Washing, shearing, blowing and grabbing foreign matter removing method and system for overhead line system

Abstract

The invention relates to a cleaning, shearing, blowing and grabbing foreign matter removing method and system for a contact network, and the method comprises the steps: obtaining multi-modal data through an acquisition terminal, carrying out the cross-modal fusion of the multi-modal data, and generating three-dimensional environment representation data; extracting morphological features and texture features of the target foreign body, and generating joint features; matching a clearing strategy set according to the joint features; performing action sequence optimization on the clearing strategy set to obtain an optimized clearing strategy; constructing a digital twin system of the target contact network, and performing online optimization on the optimization strategy set to obtain a target clearing strategy; according to the cleaning, shearing, blowing and grabbing foreign matter removing method and system for the overhead line system, the problems that a traditional method is low in efficiency, poor in adaptability and high in safety risk are solved through multi-modal data fusion, feature extraction and strategy matching mechanisms in combination with a dynamic optimization algorithm and a digital twinborn system, and the method and system are suitable for large-scale popularization and application. The method has the advantages of improving the clearing efficiency, enhancing the adaptability, reducing the safety risk and realizing intelligent strategy selection.

Description

Technical Field

[0001] The present application relates to the technical field of robot control, and in particular to a method and system for cleaning foreign matter by washing, shearing, blowing and grabbing a contact network. Background Art

[0002] In electrified railway systems, the removal of foreign objects from the contact network is a crucial step in ensuring the safe operation of trains. Traditional technologies for removing foreign objects from the contact network rely primarily on manual inspections or single mechanical devices, such as robotic shearing and high-pressure airflow purging. Manual removal methods require power outages, resulting in slow response times and high safety risks. Existing mechanical devices, on the other hand, generally suffer from single functions and poor adaptability, and are unable to dynamically adjust removal strategies based on the material (e.g., plastic, metal, branches, etc.) and form (e.g., entanglement, suspension, etc.) of the foreign object. At the algorithmic level, existing technologies often rely on a single sensor for foreign object detection, resulting in a high false detection rate and a lack of the ability to coordinate and optimize multiple actuators, making it difficult to cope with the complex and changing contact network environment. Summary of the Invention

[0003] In order to solve the above-mentioned defects, the present application provides a method and system for cleaning foreign matter by washing, shearing, blowing and grabbing for a contact network.

[0004] The above-mentioned invention objective of this application is achieved through the following technical solutions:

[0005] A method for cleaning foreign matter in a contact network by washing, shearing, blowing and grabbing, comprising the steps of:

[0006] Acquire multimodal data through the acquisition terminal and perform cross-modal fusion based on spatiotemporal synchronization to generate three-dimensional environmental representation data;

[0007] The morphological and texture features of the target foreign object are extracted based on the 3D environment representation data, and joint features are generated through a feature fusion mechanism based on attention weights.

[0008] Match the cleanup policy set in the pre-built policy tree based on the joint features;

[0009] The action sequence of the clearing strategy set is optimized by a preset optimization algorithm to obtain an optimized clearing strategy;

[0010] A digital twin system of the target contact network is constructed, and the optimization strategy set is optimized online through a pre-set reinforcement learning algorithm to obtain the target clearance strategy.

[0011] Furthermore, the present application also proposes that: the multimodal data is acquired through the acquisition terminal, and cross-modal fusion based on spatiotemporal synchronization is performed to generate three-dimensional environment representation data, including:

[0012] Synchronously collecting multimodal data through an acquisition terminal, wherein the multimodal data includes visual data, point cloud data, and mechanical data;

[0013] Timestamp calibration is used to align the acquisition time points of multimodal data and generate an affine transformation matrix;

[0014] The affine transformation matrix is ​​solved based on the improved spatiotemporal ICP algorithm to spatially align multimodal data.

[0015] A spatiotemporal flow alignment model is established based on the multimodal data after spatial registration, and the multimodal data are fused to generate three-dimensional environment representation data.

[0016] Furthermore, the present application also proposes that the improved spatiotemporal ICP algorithm includes a catenary dynamic deformation compensation term.

[0017] Furthermore, the present application also proposes that: the three-dimensional environment representation data includes spatial geometry information, physical property information, and dynamic characteristic information; the method of extracting morphological features and texture features of the target foreign object based on the three-dimensional environment representation data, and generating joint features through a feature fusion mechanism based on attention weights, includes:

[0018] The spatial geometric information is processed through a convolutional neural network, and the morphological features of the target foreign matter are extracted;

[0019] Combining physical property information and dynamic characteristic information to construct multi-scale texture features;

[0020] A multi-level optimization mechanism based on attention weights is used to fuse morphological and texture features.

