Adaptive Spraying Parameter Adjustment Method and System for Railway Maintenance Painting Robot

By acquiring multispectral images and environmental information to identify rust, stains, etc., a surface state feature map is constructed. Lightweight neural networks are used to generate spraying parameters, which solves the problem of unstable spraying parameters in railway maintenance painting, realizes adaptive spraying parameter adjustment, and improves spraying quality.

CN123103705APending Publication Date: 2026-11-17BEIJING YANLING JIAYE INTELLIGENT TECH CO LTD
View PDF 0 Cites 0 Cited by

Patent Information

Application Number
CN202611614190.5
Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
Filing Date
2026-10-19
Publication Date
2026-11-17

AI Technical Summary

Technical Problem

In existing railway maintenance painting operations, it is difficult to stably generate and correct the painting parameters in real time according to the differences in the condition of the surface to be painted and the changes in the painting operation environment, resulting in poor painting effect.

Method used

By acquiring multispectral image information, temperature and humidity information, and wind speed information, rust, stains, and geometric morphology information are identified, a surface condition feature map is constructed, the surface condition level is determined, and an initial spraying parameter is generated using a lightweight neural network. The parameter is then corrected by combining real-time spraying status information to achieve adaptive adjustment of the spraying parameter.

Benefits of technology

It enables automatic adjustment of spraying parameters based on the condition of the surface to be painted and environmental changes, improving the stability and adaptability of the spraying process and ensuring spraying quality.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN123103705A_ABST
    Figure CN123103705A_ABST
Patent Text Reader

Abstract

This invention discloses an adaptive spraying parameter adjustment method and system for a railway maintenance painting robot. The method includes: acquiring multispectral image information of the railway vehicle surface to be painted, as well as temperature, humidity, and wind speed information of the spraying environment; identifying rust, stains, and geometric morphology information of the surface and constructing a surface state feature map; determining the surface state level; generating predictive input information for spraying parameters by combining the surface state level, temperature, humidity, and wind speed information; generating initial spraying parameters such as spraying distance, nozzle angle, paint flow rate, and travel speed using a lightweight neural network model; correcting the initial spraying parameters based on real-time spraying state information to generate target spraying parameters, and controlling the painting robot to perform the spraying operation. This invention can generate and correct spraying parameters according to changes in the surface state and spraying environment, improving the stability of the spraying operation.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] This invention relates to the field of intelligent equipment and automated spraying technology, and in particular to an adaptive spraying parameter adjustment method and system for railway maintenance spraying robots based on surface condition perception. Background Technology

[0002] During railway vehicle maintenance, it is common to need to repair damaged areas of the car body, underframe, local metal components, or existing coatings by painting. Painting repair operations require setting parameters such as spraying distance, nozzle angle, paint flow rate, and travel speed based on the condition of the surface to be painted and the painting environment, so that the painting robot can complete the painting operation under relatively stable process conditions.

[0003] In existing railway maintenance painting operations, spraying parameters are typically determined based on preset process parameters, manual experience, or fixed control strategies. This approach can meet the needs of painting scenarios with relatively uniform surface conditions and minimal changes in environmental conditions. However, when the surface of the railway vehicle to be painted has rust, stains, or geometric variations, fixed spraying parameters cannot fully reflect the treatment requirements of different surface areas. At the same time, changes in temperature, humidity, and wind speed can also affect the atomization, adhesion, and transport of the paint, making it necessary to adjust the spraying parameters according to changes in operating conditions.

[0004] Therefore, in railway vehicle maintenance and painting scenarios, the main technical problem to be solved is how to stably generate and correct painting parameters during the painting process based on the differences in the state of the surface to be painted and the changes in the painting environment. There is an urgent need for an adaptive painting parameter adjustment method and system for railway maintenance painting robots that can sense the state of the surface to be painted and dynamically adjust the painting parameters in conjunction with environmental information. Summary of the Invention

[0005] This invention provides a method and system for adaptive spraying parameter adjustment of a railway maintenance painting robot based on surface state perception, which solves the problem in railway vehicle maintenance painting scenarios that it is difficult to stably generate and correct spraying parameters during the spraying process according to the differences in the state of the surface to be painted and the changes in the spraying operation environment.

[0006] According to a first aspect of the present invention, an adaptive spraying parameter adjustment method for a railway maintenance painting robot based on surface state perception is provided, comprising: Acquire multispectral image information of the surface of the railway vehicle to be painted, as well as temperature, humidity and wind speed information of the painting environment; Based on the multispectral image information, the rust information, stain information and geometric morphology information of the surface of the railway vehicle to be painted are identified respectively, and a surface state feature map is constructed based on the rust information, the stain information and the geometric morphology information; The surface condition level of the railway vehicle surface to be painted is determined based on the surface condition feature map. Based on the surface condition level, the temperature and humidity information, and the wind speed information, predictive input information for spraying parameters is generated; The predicted spraying parameters are input into a preset lightweight neural network model to generate initial spraying parameters such as spraying distance, nozzle angle, paint flow rate, and travel speed. The system acquires real-time spraying status information during the spraying process, determines spraying deviation information based on the real-time spraying status information, and corrects the initial spraying parameters based on the spraying deviation information to generate target spraying parameters. The painting robot is controlled to perform the painting operation according to the target painting parameters.

[0007] According to a second aspect of the present invention, an adaptive spraying parameter adjustment system for a railway maintenance painting robot based on surface state perception is provided, comprising: The information acquisition module is used to acquire multispectral image information of the surface of the railway vehicle to be painted, as well as temperature, humidity and wind speed information of the painting operation environment; The feature map construction module is used to identify the rust information, stain information and geometric shape information of the surface of the railway vehicle to be painted according to the multispectral image information, and to construct a surface state feature map based on the rust information, the stain information and the geometric shape information; A state level determination module is used to determine the surface state level of the railway vehicle surface to be painted based on the surface state feature map. The input information generation module is used to generate spraying parameter prediction input information based on the surface condition level, the temperature and humidity information, and the wind speed information; The initial parameter generation module is used to input the spraying parameter prediction input information into a preset lightweight neural network model to generate initial spraying parameters such as spraying distance, nozzle angle, paint flow rate and travel speed. The target parameter generation module is used to acquire real-time spraying status information during the spraying process, determine spraying deviation information based on the real-time spraying status information, and correct the initial spraying parameters based on the spraying deviation information to generate target spraying parameters. The spraying control module is used to control the painting robot to perform spraying operations according to the target spraying parameters.

[0008] The beneficial technical effects of this invention are as follows: by acquiring multispectral image information of the surface of the railway vehicle to be painted, as well as temperature, humidity and wind speed information of the spraying operation environment, and identifying rust information, stain information and geometric morphology information based on the multispectral image information to construct a surface state feature map, the state differences of the surface to be painted can be transformed into feature expressions that can be used for subsequent processing, providing a surface state basis for the stable generation of spraying parameters.

[0009] Furthermore, by determining the surface condition level based on the surface condition feature map, and generating spraying parameter prediction input information based on the surface condition level, temperature and humidity information, and wind speed information, and then generating initial spraying parameters such as spraying distance, nozzle angle, paint flow rate, and travel speed through a lightweight neural network model, the spraying parameter generation process can simultaneously take into account the differences in the state of the surface to be painted and the changes in the spraying operation environment, thereby improving the stability and adaptability of the initial spraying parameter generation.

[0010] Furthermore, by acquiring real-time spraying status information during the spraying process, determining spraying deviation information based on the real-time spraying status information, and correcting the initial spraying parameters based on the spraying deviation information to generate target spraying parameters, and then controlling the painting robot to perform the spraying operation based on the target spraying parameters, the spraying parameters can be corrected according to the actual state during the spraying process. This solves the problem in railway vehicle maintenance painting scenarios where it is difficult to stably generate and correct spraying parameters according to the state differences of the surface to be painted and changes in the spraying operation environment. Attached Figure Description

[0011] Figure 1 This is a schematic diagram illustrating a method for adaptive spraying parameter adjustment of a railway maintenance and painting robot based on surface state perception, as provided in an embodiment of the present invention.

[0012] Figure 2 This is a flowchart illustrating an adaptive spraying parameter adjustment method for a railway maintenance and painting robot based on surface state perception, provided as an embodiment of the present invention.

[0013] Figure 3 This is a schematic diagram of the surface state feature map construction process provided in an embodiment of the present invention.

[0014] Figure 4 This is a schematic diagram of the surface condition level determination process provided in an embodiment of the present invention.

[0015] Figure 5 This is a schematic diagram of the process for generating spraying parameter prediction input information according to an embodiment of the present invention.

[0016] Figure 6 This is a schematic diagram of the process for generating initial spraying parameters according to an embodiment of the present invention.

[0017] Figure 7 This is a schematic diagram of the closed-loop correction process for spraying parameters provided in an embodiment of the present invention.

