A smart robot and method for cleaning biofouling on deep-sea mining hoses

Abstract

This invention relates to the field of deep-sea mining hose cleaning technology, specifically to an intelligent robot and method for cleaning adsorbed organisms on deep-sea mining hoses. The robot includes two shells, with a power storage compartment inside each shell for storing power. It also includes a visual drug application device for monitoring bioadsorption on the mining hose and performing preliminary drug treatment, a vortex-flow erosion device for further treatment of adsorbed organisms, a positioning and detection device for monitoring the treated mining hose, treating stubborn residual adsorbents, and monitoring the overall condition of the mining hose, an AUV thruster, and a communication control device. The communication control device includes an ACN underwater acoustic communication module and an intelligent processing module. The invention also includes a cleaning method using this robot. This application features multi-level collaborative cleaning, significantly improving the efficiency of bioadhesion removal, intelligent closed-loop control, and fully autonomous operation, reducing labor costs.

Description

Technical Field

[0001] This invention relates to the field of deep-sea mining hose cleaning technology, specifically to an intelligent robot and method for cleaning biological deposits on deep-sea mining hoses. Background Technology

[0002] With the increasing depletion of terrestrial mineral resources, deep-sea minerals rich in metal nodules have become an important source of resources to meet the needs of current technological development. Subsea mining vehicles are key equipment for the development of deep-sea mineral resources, and their successful development has made the commercial mining of deep-sea resources possible. As the core equipment for mineral hoisting, the hoisting hose faces multiple severe challenges in long-term deep-sea operations due to the complex seabed environment. Its safe and stable operation directly determines the safety and economy of deep-sea mining operations.

[0003] In deep-sea operations, mining hoses primarily face "biofouling corrosion," which manifests itself in several ways: Deep-sea mining hoses span hundreds to thousands of meters, providing a wide attachment surface for marine organisms; their flexible design makes them prone to deformation under ocean currents and waves, leading to microcracks or coating peeling, creating initial points for biofouling; prolonged immersion allows bacteria, diatoms, and other microorganisms to rapidly colonize the hose surface, forming biofilms. The extracellular polymers secreted by these biofilms further attract larger organisms such as barnacle larvae and mussels, gradually solidifying into a hard fouling layer. The metabolic activities of deep-sea microorganisms produce acidic substances, accelerating the corrosion process of the mining hose; this corrosion further damages the hose's surface structure, creating conditions for even more biofouling, forming a vicious cycle of "attachment-corrosion."

[0004] Furthermore, during the mineral hoisting process, the high-pressure, high-speed slurry flow carrying solid particles continuously erodes and rubs against the inner wall of the hoisting hose, causing wear. The synergistic effect of biofouling corrosion and hose wear significantly reduces the structural strength of the hoisting hose, greatly increasing the risk of hose rupture. Once a rupture occurs, it will trigger multiple problems. For example, the leaked slurry may contain heavy metals or chemical treatment agents, causing serious pollution to the deep-sea ecosystem; a hoisting hose rupture will force the entire mining system to shut down, delaying production plans; the repair and replacement costs of the hoisting hose are extremely high, and high-pressure slurry leaks may damage other supporting equipment, causing additional economic losses.

[0005] Therefore, the treatment of bio-adsorption on deep-sea ore lifting hoses is particularly important. Existing technologies can only remove the bio-adsorption through underwater robots under human control. It is not possible to systematically integrate robots onto the ore lifting hoses to achieve on-the-loop operation. Cleaning the entire long-span ore lifting hose is a huge drain on manpower and resources.

[0006] The prior art CN222385522U discloses a carbon fiber tube marine organism removal support that uses scrapers on its fixed plate to scrape and clean the organisms adsorbed in the tube. However, there is a certain gap between adjacent scrapers, which cannot effectively remove the adsorbed organisms within the gap. Even if the technology connects multiple frames vertically to increase the contact area between the scraper and the outer wall of the tube to eliminate the impact of the gap, it still cannot achieve full coverage of the outer wall of the tube, resulting in low cleaning efficiency. In particular, for some small marine organisms with strong adsorption capacity, it is basically impossible to clean them by scraping with the scraper, resulting in very low cleaning efficiency.

[0007] The existing technology CN222817542U discloses a submarine pipeline cleaning device that uses a scraper and a circumferential rotating flushing device to clean large volumes of deposits and dirt on the inner wall of the pipeline. However, it also fails to treat adsorbed organisms and monitor the cleaning effect.

[0008] The technologies disclosed above all use physical methods to clean and protect subsea pipelines. However, they cannot effectively remove marine organisms with strong adsorption capacity, cannot differentiate between areas with different adsorption levels on the subsea pipeline, consume a lot of energy, and cannot obtain real-time information about the pipeline after cleaning.

[0009] Another existing technology, CN211203294U, discloses a locking and sealing structure for a pressurized plugging pipe clamp to prevent detachment during installation. This structure utilizes the interconnection of a locking mechanism and a sealing mechanism, as well as their connection to the pipe clamp housing, to prevent detachment of these mechanisms during installation. However, this existing technology can only perform single-time repairs at a fixed location and cannot achieve continuous operation. The entire construction process requires personnel and equipment assistance, consuming significant manpower and resources. Furthermore, due to the complexity of the deep-sea operating environment and the single-time, fixed-location nature of this technology, it is not suitable for deep-sea mining hose operations.

[0010] In summary, the existing technology lacks an integrated technology that combines "adsorption biological removal + real-time monitoring + damage prediction" specifically for the ore lifting hose of deep-sea mining vehicles. This technology cannot efficiently remove organisms with different adsorption intensities from the hose surface, nor can it monitor the hose status in real time and provide early warnings of rupture risks, resulting in a lack of effective guarantees for the safe operation of the ore lifting hose. Summary of the Invention

[0011] To address the problems existing in the prior art, this invention provides an intelligent cleaning robot and cleaning method for deep-sea mining hoses, which solves the problems of strong adsorption of biological residues and low efficiency in existing physical removal methods, the inability of existing technologies to operate continuously, and the inability of existing technologies to differentiate treatment based on the degree of biological adsorption.

[0012] To solve the above problems, the technical solution of the present invention is as follows: a deep-sea ore lifting hose adsorbing biological intelligent cleaning robot, comprising two rotatably connected semi-cylindrical shells that wrap the ore lifting hose, and a power storage compartment for storing power is provided inside the shell.

[0013] A visual drug application device is installed at the upper end of the shell to monitor the biosorption on the mining hose and to perform preliminary drug treatment.

