Underwater cleaning method, cleaning control device and underwater cleaning robot

Abstract

The invention provides an underwater cleaning method, a cleaning control device and an underwater cleaning robot, and the method comprises the steps: S1, recognizing the types and growth cycles of attachments on each region of a ship body in a video image through a deep learning network according to the underwater video frame image of the surface of the ship body, and recognizing the type of a coating on the surface of the ship body; s2, for each area, obtaining attachment hardness corresponding to the type of the attachment on the area and a paint fragility grade corresponding to the type of the paint on the surface of the ship body; s3, aiming at each area, activating all fuzzy rules matched with the obtained attachment hardness, the coating fragility grade and the current growth cycle of the attachment in a preset fuzzy rule base, and determining operation parameters of a corresponding cleaning execution device according to the activated fuzzy rules; and S4, according to the determined operation parameters, a cleaning execution device is controlled to conduct cleaning. By means of the technical scheme, the hull condition can be recognized, and the cleaning strategy can be adaptively adjusted.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of ship maintenance, and in particular to an underwater cleaning method, a cleaning control device and an underwater cleaning robot. BACKGROUND

[0002] During long-term sailing, a large amount of marine organisms such as algae and shellfish will attach to the underwater part of the ship, which is called "biofouling". Biofouling can significantly increase the sailing resistance of the ship, resulting in increased fuel consumption and reduced sailing speed. Therefore, it is usually necessary to regularly clean the underwater part of the ship.

[0003] At present, a ship cleaning robot is usually used to clean the underwater part of the ship. However, most of the current ship cleaning robots use fixed cleaning parameters, such as using a uniform type of brush disc and constant rotation speed and pressure for indiscriminate cleaning. However, the condition of the ship surface is very complex, for example:

[0004] Variety of attachments: from soft mucous algae to hard calcareous barnacles, the adhesion, hardness and thickness differ greatly;

[0005] Different types of paint: the ship surface is painted with different types of antifouling paint, such as relatively fragile self-polishing paint and relatively hard paint;

[0006] Different growth cycles: even the same type of attachment has completely different adhesion strengths at different growth stages.

[0007] The existing traditional "one-size-fits-all" cleaning method can cause many problems, for example:

[0008] (1) Inadequate cleaning: for stubborn attachments, fixed low-intensity cleaning cannot completely remove them;

[0009] (2) Over-cleaning / damage to paint: for light dirt or fragile paint layers, hard brush discs or excessive pressure can seriously wear or even damage the antifouling paint, not only shortening the service life of the paint, but also possibly releasing harmful substances into the ocean.

[0010] Therefore, there is an urgent need for a system that can intelligently identify the condition of the ship and adaptively adjust the cleaning strategy to achieve efficient, accurate and non-destructive cleaning. SUMMARY

[0011] In view of the above problems of the prior art, the embodiments of the present application provide an underwater cleaning method, a cleaning control device and an underwater cleaning robot.

[0012] In order to achieve the above-mentioned purpose, in one aspect, an underwater cleaning method is provided, which is applied to an underwater cleaning robot, the underwater cleaning robot comprising an image acquisition device, a cleaning execution device and a cleaning control device, comprising:

[0013] S1, the cleaning control device identifies the type and the current growth period of the fouling on each region of the underwater hull pre-divided in the underwater video image according to the underwater video frame image of the hull surface, and identifies the paint type of the hull surface using a pre-trained deep learning network, wherein the underwater video frame image is obtained by an image acquisition device;

[0014] S2, the cleaning control device obtains the hardness of the fouling corresponding to the type of the fouling on the region and the paint fragility level corresponding to the paint type of the hull surface for each region by querying a preset mapping table;

[0015] S3, the cleaning control device activates all fuzzy rules in the pre-set fuzzy rule base that match the obtained hardness of the fouling, the paint fragility level, and the current growth period of the fouling for each region, and determines the operation parameters of the corresponding cleaning execution device according to the activated fuzzy rules, the operation parameters including: the material of the cleaning execution device, the rotation speed of the cleaning execution device, the contact pressure of the cleaning execution device and the hull surface, and the moving speed of the underwater cleaning robot;

[0016] S4, the cleaning control device selects the cleaning execution device of the corresponding material corresponding to the region according to the operation parameters of the cleaning execution device determined in step S3 for each region, adjusts the rotation speed of the selected cleaning execution device, the contact pressure of the cleaning execution device and the hull surface, and the moving speed of the underwater cleaning robot to clean the region.

[0017] Preferably, in the underwater cleaning method, the input conditions of the fuzzy rules include the hardness of the fouling, the paint fragility level of the hull surface, and / or the current growth period of the fouling; and the decision output of the fuzzy rules includes the material of the cleaning execution device, the rotation speed level of the cleaning execution device, and / or the contact pressure level of the cleaning execution device and the hull surface.

[0018] Preferably, in the underwater cleaning method, the step S3 of determining the operation parameters of the corresponding cleaning execution device according to the activated fuzzy rules includes:

[0019] using the pre-set corresponding relationship between the rotation speed level of the cleaning execution device and the rotation speed of the cleaning execution device, and the corresponding relationship between the contact pressure level of the cleaning execution device and the hull surface and the contact pressure of the cleaning execution device and the hull surface to determine the operation parameters of the corresponding cleaning execution device;

[0020] When the activated fuzzy rule is one, the rotational speed level and / or the contact pressure level of the cleaning execution device corresponding to the rotational speed and the pressure output by the fuzzy rule are taken as the operation parameters of the cleaning execution device; when the activated fuzzy rule is multiple, the corresponding rotational speed or pressure output by each of the multiple fuzzy rules is weighted and averaged according to a predetermined weight, and the weighted and averaged value is taken as the operation parameter of the cleaning execution device.

