A dredging robot cooperative positioning and cluster operation system and method

Abstract

The application provides a dredging robot cooperative positioning and cluster operation system and method, adopts a hierarchical fusion and time sequence self-correction integrated positioning architecture, the bottom layer uses IMU pre-integration for high-frequency pose prediction, the middle layer fuses satellite positioning, relative positioning and sonar data through extended Kalman filtering, the high layer introduces a time sequence deep model to model and online compensate residual errors of systematic errors, and outputs high-precision pose estimation. In the GNSS signal limited scene, through relative ranging between multiple robots, local map matching and graph optimization constraints, cluster cooperative positioning and global consistency correction are realized. Meanwhile, a robust-fault-tolerant guarantee mechanism covering the observation layer and the system layer is constructed, including Mahalanobis distance residual gating, adaptive noise estimation and lock state recovery. The application improves the positioning accuracy, system robustness and cluster operation efficiency in complex water environment, and is suitable for unmanned dredging operation in complex water environment such as urban inland rivers, narrow waterways and bridge dense areas.

Description

Technical Field

[0001] This invention relates to the field of intelligent control technology, and in particular to a collaborative positioning and cluster operation system and method for dredging robots. Background Technology

[0002] Traditional river dredging mainly relies on manually operated dredging vessels or individual dredging equipment. While this can meet basic operational needs in open waters, it is less effective in complex environments such as urban rivers, narrow waterways, and areas with a high density of bridges. These environments are characterized by turbid water, low visibility, varied riverbed topography, and a dense distribution of obstacles (bridges, docks, underwater pipelines, etc.). Manual operation is not only labor-intensive and carries high safety risks, but it also makes it difficult to accurately perceive the working environment and perform precise operations. This can easily lead to uneven dredging, repeated work, or missed areas, seriously affecting the dredging effect and project progress.

[0003] Existing dredging robot technologies mostly employ a single-robot operation mode, with their positioning systems heavily reliant on the Global Navigation Satellite System (GNSS). In complex river environments, such as under bridges, where vegetation obstructs the water surface, or where multipath interference is severe, GNSS signals are prone to loss of lock or a sharp decline in accuracy. This leads to accumulated robot positioning drift, making it impossible to maintain the accuracy requirements for long-term continuous operation. Furthermore, existing technologies often use single sensors or simple multi-sensor overlay methods for environmental perception and positioning, lacking a structured multi-source information fusion architecture. They fail to organically integrate inertial measurement, satellite positioning, sonar sensing, and communication collaboration information, making it difficult to effectively suppress systematic errors caused by inertial measurement drift, satellite multipath interference, and underwater environmental noise. Consequently, they cannot achieve accurate perception and autonomous positioning of riverbed topography, silt distribution, and obstacles. Furthermore, the existing system lacks a multi-robot collaborative localization and group constraint mechanism. When the localization conditions of a single robot are limited, it cannot achieve error complementarity and collaborative correction through information sharing and distributed optimization between clusters. Moreover, under abnormal operating conditions such as short-term sensor failure, sudden noise changes, or communication interruption, it lacks a robustness and fault tolerance guarantee mechanism that runs through the entire process of perception, fusion, and optimization. Manual intervention is often required to restart or correct the system, which seriously affects the continuity of system operation and engineering practicality.

[0004] Therefore, there is an urgent need to develop a collaborative positioning and swarm operation system and method for dredging robots, so as to improve the positioning accuracy, environmental adaptability and swarm collaboration capabilities of dredging robot systems, and meet the needs of long-term, high-precision and high-efficiency continuous operation in complex river environments. Summary of the Invention

[0005] The main objective of this invention is to provide a collaborative positioning and cluster operation system and method for dredging robots, which solves the technical problem that existing dredging operations are difficult to achieve accurate perception and autonomous positioning of the working environment. It constructs a multi-source positioning architecture that combines hierarchical fusion and time-series compensation, and structurally integrates inertial measurement, satellite positioning, sonar perception and communication cooperation information, which significantly suppresses long-term drift errors while ensuring real-time performance.

[0006] To solve the above-mentioned technical problems, the technical solution adopted by the present invention is a method for collaborative positioning and swarm operation of dredging robots, comprising the following steps: S1. Multi-source perception data acquisition and preprocessing: Each dredging robot unit synchronously acquires multi-source perception data of itself and the surrounding environment, and performs preprocessing. S2. Multi-source positioning layered fusion processing: A layered fusion architecture is adopted to process multi-source positioning information. The bottom layer uses inertial measurement data for high-frequency pose prediction, and the middle layer uses a filtering algorithm to fuse satellite positioning, relative positioning and sonar detection data in real time to obtain preliminary pose estimation. S3, Temporal Error Compensation: Input the preliminary pose estimation into the temporal depth model, perform residual modeling and online compensation for systematic errors, and output the single-machine pose estimation. S4. Cluster cooperative localization and global correction: Each robot shares observation information through a communication network and performs global consistency correction and error cooperative correction based on relative ranging, local map matching and graph optimization constraints. S5. Environmental modeling and dredging operations: Using positioning results and sonar detection information, environmental modeling and path planning are performed to execute autonomous dredging operations. S6, robust fault tolerance guarantee, performs anomaly observation suppression, adaptive noise estimation, positioning loss detection and state recovery throughout the entire process of perception, fusion and optimization to ensure positioning continuity.