[0014] Below the visual application device, inside the two shells, is a vibratory flow ablation device for further treatment of the adsorbed biological material on the mining hose;

[0015] Below the vibratory flow erosion device is a positioning and detection device that monitors the treated lifting hose, treats stubborn residual adsorbents, and monitors the condition of the lifting hose.

[0016] Several AUV thrusters are spaced apart on the bottom sidewall of the housing;

[0017] It also includes a communication control device for controlling the visual drug application device, the oscillating ablation device, the positioning detection device, the AUV propulsion device, and the signal transmission.

[0018] Furthermore, the visual drug application device includes a drug storage chamber located at the upper end of the housing for holding the drug, a plurality of first high-definition cameras are spaced apart at the top of the drug storage chamber, a plurality of turbine drug nozzles are spaced apart on the inner side wall of the drug storage chamber, and a nozzle controller for controlling the dosage and range of the turbine drug nozzles, wherein the nozzle controller is signal-connected to a communication control device.

[0019] Furthermore, the vibratory flow ablation device includes a vibrator located below the turbine agent nozzle for vibrating and relaxing the adsorbed organisms in the pipeline to reduce their adsorption capacity, and a vibration controller for controlling the vibrator. A contact-type limit sensor is provided on the surface of the vibrator to limit the vibrator and prevent the vibrator from compressing the lifting hose and causing damage to it. Several jet nozzles are arranged at intervals below the vibrator for further vibration relaxation and jet shedding treatment of the adsorbed organisms, and a jet controller for controlling the jet nozzles. The vibration controller and the jet controller are signal connected to the communication control device.

[0020] Furthermore, the positioning and detection device includes a chute located below the jet nozzle, a sliding base slidably connected to the chute, a retractable robotic arm and a second high-definition camera mounted on the sliding base, a retractable sonar detector located below the chute, a thin-film piezoresistive pressure sensor to prevent the sonar detector from damaging the pipeline on the surface of the sonar detector, and a sonar controller inside the sonar detector to record the position of the sonar detector's detection data and transmit it to the communication control device.

[0021] Furthermore, the two housings are hinged together by a central pivot, which is connected to a hydraulic motor via a connecting rod assembly, and the hydraulic motor is signal-connected to a communication control device.

[0022] Furthermore, the communication control device includes an ACN underwater acoustic communication module for transmitting detection data from the first high-definition camera, the second high-definition camera, and the sonar controller. It also includes an intelligent processing module for processing the data transmitted by the ACN underwater acoustic communication module and predicting whether the mining hose will be damaged. The intelligent processing module is also used for regulating the nozzle controller, vibration controller, jet controller, robotic arm, AUV thruster, and hydraulic motor.

[0023] A method for cleaning biological deposits adsorbed on deep-sea mining hoses, employing the aforementioned intelligent cleaning robot for cleaning biological deposits on deep-sea mining hoses, includes the following steps:

[0024] S1: Operation preparation, the robot is stored in the charging compartment of the mining vehicle and charged, and chemical agents are replenished to the storage bin;

[0025] S2: Start operation. The AUV thruster is activated and adjusted according to the image data provided by the first high-definition camera, gradually moving towards the mining hose.

[0026] S3: The housing closes precisely. The intelligent processing module controls the hydraulic motor to open the housing and gradually approach the lifting hose. When the contact limit sensor is triggered, the hydraulic motor stops running, fixing the housing to the lifting hose. At this time, the vibrator is in close contact with the lifting hose.

[0027] S4: Chemical pretreatment, the first high-definition camera captures the bio-adsorption image on the surface of the mining hose and transmits it to the intelligent processing module through the ACN underwater acoustic communication module. The intelligent processing module processes the image to identify the adhering substances, and then controls the spraying mode of the turbine chemical nozzle through the nozzle controller to spray the chemical. The spraying mode includes wide-angle atomization mode and narrow beam spray mode.

[0028] S5: Vibration relaxation, intelligent processing module adjusts the vibration intensity of the vibrator at each position through the vibration controller based on the different adsorption conditions in the image data transmitted from the first high-definition camera, thereby reducing the adsorption effect of the adsorbed organisms on the pipeline.

[0029] S6: Jet shedding. The intelligent processing module adjusts the flow rate of the jet nozzles at each location through the jet controller based on the different adsorption conditions in the image data transmitted from the first high-definition camera, so that the attached organisms are shedding by jet.

[0030] S7: Residual monitoring. The sliding base moves back and forth along the slide groove. The cleaned image captured by the second high-definition camera is transmitted to the intelligent processing module through the ACN underwater acoustic communication module. The intelligent processing module determines the location of the remaining residual adsorbent on the mining hose and whether further processing is needed based on the transmitted image data.

[0031] S8: Precise cleaning and re-inspection. The intelligent processing module controls the robotic arm to extend and approach to remove the remaining adsorbed substances. After cleaning, the robotic arm retracts and is re-inspected by the second high-definition camera. If it does not meet the standards, the cleaning is repeated. After cleaning, the intelligent processing module controls the robotic arm to reset, the sliding base resumes movement, and the inspection continues through the second high-definition camera.

[0032] S9: Wall thickness detection. After the robotic arm finishes cleaning, the intelligent processing module controls the sonar detector to extend. When the thin-film piezoresistive pressure sensor detects the contact force, the extension stops. Wall thickness detection is started and linked with the odometer of the AUV thruster to record the detection position. The information is then transmitted to the intelligent processing module through the ACN underwater acoustic communication module.

[0033] S10: Risk assessment and intelligent processing module compares the wall thickness of the mining hose with the original wall thickness data to determine the risk of damage. If the data is abnormal, the AUV thruster is controlled to stop and rotate to detect the full circumference wall thickness of the hose, and the data is transmitted to the water work vessel through the ACN underwater acoustic communication module. If the data is normal, the sonar detector is retrieved and the AUV thruster continues to work along the hose.

[0034] S11: Operation completion. When the power supply is insufficient or the operation is completed, all devices stop operating. The intelligent processing module controls the hydraulic motor to rotate and open the shell. The robot disengages from the hoisting hose. The hydraulic motor reverses, the shell closes, and the robot returns to the mining vehicle charging compartment.

[0035] Furthermore, in S4, the intelligent processing module incorporates an optimized YOLOv5 model, and the YOLOv5 model processes image data in the following steps:

[0036] S41. Input layer image preprocessing: converts the original image into standardized data that the model can process, eliminating interference from the deep-sea environment;

[0037] S42. Feature extraction is performed on the backbone network, including:

[0038] (a) The initial convolutional layer performs low-level feature extraction.

