Underwater robot pollutant detection traceability system and method based on multi-modal data

Abstract

The invention discloses an underwater robot pollutant detection traceability system and method based on multi-modal data. The method comprises the following steps: controlling an underwater robot to advance according to a preset planning path; collecting water quality parameters detected by a water quality parameter sensor, and synchronously collecting water quality images shot by an underwater high-definition camera; weighting the current water quality parameter to obtain a physical index, identifying and analyzing the current water quality image to obtain an image feature and calculating an image index, and if the physical and image index characterizes no pollution, continuing to advance along a preset planning path; if low pollution is represented, carrying out information fusion on the physical indexes and the image features to obtain the pollution heat of the current position, and continuing to advance along a preset planning path; if high pollution is represented, information fusion processing is carried out to obtain pollution heat, a local dense sampling mode is entered, the underwater robot slows down and moves near the current position, pollution heat of different positions is obtained, a pollution distribution heat map is constructed, and the robot is guided to move in the gradient rising direction of the map until a pollution source is tracked and traced.

Description

Technical Field

[0001] This invention relates to the field of underwater robot technology, and in particular to an underwater robot pollutant detection and tracing system and method based on multimodal data, applicable to complex aquatic environments such as oceans, lakes, and reservoirs. Background Technology

[0002] With the acceleration of industrialization and urbanization, water pollution has become increasingly serious, especially in industrial areas and urban peripheries. Traditional water quality monitoring methods mainly rely on fixed sampling points and laboratory analysis, but these methods have slow response times, limited coverage, and cannot achieve dynamic tracking and source tracing of pollution sources. In recent years, the rapid development of underwater robots and sensor integration technologies has provided new ideas for water environment monitoring. In particular, the maturity of underwater robot technology has made real-time, underwater pollution detection and source tracing possible. However, existing underwater monitoring solutions still face many challenges in complex aquatic environments, especially in the intelligent tracking and identification of pollution diffusion behavior, where effective technical means are still lacking.

[0003] Currently, most traditional underwater monitoring methods rely on physical sensors, such as chemical and temperature sensors. While these methods can perform preliminary detection of pollutants, they struggle to address the issue of tracing pollutant sources in dynamic waters. Furthermore, traditional underwater visual monitoring systems are limited by lighting conditions and the limitations of visual sensors, making it difficult to accurately identify pollution sources such as oil spills and discharge outlets. Therefore, there is an urgent need for an underwater pollution source tracing technology that integrates chemical sensing, visual recognition, and autonomous navigation to provide a more comprehensive and real-time water pollution monitoring capability.

[0004] A search of existing technologies revealed several solutions that are similar to this invention in application. Patent CN119058923B proposes an underwater robot integrating a pollution sensor, communication module, and image acquisition device, capable of monitoring pollutant concentrations and preliminary pollution source location underwater. However, this solution still has limitations in target recognition accuracy, pollution heat map construction, and multi-source information fusion, making it difficult to adapt to the dynamic pollution diffusion process in complex aquatic environments.

[0005] Furthermore, patent CN111781324A discloses a marine environmental monitoring platform integrating multiple sensors and data processing modules, suitable for monitoring environmental parameters in large-scale water areas. Although the system possesses certain underwater data acquisition and communication capabilities, it lacks a dedicated design for pollution source tracing scenarios. In particular, it has not been deeply optimized in areas such as real-time pollution image recognition, pollution path tracking, and intelligent target tracking, making it difficult to achieve intelligent identification and source tracing of the entire pollution event process.

[0006] In summary, although the aforementioned existing technologies have achieved certain results in the construction of underwater robot platforms, the collection of environmental parameters, and the detection of some pollution, a pollution source tracing system that combines real-time performance, intelligence, and high integration has not yet been formed. In particular, there are still technological gaps in the dynamic identification, path tracking, and causal analysis of pollution sources in complex waters.