[0007] Preferably, in step S1, the time series of the inertial measurement unit is used as a reference, and the data of the satellite positioning unit, relative positioning unit and sonar detection unit are time-aligned using the third-order Lagrange interpolation method, and all sensor coordinate systems are transformed to the robot's central body coordinate system.

[0008] Preferably, in step S2, the filtering algorithm is an extended Kalman filter; the underlying layer utilizes the inertial measurement unit pre-integration technique, within the time interval between adjacent observation frames. Internally, for high-frequency acceleration and angular velocity After pre-integration, the state transition equation is: ; The intermediate layer constructs an extended Kalman filter, and the system state vector is... The state prediction and covariance propagation equations are as follows: ; ; The observation update equation is: ; ; .

[0009] Preferably, in step S3, the intermediate layer filtering algorithm is used in the current and past... The state estimation sequence, residual sequence, and original measurement features of the inertial measurement unit within each time period are used as the time-series feature matrix. By fitting nonlinear residual mappings through long short-term memory networks or gated recurrent units temporal deep models, a systematic cumulative error compensation term is output online. The final corrected single-machine pose state is as follows: ; Preferably, in step S4, when the robot's satellite positioning unit signal is weak or lost, the robot shares the status information of other robots in the neighborhood through the communication network and uses the relative positioning unit to obtain relative observation values. They extracted local features of the riverbed topography detected by their respective sonar sensors, and performed local map matching using a normal distribution transformation algorithm to obtain relative pose constraints. A distributed backend framework based on factor graph optimization is constructed, and the optimization objective function for global consistency correction is: ; The Gauss-Newton method is used for iterative solution to achieve cluster-level error equalization suppression and global correction.

[0010] Preferably, in step S5, the sonar detection unit acquires raw underwater sonar data, which, after preprocessing, is then processed in parallel into three processing branches: The riverbed topography extraction branch extracts three-dimensional riverbed digital elevation features for terrain matching and positioning. The silt distribution identification branch identifies the silt layer thickness and physical boundaries, which is used to assist in path planning and work area division. The obstacle detection and recognition branch identifies the three-dimensional spatial boundaries of obstacles for obstacle avoidance and local path replanning. The information from the above branches is integrated to update the local environmental map in real time and fed back to the positioning and planning module.

[0011] Preferably, in step S6, the observation layer employs a residual gating technique based on Mahalanobis distance to statistically test the observation innovation of each sensor. The Mahalanobis distance calculation formula is as follows: ; in, The residual covariance matrix; Set the chi-square distribution threshold. ,like If the current observation is determined to be an outlier, the observation in that frame will be discarded immediately. like However, if the value shows a continuous increasing trend, adaptive noise estimation is triggered, which dynamically increases the observation noise covariance matrix of the sensor using the following formula. Reduce its weight coefficient in the fusion algorithm: ; in, For the filter residual, It is a forgetting factor.

[0012] Preferably, in step S6, the system layer sets a positioning loss detection and state recovery mechanism: the system maintains a sliding window for... The historical health status circular buffer is used to determine the fully autonomous positioning loss state when all external observation sensors fail within a set time threshold. The state recovery program is then initiated, and the system automatically rolls back to the last determined healthy absolute pose state node in the buffer. It then switches to a dead reckoning degraded mode consisting of pre-integration by pure inertial measurement units and riverbed fluid dynamics constraints. The system also sends a cooperation request to neighboring robots in the cluster to reconstruct the communication route topology until it regains effective relative constraints and then performs backward smooth closed-loop correction.

[0013] Preferably, multi-source sensing data acquisition includes: The inertial measurement unit acquires triaxial angular velocity and triaxial acceleration data at a high frequency of 200Hz; When there is no obstruction, the satellite positioning unit outputs absolute three-dimensional coordinates in the geodetic coordinate system at a frequency of 10Hz. and location covariance; The relative positioning unit obtains the relative distance and azimuth of other dredging robots in the vicinity at a frequency of 20Hz through ultra-wideband (UWB) wireless ranging or laser ranging. The sonar detection unit uses a multibeam sonar to acquire raw underwater acoustic images and depth point cloud data at a frequency of 5Hz.

