Search

CN-122018327-A - Multi-rotor unmanned aerial vehicle fault-tolerant flight control method based on end-to-end direct thrust control

CN122018327ACN 122018327 ACN122018327 ACN 122018327ACN-122018327-A

Abstract

The invention discloses a multi-rotor unmanned aerial vehicle fault-tolerant flight control method based on end-to-end direct thrust control, which comprises the steps of constructing a four-rotor unmanned aerial vehicle pure-analysis micro-simulation environment oriented to implicit feature observation and gradient counter-propagation, embedding continuous micro-fault injection and execution mechanism mapping mechanism, performing discretization iteration on continuous operators by adopting a four-order Longgy-Kutta method, constructing an end-to-end control network based on space-time feature structural decoupling and hidden variable direct driving, extracting and outputting four-motor action instructions through double-flow features, training the end-to-end control network in the micro-simulation environment, optimizing control network parameters based on a micro-physical engine and counter-propagation along with time, and converging the control network to an effective fault-tolerant control strategy by combining a self-adaptive composite loss function and a multi-order course learning mechanism. The method can realize the self-adaptive control of the single motor without independent fault detection and diagnosis modules, and establish a smooth transition mechanism from normal to single failure.

Inventors

  • LIU XU
  • YANG YI
  • WAN JIANGLIN

Assignees

  • 南京信息工程大学

Dates

Publication Date
20260512
Application Date
20260410

Claims (10)

