An autonomous obstacle avoidance method for unmanned underwater vehicles in a longitudinal underwater scenario

By constructing a full-link risk characteristic system and a causal reasoning decision-making model, the problem of underwater unmanned vehicles being unable to identify hidden risks and experiencing a decline in decision-making accuracy during longitudinal navigation has been solved. This has enabled the stability and continuity of underwater longitudinal navigation, allowing the vehicle to adapt to complex environments and protect the ecological environment.

CN123113996APending Publication Date: 2026-11-17BEIJING ZHONGAN INTELLIGENT INFORMATION TECH CO LTD
View PDF 0 Cites 0 Cited by

Patent Information

Application Number
CN202611212752.3
Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
Filing Date
2026-08-11
Publication Date
2026-11-17

AI Technical Summary

Technical Problem

Existing autonomous obstacle avoidance technology for underwater unmanned vehicles is not suitable for long-distance longitudinal underwater navigation, cannot identify hidden risks, has reduced decision-making accuracy, and cannot meet the requirements for stability, continuity and long range.

Method used

We construct a full-link risk feature system, build a causal reasoning decision model that embeds the physical prior laws of underwater navigation, construct a dataset and perform spatiotemporally synchronized standardized preprocessing to generate multi-dimensional causal correlation quantitative features, and conduct real-time risk assessment and control through the causal reasoning decision model.

Benefits of technology

It achieves full-link risk coverage for underwater longitudinal navigation, improves decision-making accuracy and stability, meets the safety and continuity requirements of long-distance navigation, adapts to complex flow fields and extreme terrain, and ensures a coordinated balance between marine resource development and ecological environmental protection.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN123113996A_ABST
    Figure CN123113996A_ABST
Patent Text Reader

Abstract

This invention discloses an autonomous obstacle avoidance method for unmanned underwater vehicles (UUVs) in a longitudinal underwater scenario, relating to the field of navigation control and autonomous obstacle avoidance technology for UUVs. Its key technical points are: constructing a risk feature system covering secondary risks of self-excited disturbances, latent risks of environment-navigation coupling, obstacle collision risks, secondary risks of obstacle avoidance actions, and risks of aquatic ecological disturbances; clarifying the temporal transmission paths and coupling relationships of various risks through causal correlation feature extraction; thereby achieving coverage and advance prediction from primary risks to secondary risks, and from explicit collision risks to latent instability risks. It can not only identify conventional obstacles but also provide early warnings of hidden risks that existing technologies cannot identify, such as the gradual failure of the detection system caused by wake accumulation and the attitude instability without obstacles caused by terrain-flow coupling. Simultaneously, it completes the prediction and control of secondary risks of obstacle avoidance actions before decision-making, achieving risk management for underwater navigation.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] This invention relates to the field of underwater unmanned submersible navigation control and autonomous obstacle avoidance technology, specifically an autonomous obstacle avoidance method for unmanned submersibles in a longitudinal underwater scenario. Background Technology

[0002] With the deepening of marine development activities such as marine resource exploration, three-dimensional marine environmental monitoring, subsea pipeline inspection, and nearshore ecological protection, unmanned underwater vehicles (UUVs) have become core equipment for underwater operations due to their advantages of maneuverability, wide operating range, and lack of personnel safety risks. Among these, long-distance, fixed-depth, and fixed-course navigation is the most core and widely used navigation scenario for UUVs. It is the basic navigation mode for mainstream operational tasks such as seabed topographic mapping, marine cross-section environmental monitoring, and long-distance subsea pipeline inspection, and has extremely high specific requirements for navigation stability, continuity, low disturbance, and long-range reliability.

[0003] The typical implementation process of current general autonomous obstacle avoidance technology for underwater unmanned vehicles (UUVs) is as follows: First, environmental data around the vehicle is collected in real time using onboard sensors such as forward-looking obstacle avoidance sonar and underwater vision cameras. Second, the collected raw data is subjected to simple filtering, noise reduction, and target detection to extract basic features such as the spatial position and outline size of obstacles. Third, the extracted obstacle features are input into a preset path planning model, and an obstacle avoidance detour path is generated through a general path search algorithm. Finally, the UUV is controlled to perform obstacle avoidance maneuvers, and after the obstacle is out of the detection range, the UUV is controlled to return to the preset navigation route. However, due to its inherent limitations, this approach cannot adapt to the specific characteristics of long-distance longitudinal underwater navigation and has the following drawbacks: First, it can only identify and avoid visible obstacles that are visible to the naked eye / sensors, completely ignoring the continuous scattering of sonar echoes by the wake generated by the continuous navigation of the submersible. That is, as the navigation mileage increases, the cumulative effect of the wake will cause the effective detection range of the detection system to continuously decrease, resulting in progressive failure problems such as missed obstacle detection and false alarms. At the same time, it cannot identify the hidden instability risks such as internal waves, density flow, and turbulence amplification caused by the coupling of terrain undulations and flow fields. It is prone to depth loss of control and attitude instability in the absence of obstacles, which will eventually lead to the interruption of the navigation mission and cannot meet the stability and continuity requirements of long-distance longitudinal navigation. Secondly, many models employ purely data-driven black-box neural network models or general path planning algorithms, failing to incorporate prior physical laws such as fluid dynamics and navigation dynamics in the field of underwater navigation. They can only achieve effective obstacle avoidance in the conventional scenarios covered by the training set. When encountering unknown complex flow fields, extreme terrain, or other scenarios outside the training set during longitudinal navigation, they are prone to outputting infeasible decisions that do not conform to the laws of underwater physics, such as roll instability caused by large rudder angle turns or depth control failure within the thermocline. In addition, they mainly adopt a single-frame instantaneous decision-making mode, making a single obstacle avoidance decision based only on the sensor data at the current moment, without establishing an execution feedback-online optimization closed loop throughout the entire navigation cycle. During long-distance longitudinal navigation of tens of hours, as sensors drift, environmental characteristics change, and equipment performance degrades, the model's decision accuracy will continue to decline, eventually leading to the failure of the obstacle avoidance system. Engineering practicality and long-range stability cannot be guaranteed. Summary of the Invention

[0004] To achieve the above objectives, the present invention provides the following technical solution: An autonomous obstacle avoidance method for an unmanned underwater vehicle in a longitudinal underwater scenario, comprising the following steps: Collect all data from the entire underwater longitudinal navigation cycle and perform dataset construction and spatiotemporal synchronized standardization preprocessing. The full data includes time-series reference data, multi-source environmental perception data, submersible self-excited disturbance monitoring data, aquatic organism behavior response data, and ecological protection prior data. Deep features of the longitudinal navigation scenario are extracted from the preprocessed full data to generate multi-dimensional causal correlation quantitative features corresponding to the full-link risks of the submersible. The full-link risks include secondary risks of submersible self-excited disturbances, implicit risks of environment-navigation coupling, obstacle collision risks, secondary risks of obstacle avoidance actions, and risks of aquatic ecological disturbances. An initial decision-making model based on causal reasoning embedded with prior physical laws of underwater navigation was constructed. Quantitative features were input into the initial decision-making model. Supervised training, causal alignment, and iterative optimization of the model were carried out using time-series sample data labeled with risk causal chains, disturbance mitigation effects, and ecological impact levels to generate an autonomous obstacle avoidance decision-making model. Simultaneous preprocessing and feature extraction are performed on all data collected during longitudinal navigation to generate real-time multi-dimensional causal correlation quantitative features, which are then input into the optimized autonomous obstacle avoidance decision model. The model outputs full-link risk assessment values, disturbance mitigation control parameters, ecological compliance obstacle avoidance decision parameters, and secondary risk prediction values. By combining a dual-threshold decision interval that is dynamically adapted to the current navigation scenario, a graded navigation control command matching the risk level of the navigation scenario is generated and executed; feedback data after the execution of the graded navigation control command is collected, and the autonomous obstacle avoidance decision model is subjected to online incremental optimization and dynamic calibration of control parameters. The dual-threshold decision interval includes an upper threshold and a lower threshold.

