An unmanned aerial vehicle feedforward system and method based on flexible perception and hysteresis compensation

Abstract

The application discloses a kind of unmanned plane feedforward system and method based on flexible perception and hysteresis compensation, belong to unmanned plane intelligent control and emergency rescue technical field.The present application scheme includes: tactile perception platform, setting in the contact surface of hoisting device and load, include the distributed flexible sensing array based on VO2 Nano material on flexible substrate, for real-time acquisition pressure distribution data;Airborne computing module, running hysteresis compensation and intention prediction model based on space attention mechanism and BI-LSTM network, the gravity center offset of load in future time is predicted after space weighting, hysteresis compensation to pressure distribution data;Flight control module, according to the gravity center offset of prediction, feedforward control instruction is generated in advance and superimposed to flight control system, realize attitude pre-compensation.The present application perceives living load action from source, predicts gravity center change and compensates disturbance in advance by overcoming material hysteresis, significantly improves the anti-disturbance ability and stability of unmanned plane suspension system.

Description

Technical Field

[0001] This invention relates to the field of intelligent control and emergency rescue equipment technology for unmanned aerial vehicles (UAVs), and more specifically to a UAV feedforward system and method based on flexible sensing and hysteresis compensation. Background Technology

[0002] In complex environments such as mountain search and rescue and water rescue, rotary-wing drones have become an important means of lifting and rescuing. However, unlike traditional rigid cargo lifting, the objects being lifted in rescue missions are people or objects that can move actively.

[0003] The existing technology has the following significant pain points:

[0004] Lack of awareness of disturbance sources: Current drones mainly rely on inertial measurement units (IMUs) or visual sensors carried on the fuselage. When rescued personnel struggle violently or roll around on a stretcher due to panic or pain, this disturbance needs to be transmitted to the drone through a flexible sling tens of meters long, causing a significant change in the aircraft's attitude before the flight control system can detect and intervene. This "passive feedback" mechanism suffers from severe phase lag, which can easily lead to system divergence, increased swaying, or even a crash.

[0005] Physical limitations of flexible sensors: Although flexible sensors are suitable for contact sensing by conforming to the human body surface, commonly used flexible pressure-sensitive materials generally exhibit hysteresis and creep characteristics. That is, during the process of being compressed and rebounding, the resistance change curves do not coincide, causing traditional linear filtering algorithms to fail to accurately reproduce real-time pressure values, which seriously affects control accuracy.

[0006] Environmental interference is difficult to decouple: the strong downwash airflow generated by the drone propellers will cause high-frequency vibration of the stretcher. This background noise can easily drown out the characteristic signals of human movement, leading to misjudgment.

[0007] Therefore, there is an urgent need for an active control system that can sense the movement of a living load from the source, overcome material hysteresis through algorithms, and predict the center of gravity shift in advance. Summary of the Invention

[0008] In view of the above problems, the present invention is proposed to provide a UAV feedforward system and method based on flexible sensing and hysteresis compensation that overcomes or at least partially solves the above problems.

[0009] To achieve the above objectives, the present invention adopts the following technical solution: In a first aspect, embodiments of the present invention provide a UAV feedforward system based on flexible sensing and hysteresis compensation, comprising: A tactile sensing platform is installed at the contact surface between the UAV hoisting device and the load, and includes a distributed flexible sensor array for real-time acquisition of pressure distribution data between the load and the contact surface of the hoisting device. An airborne computing module is used to receive the pressure distribution data and run a pre-trained hysteresis compensation and intent prediction model. The hysteresis compensation and intent prediction model is used to compensate for the physical hysteresis effect of the flexible sensing array based on historical pressure distribution data and to predict the center of gravity offset of the load at future moments. The flight control module is used to generate feedforward control commands and superimpose them onto the flight control system before the disturbance is transmitted to the UAV fuselage, based on the predicted center of gravity offset, so as to pre-compensate the UAV attitude.

[0010] Preferably, the flexible sensing array is a flexible strain sensor based on nanomaterials and a flexible polymer substrate, and the flexible strain sensors are arranged in a matrix.

[0011] Preferably, the system further includes a data acquisition module, which synchronously acquires the pressure distribution data at a preset frequency, and then wirelessly transmits the data to the onboard computing module after amplification and low-pass filtering by a signal conditioning circuit.