[0014] This invention provides a collaborative positioning and cluster operation system for dredging robots to perform the above-described method, comprising: multiple dredging robot units, each dredging robot unit being equipped with a dredging execution mechanism, a propulsion device and a control module; A multi-source sensing and positioning module, installed on the dredging robot unit, includes at least: a sonar detection unit, an inertial measurement unit, a satellite positioning unit, and a relative positioning unit; The multi-source positioning fusion module is used to fuse data acquired by the multi-source sensing and positioning modules using a hierarchical fusion architecture. The bottom layer uses inertial measurement data for high-frequency prediction, and the middle layer performs real-time fusion through filtering algorithms. The temporal error compensation module, connected to the multi-source localization fusion module, is used to perform residual modeling and online dynamic compensation for the systematic errors generated during the fusion process through a temporal depth model. The cluster collaborative positioning module is used to realize the sharing of observation information among multiple dredging robot units through the communication network, and to collaboratively correct the positioning results of each dredging robot unit based on relative ranging, local map matching and graph optimization constraints. The robust and fault-tolerant control module is used to suppress anomalies, restore the state, or reconstruct the positioning process in the event of sensor malfunction, communication interruption, or positioning failure, through the observation layer residual gating unit and the system layer lock-out recovery unit. The task scheduling and monitoring module is used to uniformly manage and schedule the operation status and location information of the dredging robot cluster.

[0015] Beneficial effects: 1. Significantly improves positioning accuracy in complex aquatic environments. This invention adopts a layered fusion and temporal self-correction integrated positioning architecture. The bottom layer achieves high-frequency continuous pose prediction through IMU pre-integration. The middle layer uses extended Kalman filtering to perform physical constraint fusion of multi-source sensor information. The top layer introduces a temporal depth model to perform residual modeling and online dynamic compensation for systematic errors. By coupling classical model-driven and data-driven methods, systematic errors caused by IMU drift, GNSS multipath interference, and underwater environmental noise are effectively suppressed.

[0016] 2. To enhance the system's adaptability and stability in complex river environments, this invention introduces a multi-robot cooperative localization and multimodal complementary constraint mechanism. When a single robot navigates to areas with limited GNSS signals, such as under bridges, under vegetation cover, or in narrow river channels, it shares the state information of robots with good localization conditions within the neighborhood through a communication network. Combining relative ranging and sonar local map matching, a distributed backend framework based on factor graph optimization is constructed, transforming the long-term drift error that is unobservable by a single robot into a constrained graph optimization problem for the entire group.

[0017] 3. To improve system robustness and engineering availability, this invention constructs a robust and fault-tolerant integrated positioning guarantee mechanism covering the entire chain from the perception layer to the system layer. At the observation layer, residual gating technology based on Mahalanobis distance is used to dynamically suppress abnormal observations, and sensor fusion weights are adjusted in real time through adaptive noise estimation. At the system layer, a recovery mechanism combining positioning loss detection, state rollback, and backward smoothing is designed. When sensors fail temporarily, noise changes abruptly, or communication is interrupted, the system can still maintain continuous and stable pose output, reducing the need for manual intervention and meeting the high reliability and high availability requirements of complex river dredging projects.

[0018] 4. Improved efficiency and economy of cluster operations: This invention utilizes cluster collaborative positioning and precise positioning control to make dredging path planning more rational. Combined with information on riverbed topography, silt distribution, and obstacles obtained through sonar detection, it enables grid-based dredging operations, effectively reducing repeated dredging and missed dredging. Under the same operating conditions, it significantly improves dredging efficiency per unit time, reduces energy consumption and operation and maintenance costs, and is suitable for various engineering application scenarios such as large-scale rivers, urban inland rivers, and complex water areas, demonstrating good engineering application prospects and promotional value. Attached Figure Description

[0019] The present invention will be further described below with reference to the accompanying drawings and embodiments: Figure 1 This is the overall flowchart of the method of the present invention; Figure 2 This is a diagram of the integrated positioning architecture of the present invention, which combines layered fusion and temporal self-correction. Figure 3 This is a flowchart of the sonar-assisted positioning process of the present invention. Detailed Implementation