[0005] Furthermore, the dataset construction and spatiotemporal synchronization standardization preprocessing include: Based on the cumulative mileage of the unmanned underwater vehicle's longitudinal voyage, a spatiotemporal reference coordinate system with mileage-time dual mapping is established. To address the differences in sampling frequencies of multi-source sensors, spatiotemporal misalignment caused by voyage displacement, and detection distortion caused by changes in vertical depth, mileage matching interpolation is used for time sequence alignment and spatial coordinate registration. Based on the spatiotemporal reference coordinate system, the synchronization accuracy of each sensor's data is verified, substandard data segments are removed, and the validity of the data is verified and invalid data is removed. Based on the underwater vehicle disturbance evolution model and water environment parameters, distortion correction of sensing data is performed. The corrected data is then subjected to outlier removal, null value completion, and adaptive denoising. The dataset is divided into training, validation, and test sets according to the navigation mileage segment and scenario type. The training set samples are labeled with risk causal chains, disturbance mitigation effects, and ecological impact level labels to construct a standardized time series dataset.

[0006] Furthermore, the extraction of causal correlation quantitative features corresponding to the secondary risks of self-excited disturbances of underwater vehicles and the implicit risks of environment-navigation coupling includes: Based on the preprocessed monitoring data of self-excited disturbances of the submersible and historical navigation data, the disturbance diffusion characteristics, temporal evolution laws and the impact parameters of the sensing system under different navigation states are extracted. The self-interference temporal cumulative impact coefficient and the reliability quantification parameters of the sensing system are calculated, and the causal correlation quantification characteristics corresponding to the secondary risks of self-excited disturbances of the submersible are generated. Based on the preprocessed environmental perception data, the terrain feature parameters and water environment field distribution data of the navigation path are extracted. The time-series evolution law of the environmental field is simulated based on the fluid dynamics numerical model and the hidden risk areas are identified. Combined with the underwater vehicle navigation dynamics model, the environment-navigation coupling risk interval is predicted, the coupling risk coefficient is calculated, and the causal correlation quantitative characteristics corresponding to the environment-navigation coupling hidden risks are generated.

[0007] Furthermore, the extraction of quantitative features of the causal relationship between obstacle collision risk and secondary risks of obstacle avoidance actions includes: For the preprocessed obstacle perception data, the spatial features, temporal motion features and modal features of the obstacles are extracted. The obstacle motion path is predicted based on the multi-model filtering algorithm, collision-related feature parameters are calculated, and the target type is distinguished and the collision threat weight is matched in combination with the underwater vehicle navigation parameters to generate causal correlation quantitative features corresponding to obstacle collision risk. Based on the preprocessed underwater vehicle navigation status data, the navigation parameter adjustment amounts corresponding to the preset obstacle avoidance actions are extracted by combining the underwater vehicle dynamics model and the disturbance evolution model. The secondary risk coefficient, mission adaptability parameter and coupling instability risk coefficient of different obstacle avoidance actions are calculated, and the causal correlation quantitative features corresponding to the secondary risks of obstacle avoidance actions are generated.

[0008] Furthermore, the extraction of quantitative characteristics of causal relationships corresponding to aquatic ecological disturbance risks includes: Based on the preprocessed aquatic organism behavior response data and ecological protection prior data, we extract the characteristics, activity patterns, stress thresholds of aquatic organisms around the navigation area, as well as the protection level, control boundaries, and control rules of ecologically sensitive areas. Based on the aquatic organism stress response model, we calculate the degree of disturbance to aquatic organisms and the range of habitat interference caused by the current navigation status, sensing equipment operating parameters, and preset obstacle avoidance actions, and generate single-action ecological disturbance causality coefficients. By combining species stress thresholds, hard constraint parameters for ecological compliance are defined; by integrating aquatic biological characteristics, single-action ecological disturbance causality coefficients and hard constraint parameters for ecological compliance, causal relationship characteristics of navigation actions and ecological impacts are constructed, and quantitative causal relationship characteristics corresponding to aquatic ecological disturbance risks are generated. Among them, the species stress threshold refers to the critical quantitative indicator that core protected species change from normal physiological / behavioral state to produce irreversible negative effects when subjected to external human disturbances.

[0009] Furthermore, the steps for building, training, and optimizing the autonomous obstacle avoidance decision-making model are as follows: Construct a causal inference initial decision model that includes a temporal feature input layer, a causal feature fusion layer, a physical law constraint layer, a causal inference decision layer, and a multi-objective output layer; Multi-dimensional causal correlation quantitative features are input into the time-series feature input layer. In the causal feature fusion layer, based on the time-series correlation of the risk causal chain, risk features of different dimensions are adaptively weighted and fused to obtain the risk causal transmission path. In the physical law constraint layer, the physical prior laws of fluid mechanics, navigation dynamics, disturbance evolution, and ecological response related to underwater navigation are embedded to perform physical feasibility verification and correction on the parameters output by the decision layer. A multi-dimensional composite loss function is constructed, which includes risk prediction, physical constraints, disturbance resolution, and ecological constraints. Model supervised training and causal alignment are performed through labeled samples. The generalization of the model is enhanced through boundary samples and adversarial samples. The model hyperparameters are tuned and the performance is verified through validation and test sets. In view of the computing power constraints of the embedded platform of the submarine, the trained model is lightweighted and iteratively optimized to generate a vertical scene-specific autonomous obstacle avoidance decision model.

[0010] Furthermore, the generation of end-to-end risk assessment values ​​and corresponding decision parameters includes: The real-time multi-dimensional causal correlation quantitative features are input into the autonomous obstacle avoidance decision model. The causal feature fusion layer performs time-series causal correlation weighted fusion, and the physical feasibility of the decision parameters is verified by the physical law constraint layer. Based on the current navigation mission priority, ecological control requirements, and environmental complexity, the fusion weight of risk features of each dimension is dynamically adjusted, and risk reasoning and decision optimization are performed by the causal reasoning decision layer. The multi-objective output layer outputs the full-link risk assessment value, disturbance reduction control parameters, ecological compliance obstacle avoidance decision parameters, and secondary risk prediction value. When the end-to-end risk assessment value is in the transition range of the dual-threshold decision range, or the secondary risk prediction value exceeds the preset threshold, the multi-frame temporal feature fusion re-inference process is triggered. If the decision confidence requirement is not met even after reaching the maximum number of re-inferences, the security degradation plan is triggered.