[0012] Preferably, the hysteresis compensation and intention prediction model specifically includes: The spatial attention module is used to weight the pressure distribution data to enhance the features of the core stress region and generate a weighted feature tensor. The Bi-LSTM layer is used to memorize the historical deformation state of the flexible material in the flexible sensor using a gating mechanism, and to construct an inverse hysteresis model based on the weighted feature tensor to output the compensated accurate feature vector. The prediction head module is used to predict the load center of gravity coordinates and center of gravity offset within a preset time period in the future based on the precise feature vector. Preferably, the spatial attention module uses a one-dimensional convolutional layer for feature dimensionality reduction and calculates the spatial weight distribution of each sensor node using the Softmax activation function; The Bi-LSTM layer contains two hidden layers, each with 64 hidden units; The prediction head module consists of two fully connected layers, with the hidden layer using the ReLU activation function and the output layer using the linear activation function.

[0013] Preferably, the flight control module is further configured to: employ a dynamic weight allocation mechanism based on predicted disturbance intensity to fuse the feedforward control commands with feedback control commands based on UAV inertial measurement unit (IMU) data, in order to eliminate pre-compensated residual errors. U total=α·U ff +(1-α)·U fb Among them, U ff The feedforward control command is denoted by α; α is the dynamic feedforward weight coefficient, and α∈[0,1]; U fb This is a feedback control command.

[0014] Preferably, the feedback control command U fb Generated by a fuzzy PID controller; The fuzzy PID controller takes the attitude angle error and error change rate detected in real time by the UAV's inertial measurement unit (IMU) as input. Through fuzzification processing, fuzzy inference based on a preset fuzzy rule table, and centroid method defuzzification, it outputs the real-time tuned proportional, integral, and derivative control parameters, thereby generating the feedback control command U. fb .

[0015] Secondly, based on the same inventive concept, embodiments of the present invention provide a UAV feedforward method based on flexible sensing and hysteresis compensation, comprising the following steps: Step S1: Real-time pressure distribution data between the load and the lifting device contact surface is collected by a distributed flexible sensor array set on the contact surface between the UAV lifting device and the load. Step S2: Accept the pressure distribution data, run the pre-trained hysteresis compensation and intent prediction model, the hysteresis compensation and intent prediction model is used to compensate for the physical hysteresis effect of the flexible sensing array based on historical pressure distribution data, and predict the center of gravity offset of the load at future moments. Step S3: Based on the predicted center of gravity offset, before the disturbance is transmitted to the UAV fuselage, generate feedforward control commands and superimpose them on the flight control system to pre-compensate the UAV attitude.

[0016] The beneficial effects of the above-described technical solutions provided in the embodiments of the present invention include at least the following: 1. Source sensing and hysteresis overcoming: This invention directly senses the movement of a living load through a distributed flexible sensor array and uses a deep learning model to memorize historical deformation states, effectively constructing an inverse hysteresis model, fundamentally overcoming the inherent physical hysteresis defect of flexible sensors, and effectively restoring high-precision real pressure distribution.

[0017] 2. Intent Prediction and Feedforward Compensation: This invention uses an algorithm to predict the changing trend of the load center of gravity in advance. Before the physical disturbance is transmitted to the UAV fuselage through the suspension rope, it generates feedforward control commands in advance and superimposes them on the flight control system for attitude pre-compensation, breaking the phase lag bottleneck caused by the passive feedback of the traditional flight control system.

[0018] 3. Significantly improved anti-disturbance capability: Compared with traditional feedback control, the feedforward compensation mechanism of this invention reduces the maximum attitude deviation of the system by approximately 83.2% and shortens the stable convergence time by approximately 80.9%, greatly improving the stability and safety of the UAV hoisting system in complex rescue scenarios. As can be seen from the above technical solutions, compared with the prior art, the present invention discloses a UAV feedforward system and method based on flexible perception and hysteresis compensation. It not only effectively overcomes the physical hysteresis defect of flexible sensors from the algorithm level by using deep learning models, but also achieves accurate center of gravity intention prediction and attitude pre-compensation when the physical disturbance generated by the live load is transmitted to the fuselage. This completely breaks through the phase lag bottleneck caused by the passive feedback control of traditional UAVs, and significantly improves the anti-disturbance capability and flight safety of the sling system in complex rescue scenarios. Attached Figure Description

[0019] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on the provided drawings without creative effort.