[0039] ,

[0040] In the formula, Output feature map; : Convolutional kernel weights for the c-th output channel and the k-th input channel (RGB); : Bias term; Activation function Introducing nonlinearity;

[0041] (ii) Introducing the CSP module and SE attention mechanism for feature extraction and biometric enhancement;

[0042] (iii) Multi-scale feature fusion is performed in the SPPF layer;

[0043] S43. The neck network fuses multi-scale features through the PANet structure and outputs a scale feature map.

[0044] S44. The head network predicts the bounding box, confidence level, and category of biosorption based on the scale feature map obtained in S43.

[0045] S45. Loss Calculation and Model Optimization;

[0046] S46. Post-processing, output the final result.

[0047] Furthermore, in S42,

[0048] The CSP module includes:

[0049] (a) Input features Split into (50% channel) and (50% channel);

[0050] (b) Processed by two residual blocks (convolution + skip connections) ;

[0051] (c) Fusion get ;

[0052] The SE attention module includes:

[0053] (1) Global average pooling: , This represents the global feature of channel c;

[0054] (2) Channel weight calculation: , , For the weights of the fully connected layer, Output weights of 0 to 1;

[0055] (3) Feature weighting: , These are channel attention weights, used to dynamically adjust the importance of different channels: for biometric channels, Take a larger value (i.e.) ↑), to enhance the expression of biosorption-related features; for background channels, Take a smaller value (i.e.) (↓) to suppress background noise interference.

[0056] Furthermore, the intelligent processing module in S10 incorporates a wall thickness analysis model, including:

[0057] (A) Data acquisition;

[0058] (B) Calculate the local deviation rate and the overall deviation rate;

[0059] (C) Calculate corrosion rate and predict future wall thickness;

[0060] (D) Risk level determination.

[0061] Furthermore, in the wall thickness analysis model, in step (B):

[0062] Local deviation rate Overall deviation rate ,

[0063] Where M: the number of detection points along the circumference of the hose. : No. One detection point ( =1, 2, ..., M), Original standard wall thickness of the hose : No. Measured wall thickness at each point : No. Spatial weights of each point.

[0064] Compared with the prior art, the present invention has the following beneficial effects:

[0065] I. Multi-level collaborative cleaning significantly improves the efficiency of bioattachment removal. This application employs a three-level treatment process: "pretreatment with a visual application device (targeted spraying of chemicals) → enhancement with a vibratory flow ablation device (vibration relaxation + jet shedding) → precise removal with a positioning detection device (robotic arm targeted treatment of residues)," forming a tiered removal mechanism for organisms with varying adsorption intensities. Specifically, the turbine chemical nozzle can switch between wide-angle atomization (mild) and narrow-beam spray (severe) modes based on the adsorption level. The parameters of the vibrator and jet nozzle can be dynamically adjusted. Combined with the robotic arm's targeted treatment of stubborn residues, this solves the problems of incomplete physical cleaning and blind spots in existing technologies, significantly reducing the bioattachment rate and minimizing the risk of biocorrosion at its source.

[0066] II. Intelligent closed-loop control enables fully autonomous operation and reduces labor costs. The communication control device transmits images and detection data in real time through the ACN underwater acoustic communication module. The intelligent processing module completes the adsorption level classification, equipment parameter adjustment, cleaning effect re-inspection, and risk assessment. No manual intervention is required throughout the process, realizing an automated closed loop of "monitoring-decision-execution-feedback". This reduces the reliance on manpower in deep-sea operations, improves the operational efficiency of long-span ore hoisting hoses, and adapts to the needs of large-scale deep-sea mining.

[0067] III. Precise protection design to avoid secondary damage to the hoisting hose from equipment. The contact-type limit sensor on the vibrator surface can trigger a limit when the equipment approaches the hose, preventing excessive pressure from the vibrator from causing wear on the hose coating; the thin-film piezoresistive pressure sensor on the sonar detector surface can accurately sense the contact force, ensuring that it stops extending during detection and avoiding mechanical damage to the pipe wall. This design solves the problem of existing cleaning / inspection equipment easily damaging the hose and ensures the structural integrity of the hoisting hose.

[0068] IV. Integrated design enables coordinated "cleaning-monitoring-early warning" to ensure the safe operation of the hose. The robot integrates functions such as visual monitoring, biological cleaning, and wall thickness detection. After cleaning biological deposits, it can monitor the hose wall thickness in real time through sonar detectors and compare it with the original data to assess the risk of damage. In case of abnormalities, the data is transmitted to the surface vessel in a timely manner, forming a full-chain guarantee from biological removal to structural safety early warning. This avoids risks such as slurry leakage and system shutdown in advance, significantly improving the safety and economy of deep-sea mining operations.

[0069] V. Adaptive Operation Capability, Adaptable to Long-Span Hoses and Complex Deep-Sea Environments: The two semi-cylindrical shells achieve precise opening and closing through a central pivot and hydraulic motor. Combined with the multi-angle motion control of the AUV thrusters, it can continuously move and operate along the mining hoisting hose, covering long-span hoses of hundreds to thousands of meters. The entire equipment adopts an underwater sealed design, adapting to complex environments such as deep-sea high pressure and ocean current interference, improving the robot's ability to adapt to the environment.

[0070] VI. Based on an improved YOLOv5 bioadsorption recognition algorithm, the weights of biofeature channels are dynamically enhanced through the SE attention module (suppressing noise from seawater, hose backgrounds, etc.), and a depth correction layer is used to offset image distortion caused by different water depths and pressures. This solves the problem of missed detection of small targets in low-light and blurred water flow scenarios of the original model. In actual testing, the overall recognition accuracy is stable at over 95%, providing a precise target localization foundation for subsequent graded cleaning. It achieves refined classification of bioadsorption, adapts to the algorithm requirements of tiered cleaning, and can be expanded to four categories: "biofilm, small attachments, large attachments, and mixed attachments," and uses a category reliability factor (…). ) Optimize classification loss and accurately identify attached substances.

[0071] VII. The wall thickness analysis model can accurately identify local corrosion risks and avoid missing key areas. The model uses a spatial weighting factor ( The model assigns higher weights to high-corrosion areas such as hose joints and bends, solving the fatal flaw of the original model that "the average wall thickness meets the standard but local damage has occurred". In practical applications, the sensitivity of identifying local corrosion (such as a 40% loss of wall thickness at the joint but an overall average loss of 20%) is improved by 50%, effectively avoiding the risk of slurry leakage. Attached Figure Description

[0072] Figure 1 This is a schematic diagram of the appearance of the present invention;

[0073] Figure 2 This is a schematic diagram of the present invention fixed on a mining hoisting hose;

[0074] Figure 3 This is a schematic diagram of the internal structure of the present invention.

