Unmanned aerial vehicle control method and device based on reservoir computing, equipment and medium
By combining the nominal model of rigid body dynamics and the optical computing system of deep reservoir for UAV control, the problems of attitude instability and reduced control accuracy of UAVs in complex environments are solved, and real-time, low-power, high-precision flight control is achieved.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- SHANGHAI TECH UNIV
- Filing Date
- 2026-04-04
- Publication Date
- 2026-06-12
Smart Images
Figure CN122195059A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of unmanned aerial vehicles (UAVs), and more particularly to UAV control methods, apparatus, equipment, and media based on reservoir calculations. Background Technology
[0002] Current UAV flight control is generally based on a simplified Euler-Lagrange rigid body dynamics model, which can achieve desired trajectory tracking in open spaces and weak wind conditions. However, when UAVs fly at low altitudes, near the ground, or in confined spaces, complex aerodynamic disturbances such as ground effect, ceiling effect, and eddies introduce additional nonlinear and time-varying dynamics. These effects are difficult to accurately describe using rigid body dynamics models, leading to attitude instability, trajectory deviation, or even control failure.
[0003] To address the aforementioned issues, existing technologies can be broadly categorized into two types: one is compensation algorithms based on empirical models, which estimate additional aerodynamic forces by introducing empirical equations into rigid body dynamics. However, this method struggles to accurately characterize complex effects such as unsteady flow and vortex coupling, resulting in limited accuracy and adaptability. The other type is data-driven machine learning methods, such as multilayer perceptrons and recurrent neural networks. While these methods can achieve nonlinear prediction and feedforward compensation, they suffer from high model complexity, long training times, and high computational power consumption, making them difficult to deploy in real-time on airborne computing platforms.
[0004] However, existing compensation methods based on empirical models struggle to accurately characterize complex aerodynamic disturbances, while data-driven machine learning methods, though capable of nonlinear prediction, suffer from high model complexity, long training times, and high computational power consumption, making real-time deployment difficult under conditions of limited airborne computing resources. Therefore, achieving efficient prediction and rapid compensation of complex aerodynamic disturbances while maintaining low computational overhead, thereby improving the flight stability and control accuracy of UAVs in complex environments, remains a pressing technical challenge. Summary of the Invention
[0005] This invention provides a method, apparatus, equipment, and medium for controlling unmanned aerial vehicles (UAVs) based on reservoir calculations, which can solve the technical problem of difficulty in controlling the flight accuracy of UAVs in complex environments.
[0006] This invention provides a drone control method based on reservoir calculation, comprising:
[0007] Real-time acquisition of the expected and actual operating status of the drone; Based on the expected and actual operating states of the UAV, nominal feedforward control data are calculated using a pre-set rigid body dynamics nominal model. Based on the deviation between the expected and actual operating states of the UAV, feedback control calculations are performed to obtain feedback compensation control data. The actual operating status of the current UAV is input into the deep reservoir optical computing system, which then predicts the aerodynamic interference experienced by the UAV in the current environment in real time and outputs the corresponding feedforward compensation control data. The nominal feedforward control data, the feedback compensation control data, and the feedforward compensation control data are fused to obtain the flight control data of the UAV.
[0008] In one embodiment of the present invention, the nominal feedforward control data is calculated using a rigid body dynamics nominal model:
[0009] in, For nominal feedforward control data, This represents the current actual location of the drone. For the current expected speed of drones, To accelerate the expected performance of current drones, For the quality matrix, The matrix of Coriolis force and centrifugal force. This is the gravity vector.
[0010] In one embodiment of the present invention, the step of performing feedback control calculations based on the deviation between the desired operating state and the actual operating state of the current UAV to obtain feedback compensation control data includes: Calculate the speed deviation based on the expected speed of the current UAV in its expected operating state and the actual speed in its actual operating state; The integrals of the speed deviation and the historical speed deviation are independently adjusted and then fused to calculate the feedback compensation control data.