[0007] Despite the progress made in existing underwater pollution detection and monitoring technologies, there are still many shortcomings, mainly in the following aspects:

[0008] 1) Traditional underwater monitoring systems mostly rely on single physical sensors for pollutant detection, such as chemical sensors and temperature sensors, lacking the ability to accurately identify pollution sources visually. Underwater lighting conditions are complex, and water turbidity is high, which limits the image quality acquired by visual sensors, thus affecting the accurate location and identification of pollution sources.

[0009] 2) Existing technologies typically employ fixed-point sampling or single-test methods, which make it difficult to track and trace the dynamic diffusion process of pollutants in water bodies in real time. They also fail to dynamically reflect the propagation path and changing trends of pollutants, thus limiting the comprehensive analysis and response to pollution incidents.

[0010] Therefore, this invention proposes an intelligent pollution source tracing system based on the combination of underwater robots, sensors and visual recognition technology, aiming to overcome the limitations of existing technologies and provide a more accurate and intelligent solution for underwater environmental monitoring and pollution source tracing. Summary of the Invention

[0011] This invention addresses the problems and shortcomings of existing technologies by providing an underwater robot pollutant detection and tracing system and method based on multimodal data.

[0012] The present invention solves the above-mentioned technical problems through the following technical solution:

[0013] This invention provides an underwater robot pollutant detection and tracing system based on multimodal data, characterized by including an underwater high-definition camera and a high-brightness supplementary lighting group fixed on the underwater robot, and a water quality parameter sensor fixed at the bottom;

[0014] The system also includes: a path control module for controlling the underwater robot to travel along a preset planned path;

[0015] The information acquisition module is used to acquire water quality parameters detected by the water quality parameter sensor at a set acquisition frequency, and simultaneously acquire water quality images captured by the underwater high-definition camera under the supplementary lighting of the supplementary lighting group, forming a multi-dimensional data packet containing the current acquisition time, location coordinates, water quality parameters and water quality images;

[0016] The information identification and processing module is used to weight and process the current water quality parameters to obtain physicochemical indicators. It uses a deep learning model to analyze the current water quality image for pollutant identification, obtains image features, and calculates image indicators. Based on the physicochemical and image indicators, it determines the pollution level. If the physicochemical and image indicators indicate no pollution, it calls the path control module to continue along the pre-planned path. If the physicochemical and image indicators indicate low pollution, it calls the information fusion processing module to fuse the physicochemical indicators and image features to obtain the pollution intensity at the current location and continues along the pre-planned path. If the physicochemical and image indicators indicate high pollution... If pollution is detected, the information fusion processing module is invoked to obtain pollution heat maps, and the system then enters the local dense sampling module. The underwater robot decelerates and moves near its current position, invoking the information acquisition and information recognition modules to obtain high-density physical indicators and image features of different locations. The information fusion processing module is then invoked to obtain pollution heat maps corresponding to different locations, and the tracking and tracing module is invoked to construct and update the pollution distribution heat map. Based on the gradient changes in the pollution distribution heat map, an improved concentration gradient tracking algorithm is used to guide the underwater robot to move along the gradient upward direction, dynamically adjusting the underwater robot's path until the pollution source is tracked.

[0017] The positive and progressive effects of this invention are as follows:

[0018] This invention designs an innovative intelligent system for underwater pollutant detection and source tracing. Through the collaborative work of multiple integrated modules, the system realizes a complete process, including visual identification of underwater pollutants, accurate concentration detection, information fusion processing, and accurate source tracing of pollution.

[0019] In this invention, the underwater robot can perform automatic navigation according to a preset path during operation, and can also dynamically reconstruct the motion path based on image recognition results and concentration gradient, thus possessing the characteristics of strong autonomy, high control precision, and strong environmental adaptability.

[0020] In this invention, pollutant detection and path control are linked in real time. When the detection identifies a significant increase in pollution intensity along a certain direction, the system will automatically adjust its movement path and continue moving along the intensity gradient, thereby achieving the function of tracing the source of pollution.