[0075] In the diagram: 1. Mining hose; 2. Shell; 3. Power storage compartment; 4. AUV propulsion unit; 5. Visual application device; 51. Drug storage compartment; 52. First high-definition camera; 53. Turbine agent nozzle; 6. Vibration flow erosion device; 61. Vibrator; 62. Contact limit sensor; 63. Jet nozzle; 64. Groove; 7. Positioning detection device; 71. Slide groove; 72. Sliding base; 73. Robotic arm; 74. Second high-definition camera; 75. Sonar detector; 76. Thin-film piezoresistive pressure sensor; 8. Central pivot shaft; 9. Communication control device; 91. ACN underwater acoustic communication module; 92. Intelligent processing module. Detailed Implementation

[0076] A kind of deep-sea ore lifting hose 1 adsorbs a biological intelligent cleaning robot, such as Figure 1-3 As shown, the structure is as follows:

[0077] I. Overall Structural Composition

[0078] like Figure 1 As shown, the intelligent cleaning robot of this application has an overall openable "semi-cylindrical assembly" structure, which is used to fit onto the outer surface of the deep-sea mining hose 1 and move along its axis for operation. Its core structure includes: two semi-cylindrical shells 2, a visual drug application device 5, a vibratory flow erosion device 6, a positioning detection device 7, an AUV thruster 4, and a communication control device 9. The components work together to achieve the integrated function of "biological monitoring - graded cleaning - status detection - autonomous movement".

[0079] like Figure 2As shown, the semi-cylindrical shell 2 can be made of titanium alloy. The length of a single shell 2 is 1.2m, and the inner diameter is 50mm larger than the outer diameter of the mining hose 1 (to fit a standard mining hose 1 with a diameter of 500mm). The two shells 2 are hinged by a central pivot 8 (the pivot can be made of 316L stainless steel with chrome plating for rust prevention). It can rotate around the pivot to achieve opening and closing from 0° to 180° (completely wrapping the hose when closed, and easy to detach when open). The power storage compartment 3 is embedded inside the side wall of the shell 2 and can use a deep-sea pressure-resistant lithium battery pack (capacity 200Ah, working voltage 24V) to support the robot to work continuously for 8 hours. When the battery is low (20% remaining), the return program is automatically triggered.

[0080] II. Specific Structure and Parameters of Each Functional Device

[0081] 1. Visual drug delivery device 5, such as Figure 3 As shown.

[0082] Located at the upper end of shell 2, it is used for biosorption monitoring and reagent pretreatment, specifically including:

[0083] Storage compartment 51: A 50L corrosion-resistant polyethylene container for storing isocyanurate-based non-oxidizing biocides (concentration 5%). It is connected to the turbine agent nozzle 53 via a corrosion-resistant silicone tube. The top can be equipped with a quick-fill port for docking with the charging compartment of the mining vehicle.

[0084] First high-definition camera 52: Five cameras are evenly distributed around the top circumference of each shell 2. The model can be a deep-sea pressure-resistant CCD camera (resolution 1920×1080, field of view 120°, waterproof rating IP68). The lens surface is covered with an anti-bioadhesion coating to capture images of the distribution of organisms on the surface of the hose in real time.

[0085] Turbine spray nozzles 53: Six nozzles are distributed circumferentially on the inner wall of each storage chamber 51, featuring a telescopic design (telescopic stroke 0-30mm). The spraying mode is adjusted via a nozzle controller (STM32 microcontroller).

[0086] Wide-angle atomization mode: The nozzle opening diameter is increased to 10mm, the spray angle is 120°, and the flow rate is 1L / min (suitable for areas with light adsorption).

[0087] Narrow beam spray mode: The nozzle opening is reduced to 3mm, the spray angle is 30°, and the flow rate is 3L / min (suitable for areas with severe adsorption).

[0088] 2. Vibrating flow erosion device 6, such as Figure 3 As shown.

[0089] Located below the visual drug delivery device 5, it is used for biological vibration relaxation and jet detachment, specifically including:

[0090] Vibrator 61: Four vibrators are distributed circumferentially on the inner wall of each housing 2. Miniature ultrasonic vibrators 61 (power 50W, frequency 20kHz) can be used. The inner wall is in the shape of a circular groove 64 (radius 250mm, matching the curvature of the hose). A contact limit sensor 62 (OMRON, D5V model can be used) is installed in the groove 64 by a spring (spring coefficient 5N / mm). The contact limit sensor 62 protrudes 5mm from the surface of the vibrator 61. When the micro switch is triggered, the vibrator 61 stops approaching (ensuring the contact pressure with the hose is ≤5N to avoid damage).

[0091] Vibration controller: Corresponds one-to-one with vibrator 61, and adjusts the vibration intensity (10-50W adjustable) through PWM signal. It is dynamically controlled by communication control device 9 according to the bio-adsorption thickness.

[0092] Jet nozzle 63: An annular groove 64 can be provided in the inner wall of the housing 2. The jet nozzle 63 is located in the groove 64 and the top end does not exceed the groove 64 to prevent damage to the mining hose 1. The jet nozzles 63 are distributed in two rows in each groove 64 (12 in total). High pressure rotary nozzles (maximum water pressure 10MPa) can be used. The flow rate can be adjusted by the jet controller (PLC controller) (5-15m / s adjustable).

[0093] 3. Positioning detection device 7, such as Figure 3 As shown.

[0094] Located below the oscillating erosion device 6, it is used for residual cleaning and hose condition monitoring, specifically including:

[0095] Slide 71 and sliding base 72: Slide 71 is a T-shaped guide rail set along the inner wall of housing 2. The sliding base 72 is driven by a servo motor (speed 0.1m / s) to reciprocate along slide 71 with a positioning accuracy of ±1mm.

[0096] Robotic arm 73 and second high-definition camera 74: Robotic arm 73 is a 3-degree-of-freedom telescopic arm (maximum extension length 300mm, positioning accuracy ±0.5mm), and a miniature steel wire brush (50mm in diameter) can be installed at the end; the second high-definition camera 74 (with the same parameters as the first high-definition camera 52) is fixed to the sliding base 72 to capture images after cleaning.

[0097] Sonar detector 75: retractable (travel 0-200mm), using a high-frequency ultrasonic probe (frequency 5MHz, measurement accuracy ±0.1mm), with a thin-film piezoresistive pressure sensor 76 (model FSR402, thickness 0.5mm) attached to the surface, and a built-in Wheatstone bridge circuit, which stops extending when a contact force ≥0.5N is detected; the sonar controller records the detection position (linked with the AUV positioning system, error ±5mm) and transmits the data.