[0011] In one embodiment of the present invention, the calculation formula for the feedback compensation control data is as follows:
[0012] in, The actual speed of the current drone, For the current expected speed of drones, , These are the proportional weight matrix and the integral weight matrix, respectively.
[0013] In one embodiment of the present invention, the step of inputting the current actual operating state of the UAV into the deep reservoir optical computing system, and having the deep reservoir optical computing system predict in real time the aerodynamic interference experienced by the UAV in the current environment and output corresponding feedforward compensation control data, includes: Through the input layer of the deep reservoir optical computing system, the actual operating status of the current UAV is photoelectrically converted and modulated to form multi-channel modulated optical signals. Through the reservoir layer of the deep reservoir optical computing system, high-dimensional nonlinear temporal features are extracted from the multi-channel modulated optical signals to generate reservoir features. The output layer of the deep reservoir optical computing system calculates and outputs feedforward compensation control data corresponding to the aerodynamic interference experienced by the UAV in the current environment, based on its internal parameters and the characteristics of the reservoir.
[0014] In one embodiment of the present invention, the reservoir layer includes multiple cascaded reservoirs. The reservoir layer of the deep reservoir optical computing system performs high-dimensional nonlinear temporal feature extraction on the multi-path modulated optical signals to generate reservoir features, including: For each reserve pool: Based on the received current input optical signal and historical input optical signal, the current time series feature is obtained through high-dimensional nonlinear time series feature extraction; wherein, the input of the first reservoir is a multi-channel modulated optical signal, and the input of each subsequent reservoir is a modulated optical signal obtained by processing the time series feature output by the previous reservoir through an optical injection locking method. The time-series features of all the reserve pools are concatenated to generate the reserve pool features.
[0015] In one embodiment of the present invention, the internal parameters of the output layer of the deep reservoir optical computing system are updated in real time according to a preset update function, the calculation formula of which is as follows:
[0016] in, These are the internal parameters of the output layer. As a characteristic of the reserve pool, The regularization coefficient is . It is the identity matrix. This refers to the control data obtained by fusing the nominal feedforward control data and the feedback compensation control data of the current UAV.
[0017] The present invention also provides a drone control device based on reservoir calculation, comprising: The data acquisition module is used to acquire the expected and actual operating status of the drone in real time. The feedforward data generation module is used to calculate nominal feedforward control data based on the expected and actual operating states of the current UAV and through a preset rigid body dynamics nominal model. The feedback data generation module is used to perform feedback control calculations based on the deviation between the expected operating state and the actual operating state of the current UAV, and obtain feedback compensation control data. The compensation data generation module is used to input the actual operating status of the current UAV into the deep reserve pool optical computing system, which then predicts the aerodynamic interference experienced by the UAV in the current environment in real time and outputs the corresponding feedforward compensation control data. The control data fusion module is used to fuse the nominal feedforward control data, the feedback compensation control data, and the feedforward compensation control data to obtain the flight control data of the UAV.
[0018] The present invention also provides an electronic device, including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the computer program to implement the steps of the UAV control method based on reservoir calculation.
[0019] The present invention also provides a computer-readable storage medium storing a computer program, which, when executed by a processor, implements the steps of the unmanned aerial vehicle control method based on reservoir calculation.
[0020] The beneficial effects of this invention are as follows: Compared with traditional compensation algorithms based on empirical models, this scheme, through data-driven and optical high-dimensional nonlinear mapping, can more accurately learn and predict unsteady, time-varying residual aerodynamic forces caused by ground effects, eddy current coupling, etc., thus effectively overcoming the problems of limited accuracy and adaptability of empirical models. Secondly, compared with methods based on complex digital neural networks such as multilayer perceptrons, temporal convolutional networks, and recurrent neural networks, the core computing unit of this scheme is an optical reservoir that only requires training a linear readout layer. This fundamentally avoids the stability problems of gradient vanishing or exploding, while reducing the training time to the sub-millisecond level. Furthermore, by utilizing the high bandwidth and low latency characteristics of optical computing, it meets the stringent requirements of airborne platforms for real-time performance and low power consumption, solving the problems of complex models, time-consuming training, and high computational power consumption in traditional machine learning methods, which are not conducive to real-time deployment. Attached Figure Description
[0021] The accompanying drawings, which are incorporated in and form part of this specification, illustrate embodiments consistent with this application and, together with the description, serve to explain the principles of this application. It is obvious that the drawings described below are merely some embodiments of this application, and those skilled in the art can obtain other drawings based on these drawings without any inventive effort.