[0011] Furthermore, the dynamic adaptation of the dual-threshold decision interval and the generation and execution of hierarchical navigation control commands include: The safety redundancy coefficient is determined based on the current navigation mission priority. The ecological control requirements, environmental complexity, and effective detection range of the sensing equipment in the current navigation scenario are collected as dynamic adjustment input parameters. Through a preset threshold mapping model, combined with the safety redundancy coefficient and input parameters, the upper and lower thresholds of the dual threshold decision interval are calculated. When the navigation scenario and mission priority change abruptly, the thresholds are updated and verified in real time. The end-to-end risk assessment value is matched with the dynamic dual-threshold decision range, and combined with the corresponding control parameters and predicted values, a graded navigation control command corresponding to the risk level is generated. When the end-to-end risk assessment value is greater than or equal to the upper threshold, an ecological compliance obstacle avoidance command and disturbance mitigation control are generated and executed; when the upper threshold is greater than or equal to the end-to-end risk assessment value and the lower threshold is between, a navigation warning command is generated, the risk assessment results are updated and alternative obstacle avoidance schemes are pre-generated; when the end-to-end risk assessment value is less than or equal to the lower threshold, a low-disturbance steady-state navigation maintenance command is generated and executed, the ecological response and self-excited disturbance level are continuously monitored, and navigation parameters are dynamically adjusted.

[0012] Furthermore, the method also includes sensor anomaly fault tolerance handling and navigation return after obstacle avoidance, specifically including: The system monitors the data output status of multiple sensors in real time. When a single sensor data interruption, missing data, or continuous anomalies are detected, replacement data is generated based on the historical data of that sensor and the synchronous data of other related sensors through a spatiotemporal correlation completion model. When more than a preset number of sensors are detected to be abnormal at the same time, a safety degradation command is triggered, the navigation speed is reduced, a low-disturbance, low-energy navigation mode is activated, and fault self-check and emergency obstacle avoidance plans are triggered simultaneously. When the end-to-end risk assessment value falls below the lower limit threshold and remains below the preset duration, the risk is determined to be eliminated, and the navigation return process is initiated. Based on the current position of the submersible, the preset mission path, and surrounding environmental data, with the optimization objectives of minimum path deviation, minimum disturbance impact, and minimum power consumption, a longitudinal and vertical coupled regression path is generated. Regression control is executed and disturbance resolution control is executed simultaneously. After returning to the preset path, the system switches to a low-disturbance steady-state navigation mode.

[0013] Furthermore, the online incremental optimization and full-cycle management of the autonomous obstacle avoidance decision-making model include: Data on the underwater vehicle's navigation status, environmental risk changes, and aquatic organism behavior response after the execution of graded navigation control commands are collected as feedback data. Risk prediction bias, disturbance mitigation effect bias, and ecological impact bias are extracted from the feedback data. After data cleaning and labeling, incremental training samples are constructed. Based on incremental training samples, the decision layer weights of the autonomous obstacle avoidance decision-making model are adjusted, and the fixed parameters of the feature extraction layer and physical law constraint layer of the model are frozen. The adjusted model is verified for risk prediction accuracy, decision physical feasibility and ecological compliance through a preset validation set. If the verification is passed, the model is updated; otherwise, if the verification fails, the original model parameters are retained. Based on feedback data throughout the entire navigation cycle, the threshold mapping model, feature fusion weight allocation logic and disturbance resolution control parameters are continuously optimized.

[0014] This invention provides an autonomous obstacle avoidance method for unmanned underwater vehicles in a longitudinal underwater scenario, which has the following beneficial effects: 1. Targeting the characteristics of long-distance, fixed-depth, and fixed-course underwater navigation scenarios, a spatiotemporal reference coordinate system with mileage-time dual mapping is established, using the cumulative mileage of the submersible's longitudinal navigation as the core benchmark. A mileage-matching interpolation method is used to complete the temporal alignment and spatial registration of multi-source heterogeneous data. A full-process preprocessing mechanism is simultaneously implemented, including accuracy verification, wake distortion feedforward correction, sound velocity / optical attenuation correction, and data standardization. This solves the problem of existing synchronization modes that rely solely on timestamps, eliminates the cumulative effect of spatiotemporal misalignment caused by differences in sensor sampling frequencies and continuous displacement of the submersible, and avoids issues such as obstacle position detection deviations and inaccurate topographic mapping data. This makes the system suitable for the accuracy requirements of mainstream longitudinal navigation tasks such as seabed topographic mapping, marine environmental section monitoring, and long-distance pipeline inspection, while also improving scenario adaptability and data reliability.

[0015] 2. By constructing a full-link risk feature system covering secondary risks of self-excited disturbances of underwater vehicles, latent risks of environmental-navigation coupling, obstacle collision risks, secondary risks of obstacle avoidance actions, and risks of aquatic ecological disturbances, and by extracting causal correlation features, the temporal transmission path and coupling relationship of various risks are clarified. This achieves full-dimensional coverage and advance prediction of risks from primary risks to secondary risks, and from explicit collision risks to latent instability risks. It can not only identify conventional obstacles, but also provide early warning of hidden risks that existing technologies cannot identify, such as the gradual failure of detection systems caused by wake accumulation and the attitude instability of obstacle-free areas caused by terrain-flow field coupling. At the same time, it can predict and control the secondary risks of obstacle avoidance actions before decision-making, thus achieving full-link risk management of underwater navigation.

[0016] 3. A causal reasoning decision-making model embedded with prior physical laws of underwater navigation was constructed. An independent physical law constraint layer was set in the model architecture, embedding four core physical laws: fluid dynamics control, longitudinal and vertical coupling dynamics of the submersible, wake diffusion temporal evolution, and aquatic organism stress response. At the same time, a multi-dimensional composite loss function including risk prediction, physical constraints, disturbance resolution, and ecological constraints was constructed. Then, through the full-process model optimization mechanism of adversarial sample generalization enhancement training and lightweight processing of embedded platform, the model's decisions have complete physical interpretability, thereby eliminating infeasible decision outputs that do not conform to underwater physical laws. At the same time, the generalization ability of the model in scenarios outside the training set, such as unknown complex flow fields and extreme terrain, was greatly improved. The lightweight processing also makes the model perfectly adaptable to the computing power constraints of the submersible embedded platform.

[0017] 4. By constructing causal relationship features between navigation actions and ecological impacts, the degree of disturbance to aquatic organisms caused by different navigation and obstacle avoidance actions is quantified. In the model constraint layer, the stress response patterns of aquatic organisms are embedded to verify the ecological compliance of decision parameters. During decision execution, graded obstacle avoidance instructions matching the protection level of ecologically sensitive areas are generated, prioritizing low-disturbance obstacle avoidance methods, effectively avoiding the limitations of existing technologies that "only focus on successful obstacle avoidance while ignoring ecological impacts." Thus, while ensuring the efficiency of navigation operations, the management requirements of marine ecologically sensitive areas are met, and irreversible disturbance to aquatic organism habitats and reproductive behaviors caused by navigation and obstacle avoidance actions is avoided, thereby achieving a synergistic balance between marine resource development and ecological environmental protection.

[0018] 5. Construct a full-cycle management and control system from data acquisition, feature extraction, decision generation, execution feedback to online optimization. This system can continuously collect data on the underwater vehicle's status, environmental changes, and ecological responses after the execution of control commands during navigation, and use incremental learning to complete online fine-tuning of the model and dynamic calibration of control parameters. It adapts to changes in operating conditions such as sensor drift, changes in environmental characteristics, and equipment performance degradation during long-distance navigation. At the same time, it is equipped with a full-scenario fault-tolerant mechanism for spatiotemporal completion of single-sensor data anomalies, safety degradation of multi-sensor anomalies, and smooth regression after risk resolution. This significantly improves the safety, continuity, and engineering practicality of the unmanned underwater vehicle during long-distance longitudinal navigation. Attached Figure Description

[0019] Figure 1 This is an architecture diagram of an embodiment of the present invention; Figure 2 This is a flowchart of an embodiment of the present invention. Detailed Implementation

[0020] 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, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those of ordinary skill in the art without creative effort are within the scope of protection of the present invention.