[0098] 4. AUV thruster 4 and central pivot shaft 8, such as Figure 3 As shown.

[0099] AUV thrusters 4: Three are evenly distributed around the bottom sidewall of each housing 2. They can be ducted thrusters (100W power, 50N thrust) and can be rotated 360° to adjust the direction, enabling the robot to move along the hose axially (speed 0-0.5m / s) and rotate circumferentially (angular velocity 0-30° / s).

[0100] The central pivot shaft 8 is connected to the hydraulic motor (model OMM20, working pressure 16MPa) via a connecting rod assembly. The hydraulic motor is controlled by the communication control device 9, which drives the housing 2 to achieve an opening and closing speed of 0.5° / s. When closed, it can be locked by an electromagnetic lock (suction force 1000N).

[0101] 5. Communication control device 9, such as Figure 3 As shown.

[0102] As the "central brain" of the robot, it includes:

[0103] ACN Underwater Acoustic Communication Module 91: Employs an underwater acoustic modem (communication rate 100kbps, transmission distance 1km) to transmit camera images, sonar data, and equipment status information in real time.

[0104] Intelligent processing module 92: Based on the NVIDIA Jetson Nano edge computing platform, it has a built-in bio-adsorption recognition algorithm (YOLOv5 model, recognition accuracy ≥95%) and wall thickness analysis model. It can classify the adsorption level (mild / severe) of the first / second high-definition camera images 74; send adjustment commands to each controller (nozzle, vibration, jet, etc.); compare the real-time wall thickness with the original data (preset initial wall thickness of the hose 50mm), and trigger an abnormal alarm when the detected wall thickness is ≤30mm.

[0105] III. Specific Implementation Steps of the Cleanup Method

[0106] S1: Homework Preparation

[0107] The robot is stored in the charging compartment of the mining vehicle (water depth 5000m) and the lithium battery pack is charged through the charging interface. At the same time, isocyanurate agents are added to the storage compartment 51 through the quick dosing port until it is full. The intelligent processing module 92 self-checks the status of each device (camera, sensor, thruster, etc.) and enters the standby state after confirming that there are no problems.

[0108] S2: Robot Navigation

[0109] The watercraft sends a start command, which is received by the robot via the ACN underwater acoustic communication module 91. The robot then starts the AUV thruster 4 (initial thrust 20N), and the first high-definition camera 52 captures real-time images of the surrounding environment. The intelligent processing module 92 recognizes the outline of the hoisting hose 1 through the images, calculates the relative position, and adjusts the direction of the thruster (heading angle error ≤5°) so that the robot gradually approaches the hose (the thrusting speed is reduced to 0.1m / s when it is 1m away from the hose).

[0110] S3: Housing 2 precisely closed

[0111] The intelligent processing module 92 controls the hydraulic motor to drive the housing 2 to open to 90°, and the pusher fine-tunes the robot position so that the hose is between the two housings 2; then the housing 2 is controlled to close slowly (speed 0.2° / s). When the contact limit sensor 62 on the vibrator 61 touches the surface of the hose and triggers the micro switch, the hydraulic motor stops immediately, the electromagnetic lock starts to lock the housing 2 (at this time, the coaxiality error between the housing 2 and the hose is ≤10mm), and the contact pressure between the vibrator 61 and the hose surface is stabilized at 3N.

[0112] S4: Chemical Pretreatment

[0113] The first high-definition camera 52 captures images of the hose surface (10fps) and transmits them to the intelligent processing module 92 via the ACN module. The intelligent processing module 92 has a built-in optimized YOLOv5 model and identifies the bio-adsorption area through algorithms. The calculation steps are as follows:

[0114] S41. Input layer image preprocessing: Converting the original image into standardized data that the model can process, eliminating interference from the deep-sea environment, including:

[0115] S411. Image Standardization: The original image pixel values ​​range from [0, 255] and need to be normalized to [0, 1]. Formula: ;

[0116] Original image in coordinates ,aisle (RGB three-channel) pixel values, , Training Concentration Channel The mean and standard deviation (pre-calculated, such as...) =0.485, =0.229);

[0117] S412, Mosaic Data Augmentation: Randomly select 4 images (including positive samples with biosorption and negative samples without biosorption), and stitch them together in a 2×2 grid to form a 640×640 input image. Formula: ,

[0118] In the formula, , , , 4 random images

[0119] S413. Depth Correction: Based on the depth sensor data z (unit: m) inside the AUV thruster 4, correct image distortion (such as edge blurring) caused by water pressure. Formula: ,

[0120] In the formula, Correction factor (empirical value 0.1). Reference depth (5000m) Maximum operating depth (6000m).

[0121] S42. Feature extraction is performed on the backbone network, including:

[0122] (a) The initial convolutional layer performs low-level feature extraction.

[0123] ,

[0124] In the formula, Output feature map; : Convolutional kernel weights for the c-th output channel and the k-th input channel (RGB); : Bias term; Activation function Introducing nonlinearity;

[0125] (ii) Introducing the CSP module and SE attention mechanism for feature extraction and biometric enhancement.

[0126] The CSP module includes:

[0127] (a) Input features Split into (50% channel) and (50% channel);

[0128] (b) Processed by two residual blocks (convolution + skip connections) ;

[0129] (c) Fusion get (320×320×128);

[0130] The SE attention module includes:

[0131] (1) Global average pooling: , This represents the global feature of channel c;

[0132] (2) Channel weight calculation: , , For the weights of the fully connected layer, Output weights of 0 to 1;

[0133] (3) Feature weighting: , These are channel attention weights, used to dynamically adjust the importance of different channels: for biometric channels, Take a larger value (i.e.) ↑), to enhance the expression of biosorption-related features; for background channels, Take a smaller value (i.e.) (↓) to suppress background noise interference;

[0134] (iii) The SPPF layer performs multi-scale feature fusion, and performs high-level feature fusion. Multi-scale pooling is performed using (10×10×1024) as follows: By integrating the pooling results of 1×1, 5×5, 9×9, and 13×13, the characteristics of biosorption of different sizes (such as small biofilms and large barnacles) were extracted.

[0135] S43. The neck network fuses multi-scale features through the PANet structure to output a scale feature map, solving the difficulty of identifying small targets (such as early biofilms), including the following steps:

[0136] S431, Upsampling fusion:

[0137] High-level characteristics (10×10×1024) upsampled to 20×20, compared with mid-layer features (20×20×512) Fusion: ,

[0138] In the formula, : 2x upsampling (interpolation method) Channel splicing The CSP module (further extracts fusion features) is repeatedly upsampled to 40×40 to obtain... (40×40×256), adapted for small target detection.

