A navigation method for underwater unmanned submarine vehicle

Abstract

This invention relates to a navigation method for underwater unmanned vehicles (UUVs), comprising the following steps: acquiring data from multiple sensors and performing spatiotemporal synchronization, feature extraction, and data fusion preprocessing; defining the underwater operating area as a directed graph, with edge weights determined based on terrain complexity, obstacle risk, and water flow influence; based on the directed graph, employing an improved A* algorithm for global path planning and a fast exploratory random tree algorithm for local path planning and dynamic obstacle avoidance; assessing the environmental state considering aquatic environmental parameters, water flow characteristics, and topographic changes; and adjusting the navigation strategy using a deep reinforcement learning algorithm, with the navigation strategy adjustment objective function as the optimization goal, based on the environmental assessment results and the UUV's own state. Compared with existing technologies, this invention offers advantages such as improved navigation accuracy, reliability, and adaptability of underwater UUVs in complex underwater environments.

Description

Technical Field

[0001] This invention relates to the field of intelligent underwater equipment, and in particular to a navigation method for unmanned underwater vehicles. Background Technology

[0002] Unmanned Underwater Vehicles (UUVs) play a crucial role in marine resource exploration, environmental monitoring, and military reconnaissance. The accuracy and reliability of their navigation directly impact the success of missions and the safety of the UUV itself. However, the highly complex and unique underwater environment presents significant challenges to UUV navigation, rendering traditional navigation technologies plagued by numerous insurmountable problems. Early, widely used inertial navigation systems primarily relied on inertial devices such as gyroscopes and accelerometers to measure the UUV's motion parameters and calculate its position and attitude through integration. However, over time, errors from these inertial devices accumulate, leading to a sharp decline in navigation accuracy. In long-duration, long-distance underwater operations, this error accumulation can cause the UUV to deviate from its planned route by thousands of meters or even further, failing to accurately reach the target area. Moreover, inertial navigation systems are susceptible to underwater currents and the UUV's own maneuvers, further exacerbating errors.

[0003] Satellite navigation systems provide high-precision positioning services above the water surface, but are virtually unusable underwater. This is because satellite signals are severely attenuated in water, resulting in a very short propagation distance and an inability to penetrate deep enough to provide positioning information for UUVs. Therefore, UUVs operating underwater must rely on other navigation methods. Underwater acoustic positioning and navigation technology is one of the commonly used methods for underwater navigation. It uses underwater sonar equipment to transmit and receive sound wave signals, utilizing the propagation characteristics of sound waves in water to achieve positioning. For example, long baseline (LBL), short baseline (SBL), and ultra-short baseline (USBL) positioning systems require the deployment of multiple positioning beacons as reference points underwater. However, these beacons are not only costly to install and maintain, but also complex to deploy and have stringent environmental requirements. Furthermore, the speed of sound propagation in water is affected by various factors such as water temperature, salinity, and depth, leading to significant positioning errors. In addition, complex underwater topography and environmental noise can interfere with the propagation of sound waves, reducing the accuracy and reliability of positioning. Visual navigation technology uses underwater cameras to acquire image information for navigation. However, the underwater environment is dimly lit, and the water contains a large number of suspended particles and impurities, which significantly reduces image clarity and contrast, posing great difficulties for image recognition and processing. Moreover, visual navigation has extremely high requirements for computing resources, and when processing large amounts of image data in real time, computational delays often occur, affecting the UUV's timely response capabilities.

[0004] CN116954225A discloses a system and method for obstacle avoidance of underwater vehicles (UVs) in urban pipeline environments based on multibeam sonar. Based on sensor data from IMU, DVL, cameras, and ten-beam sonar, a multi-beam fusion model is established. The model is optimized to obtain the UV's attitude information, real-time velocity, image information of the pipeline ahead, and the UV's position and heading within the pipeline environment. Based on the image information of the pipeline ahead, the system detects and identifies the position, shape, and size of the pipeline and obstacles, generating an environmental map. Based on the environmental map, the UV's attitude information, real-time velocity, ranging information from six sets of sonar, and the UV's position and heading within the pipeline environment, a safe path to avoid the pipeline and obstacles is planned. The UV dynamically adjusts its attitude and control strategy according to the safe path to achieve trajectory tracking. However, this method is only applicable to urban pipeline environments, which are relatively simple. In complex environments such as the ocean, due to the complexity of terrain and environmental factors, directly planning UUV navigation without considering the influence of water depth, complex topography, and environmental noise is difficult to adapt to actual usage scenarios. Furthermore, with the increasing demands of marine development and applications, UUV operations are becoming increasingly complex, placing higher demands on the autonomy, adaptability, and intelligence of navigation systems. Traditional navigation technologies, due to their inherent limitations, are no longer sufficient to meet the needs of modern underwater operations. Therefore, there is an urgent need to research and develop a more advanced and intelligent navigation method to improve the navigation performance of UUVs in complex underwater environments. Summary of the Invention

[0005] The purpose of this invention is to provide a navigation method for underwater unmanned vehicles (UUVs) to improve their navigation accuracy, autonomy, and adaptability in complex underwater environments.

[0006] The objective of this invention can be achieved through the following technical solutions:

[0007] A navigation method for underwater unmanned submersibles includes the following steps:

[0008] Acquire motion state data, multibeam sonar data, underwater vision data, and aquatic environment data of the underwater unmanned submersible collected by various sensors at their respective frequencies;

[0009] The data collected by each sensor are preprocessed by spatiotemporal synchronization, feature extraction, and data fusion.

[0010] The underwater operation area is defined as a directed graph, where the nodes of the directed graph represent locations with specific significance, the edges represent feasible paths between nodes, and the edge weights are determined based on terrain complexity, obstacle risk, and water flow impact.

[0011] Based on the constructed directed graph, an improved A* algorithm is used for global path planning. The improved A* algorithm introduces correction coefficients based on terrain and environment modifications in the cost estimation.