  1. 1. The multi-rotor unmanned aerial vehicle fault-tolerant flight control method based on end-to-end direct thrust control is characterized by comprising the following steps of: Step 1, constructing a pure-analysis micro-simulation environment of a four-rotor unmanned aerial vehicle facing implicit characteristic observation and gradient counter-propagation from a bottom layer by utilizing a tensor calculation frame, embedding a continuous micro-fault injection and execution mechanism mapping mechanism, performing discretization iteration processing on a continuous operator by adopting a fourth-order Dragon-Kutta method, and converting the continuous operator into a micro-discrete state transfer operator under a deep learning frame; Step 2, constructing an end-to-end control network based on space-time characteristic structural decoupling and hidden variable direct driving, taking an observed state tensor as input, carrying out double-flow characteristic extraction through a kinematic intention perception branch and a physical time sequence diagnosis perception branch, generating a hidden state based on a gating circulation unit, and outputting a four-motor action instruction; And 3, training the end-to-end control network constructed in the step 2 in the micro-simulation environment constructed in the step 1, optimizing control network parameters based on the micro-physical engine and back propagation along with time, and combining a self-adaptive composite loss function and a multi-stage course learning mechanism to perform direct gradient analysis conduction from physical constraint to network weight, so that the control network is converged to an effective fault-tolerant control strategy, and the fault-tolerant flight control of the multi-rotor unmanned aerial vehicle is realized.
  2. 2. The multi-rotor unmanned aerial vehicle fault-tolerant flight control method of claim 1, wherein the micro-simulation environment of step 1 comprises: defining network observation state tensors Comprises the following steps of position deviation under a machine body coordinate system Linear speed of machine body Angular velocity of body Tiling column vector of gesture rotation matrix Control instruction at last moment Actual motor speed observation value Wherein, the method comprises the steps of, An orthogonal rotation matrix representing the attitude relation of the current machine body coordinate system relative to the inertial coordinate system; defining a network output four-dimensional action tensor Representing the target PWM duty cycles of the four motors of the multi-rotor unmanned aerial vehicle.
  3. 3. The multi-rotor unmanned aerial vehicle fault-tolerant flight control method of claim 2, wherein the continuous and minimal fault injection and actuator mapping mechanism of step 1 comprises: Establishing a linear mapping relation from a network output space to a real physical rotating speed, and outputting an action tensor of the network Inverse normalization to physical target speed vector ; Defining a fault attenuation matrix , Is the first The health coefficients of the motors, i=1, 2,3 and 4, wherein the fault attenuation matrix is only used in a physical engine to influence the next dynamic state; the fault injection adopts a rotating speed level model, and the actual motor rotating speed is : ; Calculating the actual thrust : ; Wherein, the Is the thrust coefficient.
  4. 4. The fault-tolerant flight control method of a multi-rotor unmanned aerial vehicle of claim 3, wherein the discretizable state transfer operator of step1 is constructed as follows: Establishing continuous analytic dynamics operator The operator sequentially performs inverse normalization, fault injection, thrust calculation, force and moment distribution and system state differential equation set inside the operator to output system state derivative : ; Wherein, the In order to be the current state of the system, Representing the current actual position; four-order Longge-Kutta method is adopted to carry out Discretizing into micro-state transfer operators : ; Wherein, the For the next time system state, transferring the micro state operator In the automatic derivative calculation graph mounted on the deep learning framework, the subsequent BPTT gradient is reversely propagated along the time axis.
  5. 5. The method for fault-tolerant flight control of a multi-rotor unmanned aerial vehicle of claim 4, wherein in the end-to-end control network of step 2, the dual-stream feature extraction comprises: Receiving positional deviations by kinematic intent-aware branching Linear speed of machine body Tiling column vector of gesture rotation matrix Extracting motion characteristics through two layers of fully connected networks ; Receiving a control instruction at the last moment through physical time sequence diagnosis sensing branches Current motor speed observations Angular velocity of body Calculating a rotational speed residual error : ; Wherein, the For a pre-calibrated motor steady-state gain matrix, Represent the first Command-rotational speed linear gains for each motor, i=1, 2,3,4; representing the operation of converting the vector into a diagonal matrix, Is the minimum steady-state rotational speed vector; Residual error of the calculated rotating speed Spliced with the original input, and diagnostic features are extracted through a layer of fully-connected network 。
  6. 6. The multi-rotor unmanned aerial vehicle fault-tolerant flight control method of claim 5, wherein generating the hidden state based on the gating cycle unit of step 2 comprises: Will diagnose the characteristics Input gating cycle unit, through time step Executing a state update equation to update the hidden state : ; Wherein, the Is a time step Is used for the hidden state of the (a), The function represents a gating circulation unit for internalizing the influence of the fault attenuation matrix on the unmanned aerial vehicle dynamics into a hidden state The distribution evolves in a high-dimensional space.
  7. 7. The method for fault-tolerant flight control of a multi-rotor unmanned aerial vehicle of claim 6, wherein the outputting of the four motor action command of step 2, the processing comprises: Characterization of motion And hidden state Input fusion layer spliced into joint features Sending into a three-layer full-connection network; dynamic redistribution from input space to control quantity output space is carried out through a three-layer full-connection network, and action tensors are output 。
  8. 8. The multi-rotor unmanned aerial vehicle fault-tolerant flight control method of claim 7, wherein the three-layer fully-connected network is specifically as follows: The first layer is a full connection layer for inputting Dimension 128, calculate The output dimension is 128, and the output dimension is obtained through a nonlinear activation layer: ; The second layer is a full connection layer, and calculates The output dimension is 64, and the output dimension is obtained through a nonlinear activation layer: ; the third layer is the output layer, calculate The output dimension is 4, and normalized action tensor is obtained through compression of the activation layer : ; Wherein, the 、 、 Respectively representing weight matrixes of the first layer, the second layer and the third layer, 、 Respectively representing the output of the first and second full connection layers after the ReLU activation function, 、 、 Respectively represent the offset vectors of the first layer, the second layer and the third layer, 、 、 The linear outputs of the first layer, the second layer and the third layer are respectively represented, In order to activate the function, Is a normalization function.
  9. 9. The multi-rotor unmanned aerial vehicle fault-tolerant flight control method of claim 7, wherein step 3 trains the end-to-end control network constructed in step 2 in the micro-simulation environment constructed in step 1, and the method comprises: Step 3.1, sampling in a micro-closed loop mode, namely sampling a fault matrix according to the fault type and probability set in the current course learning stage for each training round Corresponding fault modes are expanded along the forward direction of the time step, a complete calculation diagram is reserved, and a track set is formed Wherein, the method comprises the steps of, Is the total number of time steps; Step 3.2, constructing an adaptive loss function, and total loss The definition is as follows: ; Wherein, the In order for the time-to-discount rate, The weight coefficients of the position, velocity and control smoothing terms respectively, For the purpose of a penalty in location tracking, For the purpose of the speed tracking penalty, To be based on physically induced loss of pose compromise, Punishment for instruction smoothing and actions; Step 3.3, multi-stage course learning, namely adopting three-stage progressive training, wherein no fault exists in the first stage, soft faults are introduced in the second stage, hard faults are introduced in the third stage, and the difficulty of the faults is gradually enhanced through multi-stage course learning, so that the control network is stably converged to an effective fault-tolerant strategy; and 3.4, calculating the gradient of the loss to the network parameters by using an automatic derivation mechanism of the deep learning framework, and updating the parameters by using a AdamW optimizer and assisting in gradient cutting to prevent gradient explosion.
  10. 10. The multi-rotor unmanned aerial vehicle fault-tolerant flight control method of claim 9, wherein the position tracking penalty And speed tracking penalty For forcing the unmanned aerial vehicle to perform a three-dimensional spatial intent: ; ; Wherein, the As a result of the location of the object, Is the target body linear velocity; The instruction smoothing and action punishment Severe high frequency oscillations and long term saturation boundary dwell for punishing motor duty cycle: ; Wherein, the Is a preset L2 regularization coefficient; The physical induction-based posture compromise loss Including pitch/roll error terms and adaptive yaw weight terms: ; Wherein, the 、 A pitch/roll component and a yaw component respectively, Is a fixed weight; and the yaw weight is dynamically adjusted according to the height error, so that the yaw penalty is automatically reduced when the control network falls down at the height, and the height stability is preferentially maintained.