[0021] Furthermore, the present application also proposes that: matching the clearing policy set in the pre-built policy tree according to the joint feature includes:

[0022] Based on the texture characteristics of the target foreign matter, a material classification node is established in the strategy tree, and the material strategy branches are divided;

[0023] Construct morphological classification sub-nodes based on morphological features under each material strategy branch;

[0024] Associating a cleaning action for each morphological classification sub-node, wherein the cleaning actions include flushing, cutting, blowing, and grabbing;

[0025] The weight distribution of the strategy tree is dynamically updated through the Bayesian online learning mechanism.

[0026] Furthermore, the present application also proposes that: optimizing the action sequence of the purge strategy set by a preset optimization algorithm to obtain an optimized purge strategy includes:

[0027] Constructing an action parameter model of the actuator, wherein the action parameter model includes action time, energy consumption coefficient, and space constraint;

[0028] A mixed integer linear programming framework is constructed to optimize the action sequence of the cleaning strategy set with the optimization objectives of minimizing the total cleaning time and minimizing the energy consumption.

[0029] Multi-level constraints are set and a hierarchical solution strategy is used to correct the action sequence optimization results.

[0030] Furthermore, the present application also proposes that the multi-level constraint conditions include robot arm joint kinematic constraints, contact network dynamic deformation constraints, and multi-actuator collaborative operation constraints.

[0031] The second object of the present invention is achieved through the following technical solutions:

[0032] A foreign matter removal system for a contact network comprising:

[0033] The data fusion module is used to acquire multimodal data through the acquisition terminal and perform cross-modal fusion based on spatiotemporal synchronization to generate three-dimensional environment representation data;

[0034] The feature generation module is used to extract the morphological and texture features of the target foreign object based on the 3D environment representation data, and generate joint features through a feature fusion mechanism based on attention weights;

[0035] A policy matching module, used to match the cleanup policy set in a pre-built policy tree based on the joint features;

[0036] An action sequence optimization module is used to optimize the action sequence of the clearing strategy set through a preset optimization algorithm to obtain an optimized clearing strategy;

[0037] The clearing strategy optimization module is used to build a digital twin system of the target contact network and perform online optimization of the optimization strategy set through a preset reinforcement learning algorithm to obtain the target clearing strategy.

[0038] From the above, it can be seen that the present application provides a method and system for cleaning foreign matter by washing, cutting, blowing and grabbing for contact networks. Through multimodal data fusion, feature extraction and strategy matching mechanism, combined with dynamic optimization algorithm and digital twin system, it solves the problems of low efficiency, poor adaptability and high safety risks of traditional methods. It has the advantages of improving cleaning efficiency, enhancing adaptability, reducing safety risks and realizing intelligent strategy selection. BRIEF DESCRIPTION OF THE DRAWINGS

[0039] Figure 1 This is a flow chart of an embodiment of a method for cleaning foreign matter by washing, shearing, blowing and grabbing a contact network according to the present application;

[0040] Figure 2 This is a flow chart of step S10 in an embodiment of a method for cleaning foreign matter by washing, shearing, blowing and grabbing a contact network of the present application;

[0041] Figure 3 This is a flow chart of step S20 in an embodiment of a method for cleaning foreign matter by washing, shearing, blowing, and grabbing a contact network of the present application;

[0042] Figure 4 This is a flow chart of step S30 in an embodiment of a method for cleaning foreign matter by washing, shearing, blowing and grabbing a contact network of the present application;

[0043] Figure 5 This is a flow chart of step S40 in an embodiment of a method for removing foreign matter by washing, shearing, blowing and grabbing for a contact network of the present application. DETAILED DESCRIPTION

[0044] The technical solutions of this application will be described clearly and completely below, in conjunction with the accompanying drawings. It should be understood that the described embodiments represent only a portion of the embodiments of this application, and not all of them. The components of this application, generally described and illustrated in the drawings herein, may be arranged and designed in a variety of different configurations. Therefore, the following detailed description of the embodiments of this application provided in the drawings is not intended to limit the scope of the claimed application, but rather merely represents selected embodiments of this application. All other embodiments derived by those skilled in the art based on the embodiments of this application without inventive effort are within the scope of protection of this application. It should be noted that similar reference numerals and letters represent similar items in the following drawings. Therefore, once an item is defined in one drawing, it does not need to be further defined or explained in subsequent drawings. Furthermore, in the description of this application, the terms "first," "second," etc., are used solely to distinguish descriptions and should not be construed as indicating or implying relative importance.

[0045] Existing technologies typically remove foreign objects from electrified railway contact networks through manual inspections or single-function mechanical devices. Traditional manual methods require power outages, resulting in slow response times and potential safety hazards. Existing mechanical devices are unable to dynamically adjust strategies based on material differences and morphological changes. For example, metal entanglements and plastic hangings require different treatments, but traditional systems lack material recognition and morphological analysis capabilities. At the data acquisition level, single sensor detection is susceptible to environmental interference, resulting in a high rate of false positives. At the execution level, there is a lack of optimized multi-manipulator collaborative operations, making it difficult to cope with the complex spatial structure of the contact network.