[0021] Example 1: Autonomous obstacle avoidance method for underwater longitudinal navigation: This approach is applicable to general-purpose underwater long-distance, fixed-depth, and fixed-course navigation scenarios. Taking a general-purpose observation-type unmanned underwater vehicle (UUV) as an example, the scenario assumes a longitudinal navigation distance of 80km, a fixed-depth navigation depth of 50m, an average water depth of 120m in the navigation area, gently undulating terrain with local obstacles, and the presence of a conventional thermocline and a weak current field. The mission requirements are to ensure navigation safety, maintain the stability of the preset route, and achieve autonomous obstacle avoidance throughout the entire process. The specific implementation steps are as follows: First, a dataset construction and mileage-time dual-referenced synchronous standardization preprocessing were performed. This involved collecting time-series reference data for the entire underwater longitudinal navigation cycle, multi-source environmental perception data, full-dimensional monitoring data of the submersible's self-excited disturbances, aquatic organism behavioral response data, and prior ecological protection data. The time-series reference data came from the cumulative mileage, timestamps, and spatial coordinate data output by the submersible's INS / T integrated navigation system. The multi-source environmental perception data included underwater topographic 3D time-series point cloud data acquired by multibeam forward-looking sonar and vertical full-depth water body data acquired by acoustic Doppler current profiler. The data includes: flow field time-series data, obstacle multi-frequency echo time-series data collected by forward-looking obstacle avoidance sonar, and visual image data collected by underwater low-light cameras; full-dimensional monitoring data of self-excited disturbances of the submersible, including wake-navigation dynamics coupled disturbance data collected by wake flow field monitoring sensors, operating parameters of the submersible's propulsion system and servo system, and longitudinal and vertical coupled navigation status data; aquatic organism behavior response data comes from underwater biological acoustic data collected by hydrophones and biological behavior time-series monitoring data collected by visual cameras; and ecological protection prior data consists of pre-stored basic data for ecological management around the navigation path. After data acquisition, a spatiotemporal reference coordinate system with mileage-time dual mapping is established based on the cumulative mileage of the unmanned underwater vehicle's longitudinal voyage. To address the issues of sampling frequency differences among multi-source sensors, spatiotemporal misalignment caused by voyage displacement, and detection distortion caused by vertical depth changes, a third-order spline interpolation method with mileage matching is used to complete the temporal alignment and spatial coordinate registration of all data. Then, based on the spatiotemporal reference coordinate system, the synchronization accuracy of each sensor's data is checked item by item, and data segments with substandard synchronization accuracy are eliminated. At the same time, data validity is checked, and invalid data that exceeds the sensor's range or has temporal logic contradictions are eliminated. Subsequently, based on the pre-established underwater vehicle disturbance evolution model and measured water environment parameters, distortion correction of the sensing data was completed. Based on the measured water sound velocity profile and optical attenuation coefficient, sound velocity correction of sonar data and attenuation correction of visual data were completed. The corrected data were then processed by using the 3σ criterion to remove outliers, linear interpolation based on time-series correlation to complete missing values, and adaptive wavelet denoising to smooth the data. Finally, the training set, validation set, and test set were divided into three categories according to the navigation mileage segment and scenario type in a ratio of 7:2:1. Three types of labels were labeled for the training set samples: risk causal chain, disturbance mitigation effect, and ecological impact level, to construct a standardized time-series dataset covering the entire longitudinal navigation scenario. After completing the above processing, deep feature extraction is performed specifically for the longitudinal navigation scenario. Deep feature extraction is performed on the preprocessed full data to generate multi-dimensional causal correlation quantitative features corresponding to the full-link risks of the submersible navigation. The full-link risks include secondary risks of submersible self-excited disturbances, implicit risks of environment-navigation coupling, obstacle collision risks, secondary risks of obstacle avoidance actions, and risks of aquatic ecological disturbances. Among them, the extraction of causal correlation quantitative features corresponding to the secondary risk of self-excited disturbance of underwater vehicles involves extracting the three-dimensional diffusion range of the wake, the temporal evolution law of turbulence intensity, and the sonar echo scattering coefficient and attenuation coefficient under different navigation states from the pre-processed self-excited disturbance monitoring data and historical navigation data of underwater vehicles. Combined with the wake disturbance evolution model, the temporal interference degree of the wake generated under the current navigation state on the detection equipment within the future preset navigation range is calculated, and the self-interference temporal cumulative influence coefficient is generated. Combined with the minimum signal-to-noise ratio requirement of the detection equipment, the sensor effective detection distance attenuation rate, obstacle detection false alarm rate and missed detection rate under the current wake disturbance are calculated, and the reliability quantitative parameters of the perception system are generated. Finally, the causal correlation quantitative features corresponding to the secondary risk of self-excited disturbance of underwater vehicles are integrated to generate the causal correlation quantitative features. The extraction of causal correlation quantitative features corresponding to the implicit risks of environment-navigation coupling involves extracting topographic feature parameters and water environment field distribution data of the navigation path from the preprocessed environmental perception data. Based on the Reynolds-averaged Navier-Stokes equations, a numerical model of topographic-flow field coupling is constructed to simulate the temporal evolution of the flow field under the coupling effect of topography and flow field within a preset time period. Implicit flow field risk areas are identified. The navigation motion cycle of the submersible is calculated by combining the longitudinal and vertical coupling dynamic model of the submersible. The coupling resonance interval of the topographic undulation cycle, the flow field fluctuation cycle and the submersible motion cycle is predicted. The probability of resonance and the degree of influence on the navigation state of the submersible are calculated, and the coupling risk coefficient is generated. Finally, the causal correlation quantitative features corresponding to the implicit risks of environment-navigation coupling are integrated to generate the causal correlation quantitative features. The extraction of causal correlation quantitative features corresponding to obstacle collision risk involves extracting the spatial features, temporal motion features, and modal features of obstacles from the preprocessed obstacle perception data. Based on the interactive multi-model filtering algorithm, the longitudinal and vertical coupled motion path of the obstacle within a preset time period is predicted. The nearest intersection distance, intersection time, collision probability, and collision critical window of the obstacle and the underwater vehicle's navigation path are calculated. The target type is distinguished and the collision threat weight is matched by combining the underwater vehicle's navigation cross-section size, vertical maneuver boundary, and longitudinal braking distance. The causal correlation quantitative features corresponding to obstacle collision risk are then integrated to generate the causal correlation quantitative features. The extraction of causal correlation quantitative features corresponding to the secondary risks of obstacle avoidance actions involves taking the pre-processed underwater vehicle navigation state data, combining the underwater vehicle dynamics model and disturbance evolution model, extracting the navigation parameter adjustment amount corresponding to the preset obstacle avoidance actions, calculating the new disturbance intensity, power consumption, and navigation path deviation generated by different obstacle avoidance actions, predicting the coupling effect between obstacle avoidance actions and the surrounding environment, generating the secondary risk coefficient of obstacle avoidance actions, mission adaptability parameters, and coupling instability risk coefficient, and integrating them to generate the causal correlation quantitative features corresponding to the secondary risks of obstacle avoidance actions. The extraction of causal correlation quantitative features corresponding to aquatic ecological disturbance risks involves extracting basic characteristics, stress behavior thresholds, and ecological management rules of aquatic organisms around the navigation area from preprocessed aquatic organism behavioral response data and ecological protection prior data. Based on the aquatic organism stress response model, the degree of disturbance to aquatic organisms caused by the current navigation status and preset actions is calculated, generating single-action ecological disturbance causal coefficients. Combined with ecological management requirements, navigation disturbance thresholds and obstacle avoidance action constraint rules are defined, generating ecological compliance hard constraint parameters, and integrating them to generate causal correlation quantitative features corresponding to aquatic ecological disturbance risks.

