An unmanned underwater vehicle cross-medium driving control method based on multi-modal perception

Abstract

The application discloses a kind of unmanned underwater vehicle cross-medium drive control methods based on multi-modal perception, belong to underwater unmanned vehicle motion control technical field.The method includes: real-time acquisition motion state, medium parameter, motor data and environmental image;Multi-modal cross-medium working condition identification model is constructed, multi-scale edge feature extraction is carried out to image, physical sensor and visual feature are fused, and different working condition states are distinguished;High elasticity motor drive control model is constructed, disturbance is estimated based on nonlinear disturbance observer and fed forward into sliding mode controller, and motor speed is adaptively adjusted after thrust distribution;When recognizing to the state of water, trigger software and hardware dual electrical seal protection;After recognizing to the state of water, thrust is smoothly transitioned by nonlinear soft start function.The application solves the problem of hardware safety and accurate control under cross-medium working condition through the cooperation of multi-modal working condition identification and cross-medium protection mechanism.

Description

Technical Field

[0001] This invention belongs to the field of motion control technology for underwater unmanned vehicles, specifically relating to a cross-medium drive control method for unmanned underwater vehicles based on multimodal perception. Background Technology

[0002] Unmanned underwater vehicles (UUVs) are core equipment for performing tasks such as marine exploration, deep-sea operations, and underwater reconnaissance. In complex marine environments, UUVs face two main challenges: first, the impact of strong nonlinear ocean current disturbances on precise navigation control; and second, the hardware safety risks arising from frequent crossings of the water-air interface during near-surface operations or extreme sea states. Currently, UUV motion control primarily employs proportional-integral-derivative (PID) control, sliding mode control, or adaptive control methods. These methods can achieve basic course maintenance and trajectory tracking in steady-state deep-water environments. However, under strong nonlinear ocean current disturbances, the parameter tuning of traditional PID control depends on specific operating conditions, and control accuracy decreases significantly when the flow field changes considerably. Conventional sliding mode control requires increased switching gain to suppress disturbances under strong disturbances, leading to severe thruster chattering, accelerated thruster wear, and increased energy consumption. More critically, existing control methods generally assume that the UUV is always in homogeneous water, neglecting the underlying physical limits under cross-medium conditions. When an unmanned underwater vehicle (UUV) is impacted by waves or undergoes an emergency surfacing that causes its propeller to detach from the water surface, the propulsion unit instantly loses its hydrodynamic load. If the control algorithm continues to output thrust commands according to underwater logic, this will cause the propulsion motor speed to spike beyond its limit, resulting in bearing burnout or permanent magnet demagnetization. Simultaneously, the integral term in the controller accumulates errors during sudden changes in the medium, leading to integral depth saturation and causing lag or even complete loss of control in the actuators. When the UUV re-enters the water, the high-speed rotating, unloaded propeller suddenly encounters the reverse drag torque of the high-density water. The transient reverse shear force of hundreds of Newton-meters is sufficient to break the drive shaft or shatter the blades, causing irreversible mechanical damage to the propulsion system. Current control methods lack effective active identification and protection mechanisms for these hardware damage risks under cross-medium operating conditions.

[0003] Therefore, providing a control method that can achieve high-precision anti-disturbance cruise under complex ocean current disturbances, while actively sensing medium phase changes and effectively protecting the propulsion system during cross-medium transitions, has become a technical problem that urgently needs to be solved in this field. Summary of the Invention

[0004] The purpose of this invention is to provide a cross-medium drive control method for unmanned underwater vehicles based on multimodal perception, comprising the following steps: Step S100: Real-time acquisition of motion status data, multi-source medium parameters, motor operating parameters, and environmental image data based on forward vision sensor of unmanned underwater vehicle; Step S200: Construct a multimodal cross-medium working condition recognition model, extract multi-scale edge features from the collected environmental image data, fuse physical sensor data and visual feature data, and determine in real time whether the unmanned underwater vehicle is in deep-water cruising, moment of water emergence, complete water emergence, surface navigation or moment of water entry. Step S300: Construct a highly elastic motor drive control model. In underwater cruising mode, estimate the lumped environmental disturbance in real time based on the nonlinear disturbance observer. Inject the disturbance estimate into the sliding mode controller in the form of feedforward. The controller output is then distributed with thrust to adaptively adjust the motor speed. Step S400: When the instantaneous or complete water discharge state is detected, the dual electrical blockade protection of software and hardware is triggered. The software layer freezes the controller integral and clears the thrust command, and the hardware layer blocks the PWM output. In step S500, after recognizing the instantaneous state of entering the water, the PWM lockout is released, and a nonlinear soft-start function is used to smoothly transition the thrust and restore continuous and stable navigation.

[0005] Further, step S100 includes: Step S110: Configure a multi-channel analog-to-digital converter using the microcontroller's internal direct memory access controller to synchronously acquire phase current signals and bus voltage signals from multiple brushless DC motors at a sampling frequency of not less than 10kHz. Suppress PWM switching noise using a hardware low-pass filter circuit. , , These represent the currents in each phase. Indicates the bus voltage; Step S120: Read the three-axis acceleration, three-axis angular velocity, and magnetometer data from the inertial measurement unit via the integrated circuit's built-in bus I²C or the serial peripheral interface SPI using a high-frequency timer interrupt, and combine this with the absolute pressure value of the current water depth transmitted back from the depth gauge. and gauge pressure value; In step S130, the forward vision sensor captures an image of the foreground environment in global exposure mode and transmits the original RGB image frame sequence to the task planning layer computing node via Ethernet or USB 3.0 interface.