[0046] To address these issues, those skilled in the art face three major technical bottlenecks: insufficient effectiveness of multi-sensor data fusion, difficulty in dynamically correlating foreign object features, and a lack of collaborative optimization of multiple actuators. The inventors have discovered that multimodal spatiotemporal synchronization can improve environmental perception accuracy, that feature fusion based on an attention mechanism can establish a correlation model between material and morphology, that a policy tree architecture enables dynamic policy matching, and that the combination of digital twins and reinforcement learning can overcome the risks of debugging physical devices.

[0047] Therefore, in this embodiment, the present application proposes a method for cleaning foreign matter by washing, shearing, blowing and grabbing the contact network, such as Figure 1 As shown, the steps include:

[0048] S10: Acquire multimodal data through the acquisition terminal and perform cross-modal fusion based on spatiotemporal synchronization to generate three-dimensional environment representation data;

[0049] S20: Extract the morphological and texture features of the target foreign object based on the 3D environment representation data, and generate joint features through a feature fusion mechanism based on attention weights;

[0050] S30: matching the clearing strategy set in the pre-built strategy tree according to the joint feature;

[0051] S40: Optimizing the action sequence of the clearing strategy set by using a preset optimization algorithm to obtain an optimized clearing strategy;

[0052] S50: Build a digital twin system of the target contact network, and perform online optimization on the optimization strategy set through a preset reinforcement learning algorithm to obtain the target clearance strategy.

[0053] In this embodiment, multimodal data includes visual (RGB-D), laser radar (LiDAR) point cloud and mechanical sensor data, the visual data provides texture color information (such as the color of foreign objects), the laser radar provides millimeter-level three-dimensional coordinates, and the mechanical sensor detects the tension change of the target contact network; the spatiotemporal synchronized cross-modal fusion refers to the spatial alignment of multi-source data through timestamp calibration and improved ICP algorithm, such as using dynamic deformation compensation items to eliminate the point cloud offset caused by contact network vibration; the three-dimensional environment characterization data includes spatial coordinates, material hardness and motion trend information, and specifically can generate a multi-resolution three-dimensional grid through the spatiotemporal flow alignment model; morphological features refer to the extracted visual features such as the geometric outline, size ratio, etc. of the target foreign object (such as the slender shape of the kite line); texture features refer to the extracted material features of the target foreign object; the feature fusion mechanism based on attention weight refers to the feature fusion mechanism that dynamically weights the feature contribution using a gating mechanism; the joint feature refers to the morphological The fusion result of contour and material attributes, such as the weighted fusion of geometric features and physical property features extracted by convolutional neural networks using a multi-head attention mechanism; the strategy tree refers to a decision structure with material classification nodes and morphological sub-nodes, which can be specifically constructed as a four-layer decision structure, with the top layer branching by material (metal / non-metal), the secondary layer subdivided by morphology (winding / hanging), and the final layer matching the "wash, cut, blow, grab" action combination; action sequence optimization refers to solving the optimal operation sequence through a preset optimization algorithm (such as a mixed integer programming model), such as setting the robot arm joint angle limit and energy consumption threshold as constraints; the digital twin system refers to a virtual simulation model of the contact network, which can be specifically built using the Unity3D engine to simulate the execution effects of different strategies; the preset reinforcement learning algorithm is a reinforcement learning algorithm preset for the corresponding digital twin system, which is used to perform online optimization of the optimization strategy set based on feedback, and the target clearance strategy is its output result.

[0054] Specifically, multimodal data of the target foreign object is synchronously collected through RGB-D cameras, lidars and mechanical sensors, and the collected multimodal data are aligned by timestamps; three-dimensional environmental representation data is generated by cross-modal fusion of multimodal data based on spatiotemporal synchronization; the morphological characteristics and texture characteristics of the target foreign object are extracted based on the three-dimensional environmental representation data, and the morphological characteristics and texture characteristics are adaptively weighted through the attention mechanism to generate a joint feature vector containing the morphological geometric dimensions and texture of the target foreign object; according to the joint features, several clearing strategies are matched in the pre-constructed strategy tree to form a clearing strategy set, and the clearing strategy set is optimized with respect to the action sequence through a pre-set optimization algorithm to obtain an optimized clearing strategy; after the digital twin system loads the optimized clearing strategy, it is adjusted online through a pre-set reinforcement learning algorithm to obtain the target clearing strategy.

[0055] Through the above technical solution, this application realizes the accurate identification of the material and shape of foreign objects, improves the adaptability of the removal strategy; shortens the operation time through multi-actuator collaborative optimization; uses digital twins to reduce the risk of equipment debugging; and finally completes the efficient and safe foreign object removal operation without power outage of the contact network.