[0021] Preferably, the underwater cleaning method further comprises:

[0022] The cleaning control device uses the hardness score in the predetermined value range to represent the hardness of the adherend corresponding to the type of the adherend according to the preset mapping rule.

[0023] When a region contains multiple types of adherends, the hardness scores corresponding to the hardness of each adherend are weighted and averaged according to the area ratio of each adherend type, and the hardness score obtained after the weighted and averaged is taken as the hardness score corresponding to the region.

[0024] Preferably, the underwater cleaning method further comprises:

[0025] The obtained underwater video frame image is enhanced using a cross-domain mutual learning mechanism; wherein the cross-domain mutual learning mechanism comprises: a teacher network, a student network and a turbidity enhancement module; the teacher network is used for training in a synthetic domain, and knowledge is transferred to the student network through parameter sharing and time weighted average; the student network is trained in a real domain, and knowledge is fed back to the teacher network through adversarial loss and prior loss; the turbidity enhancement module is used to dynamically generate pseudo-real image samples for nonlinear turbidity enhancement.

[0026] Preferably, the underwater cleaning method further comprises one or more of the following:

[0027] The adherend hardness is divided into different hardness levels according to the hardness score from low to high; wherein the hardness levels include: soft fouling, light hard fouling, medium hard fouling and heavy hard fouling;

[0028] The types of the ship body adherends include: biological models, enteromorpha, tubular worms, mussels, algae, barnacles and / or bryozoans;

[0029] The growth cycle includes: early stage, middle stage and mature stage;

[0030] The paint types include: self-polishing paint, hard epoxy paint and / or organic silicon non-toxic paint.

[0031] Preferably, the underwater cleaning method, the cleaning execution device comprises a brush disc; the brush disc comprises a pressure sensor for monitoring the current contact pressure of the brush disc with the hull surface; the cleaning control device comprises an active pressure control structure; the active pressure control structure comprises a pressure actuator for controlling the output pressure of the brush disc; and a pressure PID controller for controlling the output of the pressure actuator according to the contact pressure of the cleaning execution device with the hull surface and the current contact pressure of the brush disc determined in step S3.

[0032] Preferably, the underwater cleaning method, the deep learning network comprises a backbone network and two parallel detection or classification heads connected with the backbone network; wherein the two parallel detection or classification heads comprise a dirt identification head for outputting the position of the detection frame, the category of the attached object and the current growth cycle; and a paint identification head for classifying the entire image or the selected region of the image and outputting the paint type.

[0033] In another aspect, a cleaning control device is provided, comprising a memory and a processor, the memory storing at least one program, the at least one program being executed by the processor to implement the underwater cleaning method as described in any of the above.

[0034] In yet another aspect, an underwater cleaning robot is provided, comprising: an image acquisition device for acquiring underwater video frame images of the hull surface; a cleaning execution device for cleaning the hull surface; and a cleaning control device as described above.

[0035] The above technical solutions have the following technical effects:

[0036] The technical scheme of the embodiment of the present application identifies the type of the attached matter in each region pre-divided in the underwater video frame image of the ship surface, the growth period of the attached matter currently in, and the type of the paint on the ship surface by using a deep learning network; then, according to the identified result and a preset mapping table, the hardness of the attached matter contained in each region and the fragility level corresponding to the type of the paint are determined; and for each region, all fuzzy rules matched with the obtained hardness of the attached matter, the fragility level of the paint and the growth period of the attached matter currently in are activated from a pre-set fuzzy rule library, and the operation parameters of the corresponding cleaning execution device of the region are determined according to the activated fuzzy rules, and the underwater regions of the ship surface are cleaned according to the determined operation parameters; thus, intelligent identification of the ship condition and adaptive adjustment of the cleaning strategy are realized; and the above technical scheme of the embodiment of the present application can match the cleaning execution structure such as a brush disc and the appropriate cleaning strength for different ship fouling conditions by identifying and distinguishing the attached dirt and the type of the paint on the ship, avoids paint damage caused by inappropriate cleaning execution tool materials or “excessive force” and cleaning residues caused by “insufficient force”, and realizes accurate and non-destructive cleaning of the ship surface.

[0037] Further, by selecting appropriate operation parameters according to the actual condition of the attached matter, the overall operation efficiency can be greatly improved by using high-speed and low-pressure mode for fast passing through the light dirt and concentrating time on processing stubborn attached regions; and the paint life is prolonged and the maintenance cost is reduced; and the technical scheme of the embodiment of the present application has high intelligence and automation, the whole process does not need manual intervention, the dependence on the experience of the operator is reduced, and the stability and consistency of the cleaning quality are ensured; and it protects the antifouling paint, reduces the premature release of active substances in the paint into the ocean, and has positive environmental protection significance. BRIEF DESCRIPTION OF DRAWINGS

[0038] Figure 1 FIG. 1 is a flowchart of an underwater cleaning method applied to an underwater cleaning robot according to an embodiment of the present application. DETAILED DESCRIPTION

[0039] To further illustrate the embodiments, the present application provides drawings. These drawings are part of the disclosure of the present application, mainly used to illustrate the embodiments, and can be used to explain the operating principle of the embodiments in conjunction with the related description of the specification. Those of ordinary skill in the art should be able to understand other possible implementations and advantages of the present application by referring to these contents. The components in the drawings are not drawn to scale, and similar component symbols are generally used to represent similar components.