[0020] Example 1 like Figure 1 As shown, a method for swarm operation of river dredging robots based on sonar detection and multi-source localization fusion includes the following steps: S1. Multi-source sensing data acquisition and preprocessing: Deploy the dredging robot cluster and start the system, initialize the upper monitoring platform, and each dredging robot unit synchronously collects its own and the raw data of the complex river channel surrounding environment through its onboard multi-source sensing and positioning module, and performs noise reduction, correction and spatiotemporal alignment preprocessing on the raw data. S2. Multi-source positioning hierarchical fusion architecture processing: The hierarchical fusion architecture is used to process the pre-processed multi-source positioning information. The bottom layer uses inertial measurement unit (IMU) data for high-frequency continuous pose prediction. The middle layer uses the extended Kalman filter (EKF) algorithm to perform real-time fusion and state estimation of data from satellite positioning unit (GNSS), relative positioning unit and sonar detection unit. S3. Residual modeling and online compensation based on temporal depth model: The preliminary pose estimation results obtained by the fusion of intermediate layers are input into the high-level temporal depth model. Residual modeling and online dynamic compensation are performed on the systematic errors generated during the fusion of multi-source sensors to suppress inertial drift and satellite multipath interference, and output high-precision pose estimation results for single machine. S4. Cluster Distributed Cooperative Localization and Global Consistency Correction: Each dredging robot unit shares observation information through a communication network. The cluster cooperative localization module incorporates the high-precision pose estimation results of individual robots into the group consistency framework based on relative ranging between robots, local map matching, and backend graph optimization constraints, performing global consistency correction and collaborative error correction. S5. Environmental modeling, path planning and dredging operation execution: Using high-precision positioning results and riverbed topography and silt distribution information obtained by the sonar detection module, local environmental map modeling and dredging path planning are carried out. Each robot unit controls the propulsion device and dredging execution mechanism to perform precise autonomous dredging operations. S6. Full-chain robust fault tolerance guarantee and anomaly response: Throughout the perception, fusion and optimization process of the system operation, the robustness and fault tolerance control module performs anomaly observation and suppression, adaptive noise estimation, positioning loss detection and state recovery / reconstruction in real time to ensure positioning continuity under conditions such as short-term sensor failure or communication interruption. It determines whether the task is completed. If not, it returns to step S1 to continue the loop; if completed, it reports the task results and the system stands by or returns to base.

[0021] According to an embodiment of the present invention, in step S1, the multi-source sensing and localization module on each dredging robot unit independently collects data at different inherent frequencies: The inertial measurement unit acquires triaxial angular velocity and triaxial acceleration data at a high frequency of 200Hz; When there is no obstruction, the satellite positioning unit outputs absolute three-dimensional coordinates in the geodetic coordinate system at a frequency of 10Hz. and location covariance; The relative positioning unit obtains the relative distance and azimuth of other dredging robots in the vicinity at a frequency of 20Hz through ultra-wideband (UWB) wireless ranging or laser ranging. The sonar detection unit uses a multibeam sonar to acquire raw underwater acoustic images and depth point cloud data at a frequency of 5Hz.

[0022] To eliminate time-domain asynchronous errors caused by differences in hardware clock sources and inconsistent sampling rates among various sensors, the system establishes a unified Coordinated Universal Time (UTC) reference in the upper-level monitoring platform or local control module. All acquired raw data are stamped with high-precision hardware timestamps, using the highest frequency IMU time series as the reference. For low-frequency GNSS, relative positioning, and sonar data, a time alignment mechanism based on the third-order Lagrange interpolation method is adopted. In this embodiment, the synchronization error between multi-source sensing data is strictly controlled within ±2ms. Simultaneously, the sonar coordinate system, IMU coordinate system, and the spatial coordinate systems of each sensor at the center of the GNSS antenna are all uniformly transformed to the robot's central body coordinate system using an offline hand-eye calibration method, achieving strict spatiotemporal alignment.

[0023] According to an embodiment of the present invention, in step S2, the specific processing procedure of the multi-source localization hierarchical fusion architecture is as follows: The underlying layer utilizes IMU pre-integration technology to measure the time interval between two adjacent images or satellite observation frames. Internally, for high-frequency acceleration and angular velocity By performing pre-integration independent of the initial state, the high-frequency continuous pose state variables of the robot are predicted. The state transition equation is expressed as: (1); in, , and They are respectively The position, velocity, and rotation matrix of the real-time dredging robot in the world coordinate system. and These are the bias values ​​for the accelerometer and gyroscope, respectively. and The corresponding Gaussian white noise, This is the gravitational acceleration vector.

[0024] In the intermediate layer, an interpretable extended Kalman filter (EKF) is constructed to perform physical constraint fusion of multi-source sensor information. The system state vector is defined as follows: ,in It is a quaternion of attitude.

[0025] The one-step transition equation and covariance propagation equation for EKF state prediction are as follows: (2); (3); in, The nonlinear state transition function is derived from the underlying pre-integration. All are IMU measurement inputs. The Jacobian matrix of the state transition matrix. Let be the state covariance matrix. Let be the system process noise covariance matrix.

[0026] When GNSS satellite positioning units or relative positioning units have updated observation data, a standard observation equation is constructed: (4); in To account for observation noise, the filter update calculation is as follows: (5); (6); (7); in, For observation function The partial derivative matrix with respect to the state variables, This is the Kalman gain matrix. The intermediate layer outputs preliminary pose estimation state variables with physical mechanism interpretation in real time. .

