Intelligent cleaning device and intelligent cleaning method for box girder side form

Abstract

The invention relates to the technical field of bridge construction, in particular to an intelligent cleaning device and method for a box girder side formwork. The device comprises a mobile platform, a control device, a mechanical arm assembly, a cleaning device, an infrared induction sensor, a dust collection device, a multispectral visual recognition module, a 3D environment modeling module, a path planning module, a cleaning quality detection module and a self-adaptive learning database. The multispectral visual recognition module recognizes the stain type and the adhesion strength. The 3D environment modeling module is used for constructing a three-dimensional point cloud model of a working area in real time; the path planning module generates an optimal cleaning path based on a reinforcement learning algorithm; the cleaning quality detection module detects residual stains and triggers secondary cleaning; the adaptive learning database stores historical cleaning data and optimizes a cleaning strategy through a machine learning model. The full-process automatic cleaning operation of the box girder side form can be achieved, and the comprehensive functions of dirt recognition, autonomous path planning, efficient cleaning, real-time detection and the like are achieved.

Description

Technical Field

[0001] The patent of this invention relates to the field of bridge construction technology, specifically an intelligent cleaning device and intelligent cleaning method for box girder side formwork. Background Art

[0002] In bridge construction, box girder side forms are crucial for forming concrete structures. The cleanliness of their surfaces directly impacts the quality and appearance of the concrete components. Currently, cleaning box girder side forms is largely manual, primarily using brushing and water spraying to remove contaminants such as cement residue, oil stains, and rust. However, this traditional cleaning method is not only inefficient and labor-intensive, but also results in inconsistent cleaning results due to haphazard manual operation. In severe cases, this can affect the formwork's service life and even the surface quality of the finished concrete, posing a safety hazard. Furthermore, bridge construction often takes place outdoors, and manual cleaning poses significant safety risks due to the complex and changing working environment and the risks of working at height. Furthermore, different stain types require varying cleaning intensity and detergents, making it difficult to achieve a balanced cleaning efficiency with a single cleaning method. The lack of real-time monitoring during the cleaning process often leads to incomplete cleaning or over-cleaning, impacting construction progress and wasting resources. With the development of intelligent manufacturing and artificial intelligence technology, how to introduce advanced technologies such as multispectral visual recognition, 3D modeling, autonomous navigation and deep learning into the field of bridge construction and build an intelligent cleaning system integrating recognition, decision-making, execution and feedback has become a key issue that the industry urgently needs to break through.

[0003] The intelligent cleaning device for box girder side formwork in bridge construction projects proposed in the present invention is aimed at the above-mentioned practical pain points. By integrating intelligent units such as multi-spectral recognition module, 3D point cloud environment modeling, cleaning path planning, cleaning quality detection and adaptive optimization strategy, it realizes intelligent control of the entire cleaning operation process, greatly improves the efficiency, safety and intelligence level of the cleaning operation, fills the shortcomings of existing technologies in "fine identification-adaptive cleaning-intelligent feedback", and has significant engineering practical value and promotion prospects. Summary of the Invention

[0004] The purpose of the present invention is to provide an intelligent cleaning device and method for box girder side formwork in bridge construction projects. The device and method can realize the full-process automatic cleaning operation of the box girder side formwork, and have comprehensive functions such as dirt identification, autonomous path planning, efficient cleaning, real-time detection and intelligent optimization. It significantly improves the efficiency, intelligence level and environmental adaptability of the cleaning operation, reduces manual dependence, and reduces resource waste and environmental pollution. The technical solution is as follows:

[0005] An intelligent cleaning device for a box girder side formwork, comprising:

[0006] a mobile platform configured with drive wheels;

[0007] a control device mounted on the mobile platform;

[0008] A robotic arm assembly, comprising a telescopic rod, a first robotic arm, and a second robotic arm, wherein the robotic arm assembly is connected to the mobile platform via a slewing support plate, and can achieve 360-degree rotation and multi-angle adjustment;

[0009] A cleaning device, installed at the end of the second robotic arm, including an infrared sensor, a high-pressure nozzle and a cleaning rotor, wherein the cleaning rotor is driven by a motor built into the second robotic arm;

[0010] The infrared sensor is used to detect the degree of dirt on the template surface and feed it back to the control device to dynamically adjust the speed of the cleaning rotor and the spray volume of the cleaning liquid;

[0011] The dust suction device operates synchronously with the cleaning device and comprises a dust suction air inlet, a telescopic hose and a dust collection box. The dust suction air inlet is connected to the dust collection box through the telescopic hose.

[0012] Based on the above technical solution, the box girder side form intelligent cleaning device also includes:

[0013] A multispectral visual recognition module, integrated into the end of the second robotic arm, includes visible light, infrared, and ultraviolet imaging units, and is used to capture multispectral images of the template surface and identify the type and adhesion strength of the stain;

[0014] The 3D environment modeling module, mainly composed of lidar and stereo cameras, is used to build a 3D point cloud model of the work area in real time;

[0015] Path planning module, which generates the optimal cleaning path based on reinforcement learning algorithm;

[0016] The cleaning quality detection module detects residual stains through image segmentation algorithms and triggers secondary cleaning;

[0017] Adaptive learning database that stores historical cleaning data and optimizes cleaning strategies through machine learning models.

[0018] On the basis of the above technical solution, the high-pressure nozzle is respectively connected to the cleaning agent storage box and the release agent storage box through liquid inlet pipes.

[0019] Furthermore, the intelligent cleaning device for the box girder side mold also includes a debris collection port, which is arranged below the cleaning rotor and is connected to the dust collection box through a telescopic hose.

[0020] An intelligent cleaning method for a box girder side formwork intelligent cleaning device comprises the following steps:

[0021] S1 Operation Preparation and Environmental Modeling

[0022] By integrating the navigation and positioning system with the inertial measurement unit, the initial position and heading angle of the cleaning device are determined. The mobile platform adjusts its position according to the coordinates of the target area. The cleaning module is aligned with the surface of the box girder side formwork by rotating the slewing bearing plate and adjusting the telescopic rod. A 3D point cloud model is constructed through simultaneous scanning with a lidar and a binocular camera. The plane equation, center coordinates, and normal vector of the box girder side formwork are extracted through voxel grid filtering and denoising, ICP algorithm registration, and RANSAC algorithm fitting. Obstacles in non-target areas are identified and a 3D bounding box is generated, which is used as obstacle data for path planning.

[0023] S2 multispectral visual recognition and stain analysis

[0024] After completing the 3D modeling of the work area and determining the work posture, the device enters the target area pollution identification stage. The multispectral visual recognition module collects and analyzes images of the box girder side mold surface to identify the type, distribution location and adhesion strength of the stains.

[0025] S3 reinforcement learning-driven path planning and cleaning execution

[0026] The path planning module uses a reinforcement learning path optimization method to autonomously generate high-efficiency, low-energy cleaning paths. This path planning method combines the spatial distribution of pollution points, pollution intensity, cleaning costs, and operation sequence to minimize operation time and resource waste, and trains the optimal strategy network.

[0027] S4 Cleaning Quality Monitoring and Secondary Response

[0028] The cleaning quality detection module is activated to perform a panoramic rescan of the cleaned area and detect residual stains using the U-Net image segmentation algorithm, enabling automatic evaluation and compensatory cleaning control.

[0029] S5 adaptive database design

[0030] Construct an adaptive learning database module, which uses a cloud database as a carrier and combines machine learning models to continuously and dynamically model the mapping relationship between stain identification, cleaning parameter configuration and environmental variables, thereby realizing intelligent optimization and automatic prediction capabilities of cleaning strategies.