[0020] Figure 1 This is a block diagram of the overall system structure provided in the embodiments of the present invention; Figure 2 This is a diagram of the algorithm network architecture topology provided in this embodiment of the invention; Figure 3 This is a flowchart of the feedforward control method provided in an embodiment of the present invention; Figure 4 This is a comparison chart of roll angle changes provided in an embodiment of the present invention. Detailed Implementation

[0021] 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 skilled in the art without creative effort are within the scope of protection of the present invention.

[0022] Example 1 This embodiment provides a UAV feedforward system based on flexible sensing and hysteresis compensation, applied to a quadcopter rescue UAV, combined with... Figure 1 The framework shown includes the following modules: The tactile sensing platform, located at the contact surface between the drone hoisting device and the load, includes a distributed flexible sensor array for real-time acquisition of pressure distribution data between the load and the hoisting device contact surface. The airborne computing module is used to receive pressure distribution data and run a pre-trained hysteresis compensation and intent prediction model. The hysteresis compensation and intent prediction model is used to compensate for the physical hysteresis effect of the flexible sensor array based on historical pressure distribution data and predict the center of gravity shift of the load at future moments. The flight control module generates feedforward control commands based on the predicted center of gravity offset and superimposes them onto the flight control system before the disturbance is transmitted to the UAV fuselage, in order to pre-compensate the UAV's attitude.

[0023] In one embodiment, the flexible sensing array is a flexible strain sensor based on nanomaterials and a flexible polymer substrate, and the flexible strain sensors are arranged in a matrix.

[0024] Specifically, the tactile sensing platform is a modified rescue stretcher. Distributed flexible sensor arrays are integrated into its surfaces that come into contact with the human body, such as the main support areas like the shoulders, waist, and legs. This sensor array uses a matrix arrangement of VO2 / PET flexible strain sensors, with a single-point size of 2×2cm, exhibiting high sensitivity (GF≥30). The sensors are fabricated using a screen printing process, covered with a polyimide film for waterproofing and hermetic sealing, and utilize a low-stiffness, high-damping flexible substrate to physically attenuate high-frequency vibration interference from the propeller downwash airflow.

[0025] In one embodiment, a data acquisition module is also included, which synchronously acquires analog signals from nine (or other) sensors at a frequency of 1 kHz using a multi-channel ADC to generate pressure distribution data. After amplification and low-pass filtering by a signal conditioning circuit, the data is transmitted to the onboard computing module via a wireless transmission module such as Wi-Fi or a proprietary protocol.

[0026] In one embodiment, the hysteresis compensation and intent prediction model specifically includes: The spatial attention module is used to weight the pressure distribution data to enhance the features of the core stress region and generate a weighted feature tensor. Bi-LSTM layers are used to memorize the historical deformation state of flexible materials in flexible sensors using a gating mechanism, and to construct an inverse hysteresis model based on weighted feature tensors to output the compensated accurate feature vector. The prediction head module is used to predict the load center of gravity coordinates and load center of gravity offset within a preset time period in the future, based on accurate feature vectors. Specifically, the onboard computing module, located on the drone, such as using the Jetson Nano embedded artificial intelligence computing platform, is used to receive and process stress data.

[0027] Combination Figure 2 As shown below, the processing flow of the hysteresis compensation and intention prediction model is explained in detail: The hysteresis compensation and intent prediction model is a deep learning model based on spatial attention mechanism and bidirectional long short-term memory network (Bi-LSTM). Its specific network architecture, inter-layer connections, and activation functions are as follows: Data Tensor Construction: Pressure data from 9 sensors is collected in real time, and a sliding time window (Window Size=50ms) is constructed to form an input tensor with dimensions (Batch, 50, 9).

[0028] Spatial Attention Module: This layer uses a one-dimensional convolutional layer for feature dimensionality reduction and calculates the spatial weight distribution of each sensor node using the Softmax activation function. The input tensor is multiplied element-wise with the weight matrix to enhance the features of key stressed areas and suppress sensor noise and common-mode interference from propeller airflow in unstressed areas, resulting in a weighted feature tensor.