[0012] A fast exploratory random tree algorithm is used for local path planning and dynamic obstacle avoidance. The fast exploratory random tree algorithm adds obstacle risk to the distance metric function between nodes and uses a risk assessment function to evaluate whether to add a new node to the tree.

[0013] The environmental status is assessed by considering water body environmental parameters, water flow characteristics, and changes in topography.

[0014] Based on the environmental assessment results and the underwater unmanned vehicle's own state, the navigation strategy is adjusted using a deep reinforcement learning algorithm, with the navigation strategy adjustment objective function as the optimization goal, to control the underwater unmanned vehicle to travel along the planned path.

[0015] The motion data of the underwater unmanned vehicle is acquired through a high-precision inertial measurement unit, including three-dimensional acceleration data and three-dimensional angular velocity data; the multibeam sonar data is acquired through multibeam sonar at a fixed scanning frequency, and the set of point cloud data acquired in one scan is S. t ={s t1 ,s t2 ,…,s tn}, where s ti =(x ti ,y ti ,z ti ,r ti ), representing the i-th point cloud data collected at time t, (x ti ,y ti ,z ti ) represents spatial coordinates, r ti The echo intensity is n, and the number of point cloud data is n. The underwater visual data is acquired by an underwater camera. The water environment data is acquired by a temperature, salinity, and depth sensor, including temperature, salinity, and depth data.

[0016] The spatiotemporal synchronization specifically includes:

[0017] Time synchronization: Using the high-precision inertial measurement unit as a reference, the data from other sensors are synchronized to the acquisition time of the high-precision inertial measurement unit;

[0018] Spatial alignment: A unified world coordinate system is established, and the data from each sensor is transformed into this coordinate system. Specifically, for motion state data, the motion state data is transformed into the world coordinate system through rotation and translation matrices based on the installation position and attitude parameters of the high-precision inertial measurement unit. For multibeam sonar data and underwater vision data, the transformation relationship is constructed using calibration parameters to achieve spatial alignment.

[0019] The calculation of the edge weights in the directed graph includes the following steps:

[0020] Calculating terrain complexity: Constructing a terrain model based on multibeam sonar data and calculating path e. ij The terrain undulations of the area traversed are given by σ, where the standard deviation of terrain elevation is σ. h If the terrain complexity weight coefficient is α1, then the contribution of terrain complexity to the edge weight is α1·σ. h ;

[0021] Calculate obstacle risk: Based on obstacle distribution information obtained from multi-source data fusion, calculate path e. ij The proximity to the obstacle, let path e be... ij The distance to the nearest obstacle is d obstacle If the obstacle risk weight coefficient is α2, then the contribution of obstacle risk to the edge weight is...

[0022] Calculate the impact of water flow: Using water environment data and water flow models, calculate path e ij Water flow speed The angle θ between the underwater unmanned vehicle and its desired direction of travel, and the magnitude of the water flow velocity. Let the weighting coefficient for the influence of water flow be α3, then the contribution of the influence of water flow to the edge weights is:

[0023] Taking into account terrain complexity, obstacle risk, and water flow impact, determine edge e. ij weight w ij for:

[0024]

[0025] Among them, α1, α2, and α3 are weight coefficients obtained during training, used to balance the influence of different factors on path selection.

[0026] The estimated cost is expressed as:

[0027] h(v)=β·d euclidean (v,v goal )

[0028] Where h(v) is the estimation cost, β is the correction coefficient determined through training based on terrain and environment modifications, and d euclidean (v,v goal () represents the distance from node v to target node v goal Euclidean distance.

[0029] The distance metric function is expressed as:

[0030]

[0031] Among them, (x q ,y q ,z q Let (x) be the position coordinates of node q, and (x) be the position coordinates of node q. rand ,y rand ,z rand ) represents the random sampling node q rand The location coordinates, γ is the obstacle risk weight coefficient, ObstacleRisk(q,q) rand ) represents the distance from node q to q rand The obstacle risk on the path is calculated based on obstacle information obtained from the fusion of multi-source data.

[0032] The risk assessment function is:

[0033]

[0034] Where O represents the set of obstacles, δ(o) is the danger coefficient of obstacle o, which is determined according to the type and size of the obstacle; d(q new ,o) is the new node q new The distance to obstacle o; only if the risk assessment function R(q) new Less than the set threshold R th At that time, the new node q new Add to the tree.

[0035] The environmental status assessment includes water body environmental parameter assessment, water flow characteristic assessment, and topographic and geomorphological change assessment. The water body environmental parameter assessment obtains the water body environmental status assessment function by weighted summation of temperature, salinity, depth, and water turbidity. The water flow characteristic assessment is based on the obtained water flow velocity. The angle θ between the desired direction of travel of the underwater unmanned vehicle and the direction of the water flow determines the water flow characteristic evaluation function. Wherein, α1 and α2 are weight coefficients obtained through training; the terrain and landform change assessment is based on multibeam sonar data to construct an underwater terrain model, calculate the terrain slope change rate and terrain complexity index on the underwater unmanned vehicle's navigation path, and obtain a weighted terrain and landform assessment function; the environmental assessment result is obtained by weighted summation of the water body environmental parameter assessment function, the water flow characteristic assessment function, and the terrain and landform change assessment function.

[0036] The objective function J for adjusting the navigation strategy is:

[0037]

[0038] Among them, Distance(q) cur ,q goal) represents the current position q of the underwater unmanned vehicle. cur to target position q goal The distance, E is the battery power of the underwater unmanned vehicle, S i For the working state of the i-th underwater unmanned vehicle, F env For the environmental assessment results, ω1, ω2, ω3, and ω4 are weighting coefficients.

[0039] The state space of the deep reinforcement learning algorithm is defined as: based on the environment evaluation result F. env Location of the underwater unmanned submersible (UUV) q cur Power E, Device operating status vector The state vector The action space is defined as: adjusting speed, changing course, and adjusting depth; the reward function is defined as: R = R1 + R2 + R3, where, if the decrease in distance to the target position of the underwater unmanned vehicle within one time step is Δd, then the reward R1 = μ1·Δd; if a device malfunction is detected, then R2 = -μ2; if the environmental assessment result F env Exceeding threshold F th Then the reward R3 = -μ3·(F env -F th μ1, μ2, and μ3 are reward coefficients.