[0031] Based on the above technical solution, step S1 is specifically as follows:

[0032] S101 system startup and positioning initialization

[0033] After receiving the work instruction, the control device activates the main control chip, which in turn wakes up the mobile platform, various functional modules, and the robotic arm control unit. It then establishes a remote data link with the central dispatch platform via the 5G communication module to obtain the latest task list and work area number from the construction site. Subsequently, the four-wheel drive mobile platform starts and moves toward the target area under the guidance of the navigation control module. The navigation module obtains the device's current position information and motion status through the integration of the navigation positioning system and the inertial measurement unit, forming a basic navigation state vector:

[0034] X t =[x t ,y t ,θ t ] T

[0035] Among them, x t ,y t is the position coordinate of the platform center in the ground coordinate system; θ t is the heading angle the platform is facing. After comparing the current position of the platform with the coordinates of the target area, the platform gradually moves into the preset work area along the shortest path;

[0036] S102 working posture adjustment and alignment

[0037] When the platform reaches the front of the box beam side formwork, it stops automatically and enters the fine adjustment stage. In order to make the cleaning device face the formwork surface, the slewing bearing plate (12) on the top of the platform starts the rotation function to adjust the direction of the upper structure. The rotation angle is calculated by the following formula:

[0038]

[0039] Among them, (x c ,y c ) is the reference point coordinate of the target template, (x p ,y p ) is the current coordinate of the platform. θ d Indicates the platform's rotation angle toward the template. After turning, the platform activates the telescopic rod and lifts the cleaning module to the appropriate operating height based on the measured template height. The first and second robotic arms use the current coordinate system as a reference and use the inverse kinematics algorithm to align the end cleaning head with the center of the template surface, adjusting the angle to maintain perpendicularity to the contact surface.

[0040] S103 3D Mapping and Template Recognition

[0041] The system starts the laser radar (LiDAR) and binocular stereo camera module to perform synchronous scanning and build a three-dimensional scene point cloud model P = {p i}, p i =(x i ,y i , z i, I i ), i is the point cloud data index, I i Represents the laser reflection intensity. The original point cloud data is filtered and denoised by voxel grid to form a compressed point cloud set P′, P′=voxelGrid(P, l), where l is the voxel size, which determines the compression accuracy. The ICP (Iterative Closest Point) algorithm is used to align the multi-frame point clouds to construct a stable and continuous 3D scene model. Combined with the pre-imported template standard size parameter width W t , height H t , the long plane in the point cloud is fitted by the RANSAC algorithm to extract the accurate position and posture of the box girder side form. The plane equation is as follows:

[0042] Ax+By+Cz+D=0

[0043] This plane is the candidate template surface to be identified. A, B, C, and D are plane coefficients. The system compares them with the standard template width W. t , height H t For comparison:

[0044] |W o -W t |<ε w ,|H o -H t |<ε h

[0045] Among them, W o , H o is the actual width and height of the fitting plane, ε w , ε h is the error tolerance. If the tolerance condition is met, the plane is set as the cleaning target area R c , and extract its center coordinates P c =(x c ,y c , z c ) and normal vector n, which will be used to calculate the end-of-arm working posture. In order to align the robotic arm with the center of the template, the platform needs to be rotated to the orientation angle θ d Otherwise, it will be regarded as an abnormal structure or misidentified plane, and the system will prompt manual confirmation or rescan;

[0046] Calculate the end position that the robot arm should reach: X c =P c +n·d safe

[0047] Among them, d safe Indicates the safe distance between the end and the template to ensure that no mechanical interference occurs during visual identification or cleaning.

[0048] Based on the above technical solution, step S1 further includes the following steps:

[0049] S104 Obstacle Identification and Avoidance Mapping

[0050] In order to avoid collisions in the subsequent path planning process, the system performs clustering and bounding box extraction on the point cloud part of the non-target area, identifies potential obstacles, and uses the Euclidean clustering method to extract cluster units C from the point cloud. k , generate the minimum three-dimensional bounding box B for each cluster unit k , and record its spatial position in the local map, all {B k The obstacle data will be stored in the system map and used as a path loss factor to influence the clean path optimization strategy;

[0051] S105 multi-source perception data output

[0052] After completing the above environmental modeling, the system will package and output the following data for the next stage of multispectral recognition module to call: template area boundary size and orientation information, template center point P c With the normal vector n, cleaning task posture target point X c , 3D point cloud model of the working area, obstacle list B k , current working coordinate frame X t .

[0053] Based on the above technical solution, step S2 includes the following steps:

[0054] S201 multispectral image acquisition

[0055] The multispectral visual recognition module is installed at the end of the second robotic arm and includes three types of imaging units: visible light camera Capturing color and texture; infrared thermal imager Reflecting surface temperature distribution; UV imaging equipment Enhance the detection capability of metal rust and chemical residues. To ensure the stability of the image and the coverage of the field of view, the system uses the cleaning target center point P c As the center of the circle, set the scanning radius r s With the angular resolution Δθ, multi-view images are collected along the arc trajectory to determine the multispectral shooting path;

[0056] θ i =θ d +i·Δθ,i=-N,...,N

[0057] P i =P c +R(θ i )·n·d img

[0058] Among them, R(θ i ) is a two-dimensional rotation matrix; d img is the shooting distance, and d safe Equal; P i Indicates the imaging position of each frame; N represents half of the number of shooting frames. This method ensures that multi-angle and multi-spectral images cover the complete template area R c ;

[0059] The device collects multi-channel images in sequence along the shooting path:

[0060]

[0061] These images serve as the input basis for pollution identification analysis;

[0062] S202 Multispectral Image Fusion and Feature Extraction

[0063] The image I collected in S201 i , in order to unify the feature scale of multi-channel images, the system performs i Perform image registration and feature fusion operations.

[0064] Specifically, since visible light, infrared, and ultraviolet imaging devices have different viewing angles and resolutions, the system must first spatially align the three images to ensure semantic consistency at corresponding pixels. The system uses an affine registration model for alignment:

[0065]

[0066] Among them, A ir and A uv is the image affine matrix, b ir and b uv is the translation vector. and is the image after registration.

[0067] Secondly, perform image channel normalization. The three images have different imaging mechanisms, and their brightness ranges and response frequencies vary greatly, so each image channel needs to be normalized uniformly:

[0068]

[0069] In the formula, k∈{vis, ir, uv} represents the channel type; μ k and σ k Represents the pixel mean and standard deviation of channel k, I′ k (x, y) is the final normalized image. Normalization ensures that the contributions of different channels to the fusion feature tensor are balanced.

[0070] Finally, multi-spectral fusion feature construction is performed. The three normalized images are stacked to construct the fusion feature map F i :

[0071]

[0072] in, and Represents the image gradient and performs edge enhancement on the image. The final F i Will be used as input for subsequent CNN models;

[0073] S203 3D stain mapping and spatial positioning

[0074] First, the fusion feature map F obtained in S202 is i Input to the feature extraction network f θ (·), and obtain the high-dimensional semantic feature representation of the image:

[0075] z=f θ (F i )

[0076] Here, θ represents the model parameters of the neural network, and z is the feature vector. The system identifies the stain type c (oil stain, cement residue, and rust) through the classification output branch. It uses a fully connected network and the Softmax function to build a multi-class discriminant model and output the probability distribution of each stain type:

[0077] P p =softmax(W p z+b p )

[0078] Among them, P p Represents the probability vector of stain classification, W p and b p are the weight and bias parameters of the classification layer respectively. After the Softmax output, the system determines the stain type c of the area based on the category label corresponding to the maximum probability value:

[0079] c=argmaxP p

[0080] At the same time, the adhesion strength estimation branch adopts a regression structure and uses the Sigmoid activation function to normalize the output result to obtain the predicted value of the stain adhesion strength:

[0081] S s =σ(W s z+b s )

[0082] Where W s and b sis the weight and bias of the regression layer, S s ∈[0, 1] represents the normalized adhesion strength. The larger the value, the more difficult it is to clean. After completing the joint identification of the stain type and adhesion strength, the system dynamically generates the cleaning parameter instructions for the corresponding area based on the identification results. The system has a built-in control parameter mapping function β, whose input is the stain type c and adhesion strength S s , the output is a triplet of cleaning parameters.