[0040] The present application will be further described in conjunction with the drawings and specific embodiments.

[0041] Embodiment one:

[0042] Figure 1 Flowchart of the underwater cleaning method for the underwater cleaning robot according to an embodiment of the present application. The method according to the embodiment is used for cleaning the surface of a ship body underwater. As shown in FIG. 1, the method according to the embodiment includes the following steps performed by a cleaning control device: Figure 1

[0043] S1, according to an underwater video frame image of the surface of a ship body, using a pre-trained deep learning network to identify the type of the attached object and the current growth period of the attached object in each region of the underwater ship body pre-divided in the underwater video image, and to identify the type of the paint on the surface of the ship body, wherein the underwater video frame image is obtained by an image acquisition device;

[0044] S2, for each region, by querying a preset mapping table, obtaining the hardness of the attached object corresponding to the type of the attached object on the region and the paint fragility level corresponding to the type of the paint on the surface of the ship body; wherein the mapping table pre-stores the corresponding relationship between the hardness of the attached object corresponding to the type of the attached object and the paint fragility level corresponding to the type of the paint on the surface of the ship body;

[0045] S3, for each region, all fuzzy rules in a pre-set fuzzy rule base that match the obtained hardness of the attached object, the paint fragility level and the current growth period of the attached object are activated, and the operation parameters of the corresponding cleaning execution device are determined according to the activated fuzzy rules, the operation parameters including: the material of the cleaning execution device, the rotating speed of the cleaning execution device, the contact pressure of the cleaning execution device and the ship body surface and the moving speed of the underwater cleaning robot;

[0046] S4, according to the operation parameters of the cleaning execution device determined in step S3, for each region, selecting the cleaning execution device of the corresponding material corresponding to the region, adjusting the rotating speed of the selected cleaning execution device, the contact pressure of the cleaning execution device and the ship body surface and the moving speed of the underwater cleaning robot, to clean the region.

[0047] Embodiment two:

[0048] The underwater cleaning method according to the embodiment of the present application includes an image recognition step and a cleaning strategy matching step.

[0049] (A) Image recognition

[0050] In this embodiment of the present application, a multi-stage convolutional neural network (Multi-Stage CNN) is used to realize image recognition. The image recognition according to the embodiment of the present application includes a step of pre-processing the underwater video frame image collected by an image acquisition device such as a camera. The image recognition process according to the embodiment of the present application is described below. ​

[0051] Phase one, data preprocessing:

[0052] (1) Underwater image enhancement: To solve the problem of image quality degradation such as blurring, color deviation and low contrast caused by light scattering and absorption in underwater imaging, the embodiment of the present application realizes underwater video enhancement processing on the collected video frames based on a cross-domain mutual learning mechanism. This mechanism uses synthetic image data (with clear original images as a reference) and real underwater image data for collaborative training to improve the image enhancement effect of the model in the real turbid environment.

[0053] The cross-domain mutual learning mechanism (Mutual Learning for Domain Adaptation) mainly consists of a teacher network, a student network and a turbidity enhancement module. The teacher network (Teacher) performs supervised learning on the synthetic dataset containing "turbid image-clear image" pairs, i.e. the synthetic domain, thereby learning the mapping knowledge from turbid to clear, and transferring knowledge to the student network through parameter sharing and time weighted average. The student network (Student) is trained on the real underwater dataset, i.e. the real domain, and receives knowledge transferred by the teacher network through parameter sharing and time weighted average (Exponential Moving Average, EMA) strategies, while feeding back the knowledge of the real domain to the teacher network through adversarial loss and prior loss. Through the turbidity enhancement module, pseudo-real image samples with different turbidity levels are dynamically generated for nonlinear turbidity enhancement, which narrows the domain gap and enhances the diversity of training data to help the teacher network better adapt to the real environment. Breaking the traditional one-way enhancement, cross-domain online collaborative optimization is realized, avoiding complex multi-stage or multi-field conversion training process.

[0054] This two-way knowledge transfer and collaborative optimization between the teacher and the student form a "flywheel effect", which promotes mutual improvement in training and improves the performance of image enhancement for both networks.

[0055] In one specific implementation of this mechanism, the following formulas are used to describe the network structure or loss function:

[0056] 1. Formula describing the residual learning module in the network:

[0057] Y = X + Conv(W2, b2)(ELU(Conv(W1, b1)(X)))

[0058] Wherein, Y represents the output feature map of the module; X represents the input feature map of the module; Conv represents convolution operation; W1, W2 represent the weights (convolution kernel) of the first and second convolution layers respectively; b1, b2 represent the bias terms of the first and second convolution layers respectively; ELU represents Exponetial Linear Unit, which is an activation function.

[0059] 2, describes the loss function L used in student network training rc :

[0060] L rc = L ca (G r→c (x rc ),y rc )+λ HDA L ca (G r→c (x cr ),y cr )

[0061] Wherein, L rc represents the partial loss function of the student network, L ca represents the content loss function (for example, Charbonnier loss), which is used to measure the consistency of the generated image and the target image in content.