[0006] Furthermore, step S200 involves multi-scale edge feature extraction of the environmental image data, including: Step S210 involves multi-scale edge information selection feature extraction from the environmental image data, including extracting the enhanced image... With a set of different scale factors Convolution with a two-dimensional Gaussian kernel function generates a multi-scale smooth image space. Where the scale levels i = 1, 2, ..., N, N ≥ 3; within each scale space, a manipulable filter is used to calculate the gray-level gradient magnitudes in the horizontal, vertical, and diagonal directions respectively. With gradient direction angle An adaptive weight allocation strategy based on information entropy is adopted to sum the edge response amplitudes at each scale using weighted averages, resulting in an enhanced multi-scale edge feature map. ; Among them, weight Local information entropy of edge response at this scale Proportional; Step S220: The physical sensor data is fused with the visual feature data extracted in step S210. A hidden Markov model is used for probability fusion and state estimation. The physical sensor data includes the depth gauge absolute pressure value. and its differential rate of change Z-axis acceleration of the inertial measurement unit; root mean square value of multi-channel phase current of the motor; Step S230: Based on the fused multimodal data, determine the operating condition according to the following stringent criteria. Condition A, Deep-water cruise: Depth gauge absolute pressure value Greater than the safety threshold The root mean square value of the motor phase current remained within the normal fluid load range, and no water-air interface contour was detected in the visual edge feature map. Condition B, instant of water exit: rate of change of absolute pressure difference in depth gauge The indicator rises rapidly, and the thruster motor phase current is within the time window. Internal drop exceeds threshold Furthermore, the outline of the water vapor interface in the visual image crosses the center pixel line within the vertical field of view. Condition C, complete water discharge: The state of water discharge lasts for more than the lag time at the moment of discharge. Afterward, the motor phase current drops to the no-load maintenance current baseline, and the depth gauge pressure approaches zero; Condition D, instant of water entry: The inertial measurement unit detects a short-duration high-frequency shock wave peak in the Z-axis acceleration. When the depth gauge pressure jumps from zero, the visual sensor image experiences high-frequency texture distortion or a step drop in global illumination. Condition E, Surface navigation: The depth gauge pressure exhibits periodic sinusoidal oscillations near zero, the pitch angle fluctuation of the inertial measurement unit exceeds the preset pitch threshold, and the roll angle fluctuation exceeds the preset roll threshold.

[0007] Further, step S300 includes: Step S310: Establish the six-degree-of-freedom dynamic differential equations of the unmanned underwater vehicle in both the body coordinate system and the geodetic coordinate system. ; In the formula, Let be the vectors of linear velocity and angular velocity of the unmanned underwater vehicle in its body coordinate system; Let M be the position and Euler angle attitude vector of the unmanned underwater vehicle in the geodetic coordinate system; M is the system inertial matrix, which includes the rigid body mass inertial matrix and the hydrodynamic added mass matrix, satisfying... ; The matrix of the Coriolis force and the centrifugal force satisfies the antisymmetric property. ; The hydrodynamic damping matrix includes linear and nonlinear damping, and satisfies the following conditions: ; This is the vector of restoring force and restoring moment caused by the mismatch between gravity and buoyancy; To control the input generalized force and torque vector; This represents the lumped environmental disturbance vector; Step S320: Real-time estimation based on nonlinear disturbance observer Define intermediate auxiliary state vectors ,in For environmental disturbance The estimated value, Define the nonlinear auxiliary vector function to be designed; define the observer gain matrix. By configuring a positive definite symmetric gain matrix To ensure observation error It converges to zero at an exponential rate; Step S330, define the system speed tracking error ,in To obtain the desired velocity command, design an integral sliding surface function. ; in Given a positive definite integral gain matrix; construct a hybrid robust control law that includes equivalent control, switching control, and feedforward compensation from a nonlinear disturbance observer: ; In the formula, Used to precisely offset the low-frequency dominant components of ambient ocean currents; For terms that approximate the proportion; For robust switching, For boundary layer saturation function, This refers to the boundary layer thickness parameter; Step S340, configure the matrix via thrust. Where n is the total number of thrusters, and the generalized control input is... With the target thrust vector of each motor Establish mapping relationships, ; The optimal thrust of each motor was calculated using the pseudo-inverse matrix method.

[0008] Further, step S400 includes: The software-layer anti-saturation compensator calculates the deviation between the current commanded thrust and the physical limit of the thruster in reverse. After multiplying the deviation by the convergence gain, it is injected into the controller's integral state equation, forcibly freezing the accumulation evolution of the integral term, and simultaneously forcibly resetting the target thrust command sequence to zero. After receiving a hard real-time interrupt signal indicating a sudden drop in phase current, the hardware-level microcontroller immediately utilizes the braking circuit breaker feature of the advanced timer to force the PWM output of all bridge arms to a preset low-level or high-impedance safe state within a single instruction cycle.

[0009] Furthermore, the nonlinear soft-start function in step S500 is: ; In the formula The thrust command varies over time. Let t be the target thrust, and t be the continuous recovery time calculated from the triggering of the water entry flag. To determine the relative vertical velocity at the moment of entry into the water The time constant is adaptively adjusted; the greater the impact velocity upon entering the water, Automatic increase makes the thrust build-up curve smoother; when Restored to no less than When the water level reaches 95%, the system switches back to the deep-water anti-interference cruise mode in step S300.

[0010] Furthermore, in step S230, a switching dead zone based on the principle of hysteresis comparator is set between adjacent operating conditions. The high-frequency spikes of the sensor signal at the threshold boundary between water outlet and water inlet are filtered out by using a historical time window counter to avoid high-frequency oscillation switching of the state machine. If the contrast of the feature map of the visual sensor is lower than the effective confidence threshold due to extreme turbidity or no light at night, the system automatically cuts off the visual feature weights and relies only on the cross-verification of the Z-axis impulse feature of the inertial measurement unit and the change of motor phase current to maintain the closed loop of operating condition judgment.

[0011] Furthermore, it also includes step S600, The actuator driver layer parses instructions from the upper-level operating system based on the underlying microcontroller and uses a field-oriented control algorithm to control the brushless motor of the thruster: extracting the two-phase current and transforming it into a stationary orthogonal coordinate system via Clark transformation. and The rotor electrical angles are converted into quadrature-axis currents in a rotating coordinate system using Park transformation. and direct-axis current Adjust according to desired thrust feedback The target value is achieved by using a cascaded proportional-integral controller to realize closed-loop control of the inner current loop and the outer speed loop. Six PWM waveforms driving the inverter are generated through inverse Park transformation and space vector pulse width modulation. The underlying hard real-time response cycle is controlled within 100 microseconds. Steps S100 to S500 are executed cyclically in the next control cycle.