[0056] In one embodiment, if Figure 2 As shown, step S10 includes:

[0057] S11: synchronously collecting multimodal data through an acquisition terminal, wherein the multimodal data includes visual data, point cloud data, and mechanical data;

[0058] S12: aligning the acquisition time points of the multimodal data using timestamp calibration and generating an affine transformation matrix;

[0059] S13: Solve the affine transformation matrix based on the improved spatiotemporal ICP algorithm to spatially align multimodal data;

[0060] S14: A spatiotemporal flow alignment model is established based on the multimodal data after spatial registration, and the multimodal data is fused to generate three-dimensional environment representation data.

[0061] In this embodiment, visual data refers to two-dimensional / three-dimensional image information acquired by an optical imaging device, which in this case refers to color images and depth maps synchronously acquired by an RGB-D camera. Point cloud data refers to a set of three-dimensional spatial coordinates acquired by LiDAR scanning, with each point containing geometric coordinates (accurately describing the three-dimensional structure of the foreign object and the contact network) and reflection intensity (reflecting the material's reflective properties to laser light). Mechanical data refers to physical quantities measured in real time by mechanical sensors, including three-dimensional force (the magnitude of the force exerted on the robotic arm when it contacts the foreign object) and three-dimensional torque (reflecting the tightness of the foreign object's entanglement). Timestamp calibration refers to triggering the synchronous data acquisition of multimodal sensors through hardware synchronization signals. Specifically, this can be achieved using a network time protocol or hardware trigger pulses to eliminate time deviations between different sensors. The affine transformation matrix refers to the linear transformation parameters that describe the spatial posture relationship of multimodal data. Specifically, it can be solved jointly by sensor calibration data and dynamic motion models to achieve cross-modal coordinate system alignment; the improved space-time ICP algorithm is an improved point cloud matching technology. It adds dynamic environment adaptability on the basis of the traditional iterative nearest point algorithm. The improved space-time ICP algorithm can automatically compensate for the point cloud deformation caused by the swing of the contact network, and filter out interference objects such as flying birds and fallen leaves to ensure that the alignment accuracy meets the requirements of high-speed rail scenarios; the space-time flow alignment model refers to a mathematical model that characterizes the dynamic characteristics of the contact network. By establishing a space-time correlation function, the multimodal data collected at different times are mapped to a unified space-time coordinate system. The space-time flow alignment model can predict the swing trajectory of the contact network and provide a dynamic reference benchmark for clearing operations.

[0062] Specifically, visual data is collected through RGB-D cameras to capture images of the contact network area, point cloud data is obtained through lidar to obtain three-dimensional geometric information, and mechanical data is measured through mechanical sensors to measure the force state of the end of the robotic arm; during the timestamp calibration process, the multimodal sensor receives a synchronous trigger signal to start data acquisition, ensuring that the visual frame, point cloud frame and mechanical sampling point have a unified time reference; the affine transformation matrix measures the relative position of each sensor through the calibration plate, and dynamically updates the spatial transformation parameters in combination with the robotic arm kinematic model; the improved spatiotemporal ICP algorithm can automatically compensate for the point cloud deformation caused by the swing of the contact network, and synchronously calculate the elastic deformation of the contact line under the action of load during point cloud alignment, thereby improving the alignment accuracy in dynamic environments; the spatiotemporal flow alignment model projects the aligned multimodal data onto a unified three-dimensional grid, and generates environmental representation data containing spatial geometry, material properties and dynamic characteristics through feature-level fusion.

[0063] Through the above technical solution, this application solves the problem of environmental modeling errors caused by spatiotemporal inaccuracies in multi-sensor data, and realizes accurate three-dimensional characterization of foreign objects in the contact network; multimodal data fusion enhances the perception ability of foreign material and geological characteristics, providing data support for subsequent material classification and removal strategy selection, and effectively improving the reliability and adaptability of foreign object removal in complex environments.

[0064] In one embodiment, the improved spatiotemporal ICP algorithm includes a catenary dynamic deformation compensation term.

[0065] Among them, the spatiotemporal ICP algorithm refers to an algorithm that realizes the spatial registration of multimodal data based on the iterative nearest point principle. It can be achieved by iteratively calculating the optimal rigid body transformation matrix, which is used to eliminate the spatial posture differences between multi-sensor data; the dynamic deformation compensation item of the contact network refers to the correction item that incorporates the real-time deformation parameters of the contact network wire under wind and temperature changes. It can be achieved by constructing a deformation model based on dynamic displacement measurement data of strain sensors, which is used to offset the impact of the dynamic deformation of the contact network on the accuracy of spatial registration.