[0021] This invention, through a path planning method based on multimodal fusion and gradient feedback control, enables the system to have a strong pollution source tracing capability, and is particularly suitable for water pollution scenarios with multiple superimposed sources, dynamic changes, and frequent disturbances. Attached Figure Description

[0022] Figure 1 This is a schematic diagram of the underwater robot pollutant detection and tracing system according to a preferred embodiment of the present invention.

[0023] Figure 2This is a flowchart of a preferred embodiment of the underwater robot pollutant detection and tracing method of the present invention. Detailed Implementation

[0024] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0025] For ease of description, only the parts relevant to the present invention are shown in the accompanying drawings. The terms "first," "second," etc., used in this invention are merely for the convenience of describing the technical solutions of the invention and do not have a specific limiting effect; they are all general references and do not constitute a limitation on the technical solutions of the present invention.

[0026] like Figure 1 As shown, this embodiment of the invention provides an underwater robot pollutant detection and tracing system based on multimodal data. The system includes a water quality parameter sensor 1 fixed to the bottom of the underwater robot, an underwater high-definition camera 2 and a high-brightness supplementary lighting group 3 fixed to the underwater robot, a path control module 4, an information acquisition module 5, an information recognition and processing module 6, an information fusion processing module 7, a local dense sampling module 8, and a tracking and tracing module 9. The path control module 4, information acquisition module 5, information recognition and processing module 6, information fusion processing module 7, local dense sampling module 8, and tracking and tracing module 9 can be integrated into a host computer that communicates with the underwater robot via a wired optical cable.

[0027] Water quality parameter sensor 1 is used to continuously detect the water quality parameters at the current location of the water body. The water quality parameter sensor includes various water quality parameter sensors such as pH, conductivity, chemical oxygen demand (COD), ammonia nitrogen, and heavy metal ions (such as lead, cadmium, and mercury). For example, there is a pH sensor for collecting pH value, a conductivity sensor for collecting conductivity, an optical sensor for collecting COD, and an ion sensor for collecting harmful ions. All sensors maintain a stable flow state through a water flow conduction design to avoid measurement interference and are connected to the information acquisition module 5 to realize the synchronous recording of multiple parameters. Each sensor supports temperature compensation, self-calibration, and real-time data upload, and the measurement accuracy is stable.

[0028] Specifically, the pH and conductivity measurements utilize an electrochemical method to construct the sensor system. pH measurement relies on the response of the glass electrode to hydrogen ion activity; the output potential is negatively correlated with the acidity or alkalinity of the water, and the conversion process follows the Nernst equation. The system automatically compensates for this change based on ambient temperature. Conductivity measurement involves measuring the resistance change between two platinum electrodes and, combined with the electrode constants, estimating the concentration of dissolved salts and total ions in the water. Both types of sensors integrate temperature probes and feature automatic calibration to ensure stability and accuracy during long-term operation.

[0029] For measuring chemical oxygen demand (COD), the device employs an optical sensor system based on ultraviolet absorption, emitting a 254nm ultraviolet light source. The absorbance of the water sample is measured through a colorimetric cell, and combined with empirical coefficients obtained beforehand by fitting standard solutions, the concentration of oxidizable organic matter in the water is calculated in real time. Due to the complexity of organic pollutants, this module also supports a self-learning fitting algorithm to dynamically correct the measurement results during long-term operation, further improving the reliability of the measured values.

[0030] Monitoring of ammonia nitrogen and heavy metal ions utilizes ion-selective electrode (ISE) technology. The electrode tip is covered with a membrane highly selective for target ions. When a target ion enters the membrane and forms a potential difference with the internal reference solution, the system immediately records this potential difference and converts it into the mass concentration of that ion using the Nernst equation and a calibration curve. In typical applications, the device can detect various harmful ions in water, such as ammonia nitrogen, lead, cadmium, and mercury, with concentration detection accuracy down to the microgram per liter (μg / L) level.