[0029] Bi-LSTM Layer: Weighted feature tensors are input to the Bi-LSTM network. The Bi-LSTM network contains two hidden layers, each with 64 hidden units. The Bi-LSTM utilizes its internal input gate, forget gate, and output gate mechanisms (using the Sigmoid activation function to calculate the gate weights and the Tanh activation function to update candidate memories) to remember the historical deformation state of the flexible material in the flexible sensor over the past 50ms, thereby constructing an inverse hysteresis model and outputting an accurate feature vector that compensates for hysteresis.

[0030] Prediction Head Module: After the precise feature vector output by the Bi-LSTM is expanded, it is fed into a prediction head consisting of two fully connected layers (32 and 2 nodes respectively). The hidden layer uses the ReLU activation function to prevent gradient vanishing, and the output layer uses a linear activation function. It regresses and predicts the offset (Δx) of the load center of gravity relative to the initial equilibrium center point at the next 80ms time. t+80 ,Δy t+80 ).

[0031] In one embodiment, the offline training process for the hysteresis compensation and intent prediction model is as follows: To ensure that the aforementioned hysteresis compensation and intention prediction models can accurately perform hysteresis compensation and center of gravity prediction, this invention designs a complete offline training process: Training data sources and acquisition: A high-precision simulation test bench was built in the laboratory, and standard mannequins were used to simulate the movements of injured personnel to dynamically load and unload the flexible sensor array. Simultaneously, an ATI six-dimensional force sensor was used to record the actual forces and center of gravity coordinates in real time, serving as the true labels for training.

[0032] Data preprocessing: The collected input pressure simulation signals are subjected to mean removal and Min-Max normalization to scale them to the [0,1] interval to accelerate model convergence.

[0033] Loss function design: The mean squared error (MSE) is used as the loss function for the regression task to calculate the deviation between the predicted centroid coordinates and the true centroid coordinates.

[0034] Algorithm selection: The Adam optimizer is adopted. This algorithm can adaptively adjust the learning rate and effectively deal with the non-stationary objective function optimization problem caused by the hysteresis characteristics of the sensor.

[0035] The key hyperparameter settings for further training and deployment are as follows: The selection criteria and optimization strategies for key hyperparameters during model training and deployment are as follows: Time window length: set to 50ms. If the window is too small, the hysteresis loop characteristics of flexible materials cannot be fully captured; if the window is too large, it will lead to increased memory consumption and reduced system real-time performance.

[0036] Prediction lead time: set at 80ms. This lead time is determined by comprehensively measuring system communication delay (approximately 10ms), onboard computing module processing delay (approximately 20ms), and flight control and motor response delay (approximately 50ms). The 80ms lead time accurately covers the aforementioned physical and system delays.

[0037] Learning rate: The initial learning rate is set to 0.001 and a cosine annealing strategy is used to dynamically decay the learning rate in order to avoid getting trapped in local optima.

[0038] Batch size and regularization: The batch size is set to 128, and a Dropout layer with a dropout rate of 0.2 is added between the Bi-LSTM layers to prevent the model from overfitting.

[0039] In one embodiment, the flight control module is further configured to: employ a dynamic weight allocation mechanism based on predicted disturbance intensity to fuse feedforward control commands with feedback control commands based on UAV inertial measurement unit (IMU) data, in order to eliminate pre-compensated residual errors. U total =α·U ff +(1-α)·U fb Among them, U ff U is the feedforward control command; α is the dynamic feedforward weight coefficient, and α∈[0,1]; fb This is a feedback control command.

[0040] Furthermore, feedback control command U fbGenerated by a fuzzy PID controller; The fuzzy PID controller takes the attitude angle error and error change rate detected in real time by the UAV's inertial measurement unit (IMU) as input. Through fuzzification processing, fuzzy inference based on a preset fuzzy rule table, and centroid method defuzzification, it outputs the real-time tuned proportional, integral, and derivative control parameters, and then generates the feedback control command U. fb .