[0066] Specifically, during the spatial registration of multimodal data, the real-time deformation of the contact network conductor under the influence of the external environment will cause dynamic deviations between the point cloud data and the visual data. By introducing deformation compensation terms in the transformation matrix solution process of the traditional ICP algorithm, the deformation displacement field is constructed using the real-time strain data of the contact network conductor, and the spatial coordinates of the point cloud data are dynamically corrected. For example, when the contact network conductor is bent by wind, the deformation compensation term performs reverse displacement compensation on the point cloud coordinates according to the displacement field model, so that the visual data and the point cloud data can still maintain spatial registration consistency in a dynamic environment.

[0067] Through the above technical solution, the present application can eliminate the multi-sensor data registration error caused by the dynamic deformation of the contact network, improve the spatial consistency of the three-dimensional environment characterization data, and provide an accurate spatial geometric information basis for subsequent foreign body feature extraction.

[0068] In one embodiment, the three-dimensional environment representation data includes spatial geometric information, physical attribute information, and dynamic characteristic information, such as Figure 3 As shown, step S20 includes the steps of:

[0069] S21: Process the spatial geometric information through a convolutional neural network and extract the morphological features of the target foreign body;

[0070] S22: Combine physical property information and dynamic characteristic information to construct multi-scale texture features;

[0071] S23: A multi-level optimization mechanism based on attention weights is used to fuse morphological features and texture features.

[0072] In this embodiment, spatial geometric information refers to the shape, size and spatial position data of the foreign body obtained by a three-dimensional scanning device, which can be implemented by a laser radar or a structured light sensor to determine the geometric distribution characteristics of the foreign body in the contact network; physical property information refers to the material density, elastic modulus and surface friction coefficient of the foreign body collected by the sensor, which can be implemented by a multi-spectral imager or a contact mechanical sensor to determine the material category of the foreign body; dynamic characteristic information refers to the vibration frequency, displacement fluctuation and deformation trend of the foreign body obtained by time series data analysis, which can be implemented by an inertial measurement unit or a high-speed camera to capture the dynamic behavior of the foreign body under the action of wind; convolution A neural network refers to a deep learning model with a residual connection structure, which can be implemented using the VGGNet or ResNet architecture to extract morphological features with translation invariance from three-dimensional point cloud data; multi-scale texture features refer to the fusion of texture, hardness and adhesion information at different resolutions, which can be implemented through wavelet transform or multi-layer feature pyramid construction to enhance the robustness of foreign material characterization; a multi-level optimization mechanism based on attention weights refers to the dynamic allocation of importance weights of feature channels through the self-attention module, which can be implemented using the multi-head attention structure in the Transformer model to eliminate redundant information and strengthen the relevance of key features.

[0073] Specifically, in the process of foreign body feature extraction, the three-dimensional point cloud data collected by the lidar is first input into the convolutional neural network, and the contour, volume and spatial distribution characteristics of the foreign body are extracted through the convolution layer and pooling layer; at the same time, the material parameters obtained by the mechanical sensor and the dynamic vibration data captured by the high-speed camera are used to generate texture features including hardness, elasticity and motion trajectory through the multi-scale feature extraction module; then, the morphological features and texture features are input into the attention weight calculation module, and the fusion weight is automatically assigned according to the correlation between the feature channels, and finally a joint feature vector is generated that can simultaneously reflect the geometric shape, physical properties and dynamic behavior of the foreign body.

[0074] Through the above technical solution, this application solves the problem of misjudgment of foreign body materials caused by insufficient data from a single sensor in the existing technology, and improves the ability to capture the dynamic deformation of entangled foreign bodies; by combining the multi-dimensional feature fusion of spatial geometry, physical properties and dynamic behavior, the subsequent strategy matching module can accurately distinguish the removal requirements of foreign bodies of different materials and shapes, and provide reliable feature input for the optimization of action sequences such as flushing and shearing.

[0075] In one embodiment, if Figure 4 As shown, step S30 includes the steps of:

[0076] S31: establishing a material classification node in the strategy tree based on the texture characteristics of the target foreign matter, and dividing the material strategy branches;

[0077] S32: constructing a morphological classification sub-node based on morphological features under each material strategy branch;

[0078] S33: Associating a cleaning action with each morphological classification sub-node, wherein the cleaning action includes flushing, cutting, blowing, and grabbing;

[0079] S34: Dynamically update the weight distribution of the strategy tree through the Bayesian online learning mechanism.