[0012] Furthermore, steps S100 to S600 are deployed in a hierarchical control architecture that includes a task planning layer, a motion control layer, and an actuator driving layer; the task planning layer runs the MOOS operating system and is responsible for multimodal data fusion and cross-media condition identification; the motion control layer is responsible for nonlinear disturbance observation and robust control law calculation; the actuator driving layer is based on a microcontroller and is responsible for motor field orientation control and hardware-level safety protection.

[0013] Furthermore, a state transition dead zone based on the hysteresis comparator principle is set between steps S400 and S500, and a window counter is used to prevent the microcontroller from switching between PWM blocking and nonlinear soft start at high frequency.

[0014] Compared with the prior art, the present invention has the following advantages: (1) By constructing a multimodal cross-medium operating condition recognition model, the visual multi-scale edge features and physical sensor data are fused using hidden Markov probability, achieving millisecond-level accurate discrimination of five operating conditions: deep-water cruising, instant of water exit, complete water exit, surface navigation, and instant of water entry. When the visual sensor fails in extremely turbid water or under no-light conditions at night, the system automatically switches to an adaptive degradation fault-tolerant mode, relying solely on the cross-verification of the high-frequency impulse characteristics of the inertial measurement unit and the changes in the motor phase current to maintain the judgment closed loop. This overcomes the defect of relying solely on the depth gauge, which is susceptible to interference from wave fluctuations and causes misjudgment of the medium, and provides a reliable triggering basis for subsequent cross-medium protection mechanisms.

[0015] (2) By constructing a highly elastic motor drive control model that integrates a nonlinear disturbance observer and an integral sliding mode controller, the lumped environmental disturbance estimated in real time by the nonlinear disturbance observer is injected into the sliding mode controller in the form of feedforward negative compensation. This accurately cancels the low-frequency dominant component in the ocean current disturbance, and the sliding mode switching term only needs to handle the residual reconstruction error of the observer. This effectively reduces the sliding mode switching gain, suppresses thruster chattering, and ensures the robustness and tracking accuracy of the system under complex time-varying flow fields. The example test shows that under strong lateral ocean current disturbance in sea state 4, the lateral deviation between the actual trajectory and the desired trajectory is controlled within 0.3 meters, overcoming the defect of traditional sliding mode control that exacerbates thruster wear due to increased switching gain under strong disturbances.

[0016] (3) By triggering a dual electrical blockade protection mechanism of software and hardware at the moment of water exit, the software layer anti-saturation compensator forcibly freezes the controller integral term to eliminate integral depth saturation, and the hardware layer uses the advanced timer braking characteristics of the microcontroller to block the PWM output at the nanosecond level, thus fundamentally eliminating the risk of overspeed runaway and inverter damage caused by loss of load after the propeller leaves the water surface. At the moment of water entry, a nonlinear soft-start function based on exponential decay is introduced, and the thrust command adaptively adjusts the recovery time constant according to the water entry impact speed, eliminating the risk of shear damage to the drive shaft and blades when the high-speed rotating propeller suddenly encounters the reverse torque of hydrodynamics. The synergistic effect of the above water exit protection and water entry recovery mechanism solves the fundamental defect of traditional UUV control methods in lack of active hardware protection under cross-medium conditions. Example tests show that the multimodal operating condition recognition model accurately determines the instantaneous state of the propeller after it leaves the water within 18 milliseconds and triggers protection. No speed over-limit phenomenon occurred during the shutdown process after leaving the water. The instantaneous recognition delay upon entering the water is about 22 milliseconds, and the thrust command smoothly recovers to the target value within 800 milliseconds. There are no abnormal impacts in the propeller current during the entire recovery process, and the torque of the drive shaft is within the safe range.

[0017] (4) By setting a switching dead zone based on the principle of hysteresis comparator between adjacent operating conditions, the high-frequency spikes in the sensor signal caused by water splashing are filtered out by the window counter, thus avoiding high-frequency oscillation switching at the threshold boundary between water outlet and water inlet, ensuring the switching life of power devices and the stability of the control system.

[0018] (5) By combining the MOOS operating system with the STM32 microcontroller, a three-layer hierarchical control architecture of task planning layer, motion control layer and actuator driving layer is established. The actuator driving layer completes closed-loop execution and status feedback in a hard real-time cycle of hundreds of microseconds with the field-oriented control algorithm. This forms a full-chain closed-loop control system from multimodal perception and robust control calculation to hardware safety protection, which significantly improves the survival autonomy and underlying hardware robustness of the unmanned underwater vehicle in extreme sea conditions.

[0019] The present invention will now be further described with reference to the accompanying drawings. Attached Figure Description

[0020] Figure 1 This is a schematic diagram of the multimodal operating condition feature processing flow in an embodiment of the present invention.

[0021] Figure 2 This is a schematic diagram of the hierarchical control and high-frequency communication architecture deployment in an embodiment of the present invention. Detailed Implementation

[0022] To make the above-mentioned objects, features and advantages of the present invention more apparent and understandable, the specific embodiments of the present invention will be described in detail below with reference to the accompanying drawings.

[0023] This invention provides a cross-medium drive control method for unmanned underwater vehicles based on multimodal perception, which is used to solve the problems of accurate control and hardware safety of unmanned underwater vehicles under complex ocean current disturbances and cross-medium conditions.

[0024] Reference Figure 2 This method is deployed in a hierarchical control architecture comprising a task planning layer, a motion control layer, and an actuator driving layer. The task planning layer runs the MOOS operating system and is responsible for multimodal data fusion and cross-media condition identification; the motion control layer is responsible for nonlinear disturbance observation and robust control law calculation; the actuator driving layer uses an STM32F103C8T6 microcontroller as its core and is responsible for motor field-oriented control and hardware-level safety protection.

[0025] Reference Figure 1 , Figure 2 This method includes the following steps that are executed sequentially in a loop.

[0026] Step S100 involves real-time acquisition of the unmanned underwater vehicle's current motion status data, multi-source medium parameters, motor operating parameters, and environmental image data based on the forward-looking vision sensor. This includes steps S110, S120, and S130.