[0041] Specifically, the flight control module communicates with the onboard computing module and the UAV's existing flight control system. The superposition of control signals can employ linear addition logic with a parallel feedforward compensation structure. That is, the final output to the motor drive system is the total control torque command U. total Forward control command U ff With feedback control command U fb The algebraic sum of U total =U ff +U fb This structure ensures that the feedforward channel primarily handles known strong disturbances caused by load center of gravity shift, while the feedback channel focuses on eliminating residual high-frequency interference unknown to the model.

[0042] However, in this embodiment, the superposition of the data onto the flight control system and the elimination of residual errors preferably employ a deep fusion method based on a dynamic weight allocation mechanism for predicting disturbance intensity. The final output is the total control torque command U to the motor drive system. total The logical expression is as follows: U total =α·U ff +(1-α)·U fb .

[0043] Where α is the dynamic feedforward weight coefficient, and α∈[0,1]. The airborne computing module calculates this weight coefficient α online in real time based on the load center of gravity offset gradient (i.e., the intensity of instantaneous movement) output by the prediction head module; when the load struggles violently and the center of gravity offset gradient is greater than the preset high threshold, the system determines that a strong disturbance is about to occur, and at this time, α is rapidly made to approach 1 according to the nonlinear mapping rule. At this time, the total control command is the feedforward control command U generated by the neural network. ff This is achieved by prioritizing rapid attitude pre-compensation for large disturbances, completely eliminating the phase lag caused by traditional feedback. Conversely, when the load action is smooth and the center of gravity offset gradient decreases below the low threshold, α dynamically and smoothly decays. System control then smoothly transitions to the feedback control channel, using IMU-based feedback control commands U... fb The system prioritizes eliminating unmodeled residual high-frequency interference to ensure extremely high hovering steady-state accuracy for the UAV. This dynamic fusion logic achieves optimal decoupling between feedforward agility and feedback stability.

[0044] Feedback control command U fb It is generated by the fuzzy PID controller built into the flight control system. The fuzzy PID controller uses the attitude angle error and error change rate detected in real time by the UAV IMU as dual inputs, and uses a preset fuzzy rule table to tune the three control parameters (proportional, integral, and derivative, Kp, Ki, Kd) online in real time, thereby adaptively outputting feedback control commands to complete the closed-loop fine-tuning of the overall feedforward system.

[0045] The parameter tuning process for a fuzzy PID controller specifically includes the following steps: First, the input variables are fuzzified. The actual physical domains of the real-time acquired attitude angle error e and error change rate ec are mapped to the standard fuzzy domain, and corresponding fuzzy subsets are divided, for example, set as {negative large (NB), negative medium (NM), negative small (NS), zero (ZO), positive small (PS), positive medium (PM), positive large (PB)}. Then, the membership degree corresponding to each input variable is calculated using a triangular or Gaussian membership function to generate a fuzzy vector. Secondly, fuzzy inference is performed based on a pre-set fuzzy rule table. The fuzzy rule table is constructed based on expert experience in UAV attitude control, and its core dynamic adjustment logic includes: when the attitude angle error e is large, a larger proportional coefficient increment ΔKp and a smaller differential coefficient increment ΔKd are selected to accelerate the system response speed and prevent differential saturation caused by high-frequency noise; when the attitude angle error e is small and the error change rate ec is moderate, a moderate ΔKp, a larger ΔKd, and a suitable integral coefficient increment ΔKi are selected to effectively suppress system overshoot and eliminate static error. Finally, defuzzification is performed. The centroid method is used to analyze the fuzzy control quantity obtained from fuzzy inference, and the determined parameter increments ΔKp, ΔKi, and ΔKd are calculated. The parameter increments are added to the set initial PID parameter reference values ​​to obtain the real-time updated proportional, integral, and derivative control parameters, and then the final feedback control command Ufb is calculated and output.

[0046] Example 2 like Figure 3 As shown, based on the same inventive concept, this invention also discloses a UAV feedforward method based on flexible sensing and hysteresis compensation, comprising the following steps: Step S1: Real-time pressure distribution data between the load and the lifting device contact surface is collected by a distributed flexible sensor array set on the contact surface between the UAV lifting device and the load. Step S2: Receive pressure distribution data and run a pre-trained hysteresis compensation and intent prediction model. The hysteresis compensation and intent prediction model is used to compensate for the physical hysteresis effect of the flexible sensor array based on historical pressure distribution data and predict the center of gravity shift of the load at future moments. Step S3: Based on the predicted center of gravity offset, before the disturbance is transmitted to the UAV fuselage, generate feedforward control commands and superimpose them on the flight control system to pre-compensate the UAV attitude.