[0018] Figure 8 This is a schematic diagram of the spraying control command generation and spraying operation execution process provided in an embodiment of the present invention.

[0019] Figure 9 This is a schematic diagram of an adaptive spraying parameter adjustment system for a railway maintenance and painting robot based on surface state perception, provided as an embodiment of the present invention. Detailed Implementation

[0020] The technical solution of the present invention will be described below with reference to the accompanying drawings and embodiments. The following embodiments are for illustrative purposes only and are not intended to limit the scope of protection of the present invention. Unless otherwise specified, the embodiments and technical features described in this specification can be combined with each other.

[0021] refer to Figure 1 In one embodiment, an adaptive spraying parameter adjustment system for a railway vehicle maintenance painting robot based on surface state perception can be deployed at the railway vehicle maintenance painting site. The system may include a multispectral vision sensor 10, an environmental sensor 20, a painting robot 30, a spraying actuator 40, and a controller 50. The multispectral vision sensor 10 is used to acquire multispectral image information of the surface of the railway vehicle to be painted. The multispectral image information can reflect the differences in reflection, color, texture, and local morphology of the surface to be painted in multiple wavelength bands. The environmental sensor 20 is used to acquire temperature, humidity, and wind speed information of the painting environment. The painting robot 30 is used to carry the spraying actuator 40 along the area to be painted. The spraying actuator 40 may include a spray gun, a nozzle, a paint supply assembly, and a flow regulation assembly. The controller 50 is used to receive multispectral image information, temperature and humidity information, wind speed information, and spraying process feedback information, and generate spraying control commands for controlling the painting robot 30 and the spraying actuator 40.

[0022] In one specific embodiment, the controller 50 can be located at the local control terminal of the painting robot 30, or at an edge computing device, or it can be implemented collaboratively by the robot's local control terminal and a host computer. The multispectral vision sensor 10 can be installed near the end effector of the painting robot 30, or it can be installed on an independent scanning mechanism. The environmental sensor 20 can be located near the painting station to collect real-time data on the painting operation environment. The controller 50 can be connected to the multispectral vision sensor 10, the environmental sensor 20, the painting robot 30, and the painting actuator 40 via wired or wireless communication. Wired communication methods can include CAN (Controller Area Network), Ethernet, or serial communication, while wireless communication methods can include wireless local area networks or industrial wireless communication methods.

[0023] In one embodiment, the controller 50 can first receive images of the railway vehicle surface to be painted from the multispectral vision sensor 10, and then receive temperature, humidity, and wind speed information from the environmental sensor 20. After processing the multispectral image information, the controller 50 generates a surface state feature map and determines the surface state level based on the surface state feature map. The controller 50 organizes the surface state level, temperature, humidity, and wind speed information into spraying parameter prediction input information and inputs it into a preset lightweight neural network model to obtain initial spraying parameters. When the painting robot 30 performs the spraying operation, the controller 50 continues to acquire real-time spraying state information and generates spraying parameter correction amounts based on the real-time spraying state information, and then outputs the target spraying parameters. The target spraying parameters can be converted into spraying control commands and sent to the painting robot 30 and the spraying actuator 40.

[0024] In this embodiment, the surface of the railway vehicle to be painted refers to the target surface that needs to be painted, touched up, coated with anti-corrosion coating, or repaired during the maintenance of railway vehicles. It can include the exterior surface of the car body, the surface of the car body underframe, the surface of the bogie peripheral components, the surface of the end connection parts, the surface of local metal parts, and the surface of areas with existing coating damage. For example, in railway vehicle maintenance operations, areas with partial paint peeling on the side walls of the car body, rusted areas on the underframe crossbeams, and areas with stains on the end connection structures can all be considered as the surface of the railway vehicle to be painted.

[0025] In this embodiment, multispectral image information refers to a dataset acquired by a multispectral vision sensor that characterizes the image response of the surface to be painted in multiple spectral bands. Multispectral image information may include visible light images, near-infrared images, and other spectral images used to distinguish between rust, stains, and differences in surface morphology. For example, multispectral image information may include red light images, green light images, blue light images, and near-infrared images; in rusted areas, the reflection response of different spectral bands typically differs from that of unrusted metal surfaces or existing coated surfaces, and therefore can be used to identify rust information.

[0026] In this embodiment, temperature and humidity information refers to the temperature and humidity information in the spraying operation environment. The temperature information can be the air temperature near the spraying station, and the humidity information can be the relative humidity of the air. For example, when the current temperature at the spraying station is 25°C and the relative humidity is 60%, corresponding temperature and humidity information can be generated. The temperature and humidity information is used to characterize the environmental conditions during the atomization, adhesion, and drying processes of the coating.

[0027] In this embodiment, wind speed information refers to information related to the airflow velocity within the spraying operation area, which may include instantaneous wind speed, average wind speed, or wind speed change rate. For example, the current wind speed at the spraying station is 1.2 m / s, or the wind speed increases from 0.8 m / s to 1.5 m / s over multiple consecutive sampling periods, both can be used as wind speed information. Wind speed information is used to reflect the risk of deviation or drift of sprayed atomized particles during the nozzle-to-surface transport process.

[0028] In this embodiment, the painting robot refers to an industrial robot used to perform surface painting operations on railway vehicles, and may include a mobile chassis, a robotic arm, an end-effector, and a control interface. The end-effector refers to the components used to complete paint output, spray formation, and spray control, and may include a spray gun, nozzle, paint supply assembly, and flow regulation assembly. The controller refers to an electronic control unit used to process image information, environmental information, painting parameters, and feedback information and output control commands, and may be implemented by an industrial computer, embedded controller, edge computing device, or robot control terminal.

[0029] refer to Figure 2 In one embodiment, the adaptive spraying parameter adjustment method for railway maintenance painting robot based on surface state perception provided by the present invention may include steps S10 to S70.

[0030] S10: Acquire multispectral image information of the surface of the railway vehicle to be painted, as well as temperature, humidity, and wind speed information of the painting environment.

[0031] In this embodiment, multispectral image information can be acquired in real time by a multispectral vision sensor, or it can be formed by simultaneous acquisition by multiple image acquisition units of different bands.

[0032] In this embodiment, the multispectral vision sensor can be configured as an integrated visible light and near-infrared acquisition device, or as a discrete multi-band acquisition device. The images acquired by the multispectral vision sensor can be matched with temperature, humidity, and wind speed information acquired by the environmental sensor according to timestamps, so as to establish a correspondence within the same spraying cycle. For example, the multispectral vision sensor acquires images of the vehicle surface at a frequency of 10 frames per second, while the environmental sensor acquires temperature, humidity, and wind speed data at a frequency of once per second. The controller combines and processes the image data and environmental data within the same control cycle.

[0033] S20, based on the multispectral image information, identify the rust information, stain information and geometric shape information of the surface of the railway vehicle to be painted, and construct a surface state feature map based on the rust information, the stain information and the geometric shape information.

[0034] In this embodiment, rust information refers to information that characterizes the corrosion state of the metal surface of the railway vehicle to be painted after corrosion or coating damage. It may include the location of the rusted area, the rusted area, the rusted color characteristics, the rusted texture characteristics, the rusted intensity characteristics, and the rusted distribution density. For example, when reddish-brown patches with rough texture appear in a local area of ​​the car body underframe, the location, area, and texture intensity of this area can constitute rust information.

[0035] In this embodiment, stain information refers to information that characterizes the adhesion status of contaminants on the surface of the railway vehicle to be painted. This information may include the location, coverage area, color difference, spectral response difference, and degree of contamination of oil stains, dust, mud traces, accumulated dust, or other contaminants. For example, when there is black oil stain covering the maintenance area at the bottom of the vehicle, the boundary, area, and degree of darkening of the oil stain area can constitute stain information.

[0036] In this embodiment, geometric topography information refers to information that can characterize the shape, contour, or local undulation of the surface of the railway vehicle to be painted. This information may include depressions, protrusions, edge variations, surface unevenness, undulations near welds, peeling edges, or local structural contour variations. For example, if there are peeling edges or local unevenness in the area to be painted, the edge position, undulation intensity, and abnormal topography areas can constitute geometric topography information.

[0037] In this embodiment, a surface state feature map refers to a multi-layered feature representation result formed by spatially associating rust information, stain information, and geometric shape information under the same surface coordinate system. The surface state feature map may include a rust feature layer, a stain feature layer, and a geometric shape feature layer. For example, for the same vehicle body surface area, if the rust feature layer shows that the area has rust, the stain feature layer shows that the area has localized oil stains, and the geometric shape feature layer shows that the area has slight undulations, then the surface state feature map can simultaneously express these three states and their spatial overlap relationships.

[0038] refer to Figure 3 In one embodiment, step S20 may include steps S210 to S240.

[0039] S210, perform band alignment processing on the multispectral image information to obtain aligned image information.