[0027] According to an embodiment of the present invention, in step S3, the residual modeling and online compensation process based on the time-series deep model is as follows: Due to the severe GNSS multipath effect and unmodeled dynamic disturbances caused by water flow impact in complex aquatic environments such as narrow rivers and bridges, classical EKF filtering suffers from systematic drift errors. This invention introduces a time-series deep model based on lightweight long short-term memory networks (LSTM) or gated recurrent units (GRUs) at higher levels.

[0028] The intermediate layer EKF in the present and past The state estimation sequence, residual sequence, and raw IMU measurement features within each time period are used as inputs, denoted as the time-series feature matrix. The time-series deep model, by fitting a nonlinear residual mapping, outputs the systematic cumulative error compensation term for the current time step online. The final online-corrected high-precision pose state of the single machine is as follows: (8); In the formula, This represents addition operations on a manifold. Through structured coupling of classical model-driven and data-driven methods, adaptive dynamic correction of underwater environmental noise and system drift is achieved.

[0029] According to an embodiment of the present invention, in step S4, the specific implementation method of cluster distributed cooperative positioning and global consistency correction is as follows: Figure 2 As shown, when the robot When the robot navigates under a bridge or into densely vegetated areas, resulting in weak or lost GNSS signal, it activates a group-based cooperative positioning mechanism using the swarm cooperative positioning module. Through a maritime wireless communication network, robots with good positioning conditions in the vicinity can share information in real time, including those in open water and with GNSS in RTK fixed solution mode. The two aircraft acquire relative observations using a relative positioning unit, providing status information. relative distance relative azimuth Simultaneously, the two sonar units extract local features of the riverbed topography detected by their respective sonar sensors, and perform local map matching using the Normal Distribution Transform (NDT) algorithm to obtain relative pose constraints. .

[0030] Based on this, a distributed backend framework based on factor graph optimization is constructed. The optimization objective function for global consistency correction is expressed as: (9); in, The set of poses to be optimized for all robots in the cluster. To absolutely locate the prior factor residuals, For single-unit active odometer / pre-integration factor residuals, The residuals of the collaborative complementary constraint factors are constructed based on cluster relative ranging and sonar local map matching. , , These are the information matrices for the corresponding factors and the inverse of the covariance matrix, respectively. The Gauss-Newton method is used to iteratively solve this equation, transforming the long-term drift error, which is unobservable on a single machine, into a graph optimization problem with mutual constraints among the population. This achieves balanced suppression and global correction of the error at the cluster level.

[0031] According to an embodiment of the present invention, in step S5, the sonar-assisted localization, path planning, and operation process are as follows: Figure 3 As shown, the sonar detection unit acquires raw underwater sonar data, which, after preprocessing such as filtering and bilateral denoising, is then processed in parallel into three processing branches: The riverbed topography extraction branch extracts the three-dimensional digital elevation features of the riverbed, inputs them into the terrain matching and positioning algorithm, and associates them with the pre-stored high-precision river map or the local map established by the neighboring robot to provide relative constraints between clusters for the relative positioning unit. After graph optimization and local map matching, the corrected robot pose is output. The silt distribution identification branch identifies the silt layer thickness and physical boundary through sound wave scattering intensity and echo characteristics, forming silt boundary constraints to assist in path planning and automatic division of work areas. The obstacle detection and recognition branch accurately identifies the three-dimensional spatial boundaries of obstacles such as bridge piers, underwater pipelines, and large reefs, which are used for obstacle avoidance and local path replanning.

[0032] By integrating the information from the above branches, the local environmental map is updated in real time and fed back to the positioning and planning module, dynamically generating an optimized gridded sea-sweeping path to guide the execution agency to operate efficiently and avoid missed excavations and repetitive work.

[0033] According to an embodiment of the present invention, in step S6, to cope with the harsh working conditions of short-term sensor failure, sudden noise changes, and intermittent communication in complex engineering sites, the present invention designs a dual fault-tolerant guarantee from the perception layer to the system layer. At the observation layer, a residual gating technique based on Mahalanobis distance is used to statistically test the novelty and innovation of each sensor. For any observation vector... The formula for calculating its Mahalanobis distance is: (10); in, Let be the residual covariance matrix.

[0034] Set the chi-square distribution threshold. .like If the current observation is determined to be an outlier, such as GNSS multipath jump noise, the robust control module will discard the observation in that frame and trigger outlier observation suppression. However, if the value shows a continuous increasing trend, adaptive noise estimation is triggered, which dynamically increases the observation noise covariance matrix of the sensor using the following formula. Reduce its weight coefficient in the fusion algorithm: (11); in, For the filter residual, It is a forgetting factor.

[0035] At the system level, a location lock loss detection and state recovery / reconstruction mechanism is defined. The system maintains a sliding window in the memory pool. The system establishes a historical health status circular buffer. When a complete communication interruption is detected, or all GNSS, sonar, and UWB external observation sensors fail within a set time threshold, the system immediately determines the state to be "fully autonomous positioning lost."