[0031] The underwater high-definition camera 2 is used to continuously monitor water quality images of the current location in the water area.

[0032] In this embodiment, image acquisition is performed simultaneously with the measurement of water physicochemical parameters to achieve visual identification and classification of pollutants. The image acquisition module mainly consists of an underwater high-definition camera equipped with a high-brightness LED supplementary light group 3. This camera supports stepless rotation within a pitch angle range of ±90°, ensuring that it can always automatically align with the target area on the bottom of the water body when the underwater robot is in different postures or when it is bypassing obstacles.

[0033] To overcome issues such as insufficient underwater lighting and light scattering, the camera employs a low-light imaging chip with strong noise resistance and a high dynamic range, ensuring clear, high-contrast images even in murky or low-light conditions. Simultaneously, the supplemental lighting system automatically adjusts brightness output, optimizing exposure parameters based on water depth, water color, and background brightness variations to achieve optimal image quality. All acquired image data is transmitted in real-time to a host computer via wired fiber optic cable, preventing data loss due to insufficient wireless bandwidth or underwater channel interference.

[0034] The path control module 4 is used to control the underwater robot to travel along a preset planned path.

[0035] In path control module 4, based on the proportional-integral-derivative (PID) control algorithm, the control output u(t) is calculated in real time according to the navigation error e(t). By continuously adjusting the thrust and direction of the underwater robot's thrusters, precise attitude adjustment and trajectory correction are achieved, controlling the underwater robot to move along the preset planned path.

[0036]

[0037] Where e(t) represents the deviation between the desired path and the current position, K p K i K d These are the proportional, integral, and differential coefficients, respectively.

[0038] The information acquisition module 5 is used to acquire water quality parameters detected by the water quality parameter sensor 1 at a set acquisition frequency, and simultaneously acquire water quality images captured by the underwater high-definition camera 2 under the supplementary lighting of the high-brightness supplementary lighting group 3, forming a multi-dimensional data packet containing the current acquisition time, current location coordinates, current water quality parameters and current water quality images.

[0039] During the movement, the system records complete water quality data (water quality parameters and water quality images) at each set collection time, and packages it synchronously with the current navigation location data to form a multi-dimensional data packet containing geographic coordinates, timestamps, and all water quality data. Each data packet is immediately sent to the information recognition and processing module 6 for subsequent recognition processing.

[0040] The information identification and processing module 6 is used to perform weighted processing on the current water quality parameters to obtain physical and chemical indicators.

[0041] The information recognition and processing module 6 is also used to perform pollutant identification and analysis on the current water quality image using a deep learning model to obtain image features and calculate image indicators.

[0042] The deep learning model used is a combination of the YOLOv5 and Unet models. The YOLOv5 model is used to identify pollutants in the current water quality image, while the Unet model performs semantic segmentation on the identified polluted areas, extracts boundary features from continuous polluted areas, and extracts image texture and color distribution information layer by layer through convolutional layers. Pixel-level classification and contour drawing are then performed on the polluted areas to obtain the image features of the current water quality image. These features include the number of pollutants, the type of pollutant, and the image confidence score. The image index of the current water quality image can be calculated by multiplying the number of pollutants and the image confidence score.

[0043] In terms of image recognition, the system integrates deep learning-based intelligent recognition algorithms, with the model primarily employing a combination of YOLOv5 and Unet network structures. The YOLOv5 model handles pollutant target detection, capable of identifying significant pollutants in images in real time, such as oil films, floating plastics, household waste, suspicious sewage outlets, and abnormally colored watermarks. This model is pre-trained using an underwater pollutant image dataset, exhibiting strong generalization capabilities, particularly effective for detecting small targets in complex scenes with occlusion and low contrast.

[0044] The Unet model is used for semantic segmentation of contaminated areas. Its structure is suitable for extracting boundary features from continuous contaminated areas (such as oil slicks or large areas of turbidity). It extracts image texture and color distribution information layer by layer through convolutional layers and performs pixel-level classification and contour drawing of contaminated areas. The recognition results include not only the number of contaminants and their category labels, but also their corresponding bounding boxes, segmentation masks, center coordinates, and confidence scores in the image.