[0040] In this embodiment, band alignment processing refers to the process of mapping pixels or image regions corresponding to the same location on the surface to be painted in different band images to a unified coordinate system. Band alignment processing may include calibration parameter reading, image coordinate transformation, scale matching, rotation correction, and translation correction. For example, when there is a two-pixel offset at the boundary position of the same rust area in the near-infrared image and the visible light image, band alignment processing can be used to make the position of the rust area consistent in the two band images.

[0041] In this embodiment, aligned image information refers to the data result formed by different band images under unified image coordinates or surface coordinates after band alignment processing. Aligned image information is used for subsequent comparison of different band responses at the same location and to determine the rust area, stain area, and geometric shape area respectively.

[0042] S220, the rust area, stain area and geometric shape area are determined respectively based on the aligned image information.

[0043] In this embodiment, the corroded area refers to the surface area that is determined to be corroded based on the aligned image information. For example, if an image area exhibits a reddish-brown characteristic in the visible light band and shows a significant difference in response compared to the surrounding normal coating area in the near-infrared band, then the image area can be identified as a corroded area.

[0044] In this embodiment, a stained area refers to a surface area determined to have contaminant adhesion based on the aligned image information. For example, if an area exhibits low brightness response across multiple wavelengths and its boundary morphology matches the characteristics of oil or dust coverage, then that area can be identified as a stained area.

[0045] In this embodiment, the geometric shape region refers to a surface area that is determined to have surface undulations, edge changes, unevenness, or abnormal contours based on the aligned image information. For example, if there is a coating peeling edge in a certain sprayed area, and the image shows continuous edge lines and local brightness variations, then this area can be identified as a geometric shape region.

[0046] S230, determine the regional positions and regional characteristic values ​​of the rusted area, the stained area, and the geometric shape area in the same surface coordinate system.

[0047] In this embodiment, the same surface coordinate system refers to a coordinate system used to uniformly describe the various positions on the surface of the railway vehicle to be painted. It can be a two-dimensional coordinate system established based on the unfolded view of the vehicle body surface, or a local coordinate system established based on the working coordinates of the painting robot. For example, the area to be painted on the side wall of the vehicle can be divided into a two-dimensional grid, and the position of each surface can be represented by the grid row and column positions.

[0048] In this embodiment, the region location refers to the spatial extent or coordinate description of the rusted area, stained area, or geometric shape area within the same surface coordinate system. This can include the region boundary, the region center point, the region area, the region's circumscribed rectangle, or the grid number to which the region belongs. For example, if a rusted area is located in the 3rd row and 5th column of the local coordinate system of the vehicle sidewall, and its boundary coordinates are the coordinates of multiple vertices, then this information can be used as the region location of the rusted area.

[0049] In this embodiment, the regional feature value refers to a numerical value or vector used to quantify the strength of the state of rusted areas, stained areas, or geometric morphology areas. It can include area proportion, color response intensity, texture intensity, degree of contamination coverage, surface undulation amplitude, or degree of morphological anomaly. For example, if the rusted area accounts for 20% of the corresponding area to be sprayed and has a high rust color response intensity, a high rusted area feature value can be generated.

[0050] S240, Based on the location of the region and the feature values ​​of the region, generate a surface state feature map containing a rust feature layer, a stain feature layer and a geometric shape feature layer.

[0051] In this embodiment, the rust feature layer refers to a layer that records the location of rusted areas and rust feature values ​​in the same surface coordinate system. For example, in the mesh on the surface of the vehicle body sidewall, meshes with rust are assigned corresponding rust feature values, while meshes without rust can be assigned zero values ​​or background values.

[0052] In this embodiment, the stain feature layer refers to a layer that records the location of stain areas and stain feature values ​​in the same surface coordinate system. For example, when there is oil stain in a certain area of ​​the vehicle chassis, that area has a high stain feature value in the stain feature layer.

[0053] In this embodiment, the geometric topography feature layer refers to a layer that records the location and topography feature values ​​of geometric topography regions under the same surface coordinate system. For example, the location of local protrusions, depressions, or coating peeling edges on the surface has corresponding topography feature values ​​in the geometric topography feature layer.

[0054] S30, determine the surface condition level of the railway vehicle surface to be painted based on the surface condition feature map.

[0055] In this embodiment, the surface condition level refers to a grade of information determined comprehensively based on rust information, stain information, and geometric morphology information, used to characterize the difficulty or severity of the surface coating treatment on the railway vehicle to be painted. For example, the surface condition level can be divided into four levels: Level 1, Level 2, Level 3, and Level 4, where Level 1 indicates a relatively light surface condition and Level 4 indicates a relatively heavy surface condition; it can also be divided into slight, moderate, relatively heavy, and severe. The surface condition level is used to generate surface condition coding information and participates in the construction of subsequent coating parameter prediction input information.

[0056] refer to Figure 4 In one embodiment, step S30 may include steps S310 to S350.

[0057] S310, extract rust feature values, stain feature values ​​and morphology feature values ​​from the rust feature layer, the stain feature layer and the geometric morphology feature layer respectively.

[0058] In this embodiment, the rust feature value refers to the feature value extracted from the rust feature layer to quantify the rust state, which can be obtained based on the rust area ratio, rust color response intensity, and rust texture intensity. For example, if a certain area to be sprayed has a large rust area and a rough texture, then the corresponding rust feature value of that area is high.

[0059] In this embodiment, the stain feature value refers to the feature value extracted from the stain feature layer to quantify the stain state, which can be obtained based on the stain coverage area, stain color difference, and stain distribution density. For example, the larger the oil stain coverage area, the darker the color, and the more concentrated the distribution, the higher the corresponding stain feature value can be.

[0060] In this embodiment, the morphological feature value refers to the feature value extracted from the geometric morphological feature layer to quantify the degree of surface geometric morphological anomalies. It can be obtained based on the undulation amplitude, edge density, local unevenness, or morphological anomaly area. For example, when the coating peeling edge is long and the local undulation is obvious, the corresponding morphological feature value can be high.

[0061] S320, based on the rust characteristic value, the stain characteristic value and the morphology characteristic value, determine the rust grade, stain grade and morphology grade respectively.

[0062] In this embodiment, the corrosion level refers to the severity level of corrosion determined based on the corrosion characteristic value. For example, a corrosion characteristic value below the first corrosion threshold can be determined as a slight corrosion level, a value between the first and second corrosion thresholds can be determined as a moderate corrosion level, and a value above the second corrosion threshold can be determined as a severe corrosion level.

[0063] In this embodiment, the stain grade refers to the severity level of a stain determined based on its characteristic values. For example, a small amount of dust in a localized area may correspond to a minor stain grade, while a large area of ​​oil stains may correspond to a severe stain grade.

[0064] In this embodiment, the morphology level refers to the level of surface geometric morphology anomaly determined based on morphology feature values. For example, slight coating edge undulations may correspond to a slight morphology level, while obvious depressions or peeling edges may correspond to a severe morphology level.

[0065] S330, when there are multiple target levels that reach a preset severity level among the rust level, the stain level and the morphology level, obtain the severity score corresponding to each target level.

[0066] In this embodiment, the preset severity level refers to a threshold used to determine whether a certain type of surface condition has reached the level requiring intensive treatment. For example, when the levels are divided into four levels, level three and above can be set as the preset severity level; when the levels are divided into slight, moderate, severe, and severe, severe and above can be set as the preset severity level. The target level is the corrosion level, stain level, or morphology level that reaches the preset severity level.

[0067] Severity scoring is used to handle conflicts when multiple target levels exist simultaneously. Severity scoring can be determined based on the area proportion of the corresponding defect type, characteristic response intensity, distribution concentration, overlap with the spraying path, or the weight of its impact on spraying quality. In a specific embodiment, the severity score Score_i can be expressed as:

[0068] Score_i=w1×A_i+w2×Intensity_i+w3×Density_i Where i represents the defect type, A_i represents the area proportion of the defect type, Intensity_i represents the feature response intensity of the defect type, Density_i represents the distribution density of the defect type, and w1, w2, and w3 represent weight parameters. This formula is used to illustrate an implementable calculation method for severity scoring.

[0069] S340, determine the highest severity level based on the severity score corresponding to each of the aforementioned target levels.

[0070] The controller 50 can compare the severity scores corresponding to each target level and determine the target level with the highest severity score as the highest severity level. When the severity scores corresponding to the target levels are the same or the difference is less than a preset score difference threshold, the highest severity level can be determined according to the preset defect type priority. The preset defect type priority can be set according to the degree of impact on the coating quality, for example, geometric anomalies are prioritized over rust, and rust is prioritized over stains; other priorities can also be set according to the maintenance process requirements.

[0071] S350, generate the surface condition level based on the highest severity level.