[0040] Compared with the prior art, the present invention has the following beneficial effects:

[0041] (1) Improve the accuracy and comprehensiveness of underwater environment perception.

[0042] This invention integrates multiple sensors on a UUV to construct a multi-source data fusion system, enabling more comprehensive and accurate perception of the underwater environment compared to traditional single-sensor navigation methods. In the data acquisition phase, multiple sensors, including a high-precision inertial measurement unit, multi-beam sonar, and underwater camera, work collaboratively to acquire multi-dimensional data such as motion status, terrain features, and visual images. In the data fusion and feature extraction phase, time synchronization, spatial alignment, and deep learning-based feature extraction algorithms are employed, along with an attention mechanism to achieve deep data fusion, overcoming interference from low underwater lighting and complex sound wave propagation.

[0043] (2) Optimize the efficiency and reliability of path planning

[0044] In terms of global path planning, this invention abstracts the underwater operation area into a directed graph and comprehensively considers factors such as terrain complexity, obstacle risk, and water flow influence to calculate path weights. It utilizes an improved A* algorithm for path planning, which, compared to traditional path planning algorithms, better adapts to complex underwater environments. In local path planning and dynamic obstacle avoidance, based on a variant of the RRT algorithm, and combined with a distance metric function and risk assessment function that consider underwater environmental factors, the UUV can avoid collisions when encountering sudden obstacles, ensuring the continuity and reliability of UUV operations.

[0045] (3) Enhance the adaptive capability of navigation decisions

[0046] This invention constructs a multi-dimensional environmental state assessment system, comprehensively considering factors such as aquatic environmental parameters, water flow characteristics, and topographic changes. It utilizes quantitative and comprehensive assessment functions to accurately quantify and analyze the underwater environmental state. Regarding navigation strategy adjustment, based on a deep reinforcement learning algorithm, guided by optimizing the objective function, it learns the optimal navigation strategy for different environments and its own state. Compared to traditional navigation decision-making methods, when the underwater environment changes (such as sudden changes in water flow velocity or complex terrain), the UUV can automatically adjust its speed, course, and depth, enhancing its autonomous adaptability in complex and variable underwater environments.

[0047] (4) Promote the intelligent development of underwater operations

[0048] This invention deeply applies deep learning technology to the field of UUV navigation, forming a complete intelligent system from environmental perception and path planning to navigation decision-making. Compared with traditional navigation technologies, this method reduces human intervention, lowers operating costs, and improves operational efficiency and safety, powerfully promoting the development of underwater operations towards intelligence and efficiency. Attached Figure Description

[0049] Figure 1 This is a flowchart of the method of the present invention. Detailed Implementation

[0050] The present invention will now be described in detail with reference to the accompanying drawings and specific embodiments. These embodiments are based on the technical solution of the present invention and provide detailed implementation methods and specific operating procedures. However, the scope of protection of the present invention is not limited to the following embodiments.

[0051] This embodiment provides a navigation method for underwater unmanned submersibles, such as... Figure 1 As shown, it includes the following steps:

[0052] S1 acquires motion state data, multibeam sonar data, underwater vision data, and aquatic environment data collected by various sensors at their respective frequencies.

[0053] To acquire sensor data, the first step is to deploy the sensor hardware, including the modification of the unmanned underwater vehicle (UUV), the installation of communication equipment, and the installation of a sensor calibration and data acquisition system. The UUV modification includes sensor layout planning, structural reinforcement and waterproofing, and computing unit integration; the communication equipment installation covers the deployment of underwater acoustic communication modules and the coordination of local positioning beacons and navigation buoys; and the sensor calibration and data acquisition system installation includes the sensor calibration process and the construction of the data acquisition system.

[0054] In one embodiment, the sensor layout is specifically planned as follows: a high-precision inertial measurement unit is installed near the center of mass of the UUV to reduce motion attitude measurement errors; a multi-beam sonar is installed below the front of the UUV to ensure that its beam can cover a large underwater area in front and on both sides; an underwater camera is arranged in a wide field of view at the head of the UUV and equipped with an adjustable angle bracket, and a high-intensity waterproof LED supplementary light is installed; a temperature, salinity, and depth sensor (CTD) is installed in a stable water flow area at the bottom of the UUV to avoid the measurement accuracy being affected by violent water flow disturbances.

[0055] In one embodiment, the structural reinforcement and waterproofing treatment specifically involves: strengthening the UUV shell using titanium alloy material with a yield strength not lower than a specific standard; and ensuring the shell's structural strength meets underwater operating pressure requirements through a combination of welding and machining. All sensor interfaces and equipment hatches are sealed with double sealing rings and coated with waterproof adhesive. Furthermore, a separate waterproof sealed chamber filled with dry gas is installed inside the UUV to house the core electronic equipment, maintaining a dry environment and preventing moisture damage to electronic components.

[0056] In one embodiment, the computing unit integration specifically involves selecting a high-performance embedded computing platform, preferably using an NVIDIA Jetson AGX Orin as the core processor, paired with 16GB of memory and a 256GB solid-state drive. To meet the heat dissipation requirements during algorithm operation, custom heat sinks are installed on the surface of the computing unit, and a micro fan is equipped. An intelligent temperature control system adjusts the fan speed to ensure the computing unit's operating temperature is maintained between 40℃ and 60℃. Simultaneously, a dedicated power management module is designed to provide stable power to the computing unit, featuring overvoltage and overcurrent protection functions to ensure safe device operation.

[0057] In one embodiment, the underwater acoustic communication module is deployed as follows: an underwater acoustic communication device based on spread spectrum technology is selected, with an operating frequency range of 10kHz-50kHz. When the communication distance is within one kilometer, the transmitting transducer of the underwater acoustic communication device is installed at the center of the bottom of the UUV, and the receiving transducers are symmetrically distributed on both sides of the transmitting transducer. These transducers are connected to the internal communication control unit of the UUV via a dedicated cable. To improve the communication's anti-interference capability, channel coding technology, such as low-density parity-check code (LDPC), is used in the communication link.