Description

Multi-rotor unmanned aerial vehicle fault-tolerant flight control method based on end-to-end direct thrust control Technical Field The invention relates to the technical field of aircraft flight control, in particular to a multi-rotor unmanned aerial vehicle fault-tolerant flight control method based on end-to-end direct thrust control. Background With the wide application of multi-rotor unmanned aerial vehicle in such fields as logistics distribution, inspection and security, the flight safety of the multi-rotor unmanned aerial vehicle is increasingly valued. Because the quadrotor unmanned aerial vehicle is an underactuated and strongly coupled nonlinear system and lacks physical redundancy, the crash accident is very easy to cause once the failure of an actuator (such as motor stalling and blade breakage) occurs. The prior fault-tolerant control technology mainly has the following problems: 1) The limitations of control architecture are that existing fault tolerant control is largely divided into Passive Fault Tolerance (PFTC) and Active Fault Tolerance (AFTC). Passive fault tolerance typically uses a single robust controller to address all conditions, and it is difficult to compromise the high performance of normal flight and stability under failure, often at the expense of normal flight maneuverability. Active fault tolerance relies on explicit Fault Detection and Diagnosis (FDD) modules, which are not only complex in design, but also present significant detection delay or risk of false positives (missed). Delays in FDD tend to cause pose divergence during the critical period of milliseconds after failure, losing rescue opportunity. 2) Training efficiency and physical consistency issues in recent years, reinforcement learning (Reinforcement Learning, RL) based end-to-end control has demonstrated potential. However, the mainstream "Model-Free" algorithm (such as PPO) adopts a random sampling error-testing mechanism, and hundreds of millions of steps of interactions are required to converge, so that training efficiency is extremely low. In addition, due to lack of physical constraints, the trained strategy often has the problems of large control quantity jitter and dissatisfaction with the dynamic smoothness requirement. 3) The limited viability under single motor failure is insufficient-single motor complete failure is one of the most challenging failure modes of a quad-rotor drone. Under the working condition, the system loses a thrust source and becomes severely underactuated, and the traditional method based on geometric control or PID control is completely ineffective because the pseudo-inverse of the control distribution matrix cannot be solved. The existing method is used for maintaining the total lift force by allowing the unmanned aerial vehicle to enter a high-speed spin state, but how to trigger the strategy switching smoothly and how to stabilize the high-degree control in the spin state is still difficult. The prior art lacks an end-to-end learning mechanism that automatically sacrifices yaw tracking accuracy in exchange for highly sustained yaw tracking when a single motor fails completely. Therefore, there is a need for an end-to-end fault tolerant control method that can address single motor fault conditions, has a very fast training capability, can implement policy adaptation, and has a "high-level" viability in extreme faults. Disclosure of Invention The invention aims to solve the problem of providing a multi-rotor unmanned aerial vehicle fault-tolerant flight control method based on end-to-end direct thrust control, which is used for realizing self-adaptive control without an independent fault detection and diagnosis module under single motor fault, solving the problem of out-of-control attitude under single motor complete failure, improving training efficiency and physical consistency of an end-to-end control strategy and establishing a smooth transition mechanism from normal to single machine failure. The invention adopts the following technical scheme that the multi-rotor unmanned aerial vehicle fault-tolerant flight control method based on end-to-end direct thrust control comprises the following steps: Step 1, constructing a pure-analysis micro-simulation environment of a four-rotor unmanned aerial vehicle facing implicit characteristic observation and gradient counter-propagation from a bottom layer by utilizing a tensor calculation frame, embedding a continuous micro-fault injection and execution mechanism mapping mechanism, performing discretization iteration processing on a continuous operator by adopting a fourth-order Dragon-Kutta method, and converting the continuous operator into a micro-discrete state transfer operator under a deep learning frame; Step 2, constructing an end-to-end control network based on space-time characteristic structural decoupling and hidden variable direct driving, taking an observed state tensor as input, carrying out double-flow characteristic ext