[0062] Gs→c represents the student network, which is used to map the real hazy image (r) to the clean image (c); x rc represents the real hazy image corresponding to the clean image; y rc represents the clean reference image corresponding to x rc ; λ HDA represents the weighting function of the loss generated by the haze density enhancement module (HDA); x cr and y cr represent the pseudo real hazy image and the corresponding pseudo clean image generated by the HAD module respectively.

[0063] 3, describes the formula of the content loss function L ca (taking Charbonnier loss as an example):

[0064]

[0065] Wherein, N represents the total number of pixels of the image; X i represents the i-th pixel value of the generated image; Y i represents the i-th pixel value of the target clean image; ∈ is a very small normal number, such as 10 -3This is used to ensure that the function is differentiable at zero and maintain numerical stability. 4. Describe the unsupervised loss function L used when the student network feeds back knowledge to the teacher network. un :

[0066] L un =L adv (G r→c (x rr ))+L DC (G r→c (x rr ))

[0067] Among them, L un Represents unsupervised loss; L adv Represents adversarial loss, used to make images generated by student networks stylistically closer to real, sharp images; L DC The Dark Channel Prior loss is a loss function that uses image statistical regularities to constrain the dehazing / de-hazing effect; x rr This represents a truly murky image that does not correspond to a clear image.

[0068] (2) Noise suppression: Advanced filtering algorithms such as non-local means are used to remove image noise caused by underwater suspended particles.

[0069] Phase Two: Feature Extraction and Recognition

[0070] (1) Model selection: YOLOv7 or EfficientNet-B7, which perform well in object detection and image classification tasks, were selected as the base models. The YOLO series is good at real-time object detection and can quickly frame different types of attachments; EfficientNet achieves a good balance between classification accuracy and computational efficiency.

[0071] (2) Training Data: Construct a large-scale, finely annotated "underwater image database of ship hulls". This database includes:

[0072] (3) Attachment samples: covering at least 20 common marine attachments, such as seaweed, tube worms, barnacles, mussels, etc., and labeling each attachment with its category and three growth cycles; in practice, the growth cycle includes: early stage, middle stage and mature stage.

[0073] (4) Coating samples: including self-polishing coatings, hard epoxy coatings, non-toxic silicone coatings, etc. with different colors, gloss levels and aging levels.

[0074] (5) Training strategy: The backbone of the Multi-Task Learning network is responsible for extracting general features.

[0075] (6) Two parallel "detection / classification heads" are connected afterwards:

[0076] ① Fouling Recognition Head: responsible for outputting the position of the bounding box, the category of the attached object (e.g., "barnacles"), the confidence score, and an additional output - "growth cycle"; in one specific implementation, 1, 2, and 3 are used to represent the initial, middle, and mature stages, respectively.

[0077] ② Coating Recognition Head: classifies the entire image or a specific region and outputs the coating type, such as "self-polishing coating".

[0078] (B) Cleaning strategy matching

[0079] The embodiment of the present application realizes cleaning strategy matching based on the intelligent decision algorithm of fuzzy logic.

[0080] Traditional IF-THEN rules are too rigid and cannot handle the uncertainty in the recognition results, for example, "80% possibility is barnacles, 20% is oysters"; while fuzzy logic can well simulate the human fuzzy judgment process.

[0081] Input variables:

[0082] (1) Hardness of attached object: based on the pre-set relationship between the category of attached object and the hardness of attached object, the hardness of the corresponding category of attached object is determined after the category of attached object is recognized; in one specific implementation, low and high are used to qualitatively distinguish the hardness of attached object, for example, algae = low, barnacles = high; preferably, the hardness of attached object is further mapped to a quantitative value, for example, to a continuous value within a predetermined value range such as [0, 10]. The higher the value, the higher the hardness. In further implementations, the above predetermined value range is further divided into multiple sub-regions, different regions correspond to different hardness levels, to more finely distinguish the hardness levels of different categories of attached objects and adopt corresponding different cleaning strategies.

[0083] Mapping mechanism

[0084] The mapping of "hardness of attached object" is the process of converting the qualitative category recognized by the robot vision system, such as "algae", into a quantitative value that can be used for engineering control. This process usually includes two core steps:

[0085] Fouling Identification: First, the enhanced underwater video frames are analyzed by a pre-trained deep learning vision model. The model identifies the main fouling species in each region of the image and outputs a classification label, such as algae, tube worms, or barnacles.

[0086] Hardness Mapping: A predefined "Fouling Hardness Level Lookup Table" is pre-set, which pre-stores the mapping relationship between fouling species labels and hardness levels, basic hardness scores. According to the classification label identified in the previous step, the table is queried, and the corresponding basic hardness score is assigned to each region of the pre-divided underwater window surface. If there are multiple fouling species in a region, a weighted average is further performed according to the percentage of the area occupied by each species, so as to obtain a continuous value in the range of [0, 10] to more finely reflect the comprehensive cleaning difficulty of mixed fouling.

[0087] A detailed "Fouling Hardness Level Mapping Example" is given below, which takes into account the different growth cycles of barnacles and other organisms.