[0139] S432, Downsampling fusion:

[0140] Will Downsampling to 20×20, and Fusion: In the formula, 3×3 convolution downsampling (stride 2) results in the output of 3 scale feature maps: (40×40, small goal) (20×20, medium target) (10×10, the big goal).

[0141] S44. The head network predicts the bounding box, confidence level, and category of biosorption based on the scale feature map obtained in S43, including the following steps:

[0142] S441. Bounding box prediction: For each grid cell of each scale feature map, 3 candidate boxes are predicted. The bounding box parameters (center coordinates (x, y), width w, height h) are calculated using the following formula: ,

[0143] In the formula , : Coordinates of the top left corner of the grid; : Grid size (40×40 scale) =16, 20×20 scale =32); , The offset predicted by the model; : Sigmoid function (to Mapped to [0, 1], confined within the grid); , : Preset anchor frame width and height (obtained based on the bio-adsorption size clustering of the training set, such as small target anchor frames) =10, =10);

[0144] S442. Confidence and Class Prediction:

[0145] Target confidence level (whether the box contains bioadsorption): , These are the model's predicted values. Mapped to [0, 1];

[0146] Category probability (e.g., biofilm, large attachment): , For the class prediction vector, SoftMax normalizes it to a probability sum of 1.

[0147] S45. Loss calculation and model optimization, as follows:

[0148] Optimize the total loss function. ,

[0149] In the formula, the bounding box loss (Including depth correction): , Intersection over Union (IoU) (measures the overlap between the predicted bounding box and the ground truth bounding box), depth correction term: =0.2, to offset the effect of depth on bounding box prediction;

[0150] Confidence loss (Attention-weighted): , : Channel weights output by the SE module (higher weights for biological regions);

[0151] Category loss (Scale Adaptive): s: scale factor (s=1.5 for small targets, to enhance the classification weight of small targets). Category reliability factor (easily confused categories) ↑);

[0152] Optimization process: The stochastic gradient descent (SGD) optimizer is used to iteratively update model parameters (convolution kernel weights, biases, etc.). , : The parameters of the t-th iteration; Learning rate (initially 0.01, dynamically decaying); to achieve minimization Continue until the model converges (loss stabilizes or the number of iterations reaches the target).

[0153] S46. Post-processing, outputting the final result, including the following steps:

[0154] S461, Non-maximum suppression (NMS)

[0155] 1. By confidence level Sorting prediction box;

[0156] 2. Select the box with the highest confidence as the candidate box and calculate the IoU with other boxes;

[0157] 3. Remove boxes with IoU > 0.5 (consider them as duplicate boxes);

[0158] 4. Repeat steps 2-3 until all non-redundant boxes are retained;

[0159] S462. Output results: coordinates of the bounding box of the biosorption region ( 1, 1, 2, 2);

[0160] Category labels (e.g., "mild adsorption - severe adsorption");

[0161] Confidence level (e.g., 0.98, reflecting the reliability of identification).

[0162] Mild adsorption (thickness < 5mm, coverage < 30%): Control the corresponding area turbine agent nozzle 53 to switch to wide-angle atomization mode, spray time 10s;

[0163] Severe adsorption (thickness ≥ 5 mm, coverage ≥ 30%): Control the corresponding area nozzle to switch to narrow beam spray mode, spray time 20 s; after the agent is sprayed, let it stand for 30 s (to allow the agent to penetrate the biofilm).

[0164] S5: Vibration Relaxation

[0165] The intelligent processing module 92 adjusts the vibration controller according to the adsorption thickness.

[0166] Slightly adsorbed area: Vibrator 61 power adjusted to 20W, vibration time 30s;

[0167] In areas with severe adsorption: the power of vibrator 61 is adjusted to 50W and the vibration time is 60s; during the vibration, the contact limit sensor 62 monitors the pressure in real time to ensure that it does not exceed 5N.

[0168] S6: Jet shedding

[0169] After the vibration ends, the jet controller activates the jet nozzle 63:

[0170] In areas with mild adsorption: adjust the flow rate to 8 m / s and the spraying time to 20 s;

[0171] Severe adsorption area: flow rate adjusted to 15m / s, spray time 40s; water flow is provided by the robot's own seawater filter pump (inlet located at the tail of shell 2, filtration accuracy 50μm).

[0172] S7: Residue Monitoring

[0173] The sliding base 72 reciprocates along the slide groove 71 at a speed of 0.1m / s. The second high-definition camera 74 captures the cleaned image and transmits it to the intelligent processing module 92. The intelligent processing module 92 identifies residual organisms (a single residual area ≥1cm² is a residual that needs to be treated) and records its coordinate position (the error relative to the robot's initial position is ≤5mm).

[0174] S8: Precision Cleaning and Re-inspection

[0175] The intelligent processing module 92 controls the sliding base 72 to move to the residual position, and the robotic arm 73 extends to a distance of 5mm from the residual, and starts the steel wire brush to rotate and clean (300rpm) for 10 seconds.

[0176] After cleaning, the robotic arm 73 retracts, and the second high-definition camera 74 takes another picture. If any residue is not removed, the cleaning is repeated (up to 3 times).

[0177] After all residues are processed, the robotic arm 73 returns to its initial position, and the sliding base 72 resumes its inspection state.

[0178] S9: Wall Thickness Detection

[0179] After the residue is cleaned up, the processing module controls the sonar detector 75 to extend (at a speed of 5 mm / s). When the thin-film piezoresistive pressure sensor 76 detects a contact force of 0.5 N, it stops immediately and starts ultrasonic detection (collecting one data point every 100 mm). The sonar controller records the detection position (linked with the odometer of the AUV thruster 4), and the data is transmitted to the processing module in real time.