[0045] Each image recognition result is associated with the current location coordinates, water quality parameters, and other information, and is uniformly encoded and stored for subsequent information fusion and pollution heat map construction.

[0046] Overall, image acquisition and recognition, as the "visual nervous system" of this invention, not only provides the ability to perceive pollutants as "visible," but also endows the system with the ability to intelligently determine the type, location, and extent of pollution through deep models, providing strong visual support for subsequent information fusion and source tracing path optimization.

[0047] The information identification and processing module 6 is also used to determine the pollution status of the current location based on the physical and image indicators of the current location. If the physical indicators are lower than the lower limit of the first set range and the image indicators are lower than the lower limit of the second set range, it indicates that the current location is unpolluted. If the physical indicators are higher than the upper limit of the first set range or the image indicators are higher than the upper limit of the second set range, it indicates that the current location is highly polluted. Otherwise, it indicates that the current location is lowly polluted.

[0048] The information recognition and processing module 6 is also used to call the path control module 4 to continue traveling along a preset planned path if the physical and image indicators indicate that the current location is free of pollution.

[0049] If the physical and image indicators indicate that the current location is low in pollution, the information fusion processing module 7 is called to fuse the physical and image features to obtain the pollution intensity of the current location, and the path control module 4 is called to continue traveling along the preset planned path.

[0050] If the physical and image indicators indicate high pollution at the current location, the information fusion processing module 7 is invoked to obtain the pollution heat of the current location, and the system enters the local dense sampling module 8. The underwater robot decelerates and moves repeatedly in a spiral or grid-like path near the current location. The information acquisition module 5 and the information recognition module 6 are invoked to obtain high-density physical and image features of different locations, improving the identification accuracy of the pollution source area. The information fusion processing module 7 is invoked to obtain the pollution heat corresponding to different locations near the current location. Then, the tracking and tracing module 9 is invoked to construct an updated pollution distribution heat map using all the currently obtained pollution heat. Based on the gradient changes in the pollution distribution heat map, an improved concentration gradient tracking algorithm is used to guide the underwater robot to move along the gradient upward direction, dynamically adjusting the underwater robot's travel path until the pollution source is tracked.

[0051] In this embodiment, when there is no pollution at the current location, the robot continues to travel along the preset planned path. When the pollution at the current location is low, the pollution heat at the current location is calculated, and the robot continues to travel along the preset planned path. When the pollution at the current location is high, the pollution heat at the current location is calculated, and a local dense sampling mode is entered near the current location to obtain the pollution heat corresponding to different locations near the current location. A pollution distribution heat map is established. Based on the gradient change in the pollution distribution heat map, the underwater robot is guided to move along the gradient upward direction, which can track and trace the pollution source.

[0052] In the information fusion processing module 7, the physical indicators, image features and location coordinates at the current acquisition time are synchronized in time and registered in space, and a weighted algorithm is used to calculate the pollution heat at the current location;

[0053] The weighted algorithm H(x,y,z) = α*C + β*P

[0054] In the formula, α and β are the weighting coefficients of the physical and chemical indicators and the image confidence, respectively. The system defaults to configurable coefficients and supports dynamic adjustment based on the environmental type (such as nearshore waters or reservoirs). C is the physical and chemical indicator, P is the image confidence, and H(x,y,z) is the pollution heat at the current location.

[0055] The multimodal information fusion stage aims to integrate different types of sensing data into a spatiotemporally consistent pollutant distribution model, thereby providing a basis for precise location and path decision-making of pollution sources. The information fusion module uses timestamps as the core index to synchronize and spatially register water quality parameters, camera recognition results, and platform location information collected at the same time. During spatial registration, pose (position + orientation) data provided by the underwater robot's attitude sensors and inertial navigation system (INS) is used to transform pollutant targets in the image from the image coordinate system to the world coordinate system and match their positions with water quality sampling points. Based on data alignment, the fusion module uses Gaussian kernel weighting to construct a pollutant heatmap. This heatmap reflects the current pollution intensity distribution of the water body in three-dimensional space.