[0088] Table 1: Detailed Fouling Hardness Level Mapping Table

[0089]

[0090] Scenario Example:

[0091] Scenario A: Identify a region of the ship body as "algae".

[0092] Mapping result: Query the table and assign a hardness value of 2.0 to this region.

[0093] Scenario B: Identify another region as "barnacles, and its growth cycle is late / heavy colony.

[0094] Mapping result: Assign a hardness value of 10.0 to this region.

[0095] In one specific implementation, after obtaining the above-mentioned quantized "hardness value" (hardness score), it is used to provide adaptive job decision-making for underwater cleaning robots, so that the robot upgrades from "executing commands" to "thinking decisions".

[0096] The main applications of hardness value include:

[0097] Dynamic adjustment of cleaning parameters: This is the most direct application. The robot can adjust its physical actions in real time according to the hardness value to achieve the best balance of energy efficiency and effect. The cleaning parameters are as follows:

[0098] Cleaning pressure: 30% standard pressure for algae with hardness 2.0, and automatically increase to 100% maximum for heavy barnacles with hardness 10.0.

[0099] Travel speed: keep 100% speed for soft stains, and automatically reduce to 50% when encountering high hardness areas to increase polishing time.

[0100] Brush disc power: dynamically adjust the motor torque and speed of the cleaning brush disc to exert greater cleaning force on hard stains.

[0101] The above specific values are only exemplary and can be adjusted adaptively according to actual conditions.

[0102] After the ship surface is cleaned by the cleaning method of the embodiment of the present application, the following steps can be further included:

[0103] Intelligent planning and strategy:

[0104] Local key re-cleaning: after completing a cleaning, if the visual feedback system detects that the hardness value of a certain area is still high, which usually means that the cleaning is not clean enough, the robot can automatically plan the path to perform a second key cleaning on the "stubborn point".

[0105] Job energy consumption and time estimation: before starting the task, the robot can perform a quick inspection to generate a "ship surface stain hardness distribution map", thereby more accurately estimating the total time and battery consumption required to complete the entire task.

[0106] Device health management:

[0107] Wear prediction: accumulate the total time of cleaning brush disc processing high hardness such as >8.0 adherends. When reaching the preset threshold, it can automatically remind "brush disc wear is serious, suggest checking or replacing", to realize predictive maintenance. Table 2 shows the content, input and output quantity, and key logic or parameters of each stage of the underwater cleaning method of an embodiment of the present application.

[0108] Table 2

[0109]

[0110]

[0111] The following explains some formulas in Table 2 in detail:

[0112] 1) Hardness mapping-mixed calculation formula

[0113] ∑(area i × hardness i) ; this formula is used to calculate the overall hardness value when there are multiple types of attachments in a region. Wherein, area i represents the area proportion of the i-th attachment in the region; hardness i represents the hardness branch corresponding to the i-th attachment. ∑ represents the summation symbol, which means adding the product of the area proportion of each type of attachment and the hardness score; through this weighted average calculation, an overall hardness score can be obtained, which can accurately reflect the overall cleaning difficulty of the mixed fouling.

[0114] 2) Dynamic parameter adjustment - pressure control formula

[0115] Pressure = 30% + 7% x hardness

[0116] This formula is an exemplary algorithm for dynamically adjusting the cleaning pressure according to the hardness value of the attachments. The purpose of the formula is to map the hardness value in the range [0, 10] to the pressure setting value in the range [30%, 100%]. 30%: is the basic pressure value, for example, the pressure used when cleaning fouling with low hardness. 7%: is a coefficient, for every 1 increase in hardness, the pressure increases by 7%. When the hardness is the lowest 0, the pressure is 30%; when the hardness is the highest 10, the pressure is 30% + 7% x 10 = 100%, that is, the maximum pressure, which is consistent with the description in the text that "for heavy barnacles with a hardness of 10.0, the pressure is automatically increased to 100% maximum". The specific percentage values used in the formula are pre-determined preferred values, and other suitable proportion values can be pre-determined according to actual scene requirements.

[0117] 3) Dynamic parameter adjustment - speed control formula

[0118] Speed = 100% - 5% x hardness

[0119] This formula is used to automatically adjust the speed of the robot according to the hardness value of the attachments; the purpose of the formula is to map the hardness value in the range [0, 10] to the speed setting value in the range [50%, 100%]; 100%: is the fastest speed of the robot, used for cleaning soft fouling areas with low hardness; 5%: is a coefficient, for every 1 increase in hardness, the speed decreases by 5%; when the hardness is 0, the speed is 100%; when the hardness is 10, the speed decreases to 100% - 5% * 10 = 50%, which is consistent with the description in the text that "when a high hardness area is encountered, the speed is automatically reduced to 50%".

[0120] 4) Dynamic parameter adjustment - power control formula

[0121] Formula: Torque ∝ Hardness value

[0122] The formula indicates that the motor torque of the cleaning brush disc is directly proportional to the hardness value of the attached objects. Here, torque refers to the moment of force output by the motor, which directly affects the cleaning power of the brush disc; ∝ is the proportional symbol; hardness value is the pollution hardness score obtained after system recognition and mapping. This means that when the robot detects higher hardness of attached objects, it will automatically increase the torque of the brush disc motor to exert greater cleaning force.