[0022] Then, the autonomous obstacle avoidance decision-making model is built and trained and optimized. An initial decision-making model for causal reasoning embedded with the physical prior laws of underwater navigation is built. The model structure includes a temporal feature input layer, a causal feature fusion layer, a physical law constraint layer, a causal reasoning decision layer and a multi-objective output layer. Multi-dimensional causal correlation quantitative features are input into the temporal feature input layer. In the causal feature fusion layer, the risk features of different dimensions are adaptively weighted and fused based on the temporal correlation of the risk causal chain to clarify the risk causal transmission path. Four types of physical prior laws are embedded in the physical law constraint layer: hydrodynamic control equations related to underwater navigation, longitudinal and vertical coupling dynamic equations of submersibles, wake diffusion time-series evolution equations, and aquatic organism stress response equations. The physical equations are existing solutions and will not be elaborated on here. The parameters output by the decision layer are physically feasible and corrected, and infeasible decision schemes that do not conform to physical laws are eliminated. A multi-dimensional composite loss function is constructed, which includes risk prediction loss, physical constraint loss, disturbance resolution loss, and ecological constraint loss. Model supervised training and causal alignment are completed through labeled training set samples, model generalization enhancement training is completed through boundary samples and adversarial samples, and hyperparameter tuning is completed through validation set and performance verification is completed through test set. To address the computational constraints of the embedded platform for underwater vehicles, the trained model is lightweighted, reducing the number of model parameters and inference time without sacrificing core decision-making capabilities. Iterative optimization generates a final autonomous obstacle avoidance decision-making model specific to the longitudinal scenario. Real-time inference and decision parameter output are then performed. During the underwater vehicle's longitudinal navigation, the real-time collected longitudinal navigation data undergoes the same preprocessing and feature extraction processes as during training, generating real-time multi-dimensional causal correlation quantitative features, which are input into the optimized autonomous obstacle avoidance decision-making model. A causal feature fusion layer performs time-series causal correlation weighted fusion of real-time features, dynamically adjusting the fusion weights of risk features across various dimensions based on current navigation mission priority, ecological control requirements, and environmental complexity. A physical constraint layer verifies the physical feasibility of the decision parameters. Finally, a multi-objective output layer outputs the comprehensive risk assessment value for the entire navigation scenario, feedforward disturbance mitigation control parameters, ecological compliance obstacle avoidance decision parameters, and secondary risk prediction values. When the overall risk assessment value of the entire link is in the transition range of the dual threshold decision range or the secondary risk prediction value exceeds the preset threshold, the re-inference process of multi-frame temporal feature fusion is triggered. If the decision confidence requirement is not met even after reaching the preset maximum number of re-inferences, the security degradation plan is triggered. Then, the dynamic adaptation of the dual-threshold decision interval and the generation and execution of hierarchical navigation control commands are performed. The safety redundancy coefficient is determined based on the current navigation mission priority. The ecological management requirements, environmental complexity, effective detection distance of sensing devices, and water visibility of the current navigation scenario are collected as dynamic adjustment input parameters. The upper and lower thresholds of the dual-threshold decision interval are calculated through the preset threshold mapping model. When there are sudden changes in the navigation scenario and mission priority, the thresholds are updated in real time and the rationality is verified. The comprehensive risk assessment value of the whole link is matched with the dynamic dual-threshold decision interval. The hierarchical navigation control commands with corresponding risk levels are generated by combining the corresponding control parameters and prediction values. When the comprehensive risk assessment value of the entire link exceeds the upper limit threshold, obstacle avoidance instructions matching the scenario requirements are generated and executed, and feedforward disturbance mitigation control parameters are executed simultaneously to avoid primary risks while controlling secondary risks and ecological disturbances within the preset threshold. When the comprehensive risk assessment value of the entire link is between the upper and lower limits, a navigation warning command is generated and executed to adjust the navigation parameters to reduce the impact of self-excited disturbances, continuously update the risk assessment results and pre-generate compliant alternative obstacle avoidance schemes. When the comprehensive risk assessment value of the entire link does not exceed the lower limit threshold, a low-disturbance steady-state navigation maintenance command is generated and executed. The environment and navigation status are continuously monitored, and navigation parameters are dynamically fine-tuned to achieve closed-loop control of disturbances. At the same time, sensor anomaly fault tolerance processing and navigation return process after obstacle avoidance are executed simultaneously, and the data output status of multi-source sensors is monitored in real time. When single sensor data is abnormal, alternative data is generated by spatiotemporal correlation to complete the model and maintain the continuity of model inference. When multiple sensors are abnormal at the same time, a safety degradation command and an emergency obstacle avoidance plan are triggered. When the comprehensive risk assessment value of the entire link falls below the lower limit threshold and continues for a preset time, the risk is determined to be eliminated and the navigation return process is initiated. Based on the current position of the submersible, the preset mission path and the surrounding environment data, a smooth longitudinal and vertical coupling return path is generated. After returning to the preset mission path, it automatically switches to a low-disturbance steady-state navigation mode. Finally, online incremental optimization and full-cycle closed-loop management are implemented. Throughout the entire navigation cycle, data on the underwater vehicle's navigation status, environmental risk changes, and aquatic organism behavior responses after the execution of graded navigation control commands are collected as feedback data. Various deviation values ​​are extracted from the feedback data, and after data cleaning and labeling, incremental training samples are constructed. Based on the incremental training samples, the decision layer weights of the autonomous obstacle avoidance decision model are fine-tuned. The parameters of the model's feature extraction layer and physical law constraint layer are frozen to avoid catastrophic amnesia. The performance of the fine-tuned model is verified through a preset validation set. After the verification is passed, the model parameters are updated online. At the same time, based on the feedback data throughout the entire navigation cycle, the threshold mapping model, feature fusion weight allocation logic, and disturbance resolution control parameters are continuously optimized to achieve full-link risk closed-loop management and continuous iterative optimization of the model throughout the entire longitudinal navigation cycle.