[0058] In one embodiment, the coordination between local positioning beacons and navigation buoys is as follows: Within the underwater operating area, a local positioning beacon is deployed approximately every 500 meters. Each beacon has a built-in high-precision positioning chip. The UUV interacts with the positioning beacons via underwater acoustic communication, receiving location information transmitted by the beacons. Simultaneously, navigation buoys are deployed on the surface. These buoys are equipped with GPS positioning modules and wireless communication modules. After establishing contact with the UUV via underwater acoustic communication, they transmit the UUV's location information to the shore-based control center in real time. The navigation buoys also have relay communication capabilities; when the distance between the UUV and the shore-based control center is relatively far, they can act as signal relay stations to ensure uninterrupted communication links.

[0059] In one embodiment, the sensor calibration process specifically involves: rigorously calibrating all sensors before the UUV is launched into the water. For the IMU, a high-precision turntable is used for attitude calibration. By collecting data at different angles, an error compensation model is established to control the zero-bias error of the accelerometer within ±0.01 m / s². 2 Within ±0.05° / h, the gyroscope drift error is controlled within ±0.05° / h; the multibeam sonar is calibrated using a standard depth target, and the sonar parameters are adjusted under different water depth conditions to ensure that the water depth measurement error does not exceed ±1%; the underwater camera is calibrated for internal and external parameters using a checkerboard calibration plate to correct lens distortion and make the image distortion rate less than 1%; the CTD sensor is compared and calibrated with a standard temperature, salinity, and depth measuring instrument to control the temperature measurement error within ±0.1℃, the salinity measurement error within ±0.01℃, and the depth measurement error within ±0.5%.

[0060] In one embodiment, the data acquisition system is constructed as follows: a high-speed data acquisition card is designed, connected to the computing unit via a PCIe interface, and supports multi-channel synchronous acquisition. An independent data acquisition channel is allocated to each sensor, and signal conditioning circuitry is integrated on the acquisition card to amplify and filter the analog signals output by the sensors. Dedicated data acquisition software is developed to achieve real-time acquisition, storage, and preliminary processing of data from each sensor. The software employs multi-threading technology to ensure the real-time performance and accuracy of data acquisition from different sensors, and stores the acquired data in binary file format on a solid-state drive for subsequent analysis and processing.

[0061] After completing the above steps, the high-precision inertial measurement unit (IMU) operates at a frequency f. IMU Data is collected, and the three-dimensional acceleration data collected at time t is denoted as... Three-dimensional angular velocity data are denoted as Used to acquire real-time motion status information of UUVs. Multibeam sonar uses a scanning frequency f... sonar The work involves collecting a point cloud dataset S in a single scan. t ={s t1 ,s t2 ,…,s tn}, where s ti =(x ti ,y ti ,z ti ,r ti ), representing the i-th point cloud data collected at time t, (x ti ,y ti ,z ti ) represents spatial coordinates, r ti The echo intensity is used to construct underwater terrain models, and n is the number of point cloud data points. The underwater camera operates at a frame rate f. cam Image acquisition, acquired image I t It is an H×W×3 matrix, where H is the image height, W is the image width, and 3 represents the three RGB color channels, used to capture underwater visual information. The CTD sensor operates at a sampling frequency f. CTD Data was collected, including the temperature T collected at time t. t Salinity S t Depth D t Data is used to reflect the characteristics of the aquatic environment.

[0062] S2 performs spatiotemporal synchronization, feature extraction, and data fusion preprocessing on the data collected by each sensor.

[0063] Data acquisition operations were conducted in various underwater environments (such as shallow sea, deep sea, coral reef areas, and straits). The sensors on the UUV continuously collected no less than one thousand hours of raw data. After acquisition, the data underwent preliminary preprocessing. First, median filtering was used to remove noise from the sensor data. For sonar point cloud data, voxel filtering was used for downsampling to control the point cloud data density within an appropriate range. For image data, enhancement processing such as grayscale transformation and histogram equalization was performed to improve image clarity. Then, the data was labeled. Professional labeling tools were used to label the sonar point cloud data with obstacle types and location information, and the image data with target object categories. After labeling, the data was divided into training, validation, and test sets in an 8:1:1 ratio. Then, the following steps were performed:

[0064] 1) Spatiotemporal synchronization

[0065] Time synchronization: Since the sensors have different acquisition frequencies, time synchronization is performed first. Using the high-precision inertial measurement unit (IMU) as a reference, the data from other sensors are synchronized to the acquisition time of the high-precision inertial measurement unit.

[0066] In one embodiment, multibeam sonar data S t For example, the data is synchronized to the IMU acquisition time using linear interpolation. Assume that during the time interval from t to t+Δt, the multibeam sonar acquires data S. t and S t+Δt For IMU time t' (t < t' < t + Δt), the synchronized data S t′ Each point s in t′i The coordinates are calculated as follows:

[0067]

[0068] Other sensor data also need to be time-synchronized. All data are based on the acquisition time of the high-precision inertial measurement unit and linear interpolation or a more suitable synchronization algorithm is used to ensure that all sensor data are consistent in the time dimension, thereby providing an accurate and reliable time alignment basis for subsequent feature extraction and data fusion.

[0069] Spatial alignment: Establish a unified world coordinate system and transform the data from each sensor to this coordinate system.

[0070] For IMU data, the measurement data is transformed to the world coordinate system using rotation and translation matrices based on its installation location and attitude parameters; for multibeam sonar data and underwater vision data, the transformation relationship is constructed using calibration parameters to achieve spatial alignment.

[0071] 2) Feature extraction: Deep learning networks are used for feature extraction.

[0072] For multibeam sonar data S t The PointNet++ network is used for processing. Let the feature vector obtained after multi-layer feature extraction be... The calculation process is as follows: In the first layer of feature extraction, for each point s ti Local features are extracted using a multilayer perceptron (MLP). Right now Then, the global features are obtained by aggregating local features through max pooling. After multiple similar operations, the feature vector is finally obtained.