[0080] In this embodiment, the material classification node refers to the strategy branch root node established based on the texture characteristics of the foreign matter. Specifically, it can be implemented by a classification model based on a support vector machine, and is used to divide the foreign matter into different material types such as metal, plastic, and fiber, so as to match the corresponding material processing strategy; the morphology classification sub-node refers to the strategy sub-node established based on the geometric morphology of the foreign matter. Specifically, it can be implemented by features such as contour and entanglement extracted by a convolutional neural network, and is used to further distinguish the foreign matter morphologies such as hanging, wrapping, and entanglement under the same material branch; clearing action association refers to binding preset physical operation instructions to each morphology sub-node. For example, the shearing action can be configured as the cutting angle and force parameters of the robotic arm, thereby achieving accurate mapping of the action strategy and the foreign matter characteristics; the Bayesian online learning mechanism refers to updating the strategy tree node weights based on the feedback data after the action is executed. Specifically, it can be implemented by the Markov chain Monte Carlo method, and is used to dynamically adjust the priority of material and morphology classification.

[0081] Exemplarily, the clearing action associated with each morphological classification sub-node includes:

[0082] The cleaning action of flushing is associated with fixed foreign objects (such as the remains of bird nests), the cleaning action of blowing is associated with fixed foreign objects (such as bird nests), the cleaning action of cutting is associated with entangled foreign objects (such as kite strings), and the cleaning action of grabbing is associated with light foreign objects (such as plastic bags).

[0083] Specifically, the strategy tree transforms the texture and morphological characteristics of foreign matter into a hierarchical decision path through a multi-level node structure. When a foreign object is detected, the system first matches the corresponding strategy branch based on its material characteristics. For example, metal material triggers the wear-resistant tool selection branch, and then selects the cutting or grasping action within the branch based on the morphological characteristics. The cleaning action associated with each morphological sub-node defines the operating mode of the actuator through a preset parameter library. For example, the purge action can be set to a specific air pressure range and nozzle angle. The Bayesian mechanism continuously collects data such as the amount of foreign matter residue and execution time after the action is executed, and adjusts the weight distribution of each node by calculating the posterior probability. For example, when a certain material classification is frequently misjudged, the priority of the node in the strategy tree is automatically reduced.

[0084] Through the above technical solution, this application realizes the refined matching and dynamic optimization of foreign body removal strategies, effectively improving the accuracy of foreign body processing with different material and shape combinations; specifically, the system can automatically select the optimal action sequence, reduce secondary removal operations caused by improper strategies, and continuously adapt to new types of foreign bodies appearing in the contact network environment through online learning, thereby reducing the frequency of manual intervention in strategy adjustments.

[0085] In one embodiment, if Figure 5 As shown, step S40 includes the steps of:

[0086] S41: constructing an action parameter model of the actuator, wherein the action parameter model includes action time, energy consumption coefficient, and space constraint;

[0087] S42: Construct a mixed integer linear programming framework to optimize the action sequence of the cleaning strategy set with the optimization objectives of minimizing the total cleaning time and minimizing the energy consumption;

[0088] S43: Set multi-level constraints and use a hierarchical solution strategy to correct the action sequence optimization results.

[0089] In this embodiment, the actuator is a device used to execute the target clearing strategy, which may include mechanical shears, blowers, grippers, flushing machines, etc., and may also include any combination of these devices; the motion parameter model refers to a mathematical model used to quantify the actuator motion efficiency and resource consumption, which can be specifically achieved by combining the spatiotemporal characteristic parameters of the robot arm joint motion trajectory with the measured data of the energy consumption sensor, and by establishing a mapping relationship between time, energy and spatial position, a quantitative basis is provided for subsequent optimization; the mixed integer linear programming framework refers to a mathematical optimization model that integrates continuous variables and discrete variables, which can be specifically achieved by combining the branch and bound algorithm with the linear relaxation technology, and can simultaneously handle the discrete selection of action sequence and the continuous optimization problem of parameter adjustment; multi-level constraints refer to a set of constraints divided by priority, and the hierarchical processing of constraints ensures that the optimization results meet the actual operation requirements.

[0090] Specifically, first, a correlation model between action time, energy consumption coefficient and spatial position is established based on the kinematic parameters of the actuator and historical operation data, such as the correspondence between the path planning time of the end effector of the robotic arm and the power consumption of the joint motor; then, a dual-objective optimization problem including the time objective function and the energy consumption objective function is constructed, and the weighted summation method is used to convert the multiple objectives into a single objective form. For example, the time weight coefficient is set to 0.6 and the energy consumption coefficient is set to 0.4 to reflect the operation priority; then, according to the physical characteristics of the constraint conditions, it is divided into a kinematic constraint layer, a structural safety layer and a collaborative operation layer. For example, in the kinematic constraint layer, the joint angular velocity is limited to not exceed the threshold, in the structural safety layer, the upper limit of the contact wire offset is set, and in the collaborative operation layer, the working distance between the purge device and the robotic arm is specified; finally, a hierarchical iterative optimization method is adopted to first solve the feasible solution that meets the high-level constraints, and then add low-level constraints layer by layer for correction.

