi-source data and runs through the processes of task determination, path travel, interface identification, docking control, wastewater unloading monitoring, and subsequent adjustments, the robot can synchronously correct its operational strategy according to changes in the field. This improves its responsiveness to temporary obstacles, interface deviations, vacuum anomalies, and insufficient energy, which is conducive to enhancing the autonomy, safety, and continuity of wastewater unloading operations. Attached Figure Description

[0034] To more clearly illustrate the technical solutions in the embodiments of this application or the prior art, the drawings used in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments recorded in this invention. For those skilled in the art, other drawings can be obtained based on these drawings.

[0035] Figure 1 The control system flowchart provided by the present invention;

[0036] Figure 2 This is a logical connection diagram of the algorithm of this invention.

[0037] Figure 3 This is a flowchart of an intelligent control method.

[0038] Figure 4 This is a flowchart for condition monitoring and maintenance decision-making. Detailed Implementation

[0039] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0040] The technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings of the embodiments of the present invention. Obviously, the described embodiments are only some embodiments of the present invention, and not related embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those of ordinary skill in the art without creative effort are within the scope of protection of the present invention.

[0041] In the description of this invention, it should be understood that the terms above, below, front, back, left, right, top, bottom, inside, outside, etc., indicate the orientation or positional relationship based on the orientation or positional relationship shown in the accompanying drawings. They are only for the convenience of describing this invention and simplifying the description, and do not indicate or imply that the device or element referred to must have a specific orientation, or be constructed and operated in a specific orientation. Therefore, they should not be construed as limiting this invention.

[0042] Example 1

[0043] The intelligent control system for railway vacuum sewage unloading robot provided by this invention Figure 1The direction of the middle arrow indicates the data transmission direction. The railway operation and maintenance management terminal sends the work plan to the cloud control layer. The cloud control layer sends control commands to the edge control layer through the data transmission link and uploads analysis and decision data. The edge control layer sends action commands to the robot execution layer through the data transmission link and collects feedback data. The sensor group collects environmental and equipment status data of the work area and feeds it back to the edge control layer. Data communication is achieved between modules in each level, forming a closed loop of control.

[0044] The cloud-based control layer, serving as the core decision-making and management hub of the system, is deployed on cloud servers. It mainly includes a data storage module, a planning management module, a global monitoring module, a maintenance decision-making module, and a hazard control module. These modules work together to form a global control closed loop. The cloud layer receives unloading operation plans issued by the railway operation and maintenance management department, reviews, breaks down, and distributes these plans. It also receives operation data and equipment status data uploaded from the edge control layer, performs global analysis and decision-making, generates maintenance suggestions and safety hazard handling plans, and enables long-term storage, querying, and traceability of various data. It supports global scheduling of multi-robot collaborative operations, solving the problems of traditional operations lacking global control, data traceability, and difficulties in multi-device collaboration.

[0045] The edge control layer, serving as a bridge between the cloud management layer and the robot execution layer, is deployed within the robot's control box. Employing an industrial-grade controller, it integrates path planning, autonomous navigation, position recognition, end-effector compliant control, workflow control, and safety protection modules. It is the core of achieving local closed-loop control. Its primary function is to receive work plans and control commands from the cloud management layer, parse them, and convert them into action commands that the robot can execute. Simultaneously, it collects real-time work and status data from the robot execution layer, performs local preprocessing, noise reduction, and filtering, and then uploads it to the cloud management layer. This prevents work interruptions during network outages, improves work response speed, and resolves issues of data transmission latency and high network dependence between the cloud and the robot execution layer.

[0046] The robot execution layer consists of the railway vacuum sewage unloading robot body adapted in this application and its supporting execution components. It is the specific execution carrier for the operation actions, including the robot walking mechanism, sewage unloading robotic arm, ground docking device, hose reel module, sensor group, LiDAR, vision sensor, force sensor, position sensor, IMU inertial measurement unit, gas sensor, etc., actuators, valves, motors, etc. Its core function is to receive action commands issued by the edge control layer and complete specific operation actions such as path walking, sewage unloading docking, suction operation, and pipeline reeling and unloading. At the same time, it comprehensively collects real-time operation data through the sensor group, such as docking accuracy, vacuum degree, operation progress, etc., and equipment status data, such as motor speed, battery power, pipeline pressure, etc., and feeds it back to the edge control layer to provide data support for upper-level management and control, solving the problems of poor robot operation autonomy and incomplete data collection.

[0047] Data transmission between different levels is achieved through 5G / industrial Ethernet, ensuring the real-time performance, security, and reliability of data transmission. Data encryption, authentication, access control, and secure communication protocols are employed to prevent data leakage and tampering. Offline operation mode is also supported, and the edge control layer can independently complete preset tasks when the network is interrupted. After the network is restored, data is automatically synchronized to the cloud management layer, further improving system reliability and adapting to scenarios where the network is unstable at railway operation sites.