[0073] For underwater visual data, i.e., underwater image I tFeature extraction is performed using a ResNet network. Let F be the feature map obtained after network processing. cam In the convolutional layers of the network, for the l-th convolutional operation, the output feature map is... The calculation formula is: Where σ is the activation function (such as ReLU), B l As a bias term, W lk K represents the kernel weights. l This is the set of feature maps from the previous layer associated with the l-th convolutional layer. After multiple convolutional and pooling operations, the feature map F is obtained. cam Then, the feature vector is obtained through global average pooling.

[0074] For IMU data and and CTD data T t S t D t Combine them into a vector Feature extraction is performed using a fully connected layer to obtain the feature vector.

[0075] 3) Data fusion

[0076] In this embodiment, an attention-based fusion method is used to construct an attention weight matrix A. For the feature vector... Calculate attention weight a sonar a cam a other The calculation method is as follows:

[0077]

[0078] in, The query vector is obtained through training. The fused feature vector is... The calculation is as follows:

[0079] This fused feature vector contains multifaceted information about the underwater environment. Essentially, it is a semantic-level integration product of multi-source data after spatiotemporal synchronization and feature extraction. By eliminating sensor heterogeneity (such as spatiotemporal bias and dimensional differences), it forms standardized input, enhancing feature complementarity while suppressing single-source noise. Although subsequent tasks focus on single-type data applications (such as terrain assessment), the fused vector has already provided a foundation for spatiotemporal registration (such as unified geographic coordinates), multi-source confidence indicators (such as sensor confidence), and cross-modal feature enhancement (such as multispectral and point cloud fusion to improve terrain classification accuracy). This preprocessing is not redundant, but a necessary step to avoid data fragmentation and ensure robustness in complex scene analysis. Its value is implicit in the entire process from data acquisition to decision-making, providing key data support for subsequent underwater environment modeling and navigation decisions.

[0080] S3 defines the underwater operation area as a directed graph.

[0081] The underwater operation area is abstracted as a directed graph G = (V, E), where the set of nodes V = {v1, v2, ..., v...} n} represents a location underwater with specific significance, such as a target detection point, resource collection point, or predetermined charging point; the edge set E = {e ij} represents node v i With v j There are feasible paths between each edge e. ij All are assigned weights w ij This weight comprehensively reflects the impact of underwater environmental factors on path passage. Weight w ij The calculation takes into account the following factors:

[0082] Terrain complexity: A terrain model is constructed based on multibeam sonar data, and path e is calculated. ij The terrain undulations of the area traversed are given by σ, where the standard deviation of terrain elevation is σ. h If the terrain complexity weight coefficient is α1, then the contribution of terrain complexity to the edge weight is α1·σ. h .

[0083] Obstacle risk: Based on obstacle distribution information obtained from multi-source data fusion, statistical path e is calculated. ij The proximity to the obstacle, let path e be... ij The distance to the nearest obstacle is d obstacle If the obstacle risk weight coefficient is α2, then the contribution of obstacle risk to the edge weight is...

[0084] Impact of water flow: Calculate path e using water environment data and water flow models. ij Water flow speed The angle θ between the underwater unmanned vehicle and its desired direction of travel, and the magnitude of the water flow velocity. Let the weighting coefficient for the influence of water flow be α3, then the contribution of the influence of water flow to the edge weights is:

[0085] Taking into account terrain complexity, obstacle risk, and water flow impact, determine edge e. ij weight w ij for:

[0086]

[0087] Among them, α1, α2, and α3 are weight coefficients obtained through training with a large amount of underwater experimental data, which are used to balance the influence of different factors on path selection.

[0088] S4 uses an improved A* algorithm for global path planning based on the constructed directed graph.

[0089] The evaluation function f(v) for node v is defined as: f(v) = g(v) + h(v). Here, g(v) represents the evaluation function from the initial node v. start The actual cost to node v is obtained by summing the weights of each edge on the path, i.e.

[0090]

[0091] Where, path(v start ,v) indicates starting from the initial node v start The path to node v; h(v) is the path from node v to the target node v. goal The estimated cost incorporates correction coefficients based on terrain and environment modifications. Specifically, it employs a heuristic function based on an underwater environment model, such as utilizing the Euclidean distance d between nodes. euclidean (v,v goal And by taking into account factors such as topography and water flow, we make corrections. Let the correction coefficient be β, then h(v) = β·d euclidean (v,v goal ), and β is also determined through training.

[0092] S5 employs a fast exploratory random tree algorithm for local path planning and dynamic obstacle avoidance.

[0093] During UUV operation, a variant of the Fast Exploratory Random Tree (RRT) algorithm is used for local path planning and dynamic obstacle avoidance. In this embodiment, the RRT algorithm adds obstacle risk to the distance metric function between nodes and uses a risk assessment function to evaluate whether to add a new node to the tree.

[0094] With the current position q of the UUV cur Construct a random tree T for the root node. In each iteration, randomly sample a point q from the workspace W. randFind the distance q in tree T rand The nearest node q near Define node q to q rand Distance metric function d(q,q) rand Considering underwater environmental factors, the following formula is used:

[0095]

[0096] Among them, (x q ,y q ,z q ) and (x rand ,y rand ,z rand ) are nodes q and q respectively. rand The coordinates; γ is the obstacle risk weight coefficient; ObstacleRisk(q,q rand ) represents the distance from node q to q rand The risk of obstacles on the path is calculated based on obstacle information obtained from the fusion of multi-source data, such as by comprehensively assessing factors such as the number, size, and proximity of obstacles on the path.