[0047] The following describes the application of this invention in a specific rescue scenario: Suppose a drone is hovering to lift a wounded person weighing 70kg. During the hovering process, the high-frequency vibrations caused by the propeller downwash are physically attenuated by the flexible substrate and suppressed by the spatial attention layer, and are regarded as background noise.

[0048] Suddenly, the wounded soldier recoiled violently to the left and backward due to pain. In that instant: Source detection: The pressure in the lower left corner area (sensors S7, S8) of the smart stretcher rises sharply, and the pressure distribution data changes drastically.

[0049] Algorithm Solution: The model in the onboard computing module rapidly processes the data sequence. The Bi-LSTM network compensates for the dynamic hysteresis caused by the rapid loading of the sensor, accurately reconstructing the pressure value. Based on the pressure change trend over the past 50ms, the prediction head successfully predicted that the injured person's center of gravity would shift 8cm to the left and 3cm backward after 80ms.

[0050] Feedforward control: Based on this predicted center of gravity shift, the flight control module immediately calculates that this will result in a "pitch-left" moment of approximately 4.5° and a "pitch-back" moment for the drone. Before the drone's IMU detects any attitude change, because the disturbance is still being transmitted along the tether, the system generates a feedforward command: Target Roll =+4.5°, Target Pitch =-1.2°. This command is directly superimposed on the flight control system, causing the drive motor to adjust its speed in advance, generating a reverse torque.

[0051] Feedback Fine-tuning: Subsequently, when the disturbance is actually transmitted to the fuselage, the attitude change detected by the UAV's IMU is very small because the feedforward control has already offset most of the impact. At this time, the system calls the internal fuzzy PID controller for feedback adjustment. This controller uses the attitude angle error and error change rate detected by the IMU as input, and uses a preset fuzzy rule table to tune the proportional, integral, and derivative (Kp, Ki, Kd) parameters online in real time, thereby adaptively generating the feedback control command U. fb This eliminates residual nonlinear errors. Ultimately, the entire hoisting system is stabilized, and the load sway amplitude is suppressed to an extremely low level.

[0052] like Figure 4As shown, to verify the practical effect of the feedforward system based on flexible sensing and hysteresis compensation of this invention, a comparative experiment on disturbance resistance during actual flight and suspension was conducted. The experimental conditions were as follows: a 70kg mannequin was suspended by a drone, and a mechanical device triggered an instantaneous shift in the mannequin's center of gravity (8cm to the left and 3cm to the back). In the control group (traditional PID feedback control), the disturbance was transmitted to the fuselage via the suspension cable, and the drone's IMU intervened only after sensing the attitude change. The system's maximum roll deviation reached 12.5°, and the time to restore stable attitude was 4.2 seconds, accompanied by severe pendulum oscillations.

[0053] Experimental group: The flight control module generates a target based on the predicted center of gravity shift before the disturbance is transmitted to the fuselage. Roll =+4.5°, Target _Pitch A pre-compensation command of -1.2° was given. Combined with data from the UAV's inertial measurement unit (IMU), a feedback control algorithm was used to eliminate the residual error after feedforward compensation. The system's maximum roll angle deviation was limited to within 2.1°, and attitude recovery to stability took only 0.8 seconds.

[0054] Experimental conclusion: The experimental data above verifies that the method of the present invention can effectively overcome the hysteresis effect of flexible sensors and achieve intention prediction. Its feedforward compensation mechanism reduces the maximum attitude deviation of the system by 83.2% and shortens the stable convergence time by 80.9%, significantly improving the anti-disturbance capability and safety of the UAV mounting system.

[0055] The various embodiments in this specification are described in a progressive manner, with each embodiment focusing on its differences from other embodiments. Similar or identical parts between embodiments can be referred to interchangeably. For the apparatus disclosed in the embodiments, since they correspond to the methods disclosed in the embodiments, the description is relatively simple; relevant parts can be referred to the method section.