[0022] In the attached diagram: Figure 1A flowchart illustrating a UAV control method based on reservoir calculation provided in an embodiment of the present invention; Figure 2 This is a schematic diagram of a deep reservoir optical computing system provided in one embodiment of the present invention; Figure 3 This is a schematic diagram of a drone control device based on reservoir calculation provided in one embodiment of the present invention; Figure 4 This is a schematic diagram of an electronic device provided in one embodiment of the present invention. Detailed Implementation
[0023] The following specific examples illustrate the implementation of the present invention. Those skilled in the art can easily understand other advantages and effects of the present invention from the content disclosed in this specification. The present invention can also be implemented or applied through other different specific embodiments. Various details in this specification can also be modified or changed based on different viewpoints and applications without departing from the spirit of the present invention. In the absence of conflict, the following embodiments and features in the embodiments can be combined with each other.
[0024] This invention discloses a UAV control method based on reservoir calculation. This method can solve the problems of attitude instability and decreased control accuracy caused by complex aerodynamic disturbances such as ground effect and eddies when existing UAVs fly at low altitudes, near the ground, or in confined spaces. This control method is applicable to scenarios with stringent requirements for flight stability and anti-disturbance, such as alleyway delivery in urban logistics, emergency detection inside buildings or complex pipelines, and continuous monitoring in near-ground environments such as jungles and valleys.
[0025] Please see Figure 1 The control method includes the following steps: S10, real-time acquisition of the expected operating status and actual operating status of the UAV.
[0026] To achieve closed-loop control of the drone's flight trajectory, it is first necessary to acquire two reference operating states in real time: the expected operating state of the drone and the actual operating state of the drone in the real environment.
[0027] The desired operational state refers to the motion state that the UAV is expected to achieve at a specific moment, determined by the UAV's trajectory planning algorithm or operator commands based on a pre-set flight mission. The desired operational state includes a series of time-continuous reference data, which may specifically include vector information such as desired position, desired velocity, desired acceleration, and desired angular velocity.
[0028] Actual operational status refers to the real-time motion state of the UAV as measured by its various onboard sensors. These sensors include inertial measurement units, global navigation satellite system receivers, visual odometry, and optical flow sensors. By fusing the raw sensor data, the UAV's actual position, actual velocity, actual acceleration, actual attitude angle, and angular velocity at the current moment can be obtained. This actual operational status is the most direct reflection of the UAV's true flight condition.
[0029] During the operation of the drone, the desired operating state and the actual operating state are acquired in parallel and at high frequency. The desired operating state comes from a preset trajectory command stream, while the actual operating state comes from the real-time measurement data stream from the sensors. After synchronizing the desired and actual operating states, subsequent error calculations and control quantity generation can be performed.
[0030] Please see Figure 1 The control method further includes the following steps: S20, based on the desired and actual operating states of the current UAV, nominal feedforward control data is calculated using a preset rigid body dynamics nominal model. The nominal feedforward control data is calculated using the rigid body dynamics nominal model.
[0031] in, For nominal feedforward control data, This represents the current actual location of the drone. For the current expected speed of drones, To accelerate the expected performance of current drones, For the quality matrix, The matrix of Coriolis force and centrifugal force. This is the gravity vector.
[0032] After obtaining the desired and actual operating states, calculations need to be performed using a pre-defined rigid body dynamics nominal model to generate a basic control command, namely nominal feedforward control data. The core function of nominal feedforward control data is to provide the main power required for the UAV to maintain baseline flight.