[0123] 5) Intelligent planning - pre-inspection scanning symbols

[0124] Symbol: Grid scanning (0.5m 2 / cell); Meaning explanation: This represents a working mode of the robot when conducting a quick hull inspection. It virtually divides the hull surface into multiple grids, with each grid cell having an area of 0.5 square meters. The robot will scan in these units to generate a "hull pollution hardness distribution map" for subsequent task estimation and path planning.

[0125] 6) Intelligent planning - task estimation formula

[0126] Formula: Total time = ∑ (Area time × Hardness coefficient)

[0127] Meaning explanation: This formula is used to estimate the total time required to complete the cleaning task based on the hardness distribution map generated by pre-inspection scanning before the cleaning task begins. Here, total time is the estimated total operation time; Area time is the basic time required to clean a standard area; Hardness coefficient is a multiplier set according to the hardness value of the area; the higher the hardness of the area, the larger the coefficient, meaning that cleaning requires more time; ∑ represents the sum of the estimated times of all areas to obtain the final total time.

[0128]

[0129] 7) Intelligent planning - effect verification steps

[0130] Steps 2.1-3.2 in the table represent the process of repeating steps 2.1 to 3.2 after cleaning to verify the cleaning effect. According to the table content, these steps include:

[0131] Visual analysis: using deep learning model to identify images of cleaned areas;

[0132] Type judgment: detecting whether there are residual attached objects.

[0133] Single query and mixed calculation: mapping the hardness of the residual objects and calculating the "residual hardness value".

[0134] This process is part of the system's closed-loop feedback, used to evaluate the cleaning effect and determine whether secondary cleaning is needed. ​

[0135] 8) Equipment management - wear accumulation formula

[0136] Formula: Wear = ∑(Time x Hardness 2 )

[0137] Meaning explanation: This formula is used to calculate and accumulate the wear of the cleaning brush disc to achieve predictive maintenance; Wear: represents the accumulated wear degree of the brush disc. Time: refers to the time length of the brush disc working on a certain hardness of the fouling; Hardness 2 : the square of the hardness value; here the square is used to represent the non-linear relationship between wear and hardness, i.e. cleaning high hardness of the attachment will cause greater wear of the brush disc; ∑: all the wear values generated by the brush disc working on different hardness of the fouling are accumulated. When the accumulated wear value calculated exceeds the preset threshold value, the system will automatically trigger a maintenance alarm to remind checking or replacing the brush disc.

[0138] (2) Fouling growth cycle: early stage = low, middle stage = medium, mature stage = high, also mapped as continuous values [0, 10].

[0139] (3) Coating fragility: according to the type of coating, such as hard coating = low, self-polishing coating = high, mapped as continuous values [0, 10].

[0140] Output variables:

[0141] (1) Brush disc hardness: [soft nylon, medium hardness polypropylene, hard steel wire]; indicates that the robot can replace the brush disc materials of different hardness. The system will make decisions through fuzzy logic algorithm according to the information such as the hardness of the identified fouling, growth cycle and hull coating fragility, and output the brush disc hardness of hard use, i.e. different hardness of the brush disc material;

[0142] (2) Brush disc rotation speed: [low, medium, high].

[0143] (3) Contact pressure: [low, medium, high].

[0144] (4) Robot moving speed: [fast, medium, slow].

[0145] In a specific implementation, the low, medium and high or fast, medium and slow of the above brush disc rotation speed, contact pressure and robot moving speed are pre-set to corresponding speed given value, contact pressure given value or moving speed given value.

[0146] Fuzzy rule base: a series of IF-THEN rules defined by expert experience.

[0147] (1) Rule Example 1: IF (hardness of attached object is high) AND (growth cycle is mature) AND (fragility of paint is low) THEN (hardness of brush disc is hard) AND (rotation speed is high) AND (pressure is high) AND (moving speed is slow).

[0148] (2) Rule Example 2: IF (hardness of attached object is low) AND (fragility of paint is high) THEN (hardness of brush disc is soft) AND (rotation speed is low) AND (pressure is low) AND (moving speed is fast).

[0149] • Defuzzification: The fuzzy inference engine calculates a weighted average value according to the values corresponding to each activated rule, and finally outputs an accurate and executable control value (e.g., rotation speed = 800 RPM, pressure = 35 N); the weighting weights are pre-set.

[0150] The following is a brief description of the exemplary workflow of the underwater cleaning method using the embodiments of the present application in actual application scenarios:

[0151] (1) The robot starts and approaches the ship body, and the image acquisition module begins to work.

[0152] (2) The image data stream is transmitted to the data processing and recognition module in real time.

[0153] (3) The CNN algorithm completes the analysis of the current picture in a short time, such as 0.1 seconds, and outputs: "Region A: barnacles, mature period; paint: hard epoxy paint".

[0154] (4) The result is sent to the parameter matching and decision-making module to determine the cleaning execution structure operation parameters corresponding to the cleaning area.

[0155] (5) The fuzzy logic algorithm reasons according to the rule base and outputs the instruction: "use No. 3 hard brush disc, set the rotation speed to 1200 RPM, the pressure to 50 N, and the robot moving speed to 0.2 meters / second".

[0156] (6) The cleaning execution control module receives the instruction and accurately adjusts the hardware parameters.

[0157] (7) After cleaning is completed, the camera scans the area again for effect evaluation; if there are still residues, the system will fine-tune the parameters, for example, slightly increase the pressure for secondary cleaning, to achieve closed-loop feedback.