[0048] Example 2

[0049] Figure 3 This is a flowchart illustrating the intelligent control method, used to demonstrate the control logic of the unmanned operation of the railway vacuum sewage unloading robot. It clearly shows the connection between each operation link, and the specific process is as follows:

[0050] The system receives the sewage unloading operation task instruction, specifies the operation compartment, operation location and operation requirements, and completes the task initialization.

[0051] Based on the target location and environmental information of the work area, the robot plans a constrained walking path from the waiting area to the work position, avoiding obstacles and dangerous areas. It then autonomously drives along the planned path and adjusts its driving posture in real time using the positioning and navigation module to ensure accurate arrival at the designated work position.

[0052] Through a multi-sensor fusion positioning and navigation module, the system automatically identifies train sewage outlets, performs three-dimensional positioning and attitude estimation, and outputs accurate alignment data.

[0053] Based on positioning data, the robotic arm, under the action of the end effector compliant control module, achieves flexible docking with the train's sewage outlet and locks it in place.

[0054] Start the vacuum sewage unloading system, monitor the sewage unloading status in real time, and after the sewage is pumped out, close the valve, unlock the docking point, and realize the closed loop of the sewage unloading operation process.

[0055] The safety protection module monitors the entire operation process and provides real-time early warnings and responses to abnormal conditions such as collisions, leaks, and communication interruptions.

[0056] When the unloading task is completed or the battery is low, the robot autonomously plans its return route to the charging station to complete automatic charging.

[0057] The status monitoring and maintenance decision-making module summarizes and analyzes the data from this operation and the equipment operating status, providing data support for subsequent maintenance optimization.

[0058] This flowchart fully covers the robotic unmanned wastewater unloading operation process, with each step connected sequentially and under closed-loop control to ensure automation, accuracy, and safety of the operation.

[0059] Example 3

[0060] See Figure 4 This demonstrates the connection between equipment health management and maintenance procedures. The specific process is as follows:

[0061] The status monitoring and maintenance decision-making module collects real-time data on the robot's core components, including the robotic arm, locomotive, vacuum system, and sensors, as well as operational parameters, work process data, and environmental status data. It then aggregates and performs preliminary preprocessing of the data. Based on the collected real-time and historical maintenance data, it constructs an equipment health model to quantitatively assess the health status, aging degree, and wear trend of the robot as a whole and its key components, outputting the health assessment results.

[0062] Based on the health assessment results, potential equipment malfunctions are identified, and risk levels are categorized for early warning. The types of malfunctions, their impact range, and development trends are clearly defined, promptly alerting maintenance personnel. Combining malfunction warning information, equipment runtime, workload, and historical maintenance data, personalized maintenance recommendations are intelligently generated, specifying maintenance content, priorities, replaceable vulnerable components, and maintenance procedures.

[0063] Based on maintenance recommendations, the system automatically generates standardized maintenance work orders, clearly specifying the work order number, maintenance tasks, completion deadline, and responsible personnel, and pushes them to the operation and maintenance management terminal to achieve standardized and automated management of maintenance work.

[0064] This flowchart fully covers the work process from equipment status monitoring to maintenance work order generation, forming a closed-loop management of data collection, health assessment, potential hazard warning, maintenance suggestions, and work order generation. It provides clear logical support for intelligent robot operation and maintenance, ensuring long-term stable and low-failure operation of equipment.

[0065] Example 4

[0066] See Figure 2 The environmental data acquisition unit collects environmental data from the work area and inputs it into the improved GBB-RRT* path planning algorithm to generate a path that meets the constraints. The path data is integrated with the multi-sensor fusion autonomous navigation algorithm and combined with the position data from the work data acquisition unit to achieve autonomous robot navigation. The fusion positioning algorithm receives the interface position from the work data acquisition unit and the robot's position data, and outputs accurate positioning results, providing data support for the fuzzy PID end-effector compliant control algorithm. The equipment status data acquisition unit collects equipment operation data and inputs it into the PHM maintenance decision algorithm to generate maintenance suggestions and the safety hazard identification algorithm to identify hazards. The output results of each algorithm are summarized into the control command output unit to generate action commands that the robot can execute, realizing a closed-loop connection between the algorithm and the execution process.

[0067] An improved GBB-RRT* path planning algorithm is used for robot dynamic obstacle avoidance and constraint-satisfying path planning. Based on the traditional RRT* algorithm, target bias sampling, path cost optimization, collision detection acceleration, and dynamic obstacle constraints are introduced to improve path search efficiency, smoothness, and real-time performance. The formulas used are as follows.