[0072] The controller 50 can directly use the highest severity level as the surface condition level, or it can generate a surface condition level by fine-tuning it after combining other severity level information that has not reached the preset severity level. For example, when the highest severity level is the rust severity level, and the stain level and morphology level are both medium, the rust severity level can be maintained as the surface condition level; when the highest severity level is the stain severity level, and the morphology level is close to the severity level, the surface condition level can be determined as a higher processing difficulty level. In this way, the surface condition level can reflect the dominant state when multiple defect types coexist.

[0073] In another optional embodiment, when there are no multiple target levels that reach the preset severity level among the rust level, stain level, and morphology level, but only one target level that reaches the preset severity level, the controller 50 can generate a surface condition level based on the target level; when there is no target level that reaches the preset severity level, the controller 50 can generate a surface condition level based on the combination result of the rust level, stain level, and morphology level.

[0074] S40, generate spraying parameter prediction input information based on the surface condition level, the temperature and humidity information and the wind speed information.

[0075] The input information for spraying parameter prediction is used to feed into the lightweight neural network model, enabling the model to output spraying parameters based on surface condition level and environmental conditions. Temperature, humidity, and wind speed information affect the atomization, adhesion, drying, and dispersion of the coating, and therefore, along with the surface condition level, serve as inputs for spraying parameter prediction.

[0076] refer to Figure 5 In one embodiment, step S40 may include steps S410 to S440.

[0077] S410, generate surface state coding information based on the surface state level.

[0078] Surface state encoding information can be obtained by converting surface state levels into numerical codes, one-hot codes, level scores, or normalized vectors. In one specific embodiment, if the surface state levels include four levels, they can be encoded as [1,0,0,0], [0,1,0,0], [0,0,1,0], and [0,0,0,1]. In another embodiment, the surface state levels can also be encoded as continuous values, such as 0.25, 0.50, 0.75, and 1.00. The specific encoding method can be determined based on the input structure of the lightweight neural network model.

[0079] S420, generate environmental state coding information based on the temperature and humidity information and the wind speed information.

[0080] Environmental status coding information can include temperature coding, humidity coding, and wind speed coding. Controller 50 can normalize temperature, humidity, and wind speed, or it can perform hierarchical coding according to the spraying process range. In one example, the temperature value T, humidity value H, and wind speed value V can be normalized to T_norm, H_norm, and V_norm, respectively, and combined into environmental status coding information E=[T_norm,H_norm,V_norm].

[0081] S430, the surface state coding information and the environmental state coding information are combined to obtain joint coding information.

[0082] Joint encoding information can be obtained through vector concatenation, weighted combination, or feature mapping. In one specific embodiment, surface state encoding information C_surface and environment state encoding information C_env are concatenated to obtain joint encoding information C_joint:

[0083] C_joint=concat(C_surface,C_env) Here, `concat()` represents the vector concatenation operation. The joint encoding information `C_joint` simultaneously expresses the surface state of the surface to be painted and the current painting environment.

[0084] S440, Generate the spraying parameter prediction input information based on the joint coding information.

[0085] The input information for spraying parameter prediction can be the joint encoding information itself, or it can be model input data obtained after normalizing, adjusting the dimensions, or converting the format of the joint encoding information. In one specific embodiment, the controller 50 converts the joint encoding information into the input tensor format required by the lightweight neural network model and generates the spraying parameter prediction input information X_input. The spraying parameter prediction input information X_input proceeds to step S50.

[0086] S50, the spraying parameter prediction input information is input into a preset lightweight neural network model to generate initial spraying parameters such as spraying distance, nozzle angle, paint flow rate and travel speed.

[0087] A lightweight neural network model is used to predict initial spraying parameters based on input information. This model can be a feedforward neural network with a small number of parameters, a lightweight convolutional network, a lightweight regression network, or other models suitable for deployment on the controller 50, edge computing devices, or the robot's local control terminal. The model can be trained based on historical spraying samples, which may include surface condition levels, temperature and humidity information, wind speed information, and corresponding spraying distance, nozzle angle, paint flow rate, and travel speed.

[0088] In one embodiment, reference Figure 6 Step S50 may include steps S510 to S530.

[0089] S510, the spraying parameter prediction input information is input into the lightweight neural network model to obtain the predicted values ​​of spraying distance, nozzle angle, paint flow rate and travel speed.

[0090] The predicted spraying distance is used to represent the target distance between the nozzle and the surface to be painted. The predicted nozzle angle is used to represent the target angle of the nozzle axis relative to the normal of the surface to be painted or the robot's working coordinate system. The predicted paint flow rate is used to represent the amount of paint output by the spraying actuator 40 per unit time. The predicted travel speed is used to represent the speed at which the painting robot 30 moves along the spraying path.

[0091] In this embodiment, the lightweight neural network model can adopt a multilayer perceptron structure, including an input layer, a first fully connected layer, a first activation layer, a second fully connected layer, a second activation layer, a third fully connected layer, and an output layer. The input layer is used to receive the spraying parameter prediction input information X_input. The first fully connected layer has an input dimension of 7 and an output dimension of 32, and the first activation layer can use the ReLU (Rectified Linear Unit) activation function; the second fully connected layer has an input dimension of 32 and an output dimension of 16, and the second activation layer can use the ReLU activation function; the third fully connected layer has an input dimension of 16 and an output dimension of 8; the output layer has a dimension of 4 and is used to output the predicted values ​​of spraying distance, nozzle angle, paint flow rate, and travel speed.

[0092] In this embodiment, the mapping process of the lightweight neural network model can be represented as follows: H1 = ReLU(W1 × X_input + B1) H2 = ReLU(W2 × H1 + B2) H3 = ReLU(W3 × H2 + B3) Y_pred=W4×H3+B4 Where X_input represents the input information for spraying parameter prediction, H1 represents the first hidden feature vector, H2 represents the second hidden feature vector, H3 represents the third hidden feature vector, Y_pred represents the parameter prediction vector, Y_pred=[D_pred,A_pred,F_pred,V_pred], D_pred represents the predicted spraying distance, A_pred represents the predicted nozzle angle, F_pred represents the predicted paint flow rate, and V_pred represents the predicted travel speed; W1, W2, W3, and W4 represent the weight matrices of each fully connected layer, and B1, B2, B3, and B4 represent the bias vectors of each fully connected layer.

[0093] In another specific embodiment, to reduce the number of model parameters, a lightweight neural network model can employ a bottleneck structure. This bottleneck structure includes an input layer, an upscaling layer, a compression layer, and an output layer. The upscaling layer is used to extract combined features between surface state encoding information and environmental state encoding information; the compression layer is used to reduce the feature dimensionality; and the output layer is used to output the predicted values ​​of four spraying parameters. In one example, the input dimension is 7, the upscaling layer output dimension is 24, the compression layer output dimension is 12, and the output layer dimension is 4. This structure reduces the number of model parameters and inference computation while maintaining the ability to predict spraying parameters.

[0094] In one specific embodiment, the training samples for the lightweight neural network model can be derived from historical railway maintenance painting operation data, painting process test data, and manually labeled data. Each training sample includes sample input and sample labels. Sample input includes sample surface condition level, sample temperature, sample humidity, and sample wind speed; sample labels include spraying distance labels, nozzle angle labels, paint flow rate labels, and walking speed labels determined under the corresponding sample input conditions. Spraying distance labels, nozzle angle labels, paint flow rate labels, and walking speed labels can be derived from qualified painting operation records or from painting parameter combinations confirmed by process engineers.

[0095] In one example, the training samples cover different rust states, different stain states, different geometric morphologies, and different temperature, humidity, and wind speed conditions. Rust states can be categorized as slight, moderate, moderate, and severe; stain states can be categorized as slight, moderate, moderate, and severe; geometric morphologies can be categorized as smooth, slightly undulating, moderately undulating, and significantly undulating; ambient temperature can cover 5°C to 40°C, ambient humidity can cover 30% to 90%, and wind speed can cover 0 m / s to 5 m / s. The above ranges are merely examples of the training sample coverage and are not intended to limit the scope of protection of this invention.

[0096] In one specific embodiment, no fewer than 1000 training samples can be collected. To enhance the model's adaptability to complex working conditions, the training samples may include single defect samples and composite defect samples. A single defect sample refers to a sample in which only one of the following three states—rust, stains, and geometric anomalies—is evident; a composite defect sample refers to a sample in which at least two of the following three states exist simultaneously: rust, stains, and geometric anomalies. The composite defect samples may also include multiple samples whose target severity level reaches a preset severity level to support the severity scoring and the determination of the highest severity level.

[0097] Before training, the sample input can be normalized. For continuous variables such as temperature, humidity, and wind speed, minimum-maximum normalization can be used.

[0098] X_norm=(X-X_min) / (X_max-X_min) Where X represents the original variable, X_min represents the minimum value of the variable in the training samples, X_max represents the maximum value of the variable in the training samples, and X_norm represents the normalized variable. For surface condition levels, one-hot encoding or level-based numerical encoding can be used. For spraying distance labels, nozzle angle labels, paint flow rate labels, and walking speed labels, normalization can also be performed to reduce the influence of different units on the training process.