[0083] (p, v, η) = β(c, S s )

[0084] In the formula, p represents the injection pressure of the high-pressure nozzle, in bar; v is the rotation speed of the cleaning brush, in RPM; η is the detergent concentration, which is the ratio of active ingredients in the cleaning solution and has a value range of [0,1].

[0085] Finally, the system associates the image area recognition results with the spatial point cloud data. Through the obtained three-dimensional projection transformation relationship, the recognition results of the corresponding pixels in the fused image are mapped to the pollution point set in the three-dimensional point cloud coordinate system. And attach the corresponding stain type, adhesion strength and cleaning parameters to each spatial point:

[0086]

[0087] This set of contaminated points will serve as the core input of the path planning and control system, and will be used in subsequent steps to perform cleaning path priority sorting, robot arm trajectory planning, and cleaning motion control processes.

[0088] Based on the above technical solution, step S5 includes:

[0089] First, after executing each round of cleaning tasks, the system will fully record all key data related to the environment, identification, control, and quality assessment, and construct a structured operation sample set D, whose basic unit is a single contaminated area cleaning data entry d i :

[0090] d i =[c,S s , T i , H i ,θ di , p i , v i , η i , P Pi ]

[0091] Among them, T i , H i These data are uploaded to the cloud database D cloud, and form a long-term accumulated data set D = [d1, d2, ..., d n ].

[0092] Secondly, the system builds a multi-objective regression model with the goal of predicting the optimal cleaning control parameters for a given stain type and operating environment:

[0093]

[0094] Among them, the input features are c, S s , T i , H i ,θ di The output is the optimal water pressure recommended for the corresponding pollution area. Brush speed Detergent concentration

[0095] Finally, in order to enhance the system's ability to respond to dynamic environmental factors in the cleaning task, especially under the influence of factors such as pollution intensity fluctuations, humidity changes, or working angle adjustments that may occur over time in the box girder side formwork area, a long short-term memory neural network is used as the strategy parameter prediction model.

[0096] Based on the above technical solution, step S5 includes:

[0097] In order to ensure that the model not only gives reasonable prediction values but also guides actual cleaning effects and resource optimization, a composite loss function is designed:

[0098]

[0099] In the formula, represents the residual rate under the model prediction conditions, To minimize the error between the predicted residual rate and the actual cleaning effect; To constrain energy consumption; Represents the cost of cleaning fluid usage. t1, t2, t3 are the adjustment weights of the three objectives

[0100] The beneficial effects of the present invention are as follows: the device reduces the time and intensity of manual operation through the automation system, and ensures the efficiency and quality of template cleaning; the main body of the device has a simple structure, is driven by four moving wheels, and can move flexibly between different working areas; the slewing support plate can achieve 360-degree rotation, and the control device can accurately control the movement of the telescopic rod and the mechanical arm to adapt to the cleaning needs of templates at different heights and angles, and has a wide range of applications; the cleaning device is equipped with an infrared sensing sensor, which can automatically adjust the rotation speed of the cleaning panel and the spray amount of the cleaning liquid according to the amount of dirt on the template surface, reducing resource waste; the dust collection device and the cleaning device operate synchronously, and can timely collect dust and debris generated during the cleaning process, reducing environmental pollution and secondary pollution. At the same time, by integrating a multispectral visual recognition module, the device has the ability to accurately identify the type and adhesion strength of stains, and can intelligently determine cleaning modes and parameters. With the help of 3D modeling and path planning modules, the device can perceive complex on-site environments and perform dynamic path optimization, prioritizing heavily polluted areas and reducing energy consumption. After cleaning is completed, the visual inspection module conducts real-time analysis of the cleaning effect to ensure that it meets preset standards. In addition, by building an adaptive learning database, the device can continuously accumulate data from different operating scenarios and dynamically adjust cleaning strategies using machine learning methods, gradually realizing a cleaning operation mode with "fewer parameter settings and strong autonomous decision-making." The overall system has comprehensive advantages such as high intelligence, high operating efficiency, excellent cleaning results, strong safety and environmental adaptability, and has broad prospects for engineering applications. BRIEF DESCRIPTION OF THE DRAWINGS

[0101] Figure 1 The overall schematic diagram of the device

[0102] Figure 2 Enlarged view of the cleaning device

[0103] Figure 3 Front view of the device

[0104] Figure 4 Top view of the device

[0105] Figure 5 Side view of the device

[0106] Figure 6 Schematic diagram of the robotic arm

[0107] Figure 7 Complete flow chart of the system

[0108] Figure 8 CNN network stain recognition architecture diagram

[0109] Figure 9 Flowchart of cleaning trajectories for reinforcement learning DETAILED DESCRIPTION

[0110] The present invention will be further described below with reference to the accompanying drawings and examples:

[0111] In the present invention, unless otherwise expressly specified or limited, terms such as "mounted," "connected," "connect," and "fixed" should be understood broadly. For example, they may refer to fixed connection, detachable connection, or integration; they may refer to direct connection or indirect connection through an intermediate medium; they may refer to internal communication between two components or interaction between two components. Those skilled in the art will understand the specific meanings of the above terms in the present invention based on specific circumstances.

[0112] In the description of the present invention, it should be understood that the terms "center", "length", "up", "down", "front", "back", "left", "right", "vertical", "horizontal", "top", "bottom", "inside", etc., indicating the orientation or positional relationship, are based on the orientation or positional relationship shown in the accompanying drawings, and are only for the convenience of describing the present invention and simplifying the description, rather than indicating or implying that the device or element referred to must have a specific orientation, be constructed and operated in a specific orientation, and therefore cannot be understood as limiting the present invention. In addition, the terms "first" and "second" are only used for descriptive purposes and cannot be understood as indicating or implying relative importance or implicitly indicating the number of the indicated technical features. Therefore, the features defined as "first" and "second" may explicitly or implicitly include one or more of the features. In the description of the present invention, unless otherwise specified, "multiple" means two or more.

[0113] In the present invention, unless otherwise expressly specified or limited, a first feature being "above" or "below" a second feature may include the first and second features being in direct contact, or may include the first and second features being in contact not directly but through another feature between them. Furthermore, a first feature being "above," "above," and "above" a second feature may include the first feature being directly above or obliquely above the second feature, or may simply mean that the first feature is higher in level than the second feature. A first feature being "below," "below," and "below" a second feature may include the first feature being directly below or obliquely below the second feature, or may simply mean that the first feature is lower in level than the second feature.

[0114] like Figures 1 to 6 As shown, an intelligent cleaning device for box girder side formwork includes:

[0115] A mobile platform 1 is provided with drive wheels 7;

[0116] A control device 2, which is installed on the mobile platform 1 and is used to coordinate the overall operation and operation process of the device;

[0117] The robotic arm assembly includes a telescopic rod 4, a first robotic arm 5, and a second robotic arm 6. The robotic arm assembly is connected to the mobile platform 1 via a slewing support plate 12, and can achieve 360-degree rotation and multi-angle adjustment;

[0118] A cleaning device is installed at the end of the second robotic arm 6, including an infrared sensor 14, a high-pressure nozzle 18 and a cleaning rotor 8. The cleaning rotor 8 is driven by a motor built into the second robotic arm 6;

[0119] The infrared sensor 14 is used to detect the degree of dirt on the template surface and feed it back to the control device 2 to dynamically adjust the speed of the cleaning rotor 8 and the spray volume of the cleaning liquid;

[0120] The dust suction device operates synchronously with the cleaning device, and includes a dust suction air inlet 9, a telescopic hose 10 and a dust collection box 11. The dust suction air inlet 9 is connected to the dust collection box 11 through the telescopic hose 10 and is used to collect dust generated during the cleaning process.