[0068] This embodiment also illustrates the generation method of the core algorithm and the job adaptation state. The core algorithm includes the improved GBB-RRT* satellite path planning algorithm, the multi-sensor fusion autonomous navigation algorithm, the fusion positioning algorithm of visual recognition, laser positioning and UWB ultra-wideband positioning, the fuzzy PID end compliant control algorithm, the PHM-based maintenance decision algorithm, and the safety hazard identification algorithm of CNN image recognition and multi-sensor threshold judgment.

[0069] The environmental data acquisition unit collects environmental data of the work area and inputs it into the path planning algorithm. The path data is integrated with the multi-sensor fusion autonomous navigation algorithm. The work data acquisition unit outputs the robot's position and posture. The identification and positioning unit outputs the interface pose. The end effector compliant control algorithm adjusts the robotic arm's movements based on docking force feedback. The PHM algorithm outputs the health status based on operating parameters. The safety hazard identification algorithm outputs the hazard level based on image, pressure, gas, and safe area data.

[0070] During job adaptation state generation:

[0071] Positive state variables can be expressed by the formula: calculate, The generated dimensionless positive state quantity can include battery capacity, interface identification confidence level, and component health level. By eliminating the differences in original physical units (such as percentage, absolute value, etc.), multi-source data of different dimensions such as battery capacity and confidence level are mapped to a unified 0-1 scale. The closer the value is to 1, the more beneficial the single state is to subsequent operations. The inverse state quantity can be calculated according to the formula: The calculations show that inverse state variables can include collision risk, leakage confidence, path congestion, and docking force deviation. For the i-th dimensionless inverse state variable generated, the output value range of the i-th dimensionless forward / inverse state variable is limited to [0,1] by the truncation function; The real-time collected measurement value or business calculation value of this positive indicator is obtained by the corresponding sensor group or business module in real time. This indicator represents the lower limit of the safe operating conditions allowed by the system. The upper limit of this indicator is determined by the station operation strategy or hardware parameter preset; clip is a truncation function. If the input calculated value is less than 0, it is directly equal to 0; if it is greater than 1, it is directly equal to 1; if it is between 0 and 1, it remains unchanged. When the actual risk... The higher it is, the closer it gets to the upper limit. At that time, the calculated The closer it gets to 0 (representing extreme insecurity); while when the risk drops to the minimum limit... hour, The value is 1 (representing complete safety and passability), thus mathematically unifying all risk indicators into safe passage / operation degree, so as to achieve homogeneous integration with positive state variables.

[0072] The truncation function locks the calculation result within a specified range. If the input value is <0, it is set to 0; if the input value is >1, it is set to 1; if the input value is between 0 and 1, it is left unchanged.

[0073] Overall job suitability can be calculated using a formula. The calculation is as follows: C represents the job fit, and m represents the number of state variables involved in the decision. Let i be the weight of the i-th state variable. Let i be the dimensionless forward / reverse state variable. The station operation strategy can be preset and meet the requirements. and By integrating previously isolated data such as power consumption, route, risk, and interface into a single global indicator C, the system dynamically guides and corrects subsequent control decisions (such as determining task sequences, travel routes, docking control boundaries, safety response levels, and energy return conditions) by determining whether C is within a safe operating range, thus achieving coordinated risk control throughout the entire unmanned operation process.

[0074] In a specific train vacuum unloading operation, the robot is preparing to perform the unloading task at a certain point. At this time, the edge control layer and sensor group collect and calculate two key indicators in real time: a single positive state quantity, namely the remaining power, and a single negative state quantity, namely the collision risk.

[0075] For the remaining power in the positive state, the system sets a lower limit for its safe operation. upper limit The robot is currently collecting real-time battery power measurements through its battery management module. We substitute these values ​​into the formula for calculating the positive state variables in the image. :

[0076] Since the calculated result 0.8 falls exactly between 0 and 1, it remains unchanged according to the definition of the cutoff function. Therefore, the final dimensionless electric state quantity is obtained. This result indicates that the robot has ample energy, which is highly conducive to continuing to perform its tasks.

[0077] For the risk of collision with inverse state variables (a larger value indicates a higher risk and lower passability), the system sets a lower limit for safety assessment. upper limit At this moment, the robot's LiDAR and visual safety protection module detected a pedestrian approaching from the front of the path. After algorithmic evaluation, the current real-time operational risk value was calculated. We substitute the numerical values ​​into the formula for calculating the inverse state variables in the image. : ;

[0078] Similarly, after processing by the truncation function, this dimensionless inverse state variable... This indicates that although there is a certain risk of collision, it is still within a controllable and safe passage range.

[0079] These two state variables are integrated into a comprehensive operational adaptability C. Based on the current operational strategy preset for the station section, the weight of the remaining power capacity is assigned... The weighting of collision risk is assigned The sum of the two weights satisfies The specification requirements. The dimensionless state quantities calculated previously. and Substitute the weight fusion formula in the image :

[0080] The final calculated overall job fit If the normal operation safety adaptation range set by the station section cloud management layer at this time is... A score of 0.74 means that the system allows the robot to continue performing the current unloading task; however, due to the collision risk lowering the score, the edge control layer will simultaneously adopt a linkage correction strategy, such as automatically limiting the robot's driving speed and the robotic arm's docking approach speed to a more secure level.