[0027] Step S110: Acquire motor electrical parameters. A multi-channel analog-to-digital converter is configured using the microcontroller's internal direct memory access controller to synchronously acquire phase current feedback signals and bus voltage signals from multiple brushless DC motors at a sampling frequency of at least 10kHz. A hardware low-pass filter circuit is used to suppress PWM switching noise. , , These represent the currents in each phase. This indicates the bus voltage.

[0028] Step S120: Acquire motion state and environmental physical parameters. Read the three-axis acceleration, three-axis angular velocity, and magnetometer data from the inertial measurement unit via I2C or SPI using a high-frequency timer interrupt, and combine this with the absolute pressure value of the current water depth transmitted back from the depth gauge. And gauge pressure value.

[0029] Step S130: Acquire forward-looking visual images. The forward-looking visual sensor captures images of the foreground environment in global exposure mode and transmits the raw RGB image frame sequence to the task planning layer computing node via Ethernet or USB 3.0 interface.

[0030] The synchronous acquisition and high-speed transmission of the aforementioned multi-source data provide a data foundation for subsequent multimodal working condition identification using physical and visual sensors.

[0031] Step S200 involves constructing a multimodal cross-medium operating condition recognition model, integrating physical sensor data and machine vision features to accurately determine the operating condition of the unmanned underwater vehicle during deep-water cruising, the moment of surfacing, full surfacing, surface navigation, or the moment of re-entry into the water. This includes steps S210, S220, and S230.

[0032] Step S210: Multi-scale edge information selection feature extraction is performed on the environmental image data.

[0033] First, image preprocessing and enhancement are performed. Considering the light scattering and absorption attenuation characteristics of the underwater environment, an underwater image restoration algorithm based on dark channel priors is adopted. This algorithm calculates the global background illumination matrix, inverts the transmittance map, and obtains a distortion-free environmental image after dehazing and color correction. .

[0034] Then, a multi-scale Gaussian pyramid is constructed. The enhanced image... With a set of different scale factors Convolution with a two-dimensional Gaussian kernel function generates a multi-scale smooth image space. The scale levels are i = 1, 2, ..., N, where N ≥ 3.

[0035] Next, multi-directional gradient operator response calculations are performed. Within each scale space, a manipulable filter is used to calculate the gray-level gradient magnitudes in the horizontal, vertical, and diagonal directions, respectively. With gradient direction angle .

[0036] Finally, edge feature weighting and fusion are performed. An adaptive weight allocation strategy based on information entropy is adopted to sum the edge response amplitudes at each scale using weights, resulting in the final enhanced multi-scale edge feature map. Weight The local information entropy of the edge response at this scale is proportional. Subsequently, nonmaximum suppression and double threshold hysteresis thresholding are applied to extract continuous and clear water-air interface or wave profiles.

[0037] Step S220: The physical sensor data and machine vision features are fused to construct a multimodal cross-medium working condition recognition model.

[0038] Physical sensor data includes depth gauge absolute pressure values. and its differential rate of change The data includes the Z-axis acceleration of the inertial measurement unit and the root mean square value of the multi-channel phase current of the motor. Visual feature data includes the water-vapor interface contour line extracted in step S210 and its coordinate changes in the vertical field of view.

[0039] Reference Figure 1 The input forward-looking visual feature data first undergoes dimensionality reduction to extract basic local features. Then, it is fed into a channel attention module for adaptive calculation and recalibration of the channel weights of multi-scale features. Finally, a dynamic feature extraction module generates high-dimensional dynamic fusion features. These features are ultimately split into two prediction branches: a classification branch outputs the probability distribution vector of the current field of view medium via a fully connected layer and a Softmax activation function, while a regression branch outputs the target state regression value.

[0040] Hidden Markov models are used to perform probabilistic fusion and state estimation of physical sensor data and visual feature data, so as to accurately capture the cross-medium phase transition boundary.

[0041] Step S230: Based on the fused multimodal data, the operating condition is determined according to the following stringent judgment conditions.

[0042] Condition A, Deep-water cruise: When the absolute pressure value of the depth gauge greater than the set safe draft threshold When the root mean square value of the multi-channel phase current of the motor is maintained within the normal fluid load mapping range, and the visual edge feature map does not detect the water-air interface contour, the state machine is locked in the deep-water cruise state.

[0043] Condition B, instant of water exit: when the absolute pressure differential change rate of the depth gauge... The indicator rises rapidly, and the phase current of any thruster motor is in continuous motion. The threshold was exceeded within the time window. The abrupt drop, coupled with the rapid decrease in coordinates of the water vapor interface outline extracted from the forward-looking visual image within the vertical field of view, crossing the center pixel line, triggers the instantaneous state of water emergence.

[0044] Condition C, complete water discharge: The state of water discharge lasts for more than the preset lag time at the moment of discharge. Subsequently, if the motor phase current drops to near the system's no-load maintenance current baseline and the depth gauge pressure approaches zero, the system switches to a fully effluent state.

[0045] Condition D, moment of water entry: When fully out of the water or navigating on the surface, the Z-axis accelerometer of the inertial measurement unit detects the peak of a short-duration, high-frequency shock wave generated due to contact with a high-density water surface. When the depth gauge pressure value jumps from zero, and the visual sensor image experiences a sudden high-frequency texture distortion or a step drop in global illumination, it triggers the instantaneous water entry state.

[0046] Condition E, Surface navigation: When the system is in the critical state of entering and leaving the water, and the depth gauge pressure is oscillating periodically with sinusoidal fluctuations near zero, and the attitude angle of the inertial measurement unit is in the range of violent pitch and roll fluctuations, the system is locked into the surface navigation state.

[0047] A switching dead zone based on the hysteresis comparator principle is set between adjacent operating states. If the environmental image or sensor judgment data undergoes high-frequency jumps at the boundary between the water outlet and water inlet thresholds due to water splashing, the state machine uses this state transition dead zone and a historical time window counter to filter out glitch signals and avoid high-frequency oscillation switching between adjacent states.