[0097] Calculated from q by the local controller near To the new node q new The direction and distance of q's motion make q new Approaching q within a certain range rand And it satisfies the kinematic constraints of the UUV. A risk assessment function R(q) is introduced. new Evaluate the new node:

[0098]

[0099] Where O represents the set of obstacles, δ(o) is the danger coefficient of obstacle o, which is determined according to the type (e.g., fishing nets, shipwrecks, etc.) and size of the obstacle; d(q new ,o) is the new node q new The distance to obstacle o. Only when R(q) new Less than the set threshold R th Only then will the new node q be... new Add tree T. When a sudden obstacle or environmental change is detected, the local path planning algorithm is restarted, using the current UUV position as the new starting point, to quickly generate a new path that avoids the obstacle. Simultaneously, based on the new path and the UUV's current state, subsequent deep reinforcement learning algorithms are used to calculate reasonable speed and attitude adjustment commands, controlling the UUV to safely pass through hazardous areas and ensuring the continuity of the operation.

[0100] S6 assesses the environmental status by considering water body environmental parameters, water flow characteristics, and changes in topography.

[0101] To comprehensively assess the underwater environment, a multi-dimensional assessment system is constructed, taking into account factors such as water environment parameters, water flow characteristics, and changes in topography.

[0102] Aquatic environmental parameter assessment: Using temperature (T), salinity (S), and depth (D) data collected by a temperature, salinity, and depth sensor (CTD), and turbidity information obtained by an underwater camera (calculated through image analysis algorithms, denoted as Turb), an aquatic environmental state vector is constructed. To quantify the impact of each parameter on navigation, a weight vector is introduced. Where w T w S w D w Turb The weighting coefficients for temperature, salinity, depth, and turbidity were determined through training with extensive underwater experimental data. A water environment state assessment function was defined. for:

[0103]

[0104] This function value reflects the overall impact of the aquatic environment on UUV navigation; the larger the value, the greater the impact of the environment on navigation.

[0105] Water flow characteristics assessment: Water flow velocity is obtained based on temperature, salinity, and depth sensor data and water flow models. and direction of water flow (using unit vectors) (Representation). Calculate the magnitude of the water flow velocity. and the UUV's desired driving direction vector Vector of water flow direction The included angle Define the water flow characteristic evaluation function

[0106]

[0107] Here, α1 and α2 are weighting coefficients obtained through training, used to balance the influence of water flow velocity and direction on navigation, respectively. The larger the value of this function, the stronger the interference of water flow on UUV navigation.

[0108] Topographic and geomorphological change assessment: An underwater topographic model was constructed based on multibeam sonar data, and the rate of change of topographic slope r along the UUV's navigation path was calculated. slope and terrain complexity index C terrain (Calculated using the standard deviation of terrain height). Define the terrain evaluation function f. terrain (r slope C terrain ):

[0109] fterrain (r slope C terrain )=β1·r slope +β2·C terrain

[0110] β1 and β2 are weighting coefficients, determined through training. The function value reflects the degree of influence of topographic changes on UUV navigation.

[0111] Based on the above three aspects, a comprehensive evaluation function for the underwater environment state is defined as follows:

[0112]

[0113] Among them, λ1, λ2, and λ3 are weighting coefficients, which are trained and optimized through a large amount of actual underwater operation data to determine the relative importance of each evaluation function in the comprehensive evaluation.

[0114] S7, based on environmental assessment results and the underwater unmanned vehicle's own state, uses a navigation strategy adjustment objective function as the optimization goal, and adjusts the navigation strategy through a deep reinforcement learning algorithm to control the underwater unmanned vehicle to travel along the planned path.

[0115] Based on the comprehensive environmental status assessment results F env And the UUV's own status (such as UUV power E, device operating status vector) Where S i The i-th device represents its operating status (1 for normal, 0 for fault), and m is the number of devices. The navigation strategy adjustment objective function J is defined as:

[0116]

[0117] Among them, Distance(q) cur ,q goal ) represents the current position q of the underwater unmanned vehicle. cur to target position q goal The distance, ω1, ω2, ω3, and ω4 are weighting coefficients, determined through training, and are used to balance the impact of distance, battery power, device status, and environmental status on the navigation strategy.

[0118] In this embodiment, a deep reinforcement learning algorithm (such as Deep Q-Network (DQN) or Proximal Policy Optimization (PPO)) is used to learn the optimal navigation strategy under different environments and the user's own state, with the objective function J as the optimization target. The specific process is as follows:

[0119] State-space definition: The environmental assessment results F... env UUV position q cur Power E, Device operating status vector Information is combined into a state vector As the state input for reinforcement learning.

[0120] Action space definition: Define the action space A of the UUV, including operations such as adjusting the driving speed (e.g., acceleration, deceleration, constant speed), changing the heading (left turn, right turn), and adjusting the depth (ascending, descending, maintaining).

[0121] Reward Function Design: Design a reward function R. A positive reward is given when the UUV moves towards the target location, the environmental condition is stable, and the equipment is working normally; a negative reward is given when obstacles are encountered, the environment deteriorates, equipment malfunctions, or the direction of deviation from the target is encountered. For example, if the decrease in distance from the target location by the UUV within one time step is Δd, then the reward R1 = μ1·Δd; if equipment malfunction is detected, the reward R2 = -μ2; if the environmental assessment result F... env Exceeding the environmental state threshold F th The reward is R3 = -μ3·(F) env -F th ), where μ1, μ2, and μ3 are reward coefficients. The total reward R = R1 + R2 + R3.

[0122] Policy learning: During the interaction between an agent and its environment, it learns policies based on the current state. Choose action a∈A, execute the action, and observe the new state. The model parameters are updated using a reinforcement learning algorithm, along with the obtained reward R, to gradually learn the optimal navigation strategy that minimizes the objective function J (or maximizes the reward), thereby achieving adaptive navigation of the UUV in complex underwater environments.

[0123] In this embodiment, a Transformer-UNet hybrid architecture is selected to construct the algorithm model of the entire system, and training is performed under the PyTorch deep learning framework. During training, the Adam optimizer is used, with an initial learning rate set to 0.001, and the learning rate decays by 0.9 every 10 epochs. To prevent overfitting, a Dropout layer is added to the model, with the Dropout probability set to 0.2. During training, the training set data is input into the model in batches of 32. After forward propagation, the prediction results are calculated, and the total loss is calculated by summing the cross-entropy loss function and the mean squared error loss function. The model parameters are then updated using the backpropagation algorithm. Performance is evaluated every 5 epochs on the validation set. Training is stopped and the optimal model parameters are saved when the loss on the validation set no longer decreases for 10 consecutive epochs.