[0099] In one specific embodiment, the lightweight neural network model is trained using supervised learning. The training objective is to make the model's output predictions of spraying distance, nozzle angle, paint flow rate, and travel speed as close as possible to their corresponding label values. The loss function can be a weighted mean squared error loss function, expressed as:

[0100] Loss=lambda_D×(D_pred-D_label)^2+lambda_A×(A_pred-A_label)^2+lambda_F×(F_pred-F_label)^2+lambda_V×(V_pred-V_label)^2 Where Loss represents the training loss, D_pred represents the predicted spraying distance, A_pred represents the predicted nozzle angle, F_pred represents the predicted paint flow rate, and V_pred represents the predicted walking speed; D_label represents the label for spraying distance, A_label represents the label for nozzle angle, F_label represents the label for paint flow rate, and V_label represents the label for walking speed; lambda_D, lambda_A, lambda_F, and lambda_V represent the loss weights corresponding to spraying distance, nozzle angle, paint flow rate, and walking speed, respectively. The loss weights can be set according to the degree of influence of different spraying parameters on spraying quality, or they can be adjusted during training based on validation set results.

[0101] In one example, when spraying distance and paint flow rate have a significant impact on coating coverage quality, lambda_D and lambda_F can be set to values ​​higher than lambda_A and lambda_V. The specific values ​​of the loss weights can be determined based on training samples, process requirements, and validation results, and are not considered necessary limitations of this invention.

[0102] In one specific embodiment, model training can employ a mini-batch gradient descent approach. Training samples can be divided into training, validation, and test sets according to a preset ratio. For example, 70% of the samples can be used as the training set, 15% as the validation set, and 15% as the test set. The training batch size can be set to 16, 32, or 64, the number of training epochs can be set to 50 to 300, and the learning rate can be set to 0.0001 to 0.01. During training, model convergence is determined based on the validation set loss, and training is stopped if the validation set loss fails to decrease for several consecutive training epochs to reduce overfitting.

[0103] In one specific embodiment, after training, the lightweight neural network model can be evaluated using a test set. Evaluation metrics may include spray distance prediction error, nozzle angle prediction error, paint flow rate prediction error, and travel speed prediction error. Alternatively, RMSE (Root Mean Square Error) can be used to evaluate the overall prediction error, expressed as:

[0104] RMSE=sqrt(avg((Y_pred-Y_label)^2)) Where Y_pred represents the parameter prediction vector output by the model, Y_label represents the corresponding parameter label vector, and avg() represents the calculation of the average value. If the prediction error on the test set meets the preset error requirement, the trained lightweight neural network model is deployed to the controller 50 or an edge computing device.

[0105] In one optional embodiment, to accommodate the computing resources of the local control unit of the painting robot, the trained lightweight neural network model can undergo model compression, parameter pruning, or quantization. Quantization can employ FP32 (Floating Point 32-bit) to INT8 (Integer 8-bit) parameter conversion to reduce model storage space and inference time. The output results after model compression or quantization are still used to generate initial painting parameters such as spraying distance, nozzle angle, paint flow rate, and travel speed.

[0106] In one specific embodiment, the deployed lightweight neural network model receives spraying parameter prediction input information and outputs a parameter prediction vector in each spraying control cycle. The controller 50 performs constraint processing on the parameter prediction vector to obtain initial spraying parameters, and then passes these initial spraying parameters to subsequent closed-loop correction steps. Through this training and deployment method, the lightweight neural network model can generate initial spraying parameters usable for spraying control at railway maintenance painting sites based on surface condition levels and environmental conditions.

[0107] In one specific embodiment, to further enhance the model's responsiveness to different surface and environmental conditions, a sample weighting mechanism can be introduced during the training phase. This mechanism assigns higher weights to composite defect samples, severe defect samples, or high-wind-speed samples, making the model more sensitive to predicting spraying parameters under complex conditions. For example, samples with wind speeds exceeding a preset threshold can be assigned higher weights to improve the model's adaptability to fluctuating wind speeds.

[0108] S520, according to the preset spraying parameter constraint range, the predicted values ​​of the spraying distance, the nozzle angle, the paint flow rate, and the travel speed are constrained.

[0109] The preset spraying parameter constraints can be determined based on the working range of the spraying actuator 40, the motion capability of the painting robot 30, the coating process requirements, and the safety operation requirements. For example, the spraying distance needs to be within the effective atomization range of the nozzle, the nozzle angle needs to be within the reachable posture range of the robot end effector, the coating flow rate needs to be within the controllable range of the coating supply component, and the walking speed needs to be within the stable motion range of the painting robot 30.

[0110] Constraint processing can include amplitude limiting, boundary clipping, outlier replacement, or smoothing. For example, for any prediction parameter P_pred, constraints can be applied as follows:

[0111] P_con=min(max(P_pred,P_min),P_max) Where P_con represents the parameter value after constraint processing, P_min represents the minimum allowable value of the corresponding parameter, and P_max represents the maximum allowable value of the corresponding parameter.

[0112] S530, the initial spraying parameters are generated based on the predicted values ​​of spraying distance, nozzle angle, paint flow rate, and travel speed after constraint processing.

[0113] The initial spraying parameters can be expressed as P0=[D0,A0,F0,V0], where D0 represents the initial spraying distance, A0 represents the initial nozzle angle, F0 represents the initial paint flow rate, and V0 represents the initial travel speed. These initial spraying parameters are the fundamental parameters for spraying control and are subsequently corrected in a closed-loop manner based on real-time spraying status information.

[0114] S60, acquire real-time spraying status information during the spraying process, determine spraying deviation information based on the real-time spraying status information, and correct the initial spraying parameters based on the spraying deviation information to generate target spraying parameters.

[0115] Real-time spraying status information can include real-time coverage status information, real-time coating status information, and spraying execution status information. Real-time coverage status information characterizes the distribution of paint coverage on the surface to be painted. Real-time coating status information characterizes coating thickness, coating continuity, or localized coating defects. Spraying execution status information characterizes the execution status of the painting robot 30 and the spraying execution mechanism 40, such as actual spraying distance, actual nozzle angle, actual travel speed, path deviation, or execution response status.

[0116] refer to Figure 7 In one embodiment, step S60 may include steps S610 to S670.

[0117] S610 acquires real-time coverage status information, real-time coating status information, and spraying execution status information during the spraying process.

[0118] Real-time coverage status information can be acquired by the multispectral vision sensor 10 during the spraying process, or by the existing vision acquisition unit at the spraying station. Real-time coating status information can be estimated by coating detection sensors, vision estimation, or process models. Spraying execution status information can be obtained from the encoder, end effector pose sensor, flow sensor, or control system status feedback of the painting robot 30.

[0119] In this embodiment, real-time coverage status information refers to information used during the spraying process to characterize the coverage of the coating on the surface of the railway vehicle to be painted. For example, the actual coverage rate of a certain sprayed area, the proportion of uncovered areas, and the uniformity of coverage distribution can all be used as real-time coverage status information.

[0120] In this embodiment, real-time coating status information refers to information used during the spraying process to characterize the coating thickness, continuity, or local defect status. For example, the estimated actual coating thickness, areas of thin coating, and local sagging can all be used as real-time coating status information.

[0121] In this embodiment, the spraying execution status information refers to the information on the actual execution status of the painting robot and the spraying execution mechanism. For example, the actual spraying distance, actual nozzle angle, actual walking speed, path deviation, and actual paint output status can all be used as spraying execution status information.

[0122] S620, determine coverage deviation information based on the real-time coverage status information, determine coating deviation information based on the real-time coating status information, and determine execution deviation information based on the spraying execution status information.

[0123] Coverage deviation information is used to characterize the difference between the actual coverage state and the target coverage state. Coating deviation information is used to characterize the difference between the actual coating state and the target coating state. Execution deviation information is used to characterize the difference between the actual robot execution state and the target execution state.

[0124] In one specific embodiment, the coverage deviation information E_cover can be determined based on the target coverage and the actual coverage: E_cover=Cover_target-Cover_actual Here, Cover_target represents the target coverage, and Cover_actual represents the actual coverage. The coating deviation information E_layer can be determined based on the target coating thickness and the actual coating thickness.

[0125] E_layer=Thick_target-Thick_actual Where Thick_target represents the target coating thickness, and Thick_actual represents the actual coating thickness. The execution deviation information E_exec can be determined based on the target execution state and the actual execution state, for example:

[0126] E_exec=abs(Pos_target-Pos_actual)+abs(Angle_target-Angle_actual) Where Pos_target represents the target spraying position or distance, Pos_actual represents the actual spraying position or distance, Angle_target represents the target nozzle angle, and Angle_actual represents the actual nozzle angle. The above expression is only used to illustrate one feasible calculation method for deviation information.