[0048] If the forward vision sensor fails to extract multi-scale edge information due to extreme conditions such as silt agitation or lack of light at night, i.e. the feature map contrast is lower than the effective confidence threshold, the system adopts an adaptive degradation fault tolerance mechanism: automatically cuts off the visual feature weights, and relies solely on the cross-verification of the high-frequency Z-axis impulse features of the inertial measurement unit and the real-time phase current changes of the motor to continue to maintain the closed loop for judging the physical phase change boundary at the moment of water exit and water entry.

[0049] The transition relationships between the above five operating states are as follows: In deep-water cruising state, when the instantaneous water exit condition is met, the state machine switches to the instantaneous water exit state; in the instantaneous water exit state, when the complete water exit condition is met, the state machine switches to the complete water exit state; in the complete water exit state or surface navigation state, when the instantaneous water entry condition is met, the state machine switches to the instantaneous water entry state; in the instantaneous water entry state, after the thrust smooth transition is completed, the state machine automatically switches to deep-water cruising state; when in the critical water exit / entry state and the surface navigation condition is met, the state machine enters the surface navigation state.

[0050] Step S300 involves constructing a highly resilient motor drive control model. Under underwater cruising conditions, environmental disturbances are estimated based on a nonlinear disturbance observer, and the target thrust and motor speed are adaptively adjusted in real time. This includes steps S310, S320, S330, and S340.

[0051] Step S310: Establish a six-degree-of-freedom kinematic and nonlinear dynamic model of the unmanned underwater vehicle. In the body coordinate system and the geodetic coordinate system, the dynamic differential equations are expressed as: ; In the formula, Let be the vectors of linear velocity and angular velocity of the unmanned underwater vehicle in its body coordinate system; It is the first derivative of the velocity vector with respect to time, i.e., the acceleration vector in the body coordinate system; Let M be the position and Euler angle attitude vector of the unmanned underwater vehicle in the geodetic coordinate system; M is the system inertial matrix, which includes the rigid body mass inertial matrix and the hydrodynamic added mass matrix, satisfying... ; The matrix of the Coriolis force and the centrifugal force satisfies the antisymmetric property. ; The hydrodynamic damping matrix is ​​a superposition of linear and nonlinear damping, satisfying the following conditions: ; This is the vector of restoring force and restoring moment caused by the mismatch between gravity and buoyancy; The generalized force and torque vectors are mapped to the center of mass of the body for the control inputs of each propeller thruster; This is a lumped environmental disturbance vector that integrates unknown and complex ocean current disturbances, model parameter perturbations, and unmodeled high-frequency dynamics.

[0052] Step S320: Real-time estimation of lumped environmental disturbances based on a nonlinear disturbance observer. .

[0053] To avoid introducing extremely noisy acceleration terms into the observer Define an intermediate auxiliary state vector z, let ,in For environmental disturbance The estimated value, Let be the nonlinear auxiliary vector function to be designed. Define the observer gain matrix. Satisfying the relationship .

[0054] A dynamic update equation for a nonlinear disturbance observer is constructed. This is achieved by configuring a positive definite symmetric gain matrix. To ensure that the dynamic equation of the observation error satisfies the asymptotic stability condition, the disturbance estimation error is reduced. The difference converges to zero at an exponential rate, thus providing a real-time and smooth approximation of external ocean current disturbances. The estimation results... A feedforward injection is used in the closed-loop controller to precisely counteract the low-frequency dominant components of ambient ocean currents.

[0055] Step S330: Construct a robust control law based on the integral sliding mode control architecture, and fuse the feedforward estimation of the nonlinear disturbance observer with the sliding mode controller.

[0056] Define system speed tracking error ,in The desired velocity command is calculated by the task planning layer. A sliding surface function containing proportional and integral terms is designed. ; in It is the positive definite integral gain matrix.

[0057] Construct a hybrid robust control law that includes equivalent control law, switching control law, and feedforward compensation from a nonlinear disturbance observer: ; In the formula, It is the first derivative of the desired velocity vector with respect to time, i.e., the desired acceleration vector in the body coordinate system; Used to precisely offset the low-frequency dominant components of ambient ocean currents; As a proportional approaching term, it drives the system state to converge toward the sliding surface; For robust switching terms used to suppress residual reconstruction errors of the observer and unmodeled high-frequency dynamics, where This is a boundary layer saturation function used to reduce thruster chattering caused by the sign function. This is the boundary layer thickness parameter.

[0058] The control signal transmission link of the above hybrid robust control law is as follows: the desired speed command given by the task planning layer. The actual speed fed back by the inertial measurement unit After comparison, the speed tracking error is obtained. Velocity tracking error via integral sliding mode surface function After processing, they were respectively processed through the proportional approximation term. and robust switching The control component is generated; simultaneously, the nonlinear disturbance observer utilizes the control input at the current moment. and actual motion state feedback Real-time estimation of lumped environmental disturbances The control law is injected in the form of feedforward compensation, and superimposed with the equivalent control component and the switching control component to generate a generalized control input. The generalized control input is mapped to the target thrust command for each thruster via the thrust allocation matrix B. The data is sent to the actuator drive layer to drive the thruster; the actual thrust output by the thruster acts on the unmanned underwater vehicle, and the inertial measurement unit collects the motion status in real time and feeds it back to the motion control layer, forming a closed loop.

[0059] Step S340: The generalized control input is distributed through the thrust allocation matrix. The target thrust space is mapped to each thruster.

[0060] Thrust configuration matrix determined by the thruster installation configuration of the unmanned underwater vehicle Where n is the total number of thrusters, and the generalized control input is... With the target thrust vector of each motor Establish mapping relationship: ; This can be achieved by solving a quadratic programming problem with a physical output upper limit constraint or by using the pseudo-inverse matrix method. The optimal required thrust to be allocated to each motor is calculated.

[0061] When the multimodal cross-medium operating condition identification model determines that the system is in a surface navigation state, it enters a semi-submersible derating control mode. In this mode, a maximum thrust limit is introduced when solving the thrust distribution matrix, setting the target thrust upper limit for each propeller to 40% to 60% of the rated thrust. At the same time, the integral term in the integral sliding mode controller is turned off, and only the proportional switching control is retained to avoid the accumulation of periodic velocity errors caused by wave fluctuations, which could lead to controller output saturation. The operating condition detection window duration is shortened to half that of the deep-water cruise state, allowing for more frequent monitoring of whether the propeller is approaching the critical state of exiting the water. Once the exit condition is triggered, the dual hardware and software electrical lockout protection of step S400 is immediately executed.