[0036] At this point, the fault-tolerant control module initiates the state recovery / reconstruction procedure. The system automatically rolls back to the last confirmed healthy absolute pose state node within the buffer, cutting off unreliable real-time update streams and switching to a robust degradation mode of dead reckoning, consisting of pure IMU pre-integration and riverbed hydrodynamic constraints. Simultaneously, it sends cooperation requests to neighboring robots within the cluster, reconstructs the communication routing topology, and performs backward smoothing closed-loop correction until valid relative constraints are regained, ensuring the continuous operational safety and high availability of the dredging operation.

[0037] Example 2 This embodiment provides an overall hardware architecture for a collaborative positioning and cluster operation system for dredging robots. Through hardware and software collaboration, this system achieves accurate perception of complex river environments, high-precision hierarchical positioning fusion, distributed group correction, and robust fault-tolerant control across the entire chain.

[0038] Each dredging robot unit includes a multi-source perception and localization module, a multi-source localization fusion module, a timing error compensation module, a cluster cooperative localization module, a robustness and fault-tolerant control module, and a task scheduling and monitoring module deployed on shore or on a main control vessel. The physical entity of each river dredging robot unit is an underwater dredging robot, which integrates a jet-suction dredging actuator driven by a servo motor at its bottom or front. It has a water jet propulsion system driven by multiple brushless DC motors inside and is equipped with a core control module based on an ARM / FPGA microcontroller.

[0039] The multi-source sensing and localization module is rigidly fixed to the robot hull, and its hardware configuration includes: Sonar detection unit: Employs a 900kHz high-frequency forward-looking sonar or multibeam echo sounder to transmit acoustic signals and receive echoes, acquiring underwater three-dimensional topography, silt layer distribution and thickness, and obstacle boundary data. Inertial Measurement Unit: Employs a high-precision six-axis MEMS inertial sensor with a built-in three-axis accelerometer and three-axis gyroscope for real-time, high-frequency acquisition of the robot's angular velocity and acceleration status; Satellite positioning unit: It adopts a dual-frequency GNSS+RTK positioning antenna and receiver, and outputs high-precision absolute three-dimensional geodetic coordinates by receiving differential information from the reference station in unobstructed waters; Relative positioning unit: Uses UWB wireless ranging base station or vehicle-mounted 3D LiDAR to measure the relative distance, azimuth and pitch angles between the robot and other robot units in the vicinity in real time.

[0040] The multi-source localization fusion module is implemented based on an embedded computing platform mounted on the robot. This module adopts a layered fusion architecture, with its internal algorithm logic divided into a bottom layer and a middle layer: The underlying pre-integration hardware logic triggers IMU data acquisition at a frequency of 200Hz through the hardware timer interrupt of the core control module. The underlying pre-integration software thread performs manifold pre-integration on the angular velocity and acceleration between adjacent satellite frames, and predicts the robot's continuous pose, velocity and rotation matrix at high frequency, providing high-speed state updates for the system. The intermediate Kalman filter processor employs an Extended Kalman Filter (EKF) hardware accelerator to construct a nonlinear state observation matrix. When effective observations are generated by low-frequency GNSS positioning units, relative positioning units, or sonar matching algorithms, the robot's pose covariance is updated and corrected online by calculating the Kalman gain, outputting preliminary pose estimation results with physical mechanism constraints.

[0041] The temporal error compensation module is connected to the EKF output of the multi-source localization fusion module. Its hardware runs on the GPU core of the computing platform and internally contains a pre-trained lightweight Long Short-Term Memory (LSTM) temporal deep model. During system operation, this module allocates a temporal sliding window memory area to collect the preliminary pose sequence, filtered residual sequence, and raw high-frequency features of the IMU output in real time. The temporal deep model captures the nonlinear temporal features of unmodeled dynamic disturbances such as river flow impact, predicts the current systematic cumulative residual of the EKF online, and compensates it to the output of the intermediate layer through manifold addition, outputting the final high-precision pose estimation result for single-machine operation, effectively suppressing long-term drift.

[0042] The cluster collaborative positioning module forms a distributed communication network through a high-power underwater wireless data transmission radio mounted on the robot unit, such as a 4G / 5G module or a dedicated local area network wireless bridge.

[0043] When the robot When the robot drives under a bridge, causing a weakening or loss of GNSS signal, this module activates the coordination mechanism. The cluster collaborative positioning module uses a data transmission radio to call other well-positioned robots in the neighborhood in real time. The system uses the pose information of the local relative positioning unit and the relative distance obtained from the relative positioning unit and the relative pose constraints obtained from sonar terrain matching to construct a local factor graph structure. A distributed backend iterative optimization is performed using a factor graph optimization solver, incorporating individual positioning errors into the overall consistency constraint framework of the cluster to achieve global consistency correction.