[0127] S630, determine a first correction direction based on the coverage deviation information and the coating deviation information.

[0128] In this embodiment, the first correction direction refers to the adjustment trend of the spraying parameters determined jointly by the coverage deviation information and the coating deviation information, which is used to characterize the correction requirements of the spraying parameters due to the spraying quality state. For example, when the coverage is insufficient and the coating is too thin, the first correction direction can be to increase the amount of paint applied; when the coverage is excessive and the coating is too thick, the first correction direction can be to reduce the amount of paint applied.

[0129] In this embodiment, the first correction direction can be a set of directions targeting at least one parameter to be corrected in the target spraying parameters, used to characterize the adjustment trend of the coverage deviation information and coating deviation information on the spraying parameters. Specifically, it can be implemented in the following way:

[0130] P1, based on the coverage deviation information, determines the correction direction for paint flow rate and the correction direction for travel speed.

[0131] In this embodiment, the correction direction for paint flow rate refers to the direction in which the paint flow rate increases, decreases, or remains constant, determined based on the coverage deviation information. For example, if the coverage is insufficient, the correction direction for paint flow rate can be determined to be increasing; if the coverage is excessive, the correction direction for paint flow rate can be determined to be decreasing.

[0132] In this embodiment, the correction direction for walking speed refers to the direction in which the walking speed of the painting robot increases, decreases, or remains constant, determined based on the coverage deviation information. For example, if the coverage is insufficient, the correction direction for walking speed can be determined to decrease; if the coverage is excessive, the correction direction for walking speed can be determined to increase.

[0133] P2, based on the coating deviation information, determines the correction direction of the spraying distance and the correction magnitude of the paint flow rate.

[0134] In this embodiment, the correction direction for the spraying distance refers to the direction in which the nozzle approaches, moves away from, or remains stationary relative to the surface of the railway vehicle to be painted, determined based on coating deviation information. For example, if the coating is too thin, the correction direction for the spraying distance can be determined to be approaching; if the coating is too thick, the correction direction for the spraying distance can be determined to be moving away.

[0135] In this embodiment, the correction range for paint flow rate refers to the amount of paint flow rate adjustment determined based on coating deviation information. For example, the larger the coating thickness deviation, the larger the correction range for paint flow rate can be; when the coating thickness deviation is within the allowable range, the correction range for paint flow rate can be zero or a small value.

[0136] P3 generates a first correction direction to characterize the coverage state and coating state based on the correction requirements corresponding to the coverage deviation information and the coating deviation information.

[0137] The controller 50 can determine the first correction direction by comprehensively considering the correction requirements corresponding to the coverage deviation information and the coating deviation information. If both the coverage deviation information and the coating deviation information point to increasing the amount of coating applied, the first correction direction is the incremental direction; if both point to decreasing the amount of coating applied, the first correction direction is the decremental direction; if there are different correction requirements, the first correction direction can be determined based on the degree of deviation between the coverage deviation information and the coating deviation information.

[0138] In one specific embodiment, the first correction direction Dir_1 can be determined based on the weighted result of the coverage deviation information and the coating deviation information: Dir_1=sign(q1×E_cover+q2×E_layer) Here, sign() represents the sign function, and q1 and q2 are weight parameters. When Dir_1 is positive, it indicates the increment direction; when Dir_1 is negative, it indicates the decrement direction; and when Dir_1 is zero, it indicates the hold direction.

[0139] S640, determine the second correction direction based on the execution deviation information.

[0140] In this embodiment, the second correction direction refers to the adjustment trend of the spraying parameters determined by the deviation of the spraying execution state, which is used to characterize the robot's execution state's need to correct the spraying parameters. For example, when the actual nozzle angle is too small and the actual walking speed is too fast, the second correction direction can reflect the adjustment trend of increasing the nozzle angle or decreasing the walking speed.

[0141] In this embodiment, the second correction direction can be a set of directions targeting at least one parameter to be corrected in the target spraying parameters, used to characterize the adjustment trend of the spraying parameters by the execution deviation information. Specifically, it can be implemented in the following way:

[0142] F1 determines the correction direction of the nozzle angle and the correction magnitude of the travel speed based on the execution deviation information.

[0143] In this embodiment, the correction direction of the nozzle angle refers to the direction in which the nozzle angle increases, decreases, or remains constant, determined based on the execution deviation information. For example, when the actual nozzle angle is less than the target nozzle angle, the correction direction of the nozzle angle can be determined to be increasing.

[0144] In this embodiment, the correction range for walking speed refers to the adjustment amount of the painting robot's walking speed determined based on the execution deviation information. For example, when the actual walking speed is higher than the target walking speed, a negative walking speed correction range can be generated.

[0145] F2 generates a second correction direction to characterize the spraying execution state based on the correction direction of the nozzle angle and the correction magnitude of the travel speed.

[0146] In this embodiment, the second correction direction used to characterize the spraying execution state refers to the result of converting the nozzle angle correction direction and travel speed correction magnitude into a direction expression that is comparable to the first correction direction. For example, the second correction direction can be represented as an incremental direction, a decremental direction, or a holding direction, so as to make a consistency judgment with the first correction direction.

[0147] S650, when the first correction direction and the second correction direction are consistent with the adjustment direction of the same parameter to be corrected, the correction amount of the spraying parameter is generated according to the preset correction coefficient.

[0148] In this embodiment, the consistency between the first correction direction and the second correction direction means that the adjustment trends proposed by the spraying quality state and the execution state for the same parameter to be corrected are the same or do not conflict. For example, when both the first correction direction and the second correction direction point to reducing the walking speed, it can be determined that they are consistent.

[0149] In this embodiment, the preset correction coefficient refers to a coefficient used to control the magnitude of the correction when the first correction direction and the second correction direction are consistent. For example, the preset correction coefficient can be set according to the paint type, spraying cycle, robot response speed, or historical spraying data.

[0150] In this embodiment, the spraying parameter correction amount refers to the parameter increment or decrement used to correct the initial spraying parameters, and may include spraying distance correction amount, nozzle angle correction amount, paint flow rate correction amount, and travel speed correction amount. For example, the spraying parameter correction amount can be expressed as DeltaP=[DeltaD,DeltaA,DeltaF,DeltaV], where DeltaP represents the spraying parameter correction amount; DeltaD represents the spraying distance correction amount; DeltaA represents the nozzle angle correction amount; DeltaF represents the paint flow rate correction amount; and DeltaV represents the travel speed correction amount.

[0151] S660, when the adjustment directions of the first correction direction and the second correction direction for the same parameter to be corrected are inconsistent, the correction amount of the spraying parameter is generated according to the priority rule.

[0152] In this embodiment, the inconsistency between the first correction direction and the second correction direction refers to the conflicting or different adjustment trends proposed by the coating quality state and the execution state for the same parameter to be corrected. For example, if insufficient coverage requires a reduction in walking speed, but the robot's execution state deviation requires an increase in walking speed, then the two can be determined to be inconsistent.

[0153] In this embodiment, the priority rule refers to the rule used to determine which type of deviation information should be used as the dominant correction basis when the first correction direction and the second correction direction are inconsistent. For example, when the coating deviation information exceeds a preset coating deviation threshold, the spraying parameter correction amount can be generated based on the coating deviation information first; when the execution deviation information exceeds a preset execution deviation threshold, the spraying parameter correction amount can be generated based on the execution deviation information first.

[0154] S670, the initial spraying parameters are corrected according to the spraying parameter correction amount to generate the target spraying parameters.

[0155] In this embodiment, correcting the initial spraying parameters according to the spraying parameter correction amount means either adding the spraying parameter correction amount to the initial spraying parameters, or updating the initial spraying parameters according to the correction rules. For example, when the initial spraying parameters are P0=[D0,A0,F0,V0] and the spraying parameter correction amount is DeltaP=[DeltaD,DeltaA,DeltaF,DeltaV], the target spraying parameters can be P_target=[D0+DeltaD,A0+DeltaA,F0+DeltaF,V0+DeltaV].

[0156] S70, control the painting robot to perform the painting operation according to the target painting parameters.

[0157] In this embodiment, controlling the painting robot to perform the painting operation means controlling the painting robot and the painting execution mechanism to adjust the painting distance, nozzle angle, paint flow rate, and travel speed according to the target painting parameters, and then painting the surface of the railway vehicle to be painted. For example, the controller can generate distance control commands, angle control commands, flow rate control commands, and travel control commands according to the target painting parameters, and send them to the painting robot and the painting execution mechanism.

[0158] In one embodiment, reference Figure 8 Step S70 can be implemented in the following way: S710, determine multiple areas to be sprayed based on the surface state feature map.

[0159] In this embodiment, the area to be sprayed refers to the spraying operation unit obtained based on the surface state feature map, robot operation path, or vehicle surface structure. For example, the side wall of the vehicle body can be divided into multiple areas to be sprayed according to the grid, or the spatial range corresponding to the rust area, the stain area, and the geometrically abnormal area can be used as the area to be sprayed.