[0062] Step S400 integrates water discharge shutdown and anti-runaway protection mechanisms, establishing an overspeed barrier at both the application software layer and the microcontroller hardware layer.

[0063] When the state machine detects the "water exit instant" flag, the software layer immediately triggers the anti-saturation compensator. The anti-saturation compensator calculates the deviation between the current commanded thrust and the thruster's physical limits, multiplies this deviation by the convergence gain, and injects it into the integral state equations of all regulators and sliding mode controllers. This forcibly freezes the cumulative evolution of integral terms, thus avoiding integral depth saturation caused by a surge in position and velocity errors across the medium. Simultaneously, the target thrust command sequence is sent to the actuator drive layer. Force reset to zero.

[0064] At the hardware level, upon receiving a hard real-time interrupt signal indicating a sudden drop in phase current, the microcontroller immediately utilizes the braking circuit breaker feature of the advanced timer without waiting for confirmation from the upper-level state machine. By configuring the dead time of the complementary PWM output and the braking register, once the trigger condition is met, the PWM output of all bridge arms is directly forced to a preset low-level or high-impedance safe state within a single instruction cycle. This achieves nanosecond-level electrical blocking of the brushless motor drive inverter, ensuring that the propeller speed safely returns to zero after kinetic energy inertia dissipation.

[0065] The aforementioned software-layer integral freezing and hardware-layer braking block constitute a dual electrical block protection mechanism. The software layer is responsible for preventing controller state degradation, while the hardware layer is responsible for implementing the highest priority physical safety disconnection.

[0066] Step S500 integrates an automatic recovery mechanism for water entry drive, enabling flexible connection and smooth transition of thrust after the unmanned underwater vehicle re-enters the water.

[0067] Once the state machine determines that it has entered the "instantaneous water entry" state, the microcontroller first unlocks the hardware-level PWM, allowing the inverter bridge to restart. The upper-level control algorithm then obtains the target thrust required for the current depth cruise. Instead of being transmitted in the form of a step signal, a thrust smoothing transition equation based on exponential decay characteristics is introduced: ; In the formula, t is the continuous recovery time calculated from the triggering of the water ingress flag. Based on the relative vertical velocity of the unmanned underwater vehicle at the moment of entry into the water Real-time adaptive dynamic adjustment of the time constant. This is applied when the water impact velocity is affected by wave impact. When it reaches its maximum value, the adaptive algorithm automatically increases it. The value of this value makes the thrust build-up curve smoother and prolongs the buffer time; if the descent into the water is slow, it reduces... To quickly restore power.

[0068] The underlying brushless motor ESC executes this non-linear thrust command, resulting in a smooth increase in thrust. The built-in thrust-speed-PWM polynomial mapping table is used to convert the thrust into speed into duty cycle commands, driving the propeller to achieve flexible connection of hydrodynamic loads. When t is much greater than... hour, Completely approaching This marks the end of the cross-medium automatic recovery mechanism, and the system seamlessly switches back to the deep-water anti-disturbance cruise mode in step S300.

[0069] A state transition dead zone is set between the water outlet shutdown protection and the water inlet soft start recovery. By utilizing the principle of hysteresis comparator and window counter, high-frequency oscillation switching of the microcontroller between PWM blocking and nonlinear soft start is avoided, thus ensuring the switching life of power devices.

[0070] In step S600, the actuator driver layer parses the instructions issued by the upper-level operating system based on the underlying microcontroller, directly controls the thruster to execute the optimal drive control instructions, and cyclically executes steps S100 to S500 in the next control cycle.

[0071] The actuator drive layer employs a field-oriented control algorithm to control the brushless motor of the propeller thruster. Specifically, the process involves extracting the two-phase current and transforming it into a stationary orthogonal coordinate system using Clark transformation. and The rotor electrical angles obtained from the encoder or back EMF observer are transformed by Park transformation into quadrature-axis currents that are directly parallel to the torque in the rotating coordinate system. and direct-axis current Adjustment based on the analyzed expected thrust feedback The target value is achieved by using a cascaded proportional-integral controller to implement closed-loop control of the inner current loop and the outer speed loop, and generating six PWM waveforms to drive the inverter through inverse Park transform and space vector pulse width modulation. The underlying hard real-time response cycle is controlled to be within 100 microseconds.

[0072] The steps S100 to S600 constitute a complete cross-medium drive control closed loop. Steps S100 and S200 achieve accurate identification of operating conditions through multimodal sensing; step S300 achieves deep-water disturbance-resistant cruise based on the fusion of nonlinear disturbance observer and sliding mode control; steps S400 and S500 implement active protection for two cross-medium hazardous operating conditions: outflow loss of load and inflow impact, respectively; step S600 completes closed-loop execution and status feedback within a hard real-time control cycle of hundreds of microseconds. Data interaction and command transmission between the steps are achieved through a hierarchical control architecture, forming a complete technology chain from sensing and decision-making to execution.

[0073] The method will be described below with reference to specific embodiments.

[0074] Example This embodiment uses a self-developed unmanned underwater vehicle (UUV) performing near-surface reconnaissance missions in sea state 4 as its application scenario. The UUV is equipped with a forward-looking camera, an inertial measurement unit, a depth gauge, and four brushless DC motor-driven propellers. The mission planning layer runs the MOOS operating system, the motion control layer deploys a nonlinear disturbance observer and an integral sliding mode controller, and the actuator driving layer uses an STM32F103C8T6 microcontroller as its core.

[0075] During the mission, the unmanned underwater vehicle encountered strong lateral ocean currents during the deep-water cruise phase. A nonlinear disturbance observer estimated the amplitude and direction of the flow field disturbances in real time, and fed these estimates into the sliding mode controller. The sliding mode controller adaptively adjusted the target thrust of each thruster, calculating the optimal thrust command for each motor using a thrust allocation matrix. The lateral deviation between the actual and desired trajectory was controlled within 0.3 meters, validating the method's ability to withstand disturbances during cruises in complex flow fields.