[0033] The nominal rigid body dynamics model is a mathematical model based on the physical parameters of an unmanned aerial vehicle (UAV) that describes the motion of the UAV under ideal conditions where complex aerodynamic disturbances (such as ground effects and eddies) are neglected. The nominal rigid body dynamics model expresses the dynamic characteristics of the UAV as functions of parameters such as the mass matrix, Coriolis force and centrifugal force matrices, and gravity vector.
[0034] In the specific calculations, the actual position, desired velocity, and desired acceleration of the UAV need to be substituted into the rigid body dynamics nominal model as inputs. The actual position is primarily used to determine the UAV's current attitude in the gravitational field, thereby accurately calculating the gravitational components. The desired velocity and desired acceleration represent the desired motion state that the UAV is expected to achieve in the near future. Based on these desired values, the rigid body dynamics nominal model calculates the ideal control force required to drive the UAV to achieve this ideal motion.
[0035] The calculation process for nominal feedforward control data is as follows: First, the desired acceleration is multiplied by the mass matrix to obtain the force required to overcome the UAV's own inertia. Next, the Coriolis force and centrifugal force terms corresponding to the desired velocity are calculated. Then, the gravity component under the current attitude is calculated based on the actual position. Finally, these force components are synthesized. The final calculated result is the nominal feedforward control data. The nominal feedforward control data is a control vector that directly corresponds to the desired output force or torque of each actuator (such as motor) of the UAV.
[0036] The purpose of nominal feedforward control data is to enable the UAV to perfectly track the desired operating state in an ideal, undisturbed environment. However, in real-world, complex environments, nominal feedforward control data alone is insufficient to counteract actual nonlinear aerodynamic disturbances, thus requiring subsequent feedback and feedforward compensation for correction.
[0037] Please see Figure 1 The control method further includes the following steps: S30, based on the deviation between the expected operating state and the actual operating state of the current UAV, perform feedback control calculations to obtain feedback compensation control data. S30 includes the following steps: Based on the deviation between the expected and actual operating states of the UAV, feedback control calculations are performed to obtain feedback compensation control data. The integrals of the speed deviation and the historical speed deviation are independently adjusted and then fused to calculate the feedback compensation control data; the calculation formula for the feedback compensation control data is as follows:
[0038] in, The actual speed of the current drone, For the current expected speed of drones, , These are the proportional weight matrix and the integral weight matrix, respectively.
[0039] After generating nominal feedforward control data, a closed-loop feedback mechanism needs to be introduced to correct the deviation between the actual operating state and the desired operating state. This step is achieved by calculating and applying feedback compensation control data, and its core is to adjust based on the error between the desired and actual operating states.
[0040] Specifically, the first step is to calculate the velocity deviation. Velocity deviation is a parameter reflecting the rate of change of the drone's motion state, directly affecting the timeliness and stability of tracking the desired operational state. Subtracting the desired velocity contained in the desired operational state from the actual velocity in the actual operational state, measured and processed by sensors, yields the velocity deviation at the current moment. This velocity deviation quantifies the instantaneous difference between the drone's current direction and speed of motion and the desired velocity.
[0041] After obtaining the speed deviation, a proportional-integral (PI) control strategy is used to calculate the feedback compensation control data. The PPI control strategy involves two parallel processing steps: proportional adjustment and integral adjustment.
[0042] The proportional control responds to the current velocity deviation by multiplying it by a pre-set proportional weight matrix. This proportional weight matrix is a diagonal matrix, and its diagonal elements determine the strength of the UAV's correction for the velocity deviation in each degree of freedom. The purpose of proportional control is to quickly generate a corrective force proportional to the magnitude of the current velocity deviation, thereby reducing instantaneous tracking errors.
[0043] Integral control focuses on the cumulative effect of speed deviation. It integrates all historical speed deviations from the start of control to the current moment, obtaining a cumulative error. This cumulative error reflects a long-standing steady-state deviation that has not been completely eliminated by proportional control. This integral result is then multiplied by a pre-defined integral weight matrix. The integral weight matrix gradually eliminates the steady-state error caused by persistent disturbances or model mismatches, ultimately bringing the speed deviation close to zero.