[0091] Through the above technical solution, the present application effectively solves the problem of single goal and constraint conflict during action sequence optimization in the prior art, and can balance operating efficiency and energy consumption while ensuring the safety of the contact network structure; by constructing a multi-level constraint system, it avoids optimization failures caused by conflicting constraints, such as automatically switching to a low-disturbance operation mode when the conductor swing amplitude approaches a threshold; this solution can adapt to the dynamically changing contact network environment, for example, when the conductor position changes during foreign object removal, it can update the constraints in real time and re-optimize the action sequence.

[0092] In one embodiment, the multi-level constraint conditions include robot arm joint kinematic constraints, contact network dynamic deformation constraints, and multi-actuator collaborative operation constraints.

[0093] In this embodiment, the kinematic constraint of the manipulator joint refers to limiting the movement angle and speed of each joint of the manipulator, which can be achieved by using an inverse kinematics algorithm combined with joint limit parameters to prevent the manipulator from exceeding the physical movement range during the clearing operation; the dynamic deformation constraint of the contact network refers to limiting the deformation amount of the contact network wire under the action of the manipulator, which can be achieved by using a finite element model to calculate the deformation threshold in real time to avoid irreversible deformation of the contact network due to external forces; the collaborative operation constraint of multiple actuators refers to limiting the spatiotemporal synchronization of the actions of multiple actuators, which can be achieved by using a time window allocation algorithm combined with a spatial obstacle avoidance model to eliminate action conflicts when multiple devices work together.

[0094] Specifically, within the framework of mixed integer linear programming, the constraints are processed step by step through a hierarchical solution strategy. First, the kinematic constraints of the manipulator joints are applied to determine the feasibility of the basic action sequence, such as calculating whether the joint motion trajectory satisfies the angle restriction through the Jacobian matrix. Then, the dynamic deformation constraints of the contact network are superimposed, such as using an elastic mechanics model to predict the deformation response of the contact network under different external forces, to ensure that the clearing action does not cause the wire deviation to exceed the safety threshold. Finally, multi-actuator collaborative operation constraints are introduced, such as coordinating the action timing of the flushing nozzle and the shearing manipulator through a spatiotemporal trajectory planning algorithm to avoid high-pressure water flow interfering with the positioning accuracy of the manipulator. The constraints at each level are coupled through weighted coefficients to form a comprehensive optimization objective function.

[0095] Through the above technical solution, this application effectively solves the problem of motion interference during collaborative operation of multiple actuators, avoids equipment failure of the robotic arm due to joint overlimit, and maintains the stability of the contact network structure through dynamic deformation constraints; the integrated optimization of multi-level constraints enables the removal strategy to not only meet the operating efficiency requirements, but also adapt to the dynamically changing working environment of the contact network, significantly improving the success rate of foreign body removal under complex working conditions.

[0096] In one embodiment, a system for cleaning, shearing, blowing, and grabbing foreign matter for a contact network is provided. The system corresponds to the method for cleaning, shearing, blowing, and grabbing foreign matter for a contact network in the above embodiment. The system comprises:

[0097] The data fusion module is used to acquire multimodal data through the acquisition terminal and perform cross-modal fusion based on spatiotemporal synchronization to generate three-dimensional environment representation data;

[0098] The feature generation module is used to extract the morphological and texture features of the target foreign object based on the 3D environment representation data, and generate joint features through a feature fusion mechanism based on attention weights;

[0099] A policy matching module, used to match the cleanup policy set in a pre-built policy tree based on the joint features;

[0100] An action sequence optimization module is used to optimize the action sequence of the clearing strategy set through a preset optimization algorithm to obtain an optimized clearing strategy;

[0101] The clearing strategy optimization module is used to build a digital twin system of the target contact network and perform online optimization of the optimization strategy set through a preset reinforcement learning algorithm to obtain the target clearing strategy.

[0102] The specific definition of a system for cleaning, shearing, blowing, and grabbing foreign matter removal for a contact network can be found in the definition of a method for cleaning, shearing, blowing, and grabbing foreign matter removal for a contact network described above and will not be repeated here. Each module in the aforementioned system for cleaning, shearing, blowing, and grabbing foreign matter removal for a contact network can be implemented in whole or in part through software, hardware, or a combination thereof. Each of these modules can be embedded in or independent of a processor in a computer device in hardware form, or stored in a computer device memory in software form, allowing the processor to call and execute the corresponding operations of each module.

[0103] The above-described embodiments are only used to illustrate the technical solutions of the present application, rather than to limit them. Although the present application has been described in detail with reference to the aforementioned embodiments, those skilled in the art should understand that they can still modify the technical solutions described in the aforementioned embodiments, or make equivalent replacements for some of the technical features therein. These modifications or replacements do not deviate the essence of the corresponding technical solutions from the spirit and scope of the technical solutions of the various embodiments of the present application, and should all be included in the scope of protection of the present application.