[0124] Model Deployment Scheme: The trained model is converted into a format suitable for embedded platforms using cross-compilation tools and deployed to the computing units of the UUV. Simultaneously, the same model is deployed on the shore-based control center server for secondary analysis and model updates of the data uploaded by the UUV. A model update mechanism is established: when the UUV collects new representative data during operation, the data is uploaded to the shore-based server. The server uses this data to incrementally train the model. After training, the updated model parameters are distributed to the UUV, achieving dynamic model optimization.

[0125] During implementation, the overall algorithm underwent performance optimization using code optimization techniques such as loop unrolling and function inlining to reduce function call overhead. Data storage was also optimized using a high-efficiency data compression algorithm to compress and store sensor-collected data at a compression ratio of over 5:1, saving storage space. Simultaneously, software memory management was optimized using intelligent memory allocation strategies to avoid memory leaks and fragmentation, ensuring the software's stability and reliability during long-term operation.

[0126] Based on the above methods, the following algorithm performance tests and optimizations were performed:

[0127] 1. Simulated environment testing

[0128] Test Environment Setup: A simulated underwater environment test tank was built in the laboratory. Water pumps and wave generators were used to simulate different water flow speeds and wave conditions. Various simulated obstacles (such as simulated reefs, shipwreck models, and fishing nets) were placed in the tank, and a high-precision positioning system was installed to monitor the UUV's position and attitude in real time, with a positioning accuracy of ±1 cm. Simultaneously, an environmental parameter simulation device was set up, allowing adjustment of water temperature, salinity, turbidity, and other water environment parameters to simulate different underwater environmental scenarios.

[0129] Functional testing process: First, the UUV's sensor functions are tested to check the accuracy and stability of sensor data acquisition. By comparing data with standard measuring equipment, the sensor measurement error is verified to be within the allowable range. Next, the path planning function is tested by setting different starting and ending points to observe whether the UUV can plan a reasonable path, and the path planning time and path length are recorded. Then, dynamic obstacle avoidance testing is conducted. Obstacles are randomly placed during UUV operation to test the obstacle avoidance reaction time and success rate. Finally, the navigation decision-making function is tested by changing simulated environmental parameters to observe whether the UUV can adaptively adjust its navigation strategy according to environmental changes. Problems discovered during testing are promptly recorded and analyzed for their causes, and targeted modifications are made to the software code or adjustments are made to the hardware parameters.

[0130] 2. Water tank test

[0131] Test Plan Design: A large outdoor water tank was selected for field testing to simulate a more realistic underwater environment. Different test areas were divided within the tank, including shallow water, deep water, and narrow passage areas, and various types of realistic obstacles (such as abandoned ships and concrete blocks) were set up. The UUV underwent continuous 24-hour operational testing in the tank, performing simulated tasks such as marine resource exploration and environmental monitoring. During the test, the shore-based control center monitored the UUV's operational status in real time, recording sensor data, navigation trajectory, and task execution information.

[0132] Performance Optimization and Adjustment: Based on the results of the pool test, the navigation system of the UUV was optimized. If unreasonable path planning was found, the parameters of the path planning algorithm (such as weight coefficients) were readjusted, or the algorithm structure was optimized; if the obstacle avoidance effect was poor, the obstacle avoidance algorithm was adjusted, the reward function parameters were adjusted, or the number of training samples was increased; for the problem of untimely navigation decisions, the computational efficiency of the environmental state assessment algorithm and the navigation strategy adjustment algorithm was optimized to reduce decision delay. After multiple rounds of testing and optimization, the navigation performance of the UUV in the pool environment met the design requirements.

[0133] 3. Sea trials and continuous optimization

[0134] Experimental Implementation: The UUV, optimized through pool testing, was deployed to actual sea areas for maritime trials. The trial area was selected in regions with typical marine environmental characteristics, such as nearshore fishing grounds and deep-sea trenches. The UUV underwent a month-long operational trial at sea, performing various practical tasks, including seabed topographic mapping, marine life monitoring, and water quality sampling. During the trial, real-time communication was maintained with the shore-based control center via satellite, transmitting collected data and operational status information to the shore base in real time. Simultaneously, a black box was installed on the UUV to record all sensor data, navigation system operation logs, and other information for subsequent detailed analysis.

[0135] Continuous System Improvement: Based on sea trial data, a comprehensive evaluation of the UUV navigation system is conducted. The navigation performance of UUVs in actual marine environments is analyzed, including positioning accuracy, path planning rationality, obstacle avoidance capabilities, and mission completion efficiency. For identified problems, the causes are investigated in depth, and targeted improvements are implemented. For example, if positioning accuracy decreases in certain sea areas, analysis suggests it may be due to underwater acoustic communication interference or local positioning beacon errors; measures such as optimizing communication algorithms or adjusting beacon layout are taken. If mission execution efficiency does not meet expectations, the navigation decision algorithm is optimized, adjusting the objective function weight coefficients or improving the reinforcement learning model. Through continuous optimization and improvement, the performance of the UUV navigation system in actual marine environments is continuously enhanced, meeting the needs of various complex underwater operations.

[0136] Through the detailed implementation methods described above, it can be ensured that the navigation method for underwater unmanned submersibles proposed in this invention can be effectively implemented in practical applications, giving full play to its technical advantages and providing reliable navigation support for underwater operations.

[0137] The preferred embodiments of the present invention have been described in detail above. It should be understood that those skilled in the art can make numerous modifications and variations based on the concept of the present invention without creative effort. Therefore, all technical solutions that can be obtained by those skilled in the art based on the concept of the present invention through logical analysis, reasoning, or limited experimentation on the basis of existing technology should be within the scope of protection defined by the claims.