[0158] In an implementation of the embodiment of the present application, the cleaning execution device comprises: a brush disc; the brush disc comprises: a pressure sensor for monitoring the current contact pressure of the brush disc and the ship surface; the cleaning control device comprises an active pressure control structure; the active pressure control structure comprises: a pressure actuator for controlling the output pressure of the brush disc; and a pressure PID controller for controlling the output of the pressure actuator according to the determined contact pressure of the cleaning execution device and the ship surface and the current contact pressure of the brush disc.

[0159] The embodiment of the present application constructs a cleaning execution device with a double closed-loop feedback structure, which gives the whole system a "tactile" feeling, wherein:

[0160] Pressure closed loop: the brush head is integrated with a pressure sensor, which monitors the actual contact pressure between the brush disc and the ship surface in real time, and feeds back the pressure feedback signal to the pressure PID controller. The controller adjusts the action of the pressure actuator in real time by comparing the difference between the target pressure and the actual pressure;

[0161] Rotational speed / torque closed loop: the motor encoder or torque sensor is integrated on the rotating motor, which monitors the actual rotational speed and load torque of the motor in real time, and feeds back the rotational speed / torque feedback signal to the rotational speed PID controller, to ensure the constancy and accuracy of the rotational speed.

[0162] The introduction of the above closed loop structure ensures that the upper decision instruction can be accurately executed; the whole system can overcome the interference of the complex underwater environment such as water flow impact and ship surface change, and always maintain the best cleaning parameters. In addition, by sensing the load torque, the stubbornness of the pollution can also be indirectly judged, which provides further optimization data for the upper decision, greatly improving the intelligent level and operation reliability of the underwater cleaning system.

[0163] In a specific implementation of the embodiment of the present application, the brush disc has a standardized replaceable brush disc interface structure. This structure not only facilitates quick physical disassembly and assembly, but also deeply binds with the parameter matching and decision module at the software level. The database of the system pre-stores the physical characteristic parameters of different types of brush discs, such as soft nylon brush, medium hardness abrasive brush and high hardness steel wire brush. It realizes the transition from a single tool structure to an adaptive modular structure; avoids the defect that the existing cleaning robot usually installs a universal brush disc which cannot be replaced, and has poor adaptability to different pollution types with large physical property differences.

[0164] In one specific implementation, when the current working condition is identified to require a change of cleaning tool, for example, a high-value silicone coating is identified, and a dedicated soft brush disc must be used, a clear change instruction is issued to the operator. The modular design of the software and hardware combination enables the robot to handle various extremely complex cleaning tasks by changing "surgical knives", and its working range and adaptability far exceed the prior art.

[0165] Embodiment three:

[0166] The application further provides a cleaning control device, which comprises a processor and a memory, and a computer program stored in the memory and executable on the processor, the processor comprising one or more processing cores, the memory being connected to the processor through a bus, and the memory being used for storing program instructions, and the processor executes the computer program to implement the steps in the above method embodiments of Embodiment One of the application.

[0167] Further, as an executable solution, the cleaning control device can be a computer unit, which can be a desktop computer, a notebook computer, a palm computer, a cloud server and the like. The computer unit can comprise, but is not limited to, a processor and a memory. Those skilled in the art can understand that the above-mentioned constituent structure of the computer unit is only an example of the computer unit, and does not constitute a limitation on the computer unit, and can comprise more or fewer components than the above, or combine certain components, or different components. For example, the computer unit can further comprise an input / output device, a network access device, a bus and the like, and the embodiments of the application do not limit this.

[0168] Further, as an executable solution, the processor can be a central processing unit (CPU), and can also be other general-purpose processors, digital signal processors (DSP), application specific integrated circuits (ASIC), field-programmable gate arrays (FPGA) or other programmable logic devices, discrete gates or transistor logic devices, discrete hardware components, etc. The general-purpose processor can be a microprocessor or the processor can also be any conventional processor and the like, and the processor is the control center of the computer unit, and connects all parts of the computer unit through various interfaces and lines.

[0169] The memory can be configured to store the computer programs and / or modules, and the processor can realize various functions of the computer unit by running or executing the computer programs and / or modules stored in the memory and calling data stored in the memory. The memory can mainly include a program storage area and a data storage area, wherein the program storage area can store an operating system and at least one application required by a function; and the data storage area can store data created according to the use of the mobile phone and the like. In addition, the memory can include a high-speed random access memory, and can also include a nonvolatile memory, for example, a hard disk, a memory, a plug-in hard disk, a smart media card (SMC), a secure digital (SD) card, a flash card, at least one disk storage device, a flash memory device, or other volatile solid-state memory devices.

[0170] Embodiment Four

[0171] The application further provides an underwater cleaning robot, comprising: an image acquisition device configured to acquire an underwater video frame image of a ship surface; a cleaning execution device configured to clean the ship surface; and the cleaning control device described above.

[0172] Although the present application has been specifically shown and described with reference to preferred embodiments, it is understood that various changes in form and details can be made therein without departing from the spirit and scope of the application as defined by the appended claims.