[0160] S720, determine the spraying sequence of each of the areas to be sprayed according to the surface condition level corresponding to each area to be sprayed.

[0161] In this embodiment, the spraying sequence refers to the order in which the painting robot performs spraying operations on multiple areas to be sprayed. For example, areas with higher surface condition levels can be sprayed first, or areas with severe corrosion or abnormal morphology can be treated first while ensuring path continuity.

[0162] In this embodiment, for each area to be sprayed, the controller 50 can determine the surface state level of the area based on the surface state feature map corresponding to the area to be sprayed, according to the processing method of step S30.

[0163] S730, generate spraying control instructions based on the target spraying parameters and the spraying sequence.

[0164] In this embodiment, the spraying control command refers to the control data used to drive the painting robot and the spraying actuator to perform the spraying action, and may include distance control command, angle control command, flow control command, and movement control command. For example, the spraying control command may instruct the painting robot to move at a target movement speed in a certain area to be sprayed, and control the nozzle to maintain the target spraying distance and target nozzle angle.

[0165] S740, according to the spraying control command, the painting robot is controlled to perform spraying operations on the multiple areas to be sprayed.

[0166] In this embodiment, performing painting operations on multiple areas to be painted means that the painting robot sequentially completes the painting process on multiple areas according to the painting sequence and painting control instructions. For example, the painting robot can first perform painting on severely rusted areas with higher coverage requirements, and then perform regular painting on areas with minor stains.

[0167] Through steps S10 to S70 described above, this embodiment can correlate the multispectral sensing results, environmental conditions, and spraying process feedback status of the railway vehicle surface to be painted. The surface condition feature map allows rust, stains, and geometric morphology information to be expressed in the same surface coordinate system; the surface condition level allows the surface condition to be converted into spraying decision information that the model can process; the lightweight neural network model allows the initial spraying parameters to be dynamically generated based on the surface and environmental conditions; and the closed-loop correction process allows the spraying parameters to be updated with the feedback from the spraying process.

[0168] When multiple defect types simultaneously reach a preset severity level, the highest severity level is determined through severity scoring, and a surface condition level is generated. This reduces the instability in spraying parameter selection caused by inconsistent level determinations when multiple defect types overlap. This approach makes the surface condition level more suitable for reflecting the surface condition that has a dominant influence on spraying control.

[0169] In the closed-loop correction process of spraying parameters, the first correction direction is determined by coverage deviation information and coating deviation information, and the second correction direction is determined by execution deviation information. This allows for the differentiation of the different impacts of spraying quality deviation and robot execution deviation on parameter adjustment. When the first and second correction directions are consistent, the spraying parameter correction amount is generated according to a preset correction coefficient; when the directions are inconsistent, the spraying parameter correction amount is generated according to a priority rule, thereby enabling the coordinated handling of the influence of multiple types of deviation information on the spraying parameters.

[0170] In the partitioned spraying implementation, multiple areas to be sprayed are determined by surface condition feature maps, and the spraying sequence is determined according to the surface condition level of each area to be sprayed. This allows the painting robot to perform differentiated spraying control according to the surface condition of different areas, adapting to the maintenance scenario of uneven surface condition of railway vehicles.

[0171] refer to Figure 9 This invention also provides an adaptive spraying parameter adjustment system for a railway maintenance painting robot based on surface state perception. The system includes an information acquisition module 100, a feature map construction module 200, a state level determination module 300, an input information generation module 400, an initial parameter generation module 500, a target parameter generation module 600, and a spraying control module 700.

[0172] The information acquisition module 100 is used to acquire multispectral image information of the surface of the railway vehicle to be painted, as well as temperature, humidity, and wind speed information of the painting environment. The information acquisition module 100 can be connected to the multispectral vision sensor 10 and the environmental sensor 20. The information acquisition module 100 can timestamp, cache, and convert the data format of the acquired data, and send the multispectral image information to the feature map construction module 200, and the temperature, humidity, and wind speed information to the input information generation module 400.

[0173] The feature map construction module 200 is used to identify rust, stain, and geometric morphology information on the surface of the railway vehicle to be painted based on the multispectral image information, and to construct a surface state feature map based on the rust, stain, and geometric morphology information. The feature map construction module 200 can perform band alignment processing, region identification, region location determination, region feature value calculation, and feature layer generation. The surface state feature map output by the feature map construction module 200 is sent to the state level determination module 300, or it can be sent to the spraying control module 700 for zone control.

[0174] The surface condition level determination module 300 is used to determine the surface condition level of the railway vehicle to be painted based on the surface condition feature map. The module 300 can extract rust feature values, stain feature values, and morphology feature values ​​from the surface condition feature map, determine the rust level, stain level, and morphology level respectively, and determine the highest severity level based on a severity score when multiple levels reach a preset severity level, thereby generating the surface condition level. The surface condition level determination module 300 sends the surface condition level to the input information generation module 400 and the spraying control module 700.

[0175] The input information generation module 400 is used to generate spraying parameter prediction input information based on the surface condition level, the temperature and humidity information, and the wind speed information. The input information generation module 400 can generate surface condition coding information and environmental condition coding information, combine them into joint coding information, and convert the joint coding information into the input format required by the lightweight neural network model.

[0176] The initial parameter generation module 500 is used to input the predicted spraying parameters into a preset lightweight neural network model to generate initial spraying parameters such as spraying distance, nozzle angle, paint flow rate, and walking speed. The initial parameter generation module 500 can deploy the lightweight neural network model and perform constraint processing on the model output results to obtain initial spraying parameters within the allowable range of the painting robot 30 and the spraying execution mechanism 40.

[0177] The target parameter generation module 600 is used to acquire real-time spraying status information during the spraying process, determine spraying deviation information based on the real-time spraying status information, and correct the initial spraying parameters based on the spraying deviation information to generate target spraying parameters. The target parameter generation module 600 can receive real-time coverage status information, real-time coating status information, and spraying execution status information, and determine coverage deviation information, coating deviation information, and execution deviation information. The target parameter generation module 600 can further determine a first correction direction and a second correction direction, and generate spraying parameter correction amounts based on direction consistency judgment and priority rules.

[0178] The spraying control module 700 is used to control the painting robot to perform spraying operations according to the target spraying parameters. The spraying control module 700 can generate distance control commands, angle control commands, flow control commands, and travel control commands based on the target spraying parameters, and send them to the painting robot 30 and the spraying execution mechanism 40. The spraying control module 700 can also determine multiple areas to be sprayed based on surface condition feature maps, and determine the spraying sequence based on the surface condition level corresponding to each area, thereby controlling the painting robot 30 to perform spraying operations on multiple areas to be sprayed.

[0179] In one embodiment, the above modules can be implemented by software programs and run in the controller 50, edge computing device, industrial computer, or robot controller. In another embodiment, the above modules can be implemented by hardware circuits, programmable logic devices, processors, and memory. In yet another embodiment, the above modules can be distributed and deployed at the equipment end of the painting robot 30, the edge end of the painting station, and the server end of the maintenance workshop; the specific deployment method does not affect the data flow and functional collaboration between the modules.

[0180] In one specific embodiment, the controller 50 may include a processor and a memory. The memory stores a computer program, which, when executed by the processor, can implement the adaptive spraying parameter adjustment method for a railway maintenance painting robot based on surface state perception, as described in any of the above embodiments. The memory may also store a lightweight neural network model, spraying parameter constraint ranges, level mapping rules, severity scoring rules, and priority rules.

[0181] In one example, the lightweight neural network model can be a regression model containing an input layer, several hidden layers, and an output layer. The input layer receives spraying parameter prediction input information, and the output layer outputs predicted values ​​for spraying distance, nozzle angle, paint flow rate, and travel speed. The model can be trained based on historical spraying samples, which may include spraying operation data under different rust conditions, different stain conditions, different geometric morphologies, different temperature and humidity conditions, and different wind speed conditions. After the model is trained, it can be deployed on the controller 50 for online prediction of initial spraying parameters.

[0182] In one specific embodiment, the lightweight neural network model can employ either a multilayer perceptron structure or a lightweight bottleneck structure. If a multilayer perceptron structure is used, its input dimension can be 7, the output dimension of the first fully connected layer can be 32, the output dimension of the second fully connected layer can be 16, the output dimension of the third fully connected layer can be 8, and the output dimension can be 4. If a lightweight bottleneck structure is used, its input dimension can be 7, the output dimension of the upscaling layer can be 24, the output dimension of the compression layer can be 12, and the output dimension of the output layer can be 4. Both structures can output predicted values ​​for spraying distance, nozzle angle, paint flow rate, and travel speed.