[0081] Example 5

[0082] An improved GBB-RRT* path planning algorithm is adopted for robot dynamic obstacle avoidance and optimal path planning. Based on the traditional RRT* algorithm, target bias sampling, path cost optimization, collision detection acceleration, and dynamic obstacle constraints are introduced to improve path search efficiency, smoothness, and real-time performance. The formulas used are as follows:

[0083] Improve the target-biased sampling mechanism, using probability To the target point Sampling, in probability Perform global random sampling: ,in Spatial target sampling points generated for the current iteration step; The three-dimensional spatial coordinates of the preset final target waste discharge point are directly obtained from the cloud-based planning management module; Rand(0,1) is an arbitrary coordinate point randomly generated within the legal three-dimensional space where driving is permitted throughout the entire station; rand(0,1) is a random decimal number uniformly distributed within the interval (0,1) generated by the system; p is a set target bias probability constant, with a value range of (0,1), which is usually preset during algorithm initialization according to the complexity of the map (e.g., a value of 0.1-0.3).

[0084] Traditional algorithms, by blindly sampling randomly in space, cause the search tree to wander aimlessly, resulting in extremely slow convergence. This formula introduces a probability p, allowing the algorithm to directly target the endpoint with a fixed proportion. This generates a strong guiding force in physical space that pulls the endpoint towards it, significantly increasing the sampling density of the target area, accelerating algorithm convergence, and reducing ineffective, random searches.

[0085] Search for the node closest to the sampling point in the random tree:

[0086] ,in: V represents the existing node in the existing random tree that is closest to the current sampling point; V is the set of nodes in the currently constructed random tree, which expands continuously with algorithm iterations; X is the coordinate of any existing node in set V. These are the target sampling points generated by the sampling mechanism; The geometric Euclidean distance between two three-dimensional coordinate points is expressed by the formula... The calculation yields (x, y, z), which represents the coordinates of a three-dimensional point, with values ​​ranging from [0, +∞).

[0087] Since path planning is based on expanding outwards from a topological tree, the algorithm searches the entire set of nodes V and selects the end of the branch that is spatially closest to the new sampling point. This ensures that subsequent path expansion can start from the nearest physical reference point, maintaining the continuity of the path in spatial geometry and the efficiency of its growth.

[0088] By step size Generate a new node: ,in: For the new node; It is the existing node in the existing random tree that is closest to the current sampling point; These are the target sampling points generated by the sampling mechanism; The single-step expansion step size preset for the algorithm is a physical constant. Its value range is determined by the physical size of the robot and the density of obstacles in the field (the value is greater than 0 and is usually set to between tens of centimeters and one meter). It is a unit direction vector pointing from the nearest point to the sampling point, with a magnitude strictly equal to 1.

[0089] Due to random sampling points The distance to an existing tree might be very far, and a direct connection could involve crossing obstacles; therefore, the algorithm does not allow direct teleportation. This formula restricts the tree to only starting from... Set off, towards Move a specified step distance in the geometric direction. This generates a completely new, controlled candidate path point.

[0090] Improve the path cost function by combining path length, smoothness, and obstacle distance:

[0091] ,in, New node generated for the current iteration step The cumulative optimal path cost; U is the cost of the cumulative optimal path. The set of existing nodes within the local neighborhood centered on a point and with a specific physical distance as the radius; Let be any node in the neighborhood set U; For nodes The accumulated path cost from the starting point is now stored and retrieved by the previous iteration steps; For nodes To the new node Single-step transfer cost between ,in, The single-step total cost between two adjacent path points; is the spatial geometric Euclidean distance between two points, with a value range of [0, +∞); The real-time edge distance between the current assessment location and the nearest obstacle is calculated in real time by an environmental data acquisition unit (such as LiDAR or depth camera), and the value range is [value range missing]. ; The safety weight coefficient is a physical constant that is greater than 0 and is manually configured by the system according to the safety protection level of the vehicle preparation workshop.

[0092] By traversing all possible connections within the neighborhood set U, the cumulative cost of traveling from the starting point to a neighboring node and the cost of traveling from a neighboring node to the new node are calculated. The sum of the single-step costs is calculated, and the neighborhood point that minimizes the sum is selected as the temporary parent node of the new node, thus achieving local optimization of the path geometric length and potential risks.

[0093] Update the parent node to achieve asymptotic optimality: Where: This formula is the core criterion for the asymptotic optimality of RRT*. U is the new node. The set of neighboring nodes; C is any node in the neighborhood set U; ) is a node The cumulative cost of connecting via the old route; The optimal cumulative cost for the new node; For new nodes to neighboring nodes The cost of a single-step transfer.