[0056] In the source tracing module 9, an improved concentration gradient tracking algorithm is implemented:

[0057] Each step involves calculating the local gradient vector.

[0058]

[0059] In conjunction with hydrodynamic constraints, the system generates the next motion direction command. To improve path stability and anti-interference capability, a path smoothing factor λ is introduced, causing the path update to follow the following iterative strategy:

[0060]

[0061] in, This represents the increment of pollution heat H in the x-direction. This represents the increment of pollution heat H in the y-direction. This represents the increment of pollution heat H in the z-direction. Indicates the current direction of movement. Indicates the direction of the next movement, where η is the step size coefficient. Using the motion direction vector from the previous step, the heat map is plotted as a two-dimensional or three-dimensional pollution map on a geographic coordinate system.

[0062] like Figure 2 As shown, this embodiment also provides a method for detecting and tracing contaminants in underwater robots based on multimodal data, including:

[0063] S1. Control the underwater robot to travel along a preset planned path.

[0064] S2. Collect water quality parameters detected by the water quality parameter sensor at a set collection frequency, and simultaneously collect water quality images captured by the underwater high-definition camera under the supplementary lighting of the supplementary lighting group, forming a multi-dimensional data packet containing the current collection time, location coordinates, water quality parameters and water quality images. The water quality parameter sensor is fixed to the bottom of the underwater robot, and the underwater high-definition camera and high-brightness supplementary lighting group are fixed on the underwater robot.

[0065] S3. Weight the current water quality parameters to obtain physical indicators, and use a deep learning model to perform pollutant identification and analysis on the current water quality image to obtain image features.

[0066] S4. Calculate image indicators based on current image features. Determine the pollution status based on physical and image indicators. If physical and image indicators indicate no pollution, continue along the pre-planned path. If physical and image indicators indicate low pollution, fuse physical and image indicators with image features to obtain the pollution intensity at the current location. If physical and image indicators indicate high pollution, fuse physical and image indicators with image features to obtain the pollution intensity at the current location and enter a local dense sampling mode. The underwater robot decelerates and moves near the current location, executing S2 and S3 to obtain high-density physical and image indicators and image features at different locations. Fuse physical and image features at different locations to obtain the corresponding pollution intensity, then update the pollution distribution heat map. Based on the gradient changes in the pollution distribution heat map, use an improved concentration gradient tracking algorithm to guide the underwater robot to move along the gradient upward direction, dynamically adjusting the underwater robot's path until the pollution source is tracked.

[0067] In this invention, the underwater robot can perform automatic navigation according to a preset path during operation, and can also dynamically reconstruct the motion path based on image recognition results and concentration gradient, thus possessing the characteristics of strong autonomy, high control precision, and strong environmental adaptability.

[0068] In this invention, pollutant detection and path control are linked in real time. When the detection identifies a significant increase in pollution intensity along a certain direction, the system will automatically adjust its movement path and continue moving along the intensity gradient, thereby achieving the function of tracing the source of pollution.

[0069] This invention integrates multiple water quality data detection methods, supporting real-time detection of various water quality indicators such as pH, dissolved oxygen, conductivity, ammonia nitrogen, and COD. This significantly improves the comprehensiveness and accuracy of the detection, and has a wider range of application adaptability compared to existing devices with only a single detection function.

[0070] This invention constructs a multimodal pollution heat map by fusing sensor data and image recognition results, which can intuitively present the spatial distribution of pollutants, realize data-driven pollution trend analysis, and improve the intelligence level of pollution location and trend judgment.

[0071] This invention employs a direction determination method based on concentration gradients, combined with an adaptive path adjustment strategy, to guide an underwater platform to search in reverse along the pollutant diffusion path to the location of potential pollution sources, breaking through the technical bottleneck of traditional detection devices that can only "detect" but not "trace the source".