[0044] Finally, the proportional term, adjusted by the proportional weight matrix, is fused with the integral term, adjusted by the integral weight matrix, through vector addition to obtain the feedback compensation control data. This feedback compensation control data, as a dynamic correction quantity, is superimposed on the nominal feedforward control data and works together with the UAV to resist disturbances and uncertainties under various conditions, driving the actual speed to converge towards the desired speed. Feedback compensation control is a continuous online adjustment process, recalculating based on the latest speed deviation in each control cycle.
[0045] Please see Figure 1 and Figure 2 The control method also includes the following steps: S40, inputting the actual operating status of the current UAV into the deep reserve pool optical computing system, which then predicts in real time the aerodynamic interference experienced by the UAV in the current environment and outputs the corresponding feedforward compensation control data.
[0046] S40 includes the following steps: S41, through the input layer of the deep reservoir optical computing system, the actual operating state of the current UAV is photoelectrically converted and modulated to form a multi-channel modulated optical signal.
[0047] While generating nominal feedforward control data and feedback compensation control data, the UAV controller needs to simultaneously input the control data obtained by fusing the current UAV's nominal feedforward control data and feedback compensation control data into the deep reservoir optical computing system to update the internal parameters of the output layer of the deep reservoir optical computing system.
[0048] To cope with complex aerodynamic disturbances, a parallel learning-based feedforward compensation channel needs to be activated. The first step of the learning-based feedforward compensation channel is to convert the current actual operating state of the UAV into a signal format suitable for processing by the deep-pool optical computing system. This step is completed through the input layer 110 of the deep-pool optical computing system, whose core function is to realize the conversion from electrical signals to optical signals and information loading.
[0049] The actual operating status, as raw electrical signals from the UAV's sensors, is first transmitted to the input layer 110 of the deep reservoir optical computing system. The actual operating status includes multiple synchronous data streams, such as actual position, actual velocity, and actual acceleration; each stream represents the UAV's motion information in a specific degree of freedom. Before entering the input layer 110 of the deep reservoir optical computing system, these electrical signals undergo preprocessing such as format alignment and normalization to ensure the validity and stability of the input information.
[0050] Input layer 110 is composed of It consists of several parallel optical channels (ch1~chn). The value is a positive integer. Each optical channel independently receives one data stream. At the core of each optical channel is a light source capable of generating a stable optical carrier, such as a single-mode laser. These optical carriers have different but fixed wavelengths, thus forming a wavelength division multiplexing (WDM) architecture, allowing multiple data streams to be transmitted independently and simultaneously in a single physical optical path without interference.
[0051] Each optical channel is equipped with an optical modulator. The pre-processed electrical signal representing the actual operating state of each channel is applied to its corresponding optical modulator. The optical modulator encodes the information of the electrical signal onto a specific physical property of the optical carrier, such as light intensity, phase, or frequency. Through this modulation process, the continuously changing electrical signal characterizing the UAV's motion state is converted into corresponding changes in the intensity or phase of the optical carrier.
[0052] After modulation, each independent electrical signal becomes a modulated optical signal carrying specific state information. Subsequently, all these modulated optical signals of different wavelengths are combined into the same transmission medium (such as optical fiber) through a wavelength division multiplexer, forming a composite multi-modulated optical signal. This multi-modulated optical signal contains a high-fidelity mapping of the UAV's current actual operating state in the optical domain, preparing it for subsequent high-dimensional nonlinear dynamic calculations injected into the optical reservoir layer 120.
[0053] S40 further includes the following step: S42, using the reservoir layer of the deep reservoir optical computing system, high-dimensional nonlinear temporal features are extracted from the multi-channel modulated optical signals to generate reservoir features. The reservoir layer comprises multiple cascaded reservoirs, and S42 includes the following steps: For each reserve pool: Based on the received current input optical signal and historical input optical signal, the current time series feature is obtained through high-dimensional nonlinear time series feature extraction; wherein, the input of the first reservoir is a multi-channel modulated optical signal, and the input of each subsequent reservoir is a modulated optical signal obtained by processing the time series feature output by the previous reservoir through an optical injection locking method. The time-series features of all the reserve pools are concatenated to generate the reserve pool features.