[0121] The cleaning rotor 8 is mainly composed of a cleaning panel 20 and cleaning brushes 19 evenly distributed on the cleaning panel 20. The high-pressure nozzle 18 passes through the center of the cleaning rotor 8. The driving wheel 7 of the mobile platform 1 is a four-wheel drive structure, and a slewing bearing 21 is provided between the mobile platform 1 and the slewing support plate 12 to support the omnidirectional rotation of the superstructure.

[0122] The control device 2 can transmit signals through the integrated line pipe 3 to coordinate the movement, cleaning and dust collection of the robot arm.

[0123] Preferably, the box girder side form intelligent cleaning device further comprises:

[0124] A multispectral visual recognition module, integrated at the end of the second robotic arm 6, comprising visible light, infrared, and ultraviolet imaging units, for collecting multispectral images of the template surface and identifying the type and adhesion strength of stains;

[0125] The 3D environment modeling module, mainly composed of lidar and stereo cameras, is used to build a 3D point cloud model of the work area in real time;

[0126] Path planning module, which generates the optimal cleaning path based on reinforcement learning algorithm;

[0127] The cleaning quality detection module detects residual stains through image segmentation algorithms and triggers secondary cleaning;

[0128] Adaptive learning database that stores historical cleaning data and optimizes cleaning strategies through machine learning models.

[0129] The high-pressure nozzle 18 is connected to the cleaning agent storage box 15 and the release agent storage box 16 through the liquid inlet pipe 17, so as to realize the switching spraying of the cleaning agent and the release agent.

[0130] Furthermore, the intelligent cleaning device for the beam side mold may further include a debris collection port 13 , which is disposed below the cleaning rotor 8 and connected to the dust collection box 11 through a telescopic hose 10 .

[0131] A fan may be built into the dust box 11 , and partitions may be provided in the dust box 11 , wherein the two partitions are respectively connected to the debris collection port 13 and the dust suction air inlet 9 , and are used for absorbing debris and dust respectively.

[0132] like Figures 7 to 9 As shown, an intelligent cleaning method for an intelligent cleaning device for a box girder side formwork comprises the following steps:

[0133] S1 Operation Preparation and Environmental Modeling

[0134] This step primarily implements initial identification, equipment deployment, and 3D environmental perception of the box girder side formwork cleaning operation area, ensuring the stable operation of the cleaning device in complex construction sites and providing basic spatial data support for subsequent identification, planning, and operations.

[0135] S2 multispectral visual recognition and stain analysis

[0136] After completing the 3D modeling of the work area and determining the work posture, the device enters the target area contamination identification phase. Using the multispectral visual recognition module, the device collects and analyzes images of the box girder side mold surface to identify the type, distribution location, and adhesion strength of the stains. The identification results provide a decision-making basis for subsequent cleaning parameter adjustment and path planning.

[0137] S3 reinforcement learning-driven path planning and cleaning execution

[0138] After identifying and estimating the parameters of each contamination point on the box girder side formwork surface, the path planning module autonomously generates a high-efficiency, low-energy cleaning path through a reinforcement learning path optimization method. This path planning method combines the spatial distribution of contamination points, contamination intensity, cleaning costs, and operation sequence to train an optimal strategy network with the goal of minimizing operation time and resource waste.

[0139] This reinforcement learning path planning module ensures that the system can not only reasonably sort the cleaning areas under irregular pollution distribution, but also dynamically adjust the execution strategy according to pollution intensity and cleaning consumption, to achieve energy-saving and efficient differentiated cleaning task planning.

[0140] S4 Cleaning Quality Monitoring and Secondary Response

[0141] After completing the cleaning path based on reinforcement learning planning, to ensure that the box girder side formwork surface meets the preset cleaning standards, the present invention will activate the system's cleaning quality detection module, perform a panoramic rescan of the cleaned area, and detect residual stains using the U-Net image segmentation algorithm to achieve automatic evaluation and compensatory cleaning control.

[0142] S5 adaptive database design

[0143] In order to improve the system's generalization ability under multiple working conditions and multiple stain types, and gradually reduce the degree of manual intervention, the present invention constructs an adaptive learning database module after all operation processes are completed. This module uses a cloud database as a carrier and combines a machine learning model to continuously dynamically model the mapping relationship between stain identification, cleaning parameter configuration and environmental variables, thereby realizing intelligent optimization and automatic prediction capabilities of cleaning strategies.

[0144] The step S1 may specifically include the following steps:

[0145] S101 system startup and positioning initialization

[0146] After receiving the work instruction, the control device 2 activates the main control chip, wakes up the mobile platform 1, various functional modules, and the robotic arm control unit, and establishes a remote data link with the central dispatch platform via the 5G communication module to obtain the latest task list and work area number of the construction site. Subsequently, the four-wheel drive mobile platform starts and moves towards the target area under the guidance of the navigation control module. The navigation module obtains the current position information and motion status of the device through the fusion of the navigation positioning system (GNSS) and the inertial measurement unit (IMU), forming a basic navigation state vector:

[0147] X t =[x t ,y t ,θ t ] T

[0148] Among them, x t ,y t is the position coordinate of the platform center in the ground coordinate system; θ t is the heading angle the platform is facing. After comparing the current position of the platform with the coordinates of the target area, the platform gradually moves into the preset work area along the shortest path;

[0149] S102 working posture adjustment and alignment

[0150] When the platform reaches the front of the box girder side formwork, it stops automatically and enters the fine adjustment stage. In order to make the cleaning device face the formwork surface, the slewing bearing plate 12 on the top of the platform starts the rotation function to adjust the direction of the superstructure. The rotation angle is calculated by the following formula:

[0151]

[0152] Among them, (x c ,y c ) is the reference point coordinate of the target template, (x p ,y p ) is the current coordinate of the platform. θ d Indicates the rotation angle of the platform facing the template. After the rotation is completed, the platform activates the telescopic rod 4 and lifts the cleaning module to the appropriate working height according to the measured template height. The first and second robotic arms 5 and 6 use the current coordinate system as a reference and use the inverse kinematic solution algorithm to align the end cleaning head with the center of the template surface. The angle is adjusted to keep it perpendicular to the contact surface to ensure uniform force during cleaning.

[0153] S103 3D Mapping and Template Recognition

[0154] The system starts the laser radar (LiDAR) and binocular stereo camera module to perform synchronous scanning and build a three-dimensional scene point cloud model P = {p i}, p i =(x i ,y i , z i , I i ), i is the point cloud data index, I i Represents the laser reflection intensity. The original point cloud data is filtered and denoised by voxel grid to form a compressed point cloud set P′, P′=voxelGrid(P, l), where l is the voxel size, which determines the compression accuracy. The ICP (Iterative Closest Point) algorithm is used to align the multi-frame point clouds to construct a stable and continuous 3D scene model. Combined with the pre-imported template standard size parameter width W t , height H t , the long plane in the point cloud is fitted by the RANSAC algorithm to extract the accurate position and posture of the box girder side form. The plane equation is as follows:

[0155] Ax+By+Cz+D=0

[0156] This plane is the candidate template surface to be identified. A, B, C, and D are plane coefficients. The system compares them with the standard template width W. t , height H t For comparison:

[0157] |W o -W t |<ε w ,|H o -H t |<ε h

[0158] Among them, W o, H o is the actual width and height of the fitting plane, ε w , ε h is the error tolerance. If the tolerance condition is met, the plane is set as the cleaning target area R c , and extract its center coordinates P c =(x c ,y c , z c ) and normal vector n, which will be used to calculate the end-of-arm working posture. In order to align the robotic arm with the center of the template, the platform needs to be rotated to the orientation angle θ d Otherwise, it will be regarded as an abnormal structure or misidentified plane, and the system will prompt manual confirmation or rescan;

[0159] Calculate the end position that the robot arm should reach: X c =P c +n·d safe

[0160] Among them, d safe Indicates the safe distance between the end and the template to ensure that no mechanical interference occurs during visual identification or cleaning;

[0161] S104 Obstacle Identification and Avoidance Mapping

[0162] In order to avoid collisions in the subsequent path planning process, the system performs clustering and bounding box extraction on the point cloud part of the non-target area, identifies potential obstacles, and uses the Euclidean clustering method to extract cluster units C from the point cloud. k , generate the minimum three-dimensional bounding box B for each cluster unit k , and record its spatial position in the local map, all {B k The obstacle data will be stored in the system map and used as a path loss factor to influence the clean path optimization strategy;

[0163] S105 multi-source perception data output

[0164] After completing the above environmental modeling, the system will package and output the following data for the next stage of multispectral recognition module to call, template area boundary size and orientation information, template center point P c With the normal vector n, cleaning task posture target point X c , 3D point cloud model of the working area, obstacle list B k , current working coordinate frame X t The cleaning device realizes core initialization tasks such as platform positioning, template alignment, working posture adjustment, environmental structure recognition and obstacle mapping, and is in a ready state.

[0165] Furthermore, step S2 includes the following steps:

[0166] S201 multispectral image acquisition

[0167] The multispectral visual recognition module is installed at the end of the second robotic arm 6 and includes three types of imaging units: visible light camera Capturing color and texture; infrared thermal imager Reflecting surface temperature distribution; UV imaging equipment Enhance the detection capability of metal rust and chemical residues. To ensure the stability of the image and the coverage of the field of view, the system uses the cleaning target center point P c As the center of the circle, set the scanning radius r s With the angular resolution Δθ, multi-view images are collected along the arc trajectory to determine the multispectral shooting path;

[0168] θ i =θ d +i·Δθ,i=-N,...,N

[0169] P i =P c +R(θ i )·n·d img

[0170] Among them, R(θ i ) is a two-dimensional rotation matrix; d img is the shooting distance, and d safe Equal; P i Indicates the imaging position of each frame; N represents half of the number of shooting frames. This method ensures that multi-angle and multi-spectral images cover the complete template area R c ;

[0171] The device collects multi-channel images in sequence along the shooting path:

[0172]

[0173] These images serve as the input basis for pollution identification analysis.

[0174] S202 Multispectral Image Fusion and Feature Extraction

[0175] The image I collected in S201 i , in order to unify the feature scale of multi-channel images, the system performs i Perform image registration and feature fusion operations.

[0176] Specifically, since visible light, infrared, and ultraviolet imaging devices have different viewing angles and resolutions, the system must first spatially align the three images to ensure semantic consistency at corresponding pixels. The system uses an affine registration model for alignment:

[0177]

[0178] Among them, A ir and A uv is the image affine matrix, b ir and b uv is the translation vector. and is the image after registration.

[0179] Secondly, perform image channel normalization. The three images have different imaging mechanisms, and their brightness ranges and response frequencies vary greatly, so each image channel needs to be normalized uniformly:

[0180]

[0181] In the formula, k∈{vis, ir, uv} represents the channel type; μ k and σ k Represents the pixel mean and standard deviation of channel k, I′ k (x, y) is the final normalized image. Normalization ensures that the contributions of different channels to the fusion feature tensor are balanced.

[0182] Finally, multi-spectral fusion feature construction is performed. The three normalized images are stacked to construct the fusion feature map F i :

[0183]

[0184] in, and Represents the image gradient, performs edge enhancement on the image, and the final F i Will be used as input for subsequent CNN models;

[0185] S203 3D stain mapping and spatial positioning

[0186] After completing the multispectral image acquisition and spatial registration processing, the system obtains the standardized fusion image tensor F i. The fusion reflects the texture, thermal characteristics and material reflection information of the template surface in multiple bands. The fused feature map is used as the perception input of this system, and further image analysis and pollution identification are performed through the convolutional neural network CNN. In order to achieve accurate identification and differentiated processing of stains on the template surface, the present invention designs a multi-task recognition architecture based on deep neural network. The architecture consists of a shared backbone feature extraction network and two parallel output branches, which are used for stain type classification and adhesion strength estimation respectively. The present invention collected a total of 12,000 multispectral image data, covering three main types of dirt: 4,000 oil stain images; 4,000 cement residue images; and 4,000 rust images. The dataset is divided into training set, validation set and test set in a ratio of 8:1:1.

[0187] First, the fusion feature map F obtained in S202 is i Input to the feature extraction network f θ (·), and obtain the high-dimensional semantic feature representation of the image:

[0188] z=f θ (F i )

[0189] Here, θ represents the model parameters of the neural network, and z is the feature vector. The system identifies the stain type c (oil stain, cement residue, and rust) through the classification output branch. It uses a fully connected network and the Softmax function to build a multi-class discriminant model and output the probability distribution of each stain type:

[0190] P p =softmax(W p z+b p )

[0191] Among them, P p Represents the probability vector of stain classification, W p and b p are the weight and bias parameters of the classification layer respectively. After the Softmax output, the system determines the stain type c of the area based on the category label corresponding to the maximum probability value:

[0192] c=argmaxP p

[0193] At the same time, the adhesion strength estimation branch adopts a regression structure and uses the Sigmoid activation function to normalize the output result to obtain the predicted value of the stain adhesion strength:

[0194] S s =σ(W s z+b s )

[0195] Where W sand b s is the weight and bias of the regression layer, S s ∈[0, 1] represents the normalized adhesion strength. The larger the value, the more difficult it is to clean. After completing the joint identification of the stain type and adhesion strength, the system dynamically generates the cleaning parameter instructions for the corresponding area based on the identification results. The system has a built-in control parameter mapping function β, whose input is the stain type c and adhesion strength S s , the output is a triplet of cleaning parameters.

[0196] (p, v, η) = β(c, S s )

[0197] In the formula, p represents the injection pressure of the high-pressure nozzle, in bar; v is the rotation speed of the cleaning brush, in RPM; η is the detergent concentration, which is the ratio of active ingredients in the cleaning solution and has a value range of [0,1].

[0198] Finally, the system associates the image area recognition results with the spatial point cloud data. Through the obtained three-dimensional projection transformation relationship, the recognition results of the corresponding pixels in the fused image are mapped to the pollution point set in the three-dimensional point cloud coordinate system. And attach the corresponding stain type, adhesion strength and cleaning parameters to each spatial point:

[0199]

[0200] This set of contaminated points will serve as the core input of the path planning and control system, and will be used in subsequent steps to perform cleaning path priority sorting, robot arm trajectory planning, and cleaning motion control processes.

[0201] Preferably, step S3 includes the following steps:

[0202] S301 Reinforcement Learning Path Optimization Modeling

[0203] The cleaning path planning task is modeled using a reinforcement learning framework. The entire cleaning area is discretized into a grid G, and the optimal path strategy is learned in this grid environment by interacting with the environment.