[0023] Example 2, Eco-friendly: The preferred implementation plan for longitudinal navigation missions in ecologically sensitive marine areas is applicable to underwater longitudinal navigation missions passing through sensitive areas such as national-level marine ecological red line zones and aquatic germplasm resource protection areas. It is specifically optimized to meet the mission requirements of prioritizing ecological protection. The specific process is as follows: To address the limitations of existing technologies in quantifying and controlling the non-lethal irreversible disturbances of navigation and obstacle avoidance maneuvers to aquatic organisms, the potential for secondary ecological impacts from obstacle avoidance maneuvers, and the inability to meet the control requirements of ecologically sensitive areas, we employed acoustic and visual stress behavior calibration datasets of core protected species during dataset construction and preprocessing. We refined prior data such as the hierarchical control boundaries of ecologically sensitive areas, prohibited periods, and species-specific stress thresholds, raising the data synchronization accuracy threshold to a time deviation ≤30ms and a spatial deviation ≤0.3m. This ensures the spatiotemporal matching accuracy of ecological disturbance-related data. The stress threshold refers to the critical quantitative indicator at which core protected species, when subjected to external human disturbance, transition from a normal physiological / behavioral state to one that produces irreversible negative impacts. This includes factors such as navigation noise, wake flow field, light illumination, and sonar detection, as shown in the table below. Table 1. Types and ranges of stress thresholds in fish During feature extraction, the weight of aquatic ecological disturbance risk features is increased to 35%–40%. Ecological disturbance features corresponding to different navigation speeds, thruster power, sonar transmission power, and lighting parameters are extracted in detail. Differentiated stress thresholds are matched for species with different protection levels. Species-specific navigation actions-ecological disturbance causal relationship features are constructed. At the same time, in the extraction of secondary risk features of obstacle avoidance actions, the ecological disturbance intensity of obstacle avoidance actions is quantified first, and actions that exceed the ecological disturbance standard are included in the prohibited list. During model building and training, the constraint weights of the aquatic organism stress response equation are strengthened in the physical law constraint layer, and the weight of ecological constraint loss is increased to 30% in the composite loss function. In model training, samples of ecologically sensitive area boundary scenarios and peak activity periods of core species are added to ensure that all decision parameters output by the model meet the hard constraints of ecological management. During real-time reasoning and control execution, a safety redundancy coefficient of 1.3 to 1.5 is set based on the priority of ecological protection. After entering an ecologically sensitive area, the upper limit of the dual-threshold decision interval is automatically lowered to improve the sensitivity of risk management. In the graded navigation control command, the low-disturbance obstacle avoidance mode of small vertical deepening and uniform slow speed is preferred. Large rudder angle turning, high horsepower speed change and other strong disturbance actions are prohibited. Unnecessary supplementary lighting equipment is turned off and sonar transmission power is reduced to keep the navigation noise radiation intensity below 100dB and the minimum avoidance distance with core protected species is kept above 500m. During optimization, the focus is on continuously optimizing the ecological disturbance prediction model and disturbance mitigation control parameters based on feedback data from the behavioral responses of aquatic organisms, so as to achieve closed-loop management of ecological disturbances throughout their entire lifecycle.

[0024] Example 3, High-precision long-range flight: This optimized implementation plan is designed for high-precision, long-range missions such as deep-sea long-distance subsea pipeline inspection and near-shore shallow-sea topographic mapping. It is suitable for underwater longitudinal navigation missions with extremely high requirements for navigation path tracking accuracy, data spatiotemporal synchronization accuracy, and long-range stability. The plan is specifically optimized to meet the requirements of high-precision, long-range missions. The specific process is as follows: To address the issues of continuous accumulation of spatiotemporal synchronization errors, insufficient accuracy in pipeline tracking / terrain mapping, and continuous decay of model decision-making accuracy during long-distance voyages, existing technologies employ a high-precision spatiotemporal reference coordinate system with continuous mileage of the voyage as the sole core anchor point during dataset construction and preprocessing. This system establishes a three-dimensional mapping of mileage, time, and space, raising the synchronization accuracy threshold to a time deviation of ≤10ms and a spatial deviation of ≤0.1m. A high-order interpolation method based on mileage matching is then used to achieve sub-meter level synchronization registration of multi-source data. For pipeline inspection / topographic mapping tasks, new pipeline location features and topographic mapping accuracy verification data have been added. The training set samples are highlighted with the cumulative pattern of synchronization error and path tracking deviation data in long-range scenarios. During feature extraction, the weight of terrain-flow field coupling latent risk and obstacle collision risk features is increased to 40% to 45%. The focus is on extracting high-precision spatial location features of obstacles around pipelines and shallow reefs within the survey strip. The feature of the impact of wake disturbance on the accuracy of depth sounding / inspection data during long voyages is refined. In the extraction of secondary risk features of obstacle avoidance actions, the path deviation mileage and data interruption duration caused by obstacle avoidance actions are quantified first. The deceleration detour and small vertical depth change obstacle avoidance methods that do not deviate from the preset route / survey strip are selected first. During model building and training, a lightweight scheme combining structured pruning and quantization training was adopted to address the computational constraints of the embedded platform of the submersible. While ensuring decision accuracy, the inference time per frame was controlled within 30ms to meet the real-time requirements of long-range voyages. In the physical law constraint layer, path tracking constraints and depth sounding accuracy constraints were strengthened, and path tracking loss terms / depth sounding accuracy loss terms were added to the composite loss function. During model training, samples of long-range gradually changing terrain and deep-sea density current scenarios were added to improve the model's adaptability to long-range voyages. During real-time reasoning and control execution, a safety redundancy coefficient of 1.1 to 1.2 is set based on the high-precision priority task, and the dual-threshold decision interval is set to an upper threshold of 0.7 to 0.8 and a lower threshold of 0.2 to 0.3 to reduce unnecessary obstacle avoidance actions to ensure the stability of the flight path. The maximum lateral deviation distance of the obstacle avoidance action is controlled within 2m. After obstacle avoidance is completed, the path with the minimum curvature smoothly returns to the preset flight path to ensure the continuity of inspection / mapping data. Finally, optimization is carried out, focusing on online incremental optimization of the model based on route tracking deviation, data accuracy deviation, and risk prediction deviation data. The model parameters are micro-calibrated every 10km of flight distance, and the synchronous correction model and feature fusion weights are continuously optimized to ensure that the decision accuracy does not decrease during long-distance flights.

[0025] Example 4, High Reliability in Extreme Environments: The preferred implementation scheme for emergency navigation missions in extreme environments such as strong turbulence and complex terrain after a typhoon is suitable for high-risk underwater longitudinal navigation missions with extremely complex flow fields, high obstacle density, and susceptibility to sensor interference. Targeted optimizations have been made to meet the high reliability requirements of missions in extreme environments. The specific process is as follows: To address the issues of insufficient risk prediction capabilities, easy system failure due to sensor anomalies, and weak emergency response capabilities in extreme and complex flow field environments, existing technologies incorporate multi-source data redundancy verification and rapid anomaly data identification during dataset construction and preprocessing. The training set samples include boundary samples and adversarial samples from extreme flow fields, multi-sensor anomalies, and sudden dynamic obstacles, accounting for no less than 30% of the total training set, significantly improving the model's generalization ability to extreme scenarios. During feature extraction, the weight of environmental-navigation coupled latent risks and obstacle collision risks is increased to over 50%. The focus is on extracting turbulence amplification, internal wave evolution, and coupled resonance risk features under extreme flow fields. The risk prediction time is increased to 500-600 seconds to allow sufficient time for emergency response. At the same time, the high-speed trajectory features of dynamic obstacles are extracted in detail to improve the ability to identify and predict sudden floating obstacles. In the extraction of secondary risk features of obstacle avoidance actions, the anti-current stability and navigation safety of obstacle avoidance actions are quantified first, and obstacle avoidance actions with strong anti-current ability and stable attitude are selected first. During model building and training, robustness enhancement training of the model is completed through a large number of adversarial examples in extreme scenarios. The dynamic safety boundary constraints of the submarine are strengthened in the physical law constraint layer to prevent the model from outputting decision parameters that exceed the safe navigation limit of the submarine. A safety boundary loss term is added to the composite loss function. At the same time, the model inference logic is optimized so that stable inference can still be completed based on historical time series data in scenarios where sensor data is missing. During real-time reasoning and control execution, the safety redundancy coefficient is set to 1.5-1.8 based on the priority of navigation safety, and the upper threshold of the dual threshold decision interval is significantly reduced to 0.4-0.5 to improve the sensitivity of risk management and trigger early warning and emergency response in advance. The three-level emergency obstacle avoidance command, emergency braking, emergency hovering and emergency ascent, are added to the hierarchical control command to deal with sudden extreme risks. In the sensor anomaly fault tolerance mechanism, the maximum duration of single sensor anomaly compensation is increased to 60s. When two or more core sensors are abnormal at the same time, the safety degradation command is immediately triggered to reduce the navigation speed to the minimum safe speed, shut down unnecessary loads, and activate the emergency obstacle avoidance plan. Finally, based on risk prediction deviations and emergency response effectiveness data, the model is incrementally optimized online, and a model safety verification is completed every 5 minutes to ensure that the model's decisions always meet the safety boundary requirements of the submersible.