[0094] This condition physically represents the dynamic pruning and route optimization of the path topology tree. When the algorithm adds a new node... Then, it will check all old nodes in the neighborhood: if it finds a new node that is being used as a gateway... If the total cost of traversing to the old node is lower than the cost of the original path to the old node, it indicates that a shorter or safer shortcut has been found. Once this condition is met, the system will sever the original parent node connection of the old node and rewiring it to... This allows the topology of the entire tree to continuously evolve towards the global optimal solution as new nodes are added.

[0095] By introducing robot kinematic constraints and dynamic obstacle velocity fields, path points are corrected in real time.

[0096] ,in: The final legal path point 3D pose vector after controlled correction; For the new node; This represents the virtual velocity field vector generated by the movement of dynamic obstacles (such as workers moving within the station area and passing vehicles). It is obtained in real-time through high-frequency calculation of the movement trajectory and absolute velocity of the dynamic obstacles by the safety protection module. k is the velocity field gain coefficient, a positive constant that ensures the planned path meets the requirements of dynamic obstacle avoidance, nonholonomic constraints, and safety boundaries. It is suitable for complex dynamic environments in stations. A multi-sensor fusion autonomous navigation algorithm is employed, integrating data from LiDAR, visual sensors, IMU inertial measurement units, and wheeled odometers. Kalman filtering is used to achieve optimal state estimation, providing the robot with high-precision and high-stability real-time positioning and path tracking capabilities. The formulas used are as follows:

[0097] System state vector This includes the robot's position, velocity, attitude, and sensor bias in the global coordinate system of the work area, represented as: In the formula: This represents the robot's three-dimensional position vector in the global coordinate system, obtained through initial positioning by an external positioning system (such as GPS, RTK, UWB, or SLAM). This represents the robot's three-dimensional velocity vector in the global coordinate system, estimated through integration of accelerometer data, differential estimation of wheeled odometry data, or multi-sensor fusion. The value range is (for indoor mobile robots). Drones or outdoor vehicles may reach ); The quaternion representing the robot's posture is used to represent the robot's rotational posture in three-dimensional space. The posture is estimated by fusing gyroscope angular velocity and accelerometer / magnetometer data. The absolute value of each element is between [0,1]. Accelerometer bias is a constant or slowly changing error caused by sensor manufacturing processes and temperature variations. Offline calibration is performed when the robot is stationary. Low-cost IMUs achieve zero bias in... Around, while high-precision tactical-grade IMUs are smaller. ; The gyroscope's bias is the primary cause of angle drift over time. Its acquisition and calibration methods are the same as those for accelerometer bias. The gyroscope bias of a low-cost IMU is... The high-precision IMU can achieve a speed of less than 1° / h.

[0098] High-frequency state prediction is performed using IMU pre-integration and odometry. The continuous-time motion model is as follows:

[0099] In the formula: The raw measurement vector from the accelerometer is generated by the IMU sensor in each sampling period (for...). Output directly. This is the vector of the raw angular velocity measurements from the gyroscope, directly output at high frequency from the IMU sensor. The white noise for accelerometer measurements is obtained by reading the noise density from the sensor manual or by analyzing static data using Allan variance calibration software. R(q) represents the white noise from the gyroscope measurement, and R(q) is the rotation matrix (coordinate transformation matrix) obtained from the attitude quaternion q. The elements of the orthogonal matrix range from [-1, 1]. The angular velocity antisymmetric matrix corresponding to the quaternion or the quaternion multiplication matrix. The gravitational acceleration vector in the global coordinate system is directly set as a constant in the algorithm (e.g., (Direction downwards). Discrete-time Kalman prediction formula:

[0100] ,in: The state prediction vector at time k (including robot 3D position, velocity, attitude quaternion, and IMU bias, etc.) for The optimal state estimate at time step 1 is provided by the iterative storage from the previous time step. F is the nonlinear state transition function. It is the linearized state transition matrix (Jacobi matrix) of the current state, which is an intrinsic constant matrix determined by the physical laws of space kinematics. Let be the state prediction covariance matrix at time k, representing the uncertainty of the predicted state. for The optimal estimated covariance matrix at time t. The process noise covariance matrix is ​​defined by the range of values ​​for each element, which is determined by the high-frequency measurement noise of the hardware itself. The control input vector at time k is composed of real-time acceleration, angular velocity, and wheel odometer data collected by the inertial measurement unit (IMU).