Claims

1. An underwater cleaning method applied to an underwater cleaning robot, the underwater cleaning robot comprising an image acquisition device, a cleaning execution device, and a cleaning control device, characterized in that, include: S1, the cleaning control device uses a pre-trained deep learning network to identify the type of attachments and their current growth cycle on each pre-divided underwater hull area in the underwater video image based on the underwater video frame image of the hull surface, and identifies the type of paint on the hull surface, wherein the underwater video frame image is acquired by an image acquisition device. S2, the cleaning control device, for each area, obtains the hardness of the attachment corresponding to the type of attachment on the area and the coating fragility level corresponding to the type of coating on the hull surface by querying a preset mapping table. S3, for each area, the cleaning control device activates all fuzzy rules in the pre-set fuzzy rule library that match the obtained hardness of the attachment, the coating fragility level, and the current growth cycle of the attachment, and determines the operation parameters of the corresponding cleaning execution device according to the activated fuzzy rules. The operation parameters include: the material of the cleaning execution device, the rotation speed of the cleaning execution device, the contact pressure between the cleaning execution device and the hull surface, and the moving speed of the underwater cleaning robot. S4, the cleaning control device selects a cleaning execution device for each area based on the operating parameters of the cleaning execution device determined in step S3, and adjusts the rotation speed of the selected cleaning execution device, the contact pressure between the cleaning execution device and the hull surface, and the moving speed of the underwater cleaning robot to clean the area.

2. The underwater cleaning method according to claim 1, characterized in that, The input conditions of the fuzzy rule include: the hardness of the attachment, the paint fragility level of the hull surface, and / or the current growth cycle of the attachment; the decision output of the fuzzy rule includes: the material of the cleaning actuator, the rotation speed level of the cleaning actuator, and / or the contact pressure level between the cleaning actuator and the hull surface.

3. The underwater cleaning method according to claim 2, characterized in that, The step S3, which determines the operating parameters of the corresponding cleaning execution device based on the activated fuzzy rules, includes: The operating parameters of the corresponding cleaning actuator are determined by using the pre-set correspondence between the rotation speed level of the cleaning actuator and the rotation speed of the cleaning actuator, and the correspondence between the contact pressure level of the cleaning actuator and the contact pressure of the cleaning actuator and the hull surface. Specifically, when there is only one activated fuzzy rule, the rotational speed and pressure corresponding to the rotational speed level of the cleaning actuator and / or the contact pressure level between the cleaning actuator and the hull surface are used as the operating parameters of the cleaning actuator. When there are multiple activated fuzzy rules, the rotational speed or pressure corresponding to the output of each of the multiple fuzzy rules is weighted and averaged according to a predetermined weight, and the weighted average value is used as the operating parameter of the cleaning actuator.

4. The underwater cleaning method according to claim 1, characterized in that, Also includes: The cleaning control device uses a hardness score within a predetermined range to represent the hardness of the type of attachment according to a pre-set mapping rule. When an area contains multiple types of attachments, the hardness scores corresponding to the hardness of each attachment type are weighted and averaged according to the area ratio of each attachment type, and the hardness score obtained after weighted averaging is used as the hardness score corresponding to the area.

5. The underwater cleaning method according to claim 1, characterized in that, After acquiring underwater video frame images of the ship's hull surface and before recognizing the underwater video images using a pre-trained deep learning network, the process also includes: The acquired underwater video frame images are enhanced using a cross-domain mutual learning mechanism, which includes a teacher network, a student network, and a turbidity enhancement module. The teacher network is trained in the synthesis domain and transfers knowledge to the student network through parameter sharing and time-weighted averaging. The student network is trained in the real domain and feeds back knowledge to the teacher network through adversarial loss and prior loss. The turbidity enhancement module is used to dynamically generate pseudo-realistic image samples for nonlinear turbidity enhancement.

6. The underwater cleaning method according to claim 1, characterized in that, It also includes one or more of the following: The hardness of the deposits is classified into different hardness grades according to the hardness score from low to high; the hardness grades include: soft dirt, light hard dirt, medium hard dirt and heavy hard dirt. The types of organisms attached to the hull include: biomass, seaweed, tube worms, mussels, algae, barnacles and / or bryozoans; The growth cycle includes: early stage, middle stage and maturity stage; The types of coatings include: self-polishing coatings, hard epoxy coatings, and / or non-toxic silicone coatings.

7. The underwater cleaning method according to claim 1, characterized in that, The cleaning device includes a brush; the brush includes a pressure sensor for monitoring the current contact pressure between the brush and the hull surface. The cleaning control device includes an active pressure control structure; the active pressure control structure includes: a pressure actuator for controlling the pressure output by the brush; and a pressure PID controller for controlling the output of the pressure actuator based on the contact pressure between the cleaning actuator and the hull surface and the current contact pressure of the brush determined in step S3.

8. The underwater cleaning method according to claim 1, characterized in that, The deep learning network includes: a backbone network and two parallel detection or classification heads connected to the backbone network; wherein the two parallel detection or classification heads include: a dirt recognition head, used to output the position of the detection box, the type of attachment and the current growth cycle; and a paint recognition head, used to classify the entire image or a selected area of ​​the image and output the paint type.

9. A cleaning control device, characterized in that, It includes a memory and a processor, the memory storing at least one program, the at least one program being executed by the processor to implement the underwater cleaning method as described in any one of claims 1 to 8.

10. An underwater cleaning robot, characterized in that, include: Image acquisition device, used to acquire underwater video frame images of the ship's hull surface; Cleaning actuators are used to clean the surface of the ship's hull. The cleaning control device as described in claim 9.

Images