[0026] The above embodiments can be implemented, in whole or in part, by software, hardware, firmware, or any other combination thereof. When implemented in software, the above embodiments can be implemented, in whole or in part, as a computer program product. Those skilled in the art will recognize that the units and algorithm steps of the various examples described in conjunction with the embodiments disclosed herein can be implemented in electronic hardware, or a combination of computer software and electronic hardware. Whether these functions are implemented in hardware or software depends on the specific application and design constraints of the technical solution.

[0027] The units described as separate components may or may not be physically separate. The components shown as units may or may not be physical units; they may be located in one place or distributed across multiple network units. Some or all of the units can be selected to achieve the purpose of this embodiment, depending on actual needs.

[0028] The above description is merely a specific embodiment of this application, but the scope of protection of this application is not limited thereto. Any changes or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in this application should be included within the scope of protection of this application.

Claims

1. An autonomous obstacle avoidance method for an unmanned underwater vehicle in a longitudinal underwater scenario, characterized in that, The method includes: Collect all data from the entire underwater longitudinal navigation cycle and perform dataset construction and spatiotemporal synchronized standardization preprocessing. The full data includes time-series reference data, multi-source environmental perception data, submersible self-excited disturbance monitoring data, aquatic organism behavior response data, and ecological protection prior data. Deep features of the longitudinal navigation scenario are extracted from the preprocessed full data to generate multi-dimensional causal correlation quantitative features corresponding to the full-link risks of the submersible. The full-link risks include secondary risks of submersible self-excited disturbances, implicit risks of environment-navigation coupling, obstacle collision risks, secondary risks of obstacle avoidance actions, and risks of aquatic ecological disturbances. An initial decision-making model based on causal reasoning embedded with prior physical laws of underwater navigation was constructed. Quantitative features were input into the initial decision-making model. Supervised training, causal alignment, and iterative optimization of the model were carried out using time-series sample data labeled with risk causal chains, disturbance mitigation effects, and ecological impact levels to generate an autonomous obstacle avoidance decision-making model. Simultaneous preprocessing and feature extraction are performed on all data collected during longitudinal navigation to generate real-time multi-dimensional causal correlation quantitative features, which are then input into the optimized autonomous obstacle avoidance decision model. The model outputs full-link risk assessment values, disturbance mitigation control parameters, ecological compliance obstacle avoidance decision parameters, and secondary risk prediction values. By combining a dual-threshold decision interval that is dynamically adapted to the current navigation scenario, a graded navigation control command matching the risk level of the navigation scenario is generated and executed; feedback data after the execution of the graded navigation control command is collected, and the autonomous obstacle avoidance decision model is subjected to online incremental optimization and dynamic calibration of control parameters. The dual-threshold decision interval includes an upper threshold and a lower threshold.

2. The autonomous obstacle avoidance method for an unmanned underwater vehicle in a longitudinal underwater scene according to claim 1, characterized in that, Dataset construction and spatiotemporal synchronization standardization preprocessing include: Based on the cumulative mileage of the unmanned underwater vehicle's longitudinal voyage, a spatiotemporal reference coordinate system with mileage-time dual mapping is established. To address the differences in sampling frequencies of multi-source sensors, spatiotemporal misalignment caused by voyage displacement, and detection distortion caused by changes in vertical depth, mileage matching interpolation is used for time sequence alignment and spatial coordinate registration. Based on the spatiotemporal reference coordinate system, the synchronization accuracy of each sensor's data is verified, substandard data segments are removed, and the validity of the data is verified and invalid data is removed. Based on the underwater vehicle disturbance evolution model and water environment parameters, distortion correction of sensing data is performed. The corrected data is then subjected to outlier removal, null value completion, and adaptive denoising. The dataset is divided into training, validation, and test sets according to the navigation mileage segment and scenario type. The training set samples are labeled with risk causal chains, disturbance mitigation effects, and ecological impact level labels to construct a standardized time series dataset.

3. The autonomous obstacle avoidance method for an unmanned underwater vehicle in a longitudinal underwater scene according to claim 1, characterized in that, The extraction of causal correlation quantitative features for secondary risks caused by self-excited disturbances of underwater vehicles and implicit risks arising from environment-navigation coupling includes: Based on the preprocessed monitoring data of self-excited disturbances of the submersible and historical navigation data, the disturbance diffusion characteristics, temporal evolution laws and the impact parameters of the sensing system under different navigation states are extracted. The self-interference temporal cumulative impact coefficient and the reliability quantification parameters of the sensing system are calculated, and the causal correlation quantification characteristics corresponding to the secondary risks of self-excited disturbances of the submersible are generated. Based on the preprocessed environmental perception data, the terrain feature parameters and water environment field distribution data of the navigation path are extracted. The time-series evolution law of the environmental field is simulated based on the fluid dynamics numerical model and the hidden risk areas are identified. Combined with the underwater vehicle navigation dynamics model, the environment-navigation coupling risk interval is predicted, the coupling risk coefficient is calculated, and the causal correlation quantitative characteristics corresponding to the environment-navigation coupling hidden risks are generated.

4. The autonomous obstacle avoidance method for an unmanned underwater vehicle in a longitudinal underwater scene according to claim 3, characterized in that, The extraction of causal correlation quantitative features corresponding to obstacle collision risk and secondary risks of obstacle avoidance actions includes: For the preprocessed obstacle perception data, the spatial features, temporal motion features and modal features of the obstacles are extracted. The obstacle motion path is predicted based on the multi-model filtering algorithm, collision-related feature parameters are calculated, and the target type is distinguished and the collision threat weight is matched in combination with the underwater vehicle navigation parameters to generate causal correlation quantitative features corresponding to obstacle collision risk. Based on the preprocessed underwater vehicle navigation status data, the navigation parameter adjustment amounts corresponding to the preset obstacle avoidance actions are extracted by combining the underwater vehicle dynamics model and the disturbance evolution model. The secondary risk coefficient, mission adaptability parameter and coupling instability risk coefficient of different obstacle avoidance actions are calculated, and the causal correlation quantitative features corresponding to the secondary risks of obstacle avoidance actions are generated.