[0101] Using lidar point cloud matching and visual feature localization results as observations, low-frequency precise correction is performed to suppress long-term drift. The Kalman filter update formula is as follows:

[0102] , , ,in: This is the Kalman-Gain matrix, which belongs to the dynamic weight allocation matrix. The observation matrix (or measurement matrix) describes the mapping relationship from the state space to the observation space. To observe the noise covariance matrix, its value range depends on the real-time measurement standard deviation of the lidar point cloud matching algorithm and the visual feature localization algorithm in the current scene. These are the measured observation vectors, directly obtained from laser SLAM pose matching or visual feature localization. The observation prediction vector is derived from the current prediction state. via measurement model The projected theoretical observation values. This is the optimal state estimation vector after correction at time k. Let be the optimal estimated covariance matrix after correction at time k. I is the identity matrix.

[0103] The optimal observation is obtained by performing variance-weighted fusion of the laser SLAM pose and visual localization results:

[0104] In the formula: This is the output position of the laser SLAM. Output position for visual positioning. The standard deviation of laser positioning. The standard deviation of visual positioning.

[0105] Example 6

[0106] The specific steps for using the algorithm in the system are as follows:

[0107] Power-on initialization: The robot remains still for 1–2 seconds to complete IMU zero-bias estimation and coordinate system alignment;

[0108] Multi-source data acquisition: simultaneous acquisition from IMU, odometer, LiDAR, and visual sensors;

[0109] Timestamp alignment: Unifying timestamps to the same point in time through interpolation;

[0110] High-frequency prediction: IMU + odometry quickly outputs pose prediction values;

[0111] Low-frequency correction: Laser SLAM / visual positioning is periodically corrected to eliminate drift;

[0112] Output navigation commands: Output optimal position (x, y) and heading angle. It drives the chassis to travel along the planned path.

[0113] A multi-source fusion positioning method combining visual recognition, laser positioning, and UWB ultra-wideband positioning is adopted to achieve high-precision identification of the train waste bin interface, the ground vacuum tube interface, and the robot's own position.

[0114] The vision sensor acquires the coordinates of the interface image through a target detection algorithm, and then converts them into a 3D positioning result in the robot coordinate system using a camera intrinsic parameter model. The lidar obtains the laser positioning result of the interface center through point cloud fitting. The UWB ultra-wideband positioning system obtains the robot's global positioning result through base station calculations. .

[0115] A variance-weighted fusion model is used to optimally fuse the three types of localization results:

[0116] In the formula, This represents the final robot position estimate (including 3D coordinates or 2D planar coordinates) after optimal fusion. , and These represent the robot's raw position observations independently output by the visual SLAM (or visual odometry), laser SLAM (or lidar odometry), and UWB (ultra-wideband) positioning systems, respectively. , and These are the weighting coefficients for the positioning results from vision, laser, and UWB, respectively, and they satisfy the normalization condition (i.e., ...). In the variance-inverse weighted model, each weight is calculated based on the reciprocal of the current positioning variance (or the trace of the covariance matrix) of the corresponding sensor. The calculation formula is as follows: ,in is the variance of the localization result of the i-th sensor. The magnitude of the variance is estimated online in real time by each subsystem based on its observation geometry, number of feature matches or signal strength (such as RSSI / geometric distribution of UWB). Therefore, the value range of these three weights is strictly limited to [0, 1]. The better the performance and the smaller the variance of the sensor, the closer the weight is to 1, and vice versa.

[0117] A fuzzy PID end-effector compliance control algorithm is employed to achieve force and position compliance control when the robotic arm's end effector docks with the waste bin interface and the ground vacuum tube interface, ensuring flexible contact, impact-free, and leak-free docking, and protecting the interfaces from damage. The fuzzy controller uses force error... and error change rate Input, Output , , Real-time correction: ; ; ; , , These are the proportional coefficient, integral coefficient, and differential coefficient. , , These are the initial PID parameters. , , This provides a correction value for fuzzy inference output.

[0118] In addition, compliance constraints are added: , For actual docking force, In order to achieve the desired connection, This is the safety force threshold.

[0119] The fault prediction and health management (PHM) algorithm based on big data is used to perform health assessment, fault warning, remaining life prediction and maintenance decision-making for the walking mechanism, robotic arm, vacuum pump, battery, sensor and actuator valve of the vacuum unloading robot.

[0120] (1) Calculation of Health Index (HI)

[0121] In the formula, For equipment health index, For the number of features, For feature weights, For real-time monitoring values, These are standard rated values.

[0122] (2) Fault prediction model

[0123]

[0124] In the formula, In the future The fault prediction value (e.g., voltage, temperature, vibration amplitude, etc.) of the device or system at the i-th state at any given time. This represents the actual measured value (current observation value) of the state at time t. This is a trend coefficient (or adjustment correction factor) used to correct prediction biases caused by nonlinear changes or environmental disturbances. It is optimized based on historical fault data through offline experiments, and its value ranges from [value range missing]. Between (when the model completely degenerates into a standard first-order Taylor extrapolation, ); This is the rate of change of the state at the current moment (i.e., the first derivative or the trend slope). The time step for prediction.