[0056] The above description of the disclosed embodiments enables those skilled in the art to make or use the invention. Various modifications to these embodiments will be readily apparent to those skilled in the art, and the general principles defined herein may be implemented in other embodiments without departing from the spirit or scope of the invention. Therefore, the invention is not to be limited to the embodiments shown herein, but is to be accorded the widest scope consistent with the principles and novel features disclosed herein.

Claims

1. A UAV feedforward system based on flexible sensing and hysteresis compensation, characterized in that, include: A tactile sensing platform is installed at the contact surface between the UAV hoisting device and the load, and includes a distributed flexible sensor array for real-time acquisition of pressure distribution data between the load and the contact surface of the hoisting device. An airborne computing module is used to receive the pressure distribution data and run a pre-trained hysteresis compensation and intent prediction model. The hysteresis compensation and intent prediction model is used to compensate for the physical hysteresis effect of the flexible sensing array based on historical pressure distribution data and to predict the center of gravity offset of the load at future moments. The flight control module is used to generate feedforward control commands and superimpose them onto the flight control system before the disturbance is transmitted to the UAV fuselage, based on the predicted center of gravity offset, so as to pre-compensate the UAV attitude.

2. The system according to claim 1, characterized in that, The flexible sensing array is a flexible strain sensor based on nanomaterials and a flexible polymer substrate, and the flexible strain sensors are arranged in a matrix.

3. The system according to claim 1, characterized in that, It also includes a data acquisition module, which synchronously acquires the pressure distribution data at a preset frequency, and then wirelessly transmits the data to the airborne computing module after amplification and low-pass filtering by a signal conditioning circuit.

4. The system according to claim 1, characterized in that, The hysteresis compensation and intent prediction model specifically includes: The spatial attention module is used to weight the pressure distribution data to enhance the features of the core stress region and generate a weighted feature tensor. The Bi-LSTM layer is used to memorize the historical deformation state of the flexible material in the flexible sensor using a gating mechanism, and to construct an inverse hysteresis model based on the weighted feature tensor to output the compensated accurate feature vector. The prediction head module is used to predict the load center of gravity coordinates and center of gravity offset within a preset time period in the future, based on the accurate feature vector.

5. The system according to claim 4, characterized in that, The spatial attention module uses a one-dimensional convolutional layer for feature dimensionality reduction and calculates the spatial weight distribution of each sensor node through the Softmax activation function. The Bi-LSTM layer contains two hidden layers, each with 64 hidden units; The prediction head module consists of two fully connected layers, with the hidden layer using the ReLU activation function and the output layer using the linear activation function.

6. The system according to claim 1, characterized in that, The flight control module is further configured to employ a dynamic weight allocation mechanism based on predicted disturbance intensity to fuse the feedforward control commands with feedback control commands based on UAV inertial measurement unit (IMU) data, thereby eliminating pre-compensated residual errors. U total =a·U ff + (1-a) U fb Among them, U ff The feedforward control command is denoted by α; α is the dynamic feedforward weight coefficient, and α∈[0,1]; U fb This is a feedback control command.

7. The system according to claim 6, characterized in that, The feedback control command U fb Generated by a fuzzy PID controller; The fuzzy PID controller takes the attitude angle error and error change rate detected in real time by the UAV's inertial measurement unit (IMU) as input. Through fuzzification processing, fuzzy inference based on a preset fuzzy rule table, and centroid method defuzzification, it outputs the real-time tuned proportional, integral, and derivative control parameters, thereby generating the feedback control command U. fb .

8. A UAV feedforward method based on flexible sensing and hysteresis compensation, applied to the system described in any one of claims 1 to 7, characterized in that, Includes the following steps: Step S1: Real-time pressure distribution data between the load and the lifting device contact surface is collected by a distributed flexible sensor array set on the contact surface between the UAV lifting device and the load. Step S2: Accept the pressure distribution data, run the pre-trained hysteresis compensation and intent prediction model, the hysteresis compensation and intent prediction model is used to compensate for the physical hysteresis effect of the flexible sensing array based on historical pressure distribution data, and predict the center of gravity offset of the load at future moments. Step S3: Based on the predicted center of gravity offset, before the disturbance is transmitted to the UAV fuselage, generate feedforward control commands and superimpose them on the flight control system to pre-compensate the UAV attitude.

Images