Claims

1. A navigation method for an underwater unmanned vehicle, characterized by, Includes the following steps: Acquire motion state data, multibeam sonar data, underwater vision data, and aquatic environment data of the underwater unmanned submersible collected by various sensors at their respective frequencies; The data collected by each sensor are preprocessed by spatiotemporal synchronization, feature extraction, and data fusion. The underwater operation area is defined as a directed graph, where the nodes of the directed graph represent locations with specific significance, the edges represent feasible paths between nodes, and the edge weights are determined based on terrain complexity, obstacle risk, and water flow impact. Based on the constructed directed graph, an improved A* algorithm is used for global path planning. The improved A* algorithm introduces correction coefficients based on terrain and environment modifications in the cost estimation. A fast exploratory random tree algorithm is used for local path planning and dynamic obstacle avoidance. The fast exploratory random tree algorithm adds obstacle risk to the distance metric function between nodes and uses a risk assessment function to evaluate whether to add a new node to the tree. The environmental status is assessed by considering water body environmental parameters, water flow characteristics, and changes in topography. Based on the environmental assessment results and the underwater unmanned vehicle's own state, the navigation strategy is adjusted using a deep reinforcement learning algorithm with the navigation strategy adjustment objective function as the optimization objective, so as to control the underwater unmanned vehicle to travel along the planned path. The calculation of edge weights in the directed graph includes the following steps: Calculate the terrain complexity: build a terrain model based on multi-beam sonar data, calculate the path The terrain undulation degree of the passed area is set as the terrain height standard deviation of the area , the terrain complexity weight coefficient is , and the contribution of the terrain complexity to the edge weight is ; Obstacle risk calculation: Based on obstacle distribution information obtained from multi-source data fusion, statistical path calculation is performed. Determine the path based on the proximity to the obstacle. The distance to the nearest obstacle is The obstacle risk weighting coefficient is Then the contribution of obstacle risk to edge weight is ; Calculate the impact of water flow: Calculate the path using water environment data and water flow models. Water flow speed Angle with the desired direction of travel of the underwater unmanned vehicle and the magnitude of the water flow speed Let the weighting coefficient for the influence of water flow be... Then the contribution of water flow to the edge weights is ; Taking into account terrain complexity, obstacle risk, and water flow impact, determine the edge weight for: The estimated cost is expressed as: in, h ( v To estimate the cost, To determine the correction coefficients based on terrain and environment through training, For nodes To the target node Euclidean distance; The distance metric function is expressed as: in, For nodes Location coordinates, For random sampling nodes Location coordinates, This is the obstacle risk weighting coefficient. Indicates from node arrive The obstacle risk on the path is calculated based on obstacle information obtained from the fusion of multi-source data.

2. The navigation method for an underwater unmanned submersible according to claim 1, characterized in that, The motion state data of the underwater unmanned vehicle is acquired through a high-precision inertial measurement unit, including three-dimensional acceleration data and three-dimensional angular velocity data; the multibeam sonar data is acquired through multibeam sonar at a fixed scanning frequency, and the set of point cloud data acquired in one scan is... ,in ,express t The first time collected at each moment i Point cloud data, These are spatial coordinates. Echo intensity n The number of point cloud data points; the underwater visual data is acquired through an underwater camera. The water environment data is collected by temperature, salinity, and depth sensors, including temperature, salinity, and depth data.

3. The navigation method for underwater unmanned submersibles according to claim 1, characterized in that, The spatiotemporal synchronization specifically includes: Time synchronization: Using the high-precision inertial measurement unit as a reference, the data from other sensors are synchronized to the acquisition time of the high-precision inertial measurement unit; Spatial alignment: A unified world coordinate system is established, and the data from each sensor is transformed into this coordinate system. Specifically, for motion state data, the motion state data is transformed into the world coordinate system through rotation and translation matrices based on the installation position and attitude parameters of the high-precision inertial measurement unit. For multibeam sonar data and underwater vision data, the transformation relationship is constructed using calibration parameters to achieve spatial alignment.

4. The navigation method for an underwater unmanned submersible according to claim 1, characterized in that, The risk assessment function is: in, A collection of obstacles. For obstacles The risk factor is determined based on the type and size of the obstacle; For new nodes To the obstacle The distance; only when the risk assessment function Less than the set threshold At that time, the new node Add to the tree.

5. A navigation method for underwater unmanned submersibles according to claim 1, characterized in that, The environmental status assessment includes water body environmental parameter assessment, water flow characteristic assessment, and topographic and geomorphological change assessment. The water body environmental parameter assessment obtains the water body environmental status assessment function by weighted summation of temperature, salinity, depth, and water turbidity. The water flow characteristic assessment is based on the obtained water flow velocity. The angle between the desired direction of travel of the underwater unmanned vehicle and the direction of the water current. Determine the water flow characteristic evaluation function : The topographic change assessment is based on multibeam sonar data to construct an underwater topographic model, calculate the topographic slope change rate and topographic complexity index along the underwater unmanned vehicle's navigation path, and obtain a topographic assessment function by weighting. The environmental assessment result is obtained by weighted summation of the water body environmental parameter assessment function, the water flow characteristic assessment function, and the topographic change assessment function.

6. A navigation method for an underwater unmanned submersible according to claim 1, characterized in that, The navigation strategy adjustment objective function J for: in, Current location of the underwater unmanned vehicle To the target location distance, E For the power of the underwater unmanned submersible, For the first i The operational status of an underwater unmanned submersible. For the environmental assessment results, , , , These are the weighting coefficients.

7. A navigation method for underwater unmanned submersibles according to claim 1, characterized in that, The state space of the deep reinforcement learning algorithm is defined as: based on the environment evaluation results. Location of underwater unmanned submersible Battery Device operating state vector The state vector ; The action space is defined as: operations such as adjusting speed, changing course, and adjusting depth; the reward function is defined as: Wherein, if the decrease in distance from the target position of the underwater unmanned vehicle within one time step is Then reward If a device malfunction is detected, then If the environmental assessment results Exceeding the threshold Then reward , , , This is the reward coefficient.

Images