5. The autonomous obstacle avoidance method for an unmanned underwater vehicle in a longitudinal underwater scene according to claim 1, characterized in that, The extraction of quantitative characteristics of causal relationships corresponding to aquatic ecological disturbance risks includes: Based on the preprocessed aquatic organism behavior response data and ecological protection prior data, we extract the characteristics, activity patterns, stress thresholds of aquatic organisms around the navigation area, as well as the protection level, control boundaries, and control rules of ecologically sensitive areas. Based on the aquatic organism stress response model, we calculate the degree of disturbance to aquatic organisms and the range of habitat interference caused by the current navigation status, sensing equipment operating parameters, and preset obstacle avoidance actions, and generate single-action ecological disturbance causality coefficients. By combining species stress thresholds, hard constraint parameters for ecological compliance are defined; by integrating aquatic biological characteristics, single-action ecological disturbance causality coefficients and hard constraint parameters for ecological compliance, causal relationship characteristics of navigation actions and ecological impacts are constructed, and quantitative causal relationship characteristics corresponding to aquatic ecological disturbance risks are generated. Among them, the species stress threshold refers to the critical quantitative indicator that core protected species change from normal physiological / behavioral state to produce irreversible negative effects when subjected to external human disturbances.

6. The autonomous obstacle avoidance method for an unmanned underwater vehicle in a longitudinal underwater scene according to claim 1, characterized in that, The steps for building, training, and optimizing the autonomous obstacle avoidance decision-making model are as follows: Construct a causal inference initial decision model that includes a temporal feature input layer, a causal feature fusion layer, a physical law constraint layer, a causal inference decision layer, and a multi-objective output layer; Multi-dimensional causal correlation quantitative features are input into the time-series feature input layer. In the causal feature fusion layer, based on the time-series correlation of the risk causal chain, risk features of different dimensions are adaptively weighted and fused to obtain the risk causal transmission path. In the physical law constraint layer, the physical prior laws of fluid mechanics, navigation dynamics, disturbance evolution, and ecological response related to underwater navigation are embedded to perform physical feasibility verification and correction on the parameters output by the decision layer. A multi-dimensional composite loss function is constructed, which includes risk prediction, physical constraints, disturbance resolution, and ecological constraints. Model supervised training and causal alignment are performed through labeled samples. The generalization of the model is enhanced through boundary samples and adversarial samples. The model hyperparameters are tuned and the performance is verified through validation and test sets. In view of the computing power constraints of the embedded platform of the submarine, the trained model is lightweighted and iteratively optimized to generate a vertical scene-specific autonomous obstacle avoidance decision model.

7. The autonomous obstacle avoidance method for an unmanned underwater vehicle in a longitudinal underwater scene according to claim 6, characterized in that, The generation of end-to-end risk assessment values ​​and corresponding decision parameters includes: The real-time multi-dimensional causal correlation quantitative features are input into the autonomous obstacle avoidance decision model. The causal feature fusion layer performs time-series causal correlation weighted fusion, and the physical feasibility of the decision parameters is verified by the physical law constraint layer. Based on the current navigation mission priority, ecological control requirements, and environmental complexity, the fusion weight of risk features of each dimension is dynamically adjusted, and risk reasoning and decision optimization are performed by the causal reasoning decision layer. The multi-objective output layer outputs the full-link risk assessment value, disturbance reduction control parameters, ecological compliance obstacle avoidance decision parameters, and secondary risk prediction value. When the end-to-end risk assessment value is in the transition range of the dual-threshold decision range, or the secondary risk prediction value exceeds the preset threshold, the multi-frame temporal feature fusion re-inference process is triggered. If the decision confidence requirement is not met even after reaching the maximum number of re-inferences, the security degradation plan is triggered.

8. The autonomous obstacle avoidance method for an unmanned underwater vehicle in a longitudinal underwater scene according to claim 1, characterized in that, The dynamic adaptation of the dual-threshold decision interval and the generation and execution of hierarchical navigation control commands include: The safety redundancy coefficient is determined based on the current navigation mission priority. The ecological control requirements, environmental complexity, and effective detection range of the sensing equipment in the current navigation scenario are collected as dynamic adjustment input parameters. Through a preset threshold mapping model, combined with the safety redundancy coefficient and input parameters, the upper and lower thresholds of the dual threshold decision interval are calculated. When the navigation scenario and mission priority change abruptly, the thresholds are updated and verified in real time. The end-to-end risk assessment value is matched with the dynamic dual-threshold decision range, and combined with the corresponding control parameters and predicted values, a graded navigation control command corresponding to the risk level is generated. When the end-to-end risk assessment value is greater than or equal to the upper threshold, an ecological compliance obstacle avoidance command and disturbance mitigation control are generated and executed; when the upper threshold is greater than or equal to the end-to-end risk assessment value and the lower threshold is between, a navigation warning command is generated, the risk assessment results are updated and alternative obstacle avoidance schemes are pre-generated; when the end-to-end risk assessment value is less than or equal to the lower threshold, a low-disturbance steady-state navigation maintenance command is generated and executed, the ecological response and self-excited disturbance level are continuously monitored, and navigation parameters are dynamically adjusted.

9. The autonomous obstacle avoidance method for an unmanned underwater vehicle in a longitudinal underwater scene according to claim 8, characterized in that, The method also includes sensor anomaly fault tolerance handling and navigation return after obstacle avoidance, specifically including: The system monitors the data output status of multiple sensors in real time. When a single sensor data interruption, missing data, or continuous anomalies are detected, replacement data is generated based on the historical data of that sensor and the synchronous data of other related sensors through a spatiotemporal correlation completion model. When more than a preset number of sensors are detected to be abnormal at the same time, a safety degradation command is triggered, the navigation speed is reduced, a low-disturbance, low-energy navigation mode is activated, and fault self-check and emergency obstacle avoidance plans are triggered simultaneously. When the end-to-end risk assessment value falls below the lower limit threshold and remains below the preset duration, the risk is determined to be eliminated, and the navigation return process is initiated. Based on the current position of the submersible, the preset mission path, and surrounding environmental data, with the optimization objectives of minimum path deviation, minimum disturbance impact, and minimum power consumption, a longitudinal and vertical coupled regression path is generated. Regression control is executed and disturbance resolution control is executed simultaneously. After returning to the preset path, the system switches to a low-disturbance steady-state navigation mode.

10. The autonomous obstacle avoidance method for an unmanned underwater vehicle in a longitudinal underwater scene according to claim 1, characterized in that, Online incremental optimization and full-cycle management of the autonomous obstacle avoidance decision-making model include: Data on the underwater vehicle's navigation status, environmental risk changes, and aquatic organism behavior response after the execution of graded navigation control commands are collected as feedback data. Risk prediction bias, disturbance mitigation effect bias, and ecological impact bias are extracted from the feedback data. After data cleaning and labeling, incremental training samples are constructed. Based on incremental training samples, the decision layer weights of the autonomous obstacle avoidance decision-making model are adjusted, and the fixed parameters of the feature extraction layer and physical law constraint layer of the model are frozen. The adjusted model is verified for risk prediction accuracy, decision physical feasibility and ecological compliance through a preset validation set. If the verification is passed, the model is updated; otherwise, if the verification fails, the original model parameters are retained. Based on feedback data throughout the entire navigation cycle, the threshold mapping model, feature fusion weight allocation logic and disturbance resolution control parameters are continuously optimized.