[0072] In summary, this invention, by integrating high-precision water quality sensors, intelligent image recognition, underwater positioning and navigation, and multimodal information fusion technologies, achieves the ability to "see, measure accurately, and locate" pollutants. The device possesses advantages such as high integration, intelligence, and strong environmental adaptability. Especially in situations involving complex water structures, non-stationary pollution sources, or dynamically spreading pollution, it can effectively achieve automatic pollutant detection and source tracing, demonstrating broad practical application value.

[0073] While specific embodiments of the present invention have been described above, those skilled in the art should understand that these are merely illustrative examples, and the scope of protection of the present invention is defined by the appended claims. Those skilled in the art can make various changes or modifications to these embodiments without departing from the principles and essence of the present invention, but all such changes and modifications fall within the scope of protection of the present invention.

Claims

1. An underwater robot pollutant detection and tracing system based on multimodal data, characterized in that it includes an underwater high-definition camera and a high-brightness supplementary lighting group fixed on the underwater robot and a water quality parameter sensor fixed at the bottom; The system also includes: The path control module is used to control the underwater robot to travel along a preset planned path; The information acquisition module is used to acquire water quality parameters detected by the water quality parameter sensor at a set acquisition frequency, and simultaneously acquire water quality images captured by the underwater high-definition camera under the supplementary lighting of the supplementary lighting group, forming a multi-dimensional data packet containing the current acquisition time, location coordinates, water quality parameters and water quality images; The information identification and processing module is used to weight and process the current water quality parameters to obtain physicochemical indicators. It uses a deep learning model to analyze the current water quality image for pollutant identification, obtains image features, and calculates image indicators. Based on the physicochemical and image indicators, it determines the pollution level. If the physicochemical and image indicators indicate no pollution, it calls the path control module to continue along the pre-planned path. If the physicochemical and image indicators indicate low pollution, it calls the information fusion processing module to fuse the physicochemical indicators and image features to obtain the pollution intensity at the current location and continues along the pre-planned path. If the physicochemical and image indicators indicate high pollution... If pollution is detected, the information fusion processing module is invoked to obtain pollution heat maps, and the system then enters the local dense sampling module. The underwater robot decelerates and moves near its current position, invoking the information acquisition and information recognition modules to obtain high-density physical indicators and image features of different locations. The information fusion processing module is then invoked to obtain pollution heat maps corresponding to different locations, and the tracking and tracing module is invoked to construct and update the pollution distribution heat map. Based on the gradient changes in the pollution distribution heat map, an improved concentration gradient tracking algorithm is used to guide the underwater robot to move along the gradient upward direction, dynamically adjusting the underwater robot's path until the pollution source is tracked.

2. The underwater robot pollutant detection and tracing system based on multimodal data as described in claim 1, characterized in that the deep learning model includes a combination of the YOLOv5 model and the Unet model; The information recognition and processing module is used to identify pollutants in the current water quality image using the YOLOv5 model. The Unet model performs semantic segmentation on the identified polluted areas, extracts boundary features in continuous polluted areas, extracts image texture and color distribution information layer by layer through convolutional layers, and performs pixel-level classification and contour drawing on the polluted areas to obtain the image features of the current water quality image. The image features include the number and type of pollutants and the image confidence score. The image index of the current water quality image is calculated by multiplying the number of pollutants and the image confidence score. The pollution level is determined based on physical and chemical indicators and image indicators. If the physical and chemical indicators are below the lower limit of the first set range and the image indicators are below the lower limit of the second set range, it indicates no pollution. If the physical and chemical indicators are above the upper limit of the first set range or the image indicators are above the upper limit of the second set range, it indicates high pollution. Otherwise, it indicates low pollution.