[0180] S10: Risk Assessment

[0181] The intelligent processing module 92 has a built-in wall thickness analysis model: the calculation steps include,

[0182] (A) Data acquisition: Sonar detector 75 collects wall thickness data at M=12 points along the circumference of the flexible tube. - For example: [48, 47, 45, 30, 49, 46, 44, 28, 47, 48, 46, 45] mm;

[0183] Historical exploration data: the time and wall thickness of the last 3 measurements (e.g., t1 = 10 days ago). =50mm; t2 = 5 days ago, =49mm; t3=0 days (current) =47mm);

[0184] Spatial weight preset: Based on historical corrosion data, points 4 and 8 are the seams (high incidence of corrosion). = =1.5, the rest of the points =0.8;

[0185] (B) Calculate the local deviation rate and the overall deviation rate. Local deviation rate Overall deviation rate ,

[0186] Where M: the number of detection points along the circumference of the hose. : No. One detection point ( =1, 2, ..., M), Original standard wall thickness of the hose : No. Measured wall thickness at each point : No. Spatial weights of each point;

[0187] Local deviation rate (in) (For example, =4) =40%;

[0188] Overall Deviation Rate: Substituting the data, we get: =28%;

[0189] (C) Calculation of corrosion rate and prediction of future wall thickness:

[0190] Based on past n detection data (times t1, t2, ..., t... n Wall thickness , ..., The current corrosion rate is predicted using the exponential smoothing method: ,

[0191] In the formula, : Current predicted corrosion rate (unit: mm / day, positive value indicates wall thickness reduction). Smoothing coefficient (0.5 < α < 1, take 0.7, more focused on recent data). The difference in wall thickness between the two most recent measurements (a negative sign indicates thinning). : The time interval between the two most recent detections (in days); : The corrosion rate predicted in the last prediction (initial value is 0);

[0192] Based on the current corrosion rate, predict the wall thickness for the next k days: ;

[0193] In the formula, Predicted wall thickness after k days (unit: mm). Prediction period (e.g., 7 days, early warning maintenance time).

[0194] The calculation steps are as follows:

[0195] Difference between the last two data points: =47 49= 2mm, time interval t3 t2 = 5 days, recent corrosion rate: 5 2= 0.4mm / day (the negative sign indicates thinning);

[0196] Last predicted rate = 0.3 mm / day (calculated based on t1 and t2);

[0197] Current corrosion rate: =0.7×( 0.4)+(1 0.7)×( 0.3)= 0.37mm / day;

[0198] Predicted wall thickness in 7 days: =47 7 × 0.37 = 47 2.59 = 44.41 mm;

[0199] (D) Risk level determination; Risk level ,

[0200] Threshold meanings: 20% / 40%: Safety / dangerous deviation thresholds based on material strength experiments (when the wall thickness loss exceeds 40%, the structural strength decreases by more than 60%); 0.01 / 0.05mm / day: Corrosion rate thresholds (exceeding 0.05mm / day indicates accelerated corrosion, and the safety threshold may be exceeded within 7 days).

[0201] current =28% (between 20% and 40%) =0.37mm / day (far greater than 0.05mm / day). According to the judgment formula, a "danger" registration is triggered, and the system sends an emergency warning to the waterborne operation vessel, recommending immediate shutdown for maintenance.

[0202] During use, the intelligent processing module 92 compares the real-time wall thickness data with the original value (50mm). If the measured wall thickness is ≥40mm (corresponding to...), the module will process the data. ≤20%), and =0.005mm / day (≤0.01mm / day); the model determines "safe", and the system executes: the AUV thruster 4 moves along the hose axis at a speed of 0.3m / s to enter the next work section for detection; if the detected value is <30mm, the intelligent processing module 92 controls the robot to stop moving, starts circumferential rotation (angular velocity 10° / s), performs wall thickness detection around the entire circumference of the position, and sends an abnormal alarm (including position coordinates and wall thickness data) to the water work vessel through the ACN module, waiting for manual instructions.

[0203] S11: Finishing the work

[0204] When the power storage tank 3's power drops to 20%, or when the watercraft sends a "work completed" command, the intelligent processing module 92 controls all devices to stop operating;

[0205] The electromagnetic lock is unlocked, the hydraulic motor drives the housing 2 to open to 120°, the AUV thruster 4 is activated, the robot disengages from the hose, and the hydraulic motor reverses to drive the housing 2 to close.

[0206] The intelligent processing module 92, based on the initial position memory, controls the robot to return to the mining vehicle's charging compartment to recharge and stand by.

[0207] The above specific embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit it. Although the present invention has been described in detail with reference to examples, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical solutions of the present invention without departing from the scope of the technical solutions of the present invention, and all such modifications or substitutions should be covered within the scope of the claims of the present invention.

Claims

1. A deep-sea ore-lifting hose adsorbing biological intelligent cleaning robot, characterized in that: It comprises two half-cylinder housings (2) hinged by a central pivot shaft (8) connected with a hydraulic motor through a linkage assembly, and a power storage compartment (3) for storing power supply in the housing (2); A visual drug application device (5) for monitoring and preliminarily treating the bio-adsorption on the mine raising hose (1) is arranged at the upper end of the housing (2); A vibration and erosion device (6) for further treating the adsorbed organisms on the mine raising hose (1) is arranged below the visual drug application device (5) and inside the two housings (2); A positioning and detecting device (7) for monitoring the treated mine raising hose (1), treating the stubborn residual adsorbed organisms and monitoring the state of the mine raising hose (1) is arranged below the vibration and erosion device (6); A plurality of AUV thrusters (4) are arranged on the side walls of the bottom of the housing (2) at intervals; It further comprises a communication control device (9) for controlling the visual drug application device (5), the vibration and erosion device (6), the positioning and detecting device (7), the AUV thrusters (4), the hydraulic motor and signal transmission; The visual drug application device (5) comprises a drug storage compartment (51) arranged at the upper end of the housing (2) for storing the drug, a plurality of first high-definition cameras (52) arranged at intervals at the top end of the drug storage compartment (51), a plurality of turbine drug spray heads (53) arranged at intervals on the inner side walls of the drug storage compartment (51), and a spray head controller for controlling the amount and range of the drug sprayed by the turbine drug spray heads (53), wherein the spray head controller is signal connected with the communication control device (9); The vibration and erosion device (6) comprises a vibrator (61) arranged below the turbine drug spray head (53) for vibrating and relaxing the adsorbed organisms on the pipeline and reducing their adsorption capacity, a vibration controller for controlling the vibrator (61), a contact type limit sensor (62) arranged on the surface of the vibrator (61) for limiting the vibrator (61) and avoiding the damage of the vibrator (61) to the mine raising hose (1), a plurality of jet spray heads (63) arranged at intervals below the vibrator (61) for further vibrating and relaxing the adsorbed organisms and jetting off the adsorbed organisms, and a jet control controller for controlling the jet spray heads (63), wherein the vibration controller and the jet control controller are signal connected with the communication control device (9); The positioning and detecting device (7) comprises a chute (71) arranged below the jet spray head (63), a sliding base (72) slidingly connected with the chute (71), a telescopic mechanical arm (73) and a second high-definition camera (74) arranged on the sliding base (72), a telescopic sonar detector (75) arranged below the chute (71), a thin film piezoresistive pressure sensor (76) arranged on the surface of the sonar detector (75) for preventing the sonar detector (75) from damaging the pipeline, and a sonar controller arranged in the sonar detector (75) for recording the detection data of the sonar detector (75) and transmitting the detection data to the communication control device (9); The communication control device (9) comprises an ACN underwater communication module (91) for transmitting detection data of the first high-definition camera (52), the second high-definition camera (74) and the sonar controller, and further comprises an intelligent processing module (92) for processing data transmitted by the ACN underwater communication module (91) and making a prediction on whether the mine-lifting hose (1) will be damaged, and the intelligent processing module (92) is further used for regulating and controlling the jet controller, the vibration controller, the jet flow controller, the mechanical arm (73), the AUV propeller (4) and the hydraulic motor.