[0125] The algorithm employs CNN image recognition combined with multi-sensor threshold judgment to identify safety hazards during wastewater unloading operations, such as leakage, odor leakage, pipeline blockage, vacuum abnormalities, docking failures, personnel intrusion, and obstacles, enabling graded early warning and automatic emergency response.

[0126] CNN Feature Extraction: In the formula, F represents the output feature map, and W represents the convolution kernel weights, which are the core that the network needs to learn. Before the model training begins, a small random number is assigned to it through random initialization (such as He initialization or Xavier initialization). During the training process, it is updated in real time through the backpropagation algorithm (BP) and gradient descent optimizer. Its value range is within the non-zero floating-point range of [-1, 1] or smaller. represents the input image or sensor data matrix (Input Data), which is directly acquired at high frequency by hardware devices such as cameras and radars and then preprocessed (such as normalization and denoising) before being input to the network. If it is standard image data, the pixel values ​​before normalization are in the range of [0, 255], and after normalization they are in the range of [0, 1] or [-1, 1]. b represents the bias term, which is a translation amount introduced to improve the model's fitting ability. It is obtained in the same way as the weight W, and is set directly to 0 or a very small positive number (such as 0.01) during initialization. It is continuously optimized and updated online during backpropagation training, and its value range is within a range of several scalar absolute values. This refers to a non-linear activation function, including ReLU, LeakyReLU, or Sigmoid, whose main function is to introduce non-linear modeling capabilities into the network. It does not have a fixed range of values, but rather determines the numerical boundaries of the output feature map F.

[0127] Risk level confidence: In the formula, The probability of a potential hazard is 0~1.

[0128] Multi-threshold classification judgment: The criteria for Level 1 hidden danger are as follows The criteria for determining a level-two hazard are: The criteria for determining a level three hazard are: or .

[0129] For leaks of dirt and odors, the system can make a joint judgment based on changes in liquid traces in visual images, gas sensor readings, and fluctuations in vacuum negative pressure. For pipe blockages, the system can make a joint judgment based on decreased flow rate, abnormal negative pressure, and increased pump current. For personnel intrusion, the system can make a joint judgment based on laser safety zones, visual detection, and the status of emergency stop buttons. The judgment results are written into the hazard management log and linked to the handling actions.

[0130] In a station application example, the robot is in the standby area and receives a task for four sewage unloading points of a train. The system first determines the operation sequence based on the train formation, station passage, and robot position. Then, it determines that the first point can be executed directly based on the operation adaptability. After the robot travels along the planned path to the target position, it identifies the sewage outlet, completes the door opening, ground vacuum pipe connection, sewage tank discharge valve connection, and suction. If temporary personnel occupy the area near the second point, the safety protection module increases the cost of the corresponding passage, and the path planning module selects an alternative path. If the battery level falls below the return threshold after the third point, the energy management module arranges to charge the battery before continuing to the fourth point.

[0131] In another maintenance example, the system found in the continuous operation log that the temperature rise rate of a certain robotic arm joint motor was higher than the historical benchmark, and the peak force repeatedly approached the safety force deviation threshold. The maintenance decision module reduced the priority of subsequent tasks for this robot, and at the same time generated maintenance suggestions to check the reducer, calibrate the force sensor, and review the end guide seal structure. After the maintenance was completed, the staff uploaded the work order results, and the cloud management layer updated the health benchmark of the component accordingly.

[0132] In this invention, the number of indicators, weights, threshold ranges, and response strategies for the job adaptation status can be configured according to the station operation procedures, train models, interface structures, robot models, and on-site safety requirements. The path planning algorithm, fusion navigation algorithm, compliant control algorithm, health assessment algorithm, and hazard identification algorithm can also be replaced with equivalent algorithms without changing the control closed-loop logic.

[0133] The above embodiments illustrate the system structure, control method, comprehensive operation adaptation analysis, core algorithm, maintenance decision-making and safety hazard handling methods of the present invention. The relevant modules establish a closed loop through task data, sensor data, control commands and result records, enabling the railway vacuum sewage unloading robot to complete planned execution, interface docking, vacuum sewage unloading, anomaly response, energy management and operation and maintenance traceability in unmanned operation scenarios.

[0134] It should be noted that, in this document, relational terms such as first and second are used only to distinguish one entity or operation from another entity or operation, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Moreover, the terms include, contain, or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such process, method, article, or apparatus.

[0135] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and variations can be made to these embodiments without departing from the principles and spirit of the invention.

[0136] The present invention and its embodiments have been described above. This description is not restrictive. The accompanying drawings are only one embodiment of the present invention. The actual structure is not limited to this. In short, if a person skilled in the art is inspired by this description and designs a similar structure and embodiment without departing from the spirit of the present invention, such design should fall within the protection scope of the present invention.