[0204] First, all pollution points are identified according to the Calculate its cleaning priority score for reinforcement learning initialization strategy guidance, and define the priority cost function as:

[0205]

[0206] in, Indicates pollution point The cleaning priority score of the cleaning process is as follows: the larger the value, the higher the cleaning priority; λ1, λ2, and λ3 represent weighting coefficients, which are used to adjust the influence ratio of type, intensity, and location factors; wc (C) represents the stain type weight function, which assigns different cleaning difficulty levels to different stain types; is the numerical representation of the pollution heat map at this point, which is derived from the cumulative results of multiple frames of images;

[0207] Set the state space S: each state s t ∈S represents the current location information of the cleaning device in the grid and the uncleaned pollution map state:

[0208] s t =(g t , h t )

[0209] Among them, g t Indicates the current grid coordinates; h t Represents the current pollution heat map, i.e., the pollution intensity matrix, which changes dynamically during the path execution;

[0210] Action space a t ∈A: The action space is a moving instruction in eight directions, A = {up, down, left, right, upper left, upper right, lower left, lower right}, each action makes the agent move from the current position g t Migrate to the adjacent grid.

[0211] State transfer function: Path planning is based on the current state of the agent by defining a function, s t+1 =T(s t , a t ). In cleaning path planning, the transfer function updates the heat map in the following way:

[0212]

[0213] Where, γ c Indicates unit cleaning efficiency.

[0214] Reward function R(s t , a t ): The reward function is designed to encourage the system to prioritize visiting highly polluted areas, avoid repeated movements, and punish invalid or redundant paths. It is defined as follows:

[0215] R(s t , a t )=β1h t (g t )-β2C(a t )-β3D(s t )

[0216] In the formula, h t (g t) represents the pollution intensity at the current location, which is used to encourage people to go to high-pollution areas; C(a t ) is the action cost, i.e., angle change and turning penalty; D(s t ) is the redundant access penalty, that is, the current position has been cleaned, then it is set to 1; β1, β2, β3 are the weighted coefficients of each item in the reward function, which are set according to the on-site optimization; R(s t , a t ) is in state s t Execute action a t Immediate rewards after

[0217] S302 Action-value function learning and trajectory generation mapping

[0218] The system learns the optimal action value function Q(s) through reinforcement learning t , a t ):

[0219] Q(s t , a t )←Q(s t , a t )+α[R(s t , a t )+δmaxQ(s t+1 , a t )-Q(s t , a t )]

[0220] In the formula, Q(s t , a t ) represents the long-term value of the state-action pair; α represents the learning rate, which controls the update speed; δ represents the discount factor, which is used to balance the current and future rewards; maxQ(s t+1 , a t ) represents the value of the optimal behavior in the next state and is used to guide policy improvement.

[0221] Finally, the optimal strategy is obtained:

[0222] π * (s) = argmaxQ(s, a)

[0223] represents the optimal action a to be taken in any state s.

[0224] Furthermore, after the agent completes the planning, it will follow the optimal strategy path π * Derive the clean grid access sequence:

[0225] ρ=[g1,g2,...,g T ]

[0226] Then, the grid points are converted into trajectory points X in the platform coordinate system through the geometric projection function. e (t), this trajectory is fed into the low-level control system to drive the movement of the robotic arm and the action of the actuator.

[0227] S303 Dynamic parameter call and execution feedback

[0228] Every time the proxy object visits a grid point, the system reads the pollution identification result associated with that point The cleaning parameters (p, v, η) are respectively transferred to the high-pressure water pump control module (to adjust the nozzle pressure p), the brush driver, and the liquid mixing controller. During the operation, the feedback data status S is collected for dynamic monitoring and strategy correction.

[0229] The step S4 is as follows: a visible light camera installed at the end of the robot arm is used to scan the working area R after the cleaning execution track is completed. c Perform multi-angle coverage shooting to construct a high-resolution stitching image. Suppose the shooting path consists of N shooting points P i Composition, the image set is represented as:

[0230] I scan ={I1, I2…, I N}

[0231] All images are spliced after SIFT feature point registration and perspective transformation to obtain the panoramic image I f , will I f As input, pixel-level residual pollution area identification is performed to generate the residual stain mask map M seg Considering that the box girder side formwork has a regular surface contour but a large amount of interference details (formwork screw holes, reinforcement rib boundaries), the system uses a multi-scale convolution kernel combination (3*3, 5*5) in the U-Net encoding path to enhance the network's perception of microstructures.

[0232] In the encoder part, the input image I f First, we enter the encoder stage, where each layer extracts the features after spatial compression:

[0233] F (0) =I f

[0234] For the l-th layer encoder, the calculation is as follows:

[0235] F (l) =MaxPool(σ(w (l) *F (l-1) +b (l) )

[0236] In the formula, F (l) represents the output feature map of the lth layer; w(l) represents the convolution kernel of the lth layer; b (l) represents the bias vector; σ is the ReLU activation function; and MaxPool is a spatial downsampling operation using a 2x2 kernel. This process extracts semantic information from images at different scales, such as the texture differences of oil stain spread, rust edges, and dried cement clumps.

[0237] In the decoder part, starting from the deepest layer l, the spatial size is restored layer by layer and shallow layer features are fused:

[0238]

[0239] Among them, Up is the upsampling operation, which serves as deconvolution; Indicates channel dimension splicing; and Represent the decoder convolution parameters respectively; Represents the decoded output of layer l. The upsampled feature map is concatenated with skip connections and then convolved to capture both edge and high-semantic features.

[0240] Furthermore, the decoder output enters the prediction layer, and a 1*1 convolution is used to map the feature map into a single-channel probability map:

[0241]

[0242] In the formula, P P (x, y)∈(0, 1) represents the predicted probability that pixel (x, y) is “residual contamination”; σ is the Sigmoid activation function. The segmentation mask is generated by the thresholding function:

[0243]

[0244] Where θ∈[0.4, 0.6] represents the decision threshold. seg It is the output binary residual pollution mask map.

[0245] Finally, the model outputs the mask M set It is directly used as the basis for subsequent contaminated area calculation, cleaning qualification judgment, back-projection point extraction and secondary cleaning path replanning, and the results are fed back to S3 to realize a closed-loop response mechanism, so that the image input and subsequent path control can achieve logical closure and engineering control.

[0246] Further, the step S5 includes:

[0247] First, after executing each round of cleaning tasks, the system will fully record all key data related to the environment, identification, control, and quality assessment, and construct a structured operation sample set D, whose basic unit is a single contaminated area cleaning data entry d i :

[0248] d i =[c,S s , T i , H i ,θ di , p i , v i , η i , P Pi ]

[0249] Among them, T i , H i These data are uploaded to the cloud database D cloud , and form a long-term accumulated data set D = [d1, d2, ..., d n ].

[0250] Secondly, the system builds a multi-objective regression model with the goal of predicting the optimal cleaning control parameters for a given stain type and operating environment:

[0251]

[0252] Among them, the input features are c, S s , T i , H i ,θ di The output is the optimal water pressure recommended for the corresponding pollution area. Brush speed Detergent concentration To ensure that the model not only gives reasonable predictions but also guides actual cleaning results and resource optimization, the system designs a composite loss function:

[0253]

[0254] In the formula, represents the residual rate under the model prediction conditions, To minimize the error between the predicted residual rate and the actual cleaning effect; To constrain energy consumption; represents the cost of cleaning fluid usage. t1, t2, and t3 are the adjustment weights of the three objectives.

[0255] Finally, to enhance the system's responsiveness to dynamic environmental factors during cleaning tasks, particularly those affecting the box girder side formwork area, such as fluctuations in contamination intensity, humidity changes, and adjustments to operating angles, a long short-term memory neural network (LSM) was employed as a strategy parameter prediction model. This network can model the dependencies between time series features, enabling adaptive prediction of cleaning control parameters based on historical trajectories.