2. A method for cleaning the adsorbed organisms on the deep-sea ore lifting hose, characterized in that the intelligent cleaning robot for the deep-sea ore lifting hose according to claim 1 is used. The method comprises the following steps: S1: preparation for operation, the robot is stored in the charging cabin of the mining vehicle and completes charging, and chemical agents are supplemented to the agent storage bin (51); S2: start operation, the AUV propeller (4) is started and the advancing direction of the AUV propeller (4) is adjusted according to image data provided by the first high-definition camera (52), and gradually moves towards the mine-lifting hose (1); S3: precise closure of the shell (2), the intelligent processing module (92) controls the hydraulic motor to open the shell (2) and gradually approach the mine-lifting hose (1), when the contact type limit sensor (62) is triggered, the hydraulic motor stops running, so that the shell (2) is fixed on the mine-lifting hose (1), and at this time, the vibrator (61) is tightly attached to the mine-lifting hose (1); S4: agent pretreatment, the first high-definition camera (52) shoots the bio-adsorption image on the surface of the mine-lifting hose (1) and transmits the image to the intelligent processing module (92) through the ACN underwater communication module (91), the intelligent processing module (92) processes and identifies the attached objects according to the image, and then controls the spraying mode of the turbine agent nozzle (53) through the jet controller to spray the agent, wherein the spraying mode comprises a wide-angle atomization mode and a narrow-beam jetting mode; S5: vibration relaxation, the intelligent processing module (92) adjusts the vibration intensity of the vibrator (61) at each position according to different adsorption conditions in the image data transmitted by the first high-definition camera (52) through the vibration controller, so as to weaken the adsorption effect of the adsorbed organisms on the pipeline; S6: jet flow shedding, the intelligent processing module (92) adjusts the flow rate of the jet flow nozzle (63) at each position according to different adsorption conditions in the image data transmitted by the first high-definition camera (52) through the jet flow controller, so as to shed the attached organisms; S7: residual monitoring, the sliding base (72) reciprocates along the sliding groove (71), the cleaned image shot by the second high-definition camera (74) is transmitted to the intelligent processing module (92) through the ACN underwater communication module (91), and the intelligent processing module (92) determines the position of the residual adsorbed objects on the mine-lifting hose (1) and whether further treatment is needed according to the transmitted image data. S8: Precise cleaning and re-inspection, the intelligent processing module (92) controls the mechanical arm (73) to extend close to remove the remaining adsorbate, after cleaning, the mechanical arm (73) retracts and is re-inspected by the second high-definition camera (74), if it does not meet the standard, it is repeated cleaning, after cleaning, the intelligent processing module (92) controls the mechanical arm (73) to reset, the sliding base (72) restores the movement and continues to patrol through the second high-definition camera (74); S9: Wall thickness detection, after the mechanical arm (73) cleaning, the intelligent processing module (92) controls the sonar detector (75) to extend, when the thin film piezoresistive pressure sensor (76) detects the contact force, stop extending, start wall thickness detection and record the detection position with the odometer of AUV thruster (4), through ACN underwater acoustic communication module (91) to the intelligent processing module (92); S10: Risk assessment, the intelligent processing module (92) compares the detection of the mining hose (1) wall thickness with the original wall thickness data to judge the damage risk, if the data is abnormal, control the AUV thruster (4) to stay and rotate to detect the wall thickness of the whole hose, and the data is transmitted to the water operation ship through the ACN underwater acoustic communication module (91), if the data is normal, the sonar detector (75) is retracted, the AUV thruster (4) continues to operate along the hose; S11: Operation ending, when the power is insufficient or the operation is completed, all devices stop running, the intelligent processing module (92) controls the hydraulic motor to rotate to open the shell (2), the robot is separated from the mining hose (1), the hydraulic motor is reversed, the shell (2) is closed and returned to the mining car charging cabin.

3. The method of claim 2, wherein: In S4, the intelligent processing module (92) is built-in optimized YOLOv5 model, the processing of the image data of the YOLOv5 model includes the following steps: S41, image preprocessing of input layer, converting the original image into standardized data that can be processed by the model, eliminating the interference of deep sea environment; S42, backbone network for feature extraction, including: (I) initial convolutional layer for low-level feature extraction, ， In the formula, : output feature map; : convolution kernel weight of the cth output channel and the kth input channel (RGB); : bias term; : activation function , introduce nonlinearity; (II) introducing CSP module and SE attention mechanism for feature extraction and biological feature enhancement; (III) SPPF layer for multi-scale feature fusion; S43, neck network fuses multi-scale features through PANet structure, and outputs scale feature map; S44, head network predicts the boundary box, confidence and category of biological adsorption according to the scale feature map obtained in S43; S45, loss calculation and model optimization; S46, post-processing, output final result.

4. The method of claim 3, wherein: In S42, The CSP module includes: (a) splitting the input features into (50% channel) and (50% channel); (b) processed through 2 residual blocks (convolution + skip connection) to ; (c) fusion obtained ; The SE attention mechanism includes: (1) Global average pooling: , is the global feature of channel c; (2) Channel weight calculation: , , is the full connection layer weight, output 0~1 weight; (3) Feature weighting: , is the channel attention weight, used to dynamically adjust the importance of different channels: for the biological feature channel, takes a larger value (i.e. ↑), to enhance the expression of biological adsorption related features; for the background channel, takes a smaller value (i.e. ↓), to suppress the interference of background noise.

5. The method of claim 4, wherein: In S10, the intelligent processing module (92) is built-in wall thickness analysis model, including, (A) data acquisition; (B) calculate local deviation rate and comprehensive deviation rate; (C) calculate corrosion rate and future wall thickness prediction; (D) risk level determination.

6. The method of cleaning of claim 5, wherein, In the wall thickness analysis model, in step (B): local deviation rate , comprehensive deviation rate ; M: number of detection points in the hose circumferential direction, : the first =1, 2, …, M), : original standard wall thickness of the hose, : the first : measured wall thickness of the first point, : the first : spatial weight of the first point.​

Images