[0076] As the unmanned underwater vehicle (UUV) surfaces near the water's surface, the propeller periodically emerges from the water due to wave impact. The multimodal operating condition identification model accurately determines the "emergence moment" state within 18 milliseconds after the propeller emerges. The software-level anti-saturation compensator immediately freezes the controller's integral term, while the hardware-level advanced timer brake blocks the PWM output at the nanosecond level, allowing the motor speed to safely decay to zero under inertia. No speed over-limit phenomena occurred during the entire emergence and shutdown process, verifying the effectiveness of the dual electrical blocking protection mechanism.

[0077] When the wave propelled the unmanned underwater vehicle (UUV) back into the water, the operational condition identification model accurately determined the "instantaneous entry" state within 22 milliseconds of contact with the water surface. The hardware-level PWM lockout was released, and the thrust smooth transition equation automatically adjusted the time constant based on the entry impact velocity, smoothly restoring the thrust command to the target value within 800 milliseconds. Throughout the entire entry recovery process, there were no abnormal surges in the thruster current, and the drive shaft torque remained within a safe range, verifying the protective effect of the nonlinear soft-start mechanism on the mechanical transmission system.

[0078] The above are merely preferred embodiments of the present invention and do not limit the scope of the patent. Any equivalent structural modifications made based on the concept of the present invention and the content of this specification, or any direct or indirect applications in other related technical fields, are included within the scope of patent protection of the present invention.

Claims

1. A multi-modal perception based unmanned underwater vehicle cross medium drive control method, characterized in that, Includes the following steps: Step S100: Real-time acquisition of motion status data, multi-source medium parameters, motor operating parameters, and environmental image data based on forward vision sensor of unmanned underwater vehicle; Step S200: Construct a multimodal cross-medium working condition recognition model, extract multi-scale edge features from the collected environmental image data, fuse physical sensor data and visual feature data, and determine in real time whether the unmanned underwater vehicle is in deep-water cruising, moment of water emergence, complete water emergence, surface navigation or moment of water entry. Step S300: Construct a highly elastic motor drive control model. In underwater cruising mode, estimate the lumped environmental disturbance in real time based on the nonlinear disturbance observer. Inject the disturbance estimate into the sliding mode controller in the form of feedforward. The controller output is then distributed with thrust to adaptively adjust the motor speed. Step S400: When the instantaneous or complete water discharge state is detected, the dual electrical blockade protection of software and hardware is triggered. The software layer freezes the controller integral and clears the thrust command, and the hardware layer blocks the PWM output. In step S500, after recognizing the instantaneous state of entering the water, the PWM lockout is released, and a nonlinear soft-start function is used to smoothly transition the thrust and restore continuous and stable navigation.

2. The multi-modal perception based cross-medium driving control method for an unmanned underwater vehicle according to claim 1, wherein, Step S100 includes: Step S110, configure the multi-channel analog-to-digital converter through the direct memory access controller inside the microcontroller to synchronously collect the phase current signals and bus voltage signals of the plurality of brushless direct current motors at a sampling frequency of no less than 10 kHz, and inhibit PWM switching noise through a hardware low-pass filter circuit, wherein 、 、 respectively represent the phase currents, represent the bus voltage; Step S120, read the three-axis acceleration, three-axis angular velocity and magnetometer data of the inertial measurement unit through the integrated circuit built-in bus I2C or serial peripheral interface SPI in high-frequency timer interrupt mode, and combine the absolute pressure value of the current water depth returned by the depth gauge and the gauge pressure value; In step S130, the forward vision sensor captures an image of the foreground environment in global exposure mode and transmits the original RGB image frame sequence to the task planning layer computing node via Ethernet or USB 3.0 interface.

3. The multi-modal perception based cross-medium driving control method for an unmanned underwater vehicle according to claim 1, wherein, Step S200 involves multi-scale edge feature extraction of the environmental image data, including: Step S210, multi-scale edge information selection feature extraction is performed on the environmental image data, including performing convolution on the enhanced image with a set of two-dimensional Gaussian kernel functions with different scale factors to generate a multi-scale smooth image space , where the scale index i = 1, 2, …, N, N ≥ 3; in each scale space, the gray gradient amplitude in the horizontal, vertical and diagonal directions is calculated respectively using a steerable filter and the gradient direction angle ; using an adaptive weight distribution strategy based on information entropy, the edge response amplitudes at each scale are weighted and summed to obtain an enhanced multi-scale edge feature map, ； Among them, weight Calculate according to the following formula, ； The local information entropy of the edge response region at the i-th scale is calculated using the standard image information entropy formula. Step S220: The physical sensor data is fused with the visual feature data extracted in step S210. A hidden Markov model is used for probability fusion and state estimation. The physical sensor data includes the depth gauge absolute pressure value. and its differential rate of change Z-axis acceleration of the inertial measurement unit; root mean square value of multi-channel phase current of the motor; Step S230: Based on the fused multimodal data, determine the operating condition according to the following stringent criteria. Condition A, Deep-water cruise: Depth gauge absolute pressure value Greater than the safety threshold The root mean square value of the motor phase current remained within the normal fluid load range, and no water-air interface contour was detected in the visual edge feature map. Condition B, instant of water exit: rate of change of absolute pressure difference in depth gauge The indicator rises rapidly, and the thruster motor phase current is within the time window. Internal drop exceeds threshold Furthermore, the outline of the water vapor interface in the visual image crosses the center pixel line within the vertical field of view. Condition C, complete water discharge: The state of water discharge lasts for more than the lag time at the moment of discharge. Afterward, the motor phase current drops to the no-load maintenance current baseline, and the depth gauge pressure approaches zero; Condition D, instant of water entry: The inertial measurement unit detects a short-duration high-frequency shock wave peak in the Z-axis acceleration. When the depth gauge pressure jumps from zero, the visual sensor image experiences high-frequency texture distortion or a step drop in global illumination. Condition E, Surface navigation: The depth gauge pressure exhibits periodic sinusoidal oscillations near zero, the pitch angle fluctuation of the inertial measurement unit exceeds the preset pitch threshold, and the roll angle fluctuation exceeds the preset roll threshold.