[0044] The robustness and fault tolerance control module, serving as a comprehensive safeguard mechanism, directly impacts the perception, fusion, and optimization layers. This module establishes a circular, bidirectional health state buffer within the microprocessor, comprising two core control subunits: The observation layer residual gating unit calculates the Mahalanobis distance of the observation innovation of each sensor in real time. If the Mahalanobis distance exceeds the set chi-square threshold, the unit directly blocks the input of that data path to suppress abnormal observations; if the residual continues to increase, the observation noise covariance matrix of that sensor is increased online through an adaptive noise estimator to reduce its fusion weight. The system-level lock-out recovery and reconstruction unit monitors the health status words of each sensor in real time. In extreme cases such as sudden communication interruption or simultaneous failure of multiple sensors, this unit issues a positioning lock-out alarm, immediately rolls back the system state to the last healthy absolute pose node in the buffer, and then uses a backward smoothing algorithm to perform closed-loop reconstruction correction to ensure that the system maintains stable and continuous positioning output without manual intervention.

[0045] The task scheduling and monitoring module serves as the upper-level monitoring platform, such as a shore-based high-performance control computer or a ruggedized main control graphics workstation, maintaining a persistent connection with the entire dredging robot cluster via a wireless network. This platform integrates a 3D Geographic Information System (GIS) and a digital twin visualization engine. During operations, each robot unit uploads in real-time the riverbed topography, silt layer distribution, obstacle boundaries extracted by its sonar detection unit, as well as its own high-precision positioning and pose. The task scheduling and monitoring module visualizes this on a global map, monitoring the cluster's operational trajectory and dredging progress in real-time. Through built-in intelligent decision-making algorithms, it automatically divides the work area for multiple robots and dynamically schedules tasks, eliminating missed dredging and duplicate dredging, achieving fully automated cluster operations.

[0046] The above embodiments are merely preferred technical solutions of the present invention and should not be considered as limitations on the present invention. The scope of protection of the present invention should be limited to the technical solutions described in the claims, including equivalent substitutions of the technical features described in the claims. That is, equivalent substitutions and improvements within this scope are also within the scope of protection of the present invention.

Claims

1. A method for collaborative positioning and swarm operation of dredging robots, characterized in that, Includes the following steps: S1. Multi-source perception data acquisition and preprocessing: Each dredging robot unit synchronously acquires multi-source perception data of itself and the surrounding environment, and performs preprocessing. S2. Multi-source positioning layered fusion processing: A layered fusion architecture is adopted to process multi-source positioning information. The bottom layer uses inertial measurement data for high-frequency pose prediction, and the middle layer uses a filtering algorithm to fuse satellite positioning, relative positioning and sonar detection data in real time to obtain preliminary pose estimation. S3, Temporal Error Compensation: Input the preliminary pose estimation into the temporal depth model, perform residual modeling and online compensation for systematic errors, and output the single-machine pose estimation. S4. Cluster cooperative localization and global correction: Each robot shares observation information through a communication network and performs global consistency correction and error cooperative correction based on relative ranging, local map matching and graph optimization constraints. S5. Environmental modeling and dredging operations: Using positioning results and sonar detection information, environmental modeling and path planning are performed to execute autonomous dredging operations. S6, robust fault tolerance guarantee, performs anomaly observation suppression, adaptive noise estimation, positioning loss detection and state recovery throughout the entire process of perception, fusion and optimization to ensure positioning continuity.

2. The method for collaborative positioning and cluster operation of dredging robots according to claim 1, characterized in that: steps In S1, the time series of the inertial measurement unit is used as the reference, and the data of the satellite positioning unit, relative positioning unit and sonar detection unit are time-aligned using the third-order Lagrange interpolation method, and all sensor coordinate systems are transformed to the robot's central body coordinate system.

3. The method for collaborative positioning and swarm operation of dredging robots according to claim 1, characterized in that: In step S2, the filtering algorithm is extended Kalman filtering; the underlying layer utilizes the inertial measurement unit pre-integration technique, within the time interval between adjacent observation frames. Internally, for high-frequency acceleration and angular velocity After pre-integration, the state transition equation is: ； The intermediate layer constructs an extended Kalman filter, and the system state vector is... The state prediction and covariance propagation equations are as follows: ； ； The observation update equation is: ； ； 。 4. The method for collaborative positioning and swarm operation of dredging robots according to claim 1, characterized in that: In step S3, the intermediate layer filtering algorithm is applied to the current and past... The state estimation sequence, residual sequence, and original measurement features of the inertial measurement unit within each time period are used as the time-series feature matrix. By fitting nonlinear residual mappings through long short-term memory networks or gated recurrent units temporal deep models, a systematic cumulative error compensation term is output online. The final corrected single-machine pose state is as follows: ；。 5. The method for collaborative positioning and swarm operation of dredging robots according to claim 1, characterized in that: In step S4, when the robot's satellite positioning unit signal is weak or lost, the robot shares the status information of other robots in the neighborhood through the communication network and uses the relative positioning unit to obtain relative observation values. They extracted local features of the riverbed topography detected by their respective sonar sensors, and performed local map matching using a normal distribution transformation algorithm to obtain relative pose constraints. A distributed backend framework based on factor graph optimization is constructed, and the optimization objective function for global consistency correction is: ； The Gauss-Newton method is used for iterative solution to achieve cluster-level error equalization suppression and global correction.