[0256] Specifically, before executing each contaminated area cleaning subtask, the system extracts k samples from the area and its adjacent historical areas and constructs the time series input matrix Y k ∈D. Input matrix Y k Input into a single-layer LSTM network to extract the sequence feature state:

[0257] l i =LSTM(Y k ; τ)

[0258] In the formula, τ represents the LSTM network weight, l i is the final hidden state of the current sequence. This hidden state is sent to the fully connected output layer and mapped to the predicted value of the control parameter of the current region:

[0259]

[0260] Among them, w out and b out The output is the weight matrix and bias term respectively. The output is the recommended cleaning nozzle pressure for the current area. Brush speed Detergent concentration

[0261] Furthermore, the system uses a structured sample set from a historical clean database for training. The loss function is the composite loss function L. After training, the model is deployed to the edge inference module and regularly fine-tuned and iterated with new datasets in the cloud.

[0262] Furthermore, after completing the S2 pollution area identification and the S3 path planning, the system calls the state vector [c, S s , T i , H i ,θ di ], and combined with adjacent historical states to form the input sequence d i , call the deployed LSTM model for prediction, and the prediction result is directly issued as a control instruction. After the cleaning process is completed, the system uses the S4 image segmentation mask M seg Calculate the residual contamination rate of the current area, and the system records the current prediction and execution results as a complete data entry d i And upload to the cloud database D cloud It is used for subsequent model accuracy evaluation and sample accumulation.

[0263] The present invention has been described above by way of examples, but the present invention is not limited to the above specific embodiments. Any changes or modifications based on the present invention fall within the scope of protection claimed by the present invention.

Claims

1. An intelligent cleaning device for box girder side formwork, characterized in that: include: A mobile platform (1) equipped with drive wheels (7); A control device (2) mounted on the mobile platform (1); A robotic arm assembly comprises a telescopic rod (4), a first robotic arm (5) and a second robotic arm (6); the robotic arm assembly is connected to a mobile platform (1) via a slewing support plate (12) and can achieve 360-degree rotation and multi-angle adjustment; A cleaning device is installed at the end of the second robotic arm (6), comprising an infrared sensor (14), a high-pressure nozzle (18) and a cleaning rotor (8), wherein the cleaning rotor (8) is driven to rotate by a motor built into the second robotic arm (6); The infrared sensor (14) is used to detect the degree of dirt on the template surface and feed back to the control device (2) to dynamically adjust the rotation speed of the cleaning rotor (8) and the spray volume of the cleaning liquid; A dust collection device operates synchronously with the cleaning device and comprises a dust collection air inlet (9), a telescopic hose (10) and a dust collection box (11); the dust collection air inlet (9) is connected to the dust collection box (11) via the telescopic hose (10).

2. The device according to claim 1, characterized in that Also includes: A multispectral visual recognition module, integrated at the end of the second robotic arm (6), comprising visible light, infrared and ultraviolet imaging units, for collecting multispectral images of the template surface and identifying the type of stain and its adhesion strength; The 3D environment modeling module, mainly composed of lidar and stereo cameras, is used to build a 3D point cloud model of the work area in real time; Path planning module, which generates the optimal cleaning path based on reinforcement learning algorithm; The cleaning quality detection module detects residual stains through image segmentation algorithms and triggers secondary cleaning; Adaptive learning database that stores historical cleaning data and optimizes cleaning strategies through machine learning models.

3. The device according to claim 1 or 2, characterized in that The high-pressure nozzle (18) is connected to the cleaning agent storage box (15) and the release agent storage box (16) respectively through the liquid inlet pipe (17).

4. The device according to claim 3, characterized in that It also includes a debris collection port (13), which is arranged below the cleaning rotor (8) and connected to the dust collection box (11) via a telescopic hose (10).

5. An intelligent cleaning method using the device according to claim 2, characterized in that: The steps include: S1 Operation Preparation and Environmental Modeling By integrating the navigation and positioning system with the inertial measurement unit, the initial position and heading angle of the cleaning device are determined. The mobile platform adjusts its position according to the coordinates of the target area. The cleaning module is aligned with the surface of the box girder side formwork by rotating the slewing bearing plate and adjusting the telescopic rod. A 3D point cloud model is constructed through simultaneous scanning with a lidar and a binocular camera. The plane equation, center coordinates, and normal vector of the box girder side formwork are extracted through voxel grid filtering and denoising, ICP algorithm registration, and RANSAC algorithm fitting. Obstacles in non-target areas are identified and a 3D bounding box is generated, which is used as obstacle data for path planning. S2 multispectral visual recognition and stain analysis After completing the 3D modeling of the work area and determining the work posture, the device enters the target area pollution identification stage. The multispectral visual recognition module collects and analyzes images of the box girder side mold surface to identify the type, distribution location and adhesion strength of the stains. S3 reinforcement learning-driven path planning and cleaning execution The path planning module uses a reinforcement learning path optimization method to autonomously generate high-efficiency, low-energy cleaning paths. This path planning method combines the spatial distribution of pollution points, pollution intensity, cleaning costs, and operation sequence to minimize operation time and resource waste, and trains the optimal strategy network. S4 Cleaning Quality Monitoring and Secondary Response The cleaning quality detection module is activated to perform a panoramic rescan of the cleaned area and detect residual stains using the U-Net image segmentation algorithm, enabling automatic evaluation and compensatory cleaning control. S5 adaptive database design Construct an adaptive learning database module, which uses a cloud database as a carrier and combines machine learning models to continuously and dynamically model the mapping relationship between stain identification, cleaning parameter configuration and environmental variables, thereby realizing intelligent optimization and automatic prediction capabilities of cleaning strategies.

6. The intelligent cleaning method according to claim 5, characterized in that: The step S1 is specifically as follows: S101 system startup and positioning initialization After the device receives the operation instruction, the control device (2) starts the main control chip, wakes up the mobile platform (1), each functional module and the robotic arm control unit in turn, and establishes a remote data link with the central dispatching platform through the 5G communication module to obtain the latest task list and operation area number of the construction site; then, the four-wheel drive mobile platform starts and moves to the target area under the guidance of the navigation control module. The navigation module obtains the current position information and motion state of the device through the fusion of the navigation positioning system and the inertial measurement unit to form a basic navigation state vector: X t =[x t ,y t ,the t ] T Among them, x t ,y t is the position coordinate of the platform center in the ground coordinate system; θ t is the heading angle the platform is facing. After comparing the current position of the platform with the coordinates of the target area, the platform gradually moves into the preset work area along the shortest path; S102 working posture adjustment and alignment When the platform reaches the front of the box beam side formwork, it stops automatically and enters the fine adjustment stage. In order to make the cleaning device face the formwork surface, the slewing bearing plate (12) on the top of the platform starts the rotation function to adjust the direction of the upper structure. The rotation angle is calculated by the following formula: Among them, (x c ,y c ) is the reference point coordinate of the target template, (x p ,y p ) is the current coordinate of the platform, θ d Indicates the rotation angle of the platform facing the template. After the turning is completed, the platform activates the telescopic rod (4) to lift the cleaning module to a suitable working height according to the measured template height. The first and second robotic arms (5, 6) use the current coordinate system as a reference and align the end cleaning head with the center of the template surface through the kinematic inverse solution algorithm, adjusting the angle to keep it perpendicular to the contact surface. S103 3D Mapping and Template Recognition The system starts the laser radar (LiDAR) and binocular stereo camera module to perform synchronous scanning and build a three-dimensional scene point cloud model P = {p i },p i =(x i ,y i ,z i ,I i ), i is the point cloud data index, I i Represents the laser reflection intensity. The original point cloud data is filtered and denoised by voxel grid to form a compressed point cloud set P′, P′=voxelGrid(P, l), where l is the voxel size, which determines the compression accuracy. The ICP (Iterative Closest Point) algorithm is used to align the multi-frame point clouds to construct a stable and continuous 3D scene model. Combined with the pre-imported template standard size parameter width W t , height H t , the long plane in the point cloud is fitted by the RANSAC algorithm to extract the accurate position and posture of the box girder side form. The plane equation is as follows: Ax+By+Cz+D=0 This plane is the candidate template surface to be identified. A, B, C, and D are plane coefficients. The system compares them with the standard template width W. t , height H t For comparison: |W o -W t |<ε w ,|H o -H t |<ε h Among them, W o , H o is the actual width and height of the fitting plane, ε w , ε h is the error tolerance. If the tolerance condition is met, the plane is set as the cleaning target area R c , and extract its center coordinates P c =(x c ,y c , z c ) and normal vector n, which will be used to calculate the end-of-arm working posture. In order to align the robotic arm with the center of the template, the platform needs to be rotated to the orientation angle θ d Otherwise, it will be regarded as an abnormal structure or misidentified plane, and the system will prompt manual confirmation or rescan; Calculate the end position that the robot arm should reach: X c =P c +n·d safe Among them, d safe Indicates the safe distance between the end and the template to ensure that no mechanical interference occurs during visual identification or cleaning.