[0183] In one specific embodiment, the training samples for the lightweight neural network model can be derived from historical railway maintenance and painting operation data, painting process test data, and manually labeled data. Each training sample includes sample input and sample labels. Sample input includes sample surface condition level, sample temperature, sample humidity, and sample wind speed; sample labels include spraying distance label, nozzle angle label, paint flow rate label, and walking speed label determined under the corresponding sample input conditions. The training samples cover different rust states, different stain states, different geometric morphologies, and different temperature, humidity, and wind speed conditions. Before training, the sample input is normalized. During training, a supervised learning method is used, and a weighted mean square error loss function is used as the training loss. After training, the lightweight neural network model is evaluated using a test set. If the prediction error of the test set meets the preset error requirements, the trained lightweight neural network model is deployed to the controller 50 or an edge computing device.

[0184] In one optional embodiment, to accommodate the computing resources of the local control unit of the painting robot, the trained lightweight neural network model can undergo model compression, parameter pruning, or quantization. Quantization can employ FP32 (Floating Point 32-bit) to INT8 (Integer 8-bit) parameter conversion to reduce model storage space and inference time. The output results after model compression or quantization are still used to generate initial painting parameters such as spraying distance, nozzle angle, paint flow rate, and travel speed.

[0185] In one specific embodiment, the deployed lightweight neural network model receives spraying parameter prediction input information and outputs a parameter prediction vector in each spraying control cycle. The controller 50 performs constraint processing on the parameter prediction vector to obtain initial spraying parameters, and then passes these initial spraying parameters to subsequent closed-loop correction steps. Through this training and deployment method, the lightweight neural network model can generate initial spraying parameters usable for spraying control at railway maintenance painting sites based on surface condition levels and environmental conditions.

[0186] The specific values, equipment models, model structures, sampling frequencies, measurement accuracy, model sizes, inference time, memory usage, thresholds, number of actuators, communication methods, weight parameters, and error results mentioned above are all examples under specific embodiments or experimental conditions, used to illustrate the feasible implementation methods and technical effects of the present invention, and should not be construed as limiting the scope of protection of the present invention.

[0187] The above description is merely a preferred embodiment of the present invention and is not intended to limit the scope of protection of the present invention. Any modifications, equivalent substitutions, improvements, or combinations made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

Claims

1. A method for adaptive spraying parameter adjustment of a railway maintenance painting robot based on surface state perception, characterized in that, include: Acquire multispectral image information of the surface of the railway vehicle to be painted, as well as temperature, humidity and wind speed information of the painting environment; Based on the multispectral image information, the rust information, stain information and geometric morphology information of the surface of the railway vehicle to be painted are identified respectively, and a surface state feature map is constructed based on the rust information, the stain information and the geometric morphology information; The surface condition level of the railway vehicle surface to be painted is determined based on the surface condition feature map. Based on the surface condition level, the temperature and humidity information, and the wind speed information, predictive input information for spraying parameters is generated; The predicted spraying parameters are input into a preset lightweight neural network model to generate initial spraying parameters such as spraying distance, nozzle angle, paint flow rate, and travel speed. The system acquires real-time spraying status information during the spraying process, determines spraying deviation information based on the real-time spraying status information, and corrects the initial spraying parameters based on the spraying deviation information to generate target spraying parameters. The painting robot is controlled to perform the painting operation according to the target painting parameters.

2. The method according to claim 1, characterized in that, The step of identifying rust, stain, and geometric morphology information on the surface of the railway vehicle to be painted based on the multispectral image information, and constructing a surface state feature map based on the rust, stain, and geometric morphology information, includes: The multispectral image information is subjected to band alignment processing to obtain aligned image information; Based on the aligned image information, the rust area, stain area, and geometric shape area are determined respectively; Determine the regional positions and regional characteristic values ​​of the rusted area, the stained area, and the geometric morphology area in the same surface coordinate system; Based on the location of the region and the region's feature values, a surface state feature map is generated, which includes a rust feature layer, a stain feature layer, and a geometric shape feature layer.

3. The method according to claim 2, characterized in that, Determining the surface condition level of the railway vehicle to be painted based on the surface condition feature map includes: Rust feature values, stain feature values, and morphology feature values ​​are extracted from the rust feature layer, the stain feature layer, and the geometric morphology feature layer, respectively. Based on the corrosion characteristic value, the stain characteristic value, and the morphology characteristic value, the corrosion grade, stain grade, and morphology grade are determined respectively. When there are multiple target levels that reach a preset severity level among the corrosion level, the stain level and the morphology level, obtain the severity score corresponding to each target level; The highest severity level is determined based on the severity score corresponding to each of the aforementioned target levels; The surface condition level is generated based on the highest severity level.

4. The method according to claim 3, characterized in that, The step of generating spraying parameter prediction input information based on the surface condition level, the temperature and humidity information, and the wind speed information includes: Surface state coding information is generated based on the surface state level; Environmental status coding information is generated based on the temperature and humidity information and the wind speed information; The surface state coding information and the environmental state coding information are combined to obtain joint coding information; The spraying parameter prediction input information is generated based on the joint coding information.

5. The method according to claim 4, characterized in that, The step of inputting the predicted spraying parameters into a preset lightweight neural network model to generate initial spraying parameters, including spraying distance, nozzle angle, paint flow rate, and travel speed, includes: The spraying parameter prediction input information is input into the lightweight neural network model to obtain the predicted values ​​of spraying distance, nozzle angle, paint flow rate and travel speed. Based on the preset spraying parameter constraint range, the predicted values ​​of the spraying distance, the nozzle angle, the paint flow rate, and the travel speed are constrained. The initial spraying parameters are generated based on the predicted values ​​of spraying distance, nozzle angle, paint flow rate, and travel speed after constraint processing.

6. The method according to claim 5, characterized in that, The process of acquiring real-time spraying status information during the spraying process, determining spraying deviation information based on the real-time spraying status information, and correcting the initial spraying parameters based on the spraying deviation information to generate target spraying parameters includes: Acquire real-time coverage status information, real-time coating status information, and spraying execution status information during the spraying process; Coverage deviation information is determined based on the real-time coverage status information, coating deviation information is determined based on the real-time coating status information, and execution deviation information is determined based on the spraying execution status information; A first correction direction is determined based on the coverage deviation information and the coating deviation information; A second correction direction is determined based on the execution deviation information; When the first correction direction is consistent with the second correction direction, the correction amount of the spraying parameters is generated according to the preset correction coefficient; When the first correction direction and the second correction direction are inconsistent, the spraying parameter correction amount is generated according to the priority rule; The initial spraying parameters are corrected according to the spraying parameter correction amount to generate the target spraying parameters.

7. The method according to claim 6, characterized in that, Determining the first correction direction based on the coverage deviation information and the coating deviation information includes: The correction direction for paint flow rate and the correction direction for travel speed are determined based on the coverage deviation information. The correction direction for the spraying distance and the correction magnitude for the paint flow rate are determined based on the coating deviation information. Based on the correction requirements corresponding to the first correction direction and the execution deviation information, a first correction direction is generated to characterize the coverage state and coating state.

8. The method according to claim 6, characterized in that, Determining the second correction direction based on the execution deviation information includes: The correction direction of the nozzle angle and the correction magnitude of the travel speed are determined based on the execution deviation information. A second correction direction is generated based on the correction direction of the nozzle angle and the correction magnitude of the travel speed to characterize the spraying execution state.

9. The method according to claim 1, characterized in that, The step of controlling the painting robot to perform the painting operation according to the target painting parameters includes: Multiple areas to be sprayed are determined based on the surface state feature map; The spraying sequence of each area to be sprayed is determined according to the surface condition level corresponding to each area to be sprayed. Generate spraying control instructions based on the target spraying parameters and the spraying sequence; The painting robot is controlled to perform painting operations on the multiple areas to be painted according to the painting control command.

10. An adaptive spraying parameter adjustment system for a railway maintenance painting robot based on surface state perception, characterized in that, include: The information acquisition module is used to acquire multispectral image information of the surface of the railway vehicle to be painted, as well as temperature, humidity and wind speed information of the painting operation environment; The feature map construction module is used to identify the rust information, stain information and geometric shape information of the surface of the railway vehicle to be painted according to the multispectral image information, and to construct a surface state feature map based on the rust information, the stain information and the geometric shape information; A state level determination module is used to determine the surface state level of the railway vehicle surface to be painted based on the surface state feature map. The input information generation module is used to generate spraying parameter prediction input information based on the surface condition level, the temperature and humidity information, and the wind speed information; The initial parameter generation module is used to input the spraying parameter prediction input information into a preset lightweight neural network model to generate initial spraying parameters such as spraying distance, nozzle angle, paint flow rate and travel speed. The target parameter generation module is used to acquire real-time spraying status information during the spraying process, determine spraying deviation information based on the real-time spraying status information, and correct the initial spraying parameters based on the spraying deviation information to generate target spraying parameters. The spraying control module is used to control the painting robot to perform spraying operations according to the target spraying parameters.