4. The method for cross-medium drive control of an unmanned underwater vehicle based on multimodal perception according to claim 1, characterized in that, Step S300 includes: Step S310: Establish the six-degree-of-freedom dynamic differential equations of the unmanned underwater vehicle in both the body coordinate system and the geodetic coordinate system. ; In the formula, Let be the vectors of linear velocity and angular velocity of the unmanned underwater vehicle in its body coordinate system; It is the first derivative of the velocity vector with respect to time, i.e., the acceleration vector in the body coordinate system; Let M be the position and Euler angle attitude vector of the unmanned underwater vehicle in the geodetic coordinate system; M is the system inertial matrix, which includes the rigid body mass inertial matrix and the hydrodynamic added mass matrix, satisfying... ; The matrix of the Coriolis force and the centrifugal force satisfies the antisymmetric property. ; The hydrodynamic damping matrix includes linear and nonlinear damping, and satisfies the following conditions: ; This is the vector of restoring force and restoring moment caused by the mismatch between gravity and buoyancy; To control the input generalized force and torque vector; This represents the lumped environmental disturbance vector; Step S320: Real-time estimation based on nonlinear disturbance observer Define intermediate auxiliary state vectors ,in For environmental disturbance The estimated value, Define the nonlinear auxiliary vector function to be designed; define the observer gain matrix. By configuring a positive definite symmetric gain matrix To ensure observation error It converges to zero at an exponential rate; Step S330, define the system speed tracking error ,in To obtain the desired velocity command, design an integral sliding surface function. ； in Given a positive definite integral gain matrix; construct a hybrid robust control law that includes equivalent control, switching control, and feedforward compensation from a nonlinear disturbance observer: ； In the formula, It is the first derivative of the desired velocity vector with respect to time, i.e., the desired acceleration vector in the body coordinate system; Used to precisely offset the low-frequency dominant components of ambient ocean currents; For terms that approximate the proportion; Let be the sliding mode scaling gain matrix, and be a positive definite diagonal matrix; For robust switching, Here, is the sliding mode robust switching gain matrix, and is a positive definite diagonal matrix. For boundary layer saturation function, This refers to the boundary layer thickness parameter; Step S340, configure the matrix via thrust. Where n is the total number of thrusters, and the generalized control input is... With the target thrust vector of each motor Establish mapping relationships, ； The optimal thrust of each motor was calculated using the pseudo-inverse matrix method.

5. The method for cross-medium drive control of an unmanned underwater vehicle based on multimodal perception according to claim 1, characterized in that, Step S400 includes: The software-layer anti-saturation compensator calculates the deviation between the current commanded thrust and the physical limit of the thruster in reverse. After multiplying the deviation by the convergence gain, it is injected into the controller's integral state equation, forcibly freezing the accumulation evolution of the integral term, and simultaneously forcibly resetting the target thrust command sequence to zero. After receiving a hard real-time interrupt signal indicating a sudden drop in phase current, the hardware-level microcontroller immediately utilizes the braking circuit breaker feature of the advanced timer to force the PWM output of all bridge arms to a preset low-level or high-impedance safe state within a single instruction cycle.

6. The method for cross-medium drive control of an unmanned underwater vehicle based on multimodal perception according to claim 1, characterized in that, The nonlinear soft-start function in step S500 is: ； In the formula The thrust command varies over time. Let t be the target thrust, and t be the continuous recovery time calculated from the triggering of the water entry flag. To determine the relative vertical velocity at the moment of entry into the water The time constant is adaptively adjusted; the greater the impact velocity upon entering the water, Automatic increase makes the thrust build-up curve smoother; when Restored to no less than When the water level reaches 95%, the system switches back to the deep-water anti-interference cruise mode in step S300.

7. The method for cross-medium drive control of an unmanned underwater vehicle based on multimodal perception according to claim 3, characterized in that, In step S230, a switching dead zone based on the principle of hysteresis comparator is set between adjacent operating conditions. The high-frequency spikes of the sensor signal at the threshold boundary between water outlet and water inlet are filtered out by the historical time window counter to avoid high-frequency oscillation switching of the state machine. If the contrast of the feature map of the visual sensor is lower than the effective confidence threshold due to extreme turbidity or no light at night, the system automatically cuts off the visual feature weights and relies only on the cross-verification of the Z-axis impulse feature of the inertial measurement unit and the change of motor phase current to maintain the closed loop of operating condition judgment.

8. The method for cross-medium drive control of an unmanned underwater vehicle based on multimodal perception according to claim 1, characterized in that, It also includes step S600, The actuator driver layer parses instructions from the upper-level operating system based on the underlying microcontroller and uses a field-oriented control algorithm to control the brushless motor of the thruster: extracting the two-phase current and transforming it into a stationary orthogonal coordinate system via Clark transformation. and The rotor electrical angles are converted into quadrature-axis currents in a rotating coordinate system using Park transformation. and direct-axis current Adjust according to desired thrust feedback The target value is achieved by using a cascaded proportional-integral controller to realize closed-loop control of the inner current loop and the outer speed loop. Six PWM waveforms driving the inverter are generated through inverse Park transformation and space vector pulse width modulation. The underlying hard real-time response cycle is controlled within 100 microseconds. Steps S100 to S500 are executed cyclically in the next control cycle.

9. The method for cross-medium drive control of an unmanned underwater vehicle based on multimodal perception according to claim 1, characterized in that, Steps S100 to S600 are deployed in a hierarchical control architecture that includes a task planning layer, a motion control layer, and an actuator driving layer. The task planning layer runs the MOOS operating system and is responsible for multimodal data fusion and cross-media condition identification. The motion control layer is responsible for nonlinear disturbance observation and robust control law calculation. The actuator driving layer is based on a microcontroller and is responsible for motor field orientation control and hardware-level safety protection.

10. The method for cross-medium drive control of an unmanned underwater vehicle based on multimodal perception according to claim 1, characterized in that, A state transition dead zone based on the principle of hysteresis comparator is set between steps S400 and S500, and a window counter is used to avoid high-frequency oscillation switching between PWM blocking and nonlinear soft start by the microcontroller.

Images