7. The intelligent cleaning method according to claim 6, characterized in that: The step S1 further includes the following steps: S104 Obstacle Identification and Avoidance Mapping In order to avoid collisions in the subsequent path planning process, the system performs clustering and bounding box extraction on the point cloud part of the non-target area, identifies potential obstacles, and uses the Euclidean clustering method to extract cluster units C from the point cloud. k , generate the minimum three-dimensional bounding box B for each cluster unit k , and record its spatial position in the local map, all {B k The obstacle data will be stored in the system map and used as a path loss factor to influence the clean path optimization strategy; S105 multi-source perception data output After completing the above environmental modeling, the system will package and output the following data for the next stage of multispectral recognition module to call: template area boundary size and orientation information, template center point P c With the normal vector n, cleaning task posture target point X c , 3D point cloud model of the working area, obstacle list B k , current working coordinate frame X t .

8. The intelligent cleaning method according to claim 6 or 7, characterized in that: The step S2 comprises the following steps: S201 multispectral image acquisition The multispectral visual recognition module is installed at the end of the second robotic arm 6 and includes three types of imaging units: visible light camera Capturing color and texture; infrared thermal imager Reflecting surface temperature distribution; UV imaging equipment Enhance the detection capability of metal rust and chemical residues. To ensure the stability of the image and the coverage of the field of view, the system uses the cleaning target center point P c As the center of the circle, set the scanning radius r s With the angular resolution Δθ, multi-view images are collected along the arc trajectory to determine the multispectral shooting path; i i =θ d +i·Δθ,i=-N,...,N P i =P c +R(θ i )·n·d img Among them, R(θ i ) is a two-dimensional rotation matrix; d img is the shooting distance, and d safe Equal; P i Indicates the imaging position of each frame; N represents half of the number of shooting frames. This method ensures that multi-angle and multi-spectral images cover the complete template area R c ; The device collects multi-channel images in sequence along the shooting path: These images serve as the input basis for pollution identification analysis; S202 Multispectral Image Fusion and Feature Extraction The image I collected in S201 i , in order to unify the feature scale of multi-channel images, the system performs i Perform image registration and feature fusion operations; Specifically, because visible light, infrared, and ultraviolet imaging devices have different viewing angles and resolutions, the system must first spatially align the three images to ensure semantic consistency at corresponding pixels. The system uses an affine registration model for alignment: Among them, A ir and A uv is the image affine matrix, b ir and b uv is the translation vector, and is the image after registration; Secondly, perform image channel normalization. The three images have different imaging mechanisms, and their brightness ranges and response frequencies vary greatly. Therefore, it is necessary to perform unified normalization on each image channel: In the formula, k∈{vis, ir, uv} represents the channel type; μ k and σ k Represents the pixel mean and standard deviation of channel k, I′ k (x, y) is the final standardized image. Standardization ensures that different channels contribute evenly to the fusion feature tensor. Finally, multi-spectral fusion feature construction is performed, and the three normalized images are stacked to construct the fusion feature map F i : in, and Represents the image gradient, performs edge enhancement on the image, and the final F i Will be used as input for subsequent CNN models; S203 3D stain mapping and spatial positioning First, the fusion feature map F obtained in S202 is i Input to the feature extraction network f θ (·), and obtain the high-dimensional semantic feature representation of the image: z=f θ (F i ) Here, θ represents the model parameters of the neural network, and z is the feature vector. The system identifies the stain type c (oil stain, cement residue, and rust) through the classification output branch. A multi-class discriminant model is constructed using a fully connected network and the Softmax function to output the probability distribution of each type of stain: P p =softmax(W p z+b p ) Among them, P p Represents the probability vector of stain classification, W p and b p are the weight and bias parameters of the classification layer. After the Softmax output, the system determines the stain type c of the area based on the category label corresponding to the maximum probability value: C=argmaxP p At the same time, the adhesion strength estimation branch adopts a regression structure and uses the Sigmoid activation function to normalize the output result to obtain the predicted value of the stain adhesion strength: S s =σ(W s z+b s ) Where W s and b s is the weight and bias of the regression layer, S s ∈[0,1] represents the normalized adhesion strength. The larger the value, the higher the cleaning difficulty. After completing the joint identification of the stain type and adhesion strength, the system dynamically generates the cleaning parameter instructions for the corresponding area based on the identification results. The system has a built-in control parameter mapping function β, whose input is the stain type c and adhesion strength S s , the output is a triplet of cleaning parameters; (p,v,η)=β(c,S s ) In the formula, p represents the spray pressure of the high-pressure nozzle, in bar; v is the rotation speed of the cleaning brush, in RPM; η is the cleaning agent concentration, which is the ratio of active ingredients in the cleaning solution, and the value range is [0,1]. Finally, the system associates the image area recognition results with the spatial point cloud data, and maps the recognition results of the corresponding pixel points in the fused image to the pollution point set in the three-dimensional point cloud coordinate system through the obtained three-dimensional projection transformation relationship. And attach the corresponding stain type, adhesion strength and cleaning parameters to each spatial point: This set of contaminated points will serve as the core input of the path planning and control system, and will be used in subsequent steps to perform cleaning path priority sorting, robot arm trajectory planning, and cleaning motion control processes.

9. The intelligent cleaning method according to claim 5, characterized in that: The step S5 comprises: First, after executing each round of cleaning tasks, the system will fully record all key data related to the environment, identification, control, and quality assessment, and construct a structured operation sample set D, whose basic unit is a single contaminated area cleaning data entry d i : d i =[c,S s ,T i ,H i ,θ di ,p i ,v i ,η i ,P Pi ] Among them, T i , H i Respectively represent the ambient temperature and humidity during operation, and these data are uploaded to the cloud database D cloud , and form a long-term accumulated data set D = [d1, d2, ..., d n ]; Secondly, the system builds a multi-objective regression model with the goal of predicting the optimal cleaning control parameters for a given stain type and operating environment: Among them, the input features are c, S s , T i , H i ,θ di , the output is the optimal water pressure recommended for the corresponding pollution area Brush speed Detergent concentration Finally, in order to enhance the system's ability to respond to dynamic environmental factors in the cleaning task, especially under the influence of factors such as pollution intensity fluctuations, humidity changes, or working angle adjustments that may occur over time in the box girder side formwork area, a long short-term memory neural network is used as the strategy parameter prediction model.

10. The intelligent cleaning method according to claim 9, characterized in that: The step S5 comprises: In order to ensure that the model not only gives reasonable prediction values but also guides actual cleaning effects and resource optimization, a composite loss function is designed: In the formula, represents the residual rate under the model prediction conditions, To minimize the error between the predicted residual rate and the actual cleaning effect; To constrain energy consumption; represents the cost of cleaning fluid usage, and t1, t2, and t3 are the adjustment weights of the three objectives.

Images