3. The underwater robot pollutant detection and tracing system based on multimodal data as described in claim 2, characterized in that the information fusion processing module is used to synchronize and register the physical indicators, image features and location coordinates at the current acquisition time in time and to calculate the pollution heat at the current location using a weighted algorithm; The weighted algorithm H(x,y,z) = α*C + β*P In the formula, α and β are the weighting coefficients of the physical index and the image confidence, respectively. The system defaults to configurable coefficients and supports dynamic adjustment based on the environment type. C is the physical index, P is the image confidence, and H(x,y,z) is the pollution heat at the current location.

4. The underwater robot pollutant detection and tracing system based on multimodal data as described in claim 1, characterized in that the local dense sampling module is used for the underwater robot to decelerate and move repeatedly in a spiral or grid-like path near the current position, and calls the information acquisition module and information recognition module to obtain high-density physical indicators and image features at different locations.

5. The underwater robot pollutant detection and tracing system based on multimodal data as described in claim 1, characterized by an improved concentration gradient tracking algorithm: Each step involves calculating the local gradient vector. In conjunction with hydrodynamic constraints, the system generates the next motion direction command. To improve path stability and anti-interference capability, a path smoothing factor λ is introduced, causing the path update to follow the following iterative strategy: in, This represents the increment of pollution heat H in the x-direction. This represents the increment of pollution heat H in the y-direction. This represents the increment of pollution heat H in the z-direction. Indicates the current direction of movement. Indicates the direction of the next movement, where η is the step size coefficient. Using the motion direction vector from the previous step, the heat map is plotted as a two-dimensional or three-dimensional pollution map on a geographic coordinate system.

6. The underwater robot pollutant detection and tracing system based on multimodal data as described in claim 1, characterized in that, The path control module is used to calculate the control output u(t) in real time based on the navigation error e(t) using a proportional-integral-derivative control algorithm. By continuously adjusting the thrust and direction of the underwater robot's thrusters, it achieves precise attitude adjustment and trajectory correction, controlling the underwater robot to move along a preset planned path. Where e(t) represents the deviation between the desired path and the current position, K p K i K d These are the proportional, integral, and differential coefficients, respectively.

7. The underwater robot pollutant detection and tracing system based on multimodal data as described in claim 1, characterized in that, Water quality parameter sensors include pH sensors for collecting pH values, conductivity sensors for collecting conductivity, optical sensors for collecting chemical oxygen demand, and ion sensors for collecting harmful ions.

8. A method for detecting and tracing contaminants in underwater robots based on multimodal data, characterized in that it includes: S1. Control the underwater robot to travel along a preset planned path; S2. Collect water quality parameters detected by the water quality parameter sensor at a set collection frequency, and simultaneously collect water quality images captured by the underwater high-definition camera under the supplementary lighting of the supplementary lighting group, forming a multi-dimensional data packet containing the current collection time, location coordinates, water quality parameters and water quality images. The water quality parameter sensor is fixed to the bottom of the underwater robot, and the underwater high-definition camera and high-brightness supplementary lighting group are fixed on the underwater robot. S3. Weight the current water quality parameters to obtain physical indicators, and use a deep learning model to perform pollutant identification and analysis on the current water quality image to obtain image features. S4. Calculate image indicators based on current image features. Determine the pollution status based on physical and image indicators. If physical and image indicators indicate no pollution, continue along the pre-planned path. If physical and image indicators indicate low pollution, fuse physical and image indicators with image features to obtain the pollution intensity at the current location. If physical and image indicators indicate high pollution, fuse physical and image indicators with image features to obtain the pollution intensity at the current location and enter a local dense sampling mode. The underwater robot decelerates and moves near the current location, executing S2 and S3 to obtain high-density physical and image indicators and image features at different locations. Fuse physical and image features at different locations to obtain the corresponding pollution intensity, then update the pollution distribution heat map. Based on the gradient changes in the pollution distribution heat map, use an improved concentration gradient tracking algorithm to guide the underwater robot to move along the gradient upward direction, dynamically adjusting the underwater robot's path until the pollution source is tracked.

Images