6. The method for collaborative positioning and swarm operation of dredging robots according to claim 1, characterized in that: In step S5, the sonar detection unit acquires raw underwater sonar data, which, after preprocessing, is then processed in parallel into three processing branches: The riverbed topography extraction branch extracts three-dimensional riverbed digital elevation features for terrain matching and positioning. The silt distribution identification branch identifies the silt layer thickness and physical boundaries, which is used to assist in path planning and work area division. The obstacle detection and recognition branch identifies the three-dimensional spatial boundaries of obstacles for obstacle avoidance and local path replanning. The information from the above branches is integrated to update the local environmental map in real time and fed back to the positioning and planning module.

7. The method for collaborative positioning and swarm operation of dredging robots according to claim 1, characterized in that: In step S6, the observation layer employs a residual gating technique based on Mahalanobis distance to statistically test the observation innovation of each sensor. The Mahalanobis distance calculation formula is as follows: ； in, The residual covariance matrix; Set the chi-square distribution threshold. ,like If the current observation is determined to be an outlier, the observation in that frame will be discarded immediately. like However, if the value shows a continuous increasing trend, adaptive noise estimation is triggered, which dynamically increases the observation noise covariance matrix of the sensor using the following formula. Reduce its weight coefficient in the fusion algorithm: ； in, For the filter residual, It is a forgetting factor.

8. The method for collaborative positioning and swarm operation of dredging robots according to claim 1, characterized in that: In step S6, the system layer sets up a positioning loss detection and state recovery mechanism: the system maintains a sliding window for... The historical health status circular buffer is used to determine the fully autonomous positioning loss state when all external observation sensors fail within a set time threshold. The state recovery program is then initiated, and the system automatically rolls back to the last determined healthy absolute pose state node in the buffer. It then switches to a dead reckoning degraded mode consisting of pre-integration by pure inertial measurement units and riverbed fluid dynamics constraints. The system also sends a cooperation request to neighboring robots in the cluster to reconstruct the communication route topology until it regains effective relative constraints and then performs backward smooth closed-loop correction.

9. The method for collaborative positioning and cluster operation of dredging robots according to claim 1, characterized in that, Multi-source sensing data acquisition includes: The inertial measurement unit acquires triaxial angular velocity and triaxial acceleration data at a high frequency of 200Hz; When there is no obstruction, the satellite positioning unit outputs absolute three-dimensional coordinates in the geodetic coordinate system at a frequency of 10Hz. and location covariance; The relative positioning unit obtains the relative distance and azimuth of other dredging robots in the vicinity at a frequency of 20Hz through ultra-wideband (UWB) wireless ranging or laser ranging. The sonar detection unit uses a multibeam sonar to acquire raw underwater acoustic images and depth point cloud data at a frequency of 5Hz.

10. A collaborative positioning and swarm operation system for dredging robots used to perform the method according to any one of claims 1 to 9, characterized in that, include: Multiple dredging robot units, each equipped with a dredging actuator, a propulsion device, and a control module; A multi-source sensing and positioning module, installed on the dredging robot unit, includes at least: a sonar detection unit, an inertial measurement unit, a satellite positioning unit, and a relative positioning unit; The multi-source positioning fusion module is used to fuse data acquired by the multi-source sensing and positioning modules using a hierarchical fusion architecture. The bottom layer uses inertial measurement data for high-frequency prediction, and the middle layer performs real-time fusion through filtering algorithms. The temporal error compensation module, connected to the multi-source localization fusion module, is used to perform residual modeling and online dynamic compensation for the systematic errors generated during the fusion process through a temporal depth model. The cluster collaborative positioning module is used to realize the sharing of observation information among multiple dredging robot units through the communication network, and to collaboratively correct the positioning results of each dredging robot unit based on relative ranging, local map matching and graph optimization constraints. The robust and fault-tolerant control module is used to suppress anomalies, restore the state, or reconstruct the positioning process in the event of sensor malfunction, communication interruption, or positioning failure, through the observation layer residual gating unit and the system layer lock-out recovery unit. The task scheduling and monitoring module is used to uniformly manage and schedule the operation status and location information of the dredging robot cluster.

Images