Claims

1. A method for cleaning foreign matter by washing, shearing, blowing and grabbing a contact network, characterized by: Including steps: Acquire multimodal data through the acquisition terminal and perform cross-modal fusion based on spatiotemporal synchronization to generate three-dimensional environmental representation data; The morphological and texture features of the target foreign object are extracted based on the 3D environment representation data, and joint features are generated through a feature fusion mechanism based on attention weights. Match the cleanup policy set in the pre-built policy tree based on the joint features; The action sequence of the clearing strategy set is optimized by a preset optimization algorithm to obtain an optimized clearing strategy; A digital twin system of the target contact network is constructed, and the optimization strategy set is optimized online through a pre-set reinforcement learning algorithm to obtain the target clearance strategy.

2. A method for cleaning foreign matter by washing, shearing, blowing and grabbing for a contact network according to claim 1, characterized in that: The multimodal data is acquired through the acquisition terminal, and cross-modal fusion based on spatiotemporal synchronization is performed to generate three-dimensional environment representation data, including: Synchronously collecting multimodal data through an acquisition terminal, wherein the multimodal data includes visual data, point cloud data, and mechanical data; Timestamp calibration is used to align the acquisition time points of multimodal data and generate an affine transformation matrix; The affine transformation matrix is ​​solved based on the improved spatiotemporal ICP algorithm to spatially align multimodal data. A spatiotemporal flow alignment model is established based on the multimodal data after spatial registration, and the multimodal data are fused to generate three-dimensional environment representation data.

3. The method for cleaning foreign matter by washing, shearing, blowing and grabbing for a contact network according to claim 2, characterized in that: The improved spatiotemporal ICP algorithm includes a catenary dynamic deformation compensation term.

4. The method for cleaning foreign matter by washing, shearing, blowing and grabbing for a contact network according to claim 1, characterized in that: The three-dimensional environment representation data includes spatial geometry information, physical property information, and dynamic characteristic information. The method of extracting morphological features and texture features of the target foreign object based on the three-dimensional environment representation data and generating joint features through a feature fusion mechanism based on attention weights includes: The spatial geometric information is processed through a convolutional neural network, and the morphological features of the target foreign matter are extracted; Combining physical property information and dynamic characteristic information to construct multi-scale texture features; A multi-level optimization mechanism based on attention weights is used to fuse morphological and texture features.

5. The method for cleaning foreign matter by washing, shearing, blowing and grabbing for a contact network according to claim 1, characterized in that: The matching of the clearing policy set in the pre-built policy tree according to the joint feature includes: Based on the texture characteristics of the target foreign matter, a material classification node is established in the strategy tree, and the material strategy branches are divided; Construct morphological classification sub-nodes based on morphological features under each material strategy branch; Associating a cleaning action for each morphological classification sub-node, wherein the cleaning actions include flushing, cutting, blowing, and grabbing; The weight distribution of the strategy tree is dynamically updated through the Bayesian online learning mechanism.

6. The method for cleaning foreign matter by washing, shearing, blowing and grabbing for a contact network according to claim 1, characterized in that: The process of optimizing the action sequence of the purge strategy set by using a preset optimization algorithm to obtain an optimized purge strategy includes: Constructing an action parameter model of the actuator, wherein the action parameter model includes action time, energy consumption coefficient, and space constraint; A mixed integer linear programming framework is constructed to optimize the action sequence of the cleaning strategy set with the optimization objectives of minimizing the total cleaning time and minimizing the energy consumption. Multi-level constraints are set and a hierarchical solution strategy is used to correct the action sequence optimization results.

7. A method for cleaning foreign matter by washing, shearing, blowing and grabbing for a contact network according to claim 6, characterized in that: The multi-level constraint conditions include robot arm joint kinematic constraints, contact network dynamic deformation constraints, and multi-actuator collaborative operation constraints.

8. A system for cleaning, shearing, blowing and grabbing foreign matter for contact wire, characterized by: include: The data fusion module is used to acquire multimodal data through the acquisition terminal and perform cross-modal fusion based on spatiotemporal synchronization to generate three-dimensional environment representation data; The feature generation module is used to extract the morphological and texture features of the target foreign object based on the 3D environment representation data, and generate joint features through a feature fusion mechanism based on attention weights; A policy matching module, used to match the cleanup policy set in a pre-built policy tree based on the joint features; An action sequence optimization module is used to optimize the action sequence of the clearing strategy set through a preset optimization algorithm to obtain an optimized clearing strategy; The clearing strategy optimization module is used to build a digital twin system of the target contact network and perform online optimization of the optimization strategy set through a preset reinforcement learning algorithm to obtain the target clearing strategy.

Images