[0137] In the description of this specification, references to terms such as "an embodiment," "example," "specific example," etc., indicate that a specific feature, structure, material, or characteristic described in connection with that embodiment or example is included in at least one embodiment or example of the invention. In this specification, illustrative expressions of the above terms do not necessarily refer to the same embodiment or example. Furthermore, the specific features, structures, materials, or characteristics described may be combined in any suitable manner in one or more embodiments or examples.

[0138] The preferred embodiments of the present invention disclosed above are merely illustrative of the invention. These preferred embodiments do not exhaustively describe all details, nor do they limit the invention to any specific implementation. Clearly, many modifications and variations can be made based on the content of this specification. This specification selects and specifically describes these embodiments to better explain the principles and practical applications of the invention, thereby enabling those skilled in the art to better understand and utilize the invention. The invention is limited only by the claims and their full scope and equivalents.

Claims

1. An intelligent control method for a railway vacuum sewage unloading robot, characterized in that, include: After acquiring multi-source data, mapping and weight fusion are performed to obtain the job adaptation status. Based on the operational adaptation status, the operational task sequence, travel path, docking control boundary, safety response level, and energy return conditions are determined. According to the work task sequence, the robot is controlled to autonomously navigate to the target sewage unloading point, identify the train's sewage outlet, and perform a smooth end docking. Multimodal data is monitored during the unloading process, and the operation adaptation status is updated based on the monitoring results to adjust subsequent operations.

2. The method according to claim 1, characterized in that, The multi-source data includes wastewater unloading operation plan, station environment data, robot status data, interface pose data, vacuum operating condition data, energy reserve data, and hazard perception data. The multimodal data includes vacuum negative pressure, pipeline pressure, flow rate, docking force, valve status, and safe zone status.

3. The method according to claim 2, characterized in that, Determining the sequence of tasks includes: Based on train formation, unloading points, operation time window, robot real-time position, remaining power, and path conflict, the unloading points to be executed are sorted, and the execution order of incomplete tasks is regenerated when temporary obstacles, equipment malfunctions, or task changes are detected.

4. The method according to claim 1, characterized in that, Determining the driving route includes: An initial path is generated based on the target unloading point and the safe area in the site map. Feasible path points are selected by combining real-time obstacles and the robot's turning ability, and the path points are smoothed so that the robot can reach the target unloading point along the smooth path.

5. The method according to claim 4, characterized in that, The operational adaptation status includes at least one of the following state quantities: task urgency, path feasibility, interface compatibility, vacuum condition stability, energy sufficiency, component health, and hazard suppression. The state quantity is obtained by normalizing the corresponding sensor data or business data.

6. The method according to claim 1, characterized in that, Identifying train sewage outlets and performing end-of-line smooth docking includes: The image location of the drain outlet or drain valve is obtained through visual recognition; The relative distance and angle of the interface are obtained through laser positioning; Obtain the robot's global position using ultra-wideband positioning; The positioning results are fused into the docking pose, and the joint angles and driving force of the robotic arm are adjusted according to the docking force error.

7. The method according to claim 6, characterized in that, During end-effector compliant docking, the edge control layer determines the force-position control quantity based on the expected docking force, the actual docking force, and the safety force threshold. When the actual docking force exceeds the safety force threshold or the interface pose deviation exceeds the allowable range, the control robot arm is decelerated, retracted, or repositioned.

8. The method according to claim 7, characterized in that, The sewage unloading process includes: controlling the robotic arm to connect with the ground vacuum tube interface, opening the train sewage outlet door, connecting the end effector with the sewage tank drain valve, opening the relevant valves in sequence, using vacuum negative pressure for suction, closing the valves after the sewage unloading conditions are met, disconnecting from the interface, closing the door and retracting the hose.

9. The method according to claim 1, characterized in that, Updating job compatibility status includes the following steps: The hazard level is determined based on at least one of the following events: collision risk, personnel intrusion, leakage characteristics, odor leakage, pipeline blockage, vacuum abnormality, motor stall, communication interruption, and battery abnormality. The response is executed according to at least one of the following strategies: warning, deceleration, suspension, valve closure, emergency stop, or return to charging.

10. An intelligent control system for a railway vacuum sewage unloading robot, used to implement the method described in any one of claims 1-9, characterized in that, It includes a cloud management layer, an edge control layer, and a robot execution layer; The cloud-based management layer is used for job planning management, collaborative scheduling, data storage, global monitoring, maintenance decision-making, and hazard control. The edge control layer is used for path planning and navigation, visual recognition and pose calculation, end-effector compliance control, workflow control, safety protection, energy management and status monitoring; The robot execution layer includes a walking mechanism, a waste unloading robotic arm, a ground docking device, a hose reel module, a sensor group, and an end effector.

Images