[0054] The first reservoir in reservoir layer 120 is the starting point of the information processing flow. It receives the multi-modulated optical signal generated by input layer 110 as its current input optical signal. This current input optical signal carries the UAV's current state information. Besides the current input optical signal, the internal dynamics of the reservoir are also influenced by its own historical state, i.e., the historical input optical signal, which can be the current input optical signal from the previous time point. The current and historical input optical signals interact within the nonlinear dynamic elements of the reservoir (e.g., a multi-mode semiconductor laser). This interaction forces the longitudinal mode amplitude and phase of the semiconductor laser to produce complex nonlinear responses. The instantaneous output intensity and spectral distribution of the semiconductor laser contain high-dimensional, time-series-dependent nonlinear transformation results. This process of fusing the current and historical input optical signals through nonlinear dynamics and mapping them to a high-dimensional state space is called high-dimensional nonlinear temporal feature extraction. The instantaneous dynamic output of the first reservoir is the current temporal feature. The calculation formula for the current temporal feature is as follows:
[0055] in, For the first The current time-series characteristics of each reserve pool It is a positive integer. For nonlinear calculations within the storage pool, , These are the weights for the input layer and the hidden layer, respectively. For the current input optical signal, Input optical signals for historical data.
[0056] To construct deeper networks and extract more abstract temporal features, this embodiment employs a cascaded reservoir structure. The current temporal features output from the first reservoir are not only fed into the final output layer 130 but also passed to the next reservoir as input for deeper processing. This transfer to the next reservoir is achieved through optical injection locking. Specifically, the current temporal features output from the first reservoir are injected into the laser of the second reservoir. By precisely controlling the intensity and frequency of the injected light, the semiconductor laser of the second reservoir is locked and tracks certain dynamic characteristics of the output signal from the first reservoir. In this way, the current temporal features of the first reservoir are effectively encoded as a modulated optical signal and passed to the second reservoir, becoming the current input optical signal of the second reservoir. Simultaneously, the second reservoir itself maintains its historical input optical signals through an internal optical feedback loop.
[0057] The second reservoir, based on the modulated optical signal output from the first reservoir and its own historical input optical signals, undergoes high-dimensional nonlinear mapping and fusion within its internal nonlinear laser dynamics to extract deeper, more abstract current temporal features. These current temporal features may reflect perturbation patterns on longer timescales or under more complex modes. This process can be repeated downstream. For example, the current temporal features of the second reservoir are also passed to the third reservoir via optical injection locking, and so on, forming a multi-layered cascaded deep processing structure. The input to each subsequent reservoir is the modulated optical signal obtained by processing the current temporal features output from the previous reservoir using optical injection locking.
[0058] Ultimately, all cascaded reservoirs operate in parallel at different depths, each generating its corresponding current temporal features. To comprehensively utilize all temporal information extracted from shallow to deep layers, these features need to be integrated. Therefore, the acquisition module of output layer 130 gathers the temporal features from all reservoirs and concatenates them in a specific order. This concatenation operation connects the high-dimensional state vectors output by each reservoir into a higher-dimensional composite vector, i.e., the final generated reservoir features. These reservoir features encode, at multiple scales and levels, all dynamic patterns related to complex aerodynamic disturbances inherent in the UAV's historical and current states, providing a rich information foundation for subsequent linear readout and accurate prediction of residual forces by output layer 130.
[0059] S40 also includes the following step: S43, Calculate and output feedforward compensation control data corresponding to the aerodynamic interference experienced by the UAV in the current environment through the output layer of the deep reservoir optical computing system, based on its internal parameters and reservoir characteristics. The internal parameters of the output layer of the deep reservoir optical computing system are updated in real time according to a preset update function, the calculation formula of which is as follows:
[0060] in, These are the internal parameters of the output layer. As a characteristic of the reserve pool, The regularization coefficient is . For transpose, It is the identity matrix. This refers to the control data obtained by fusing the nominal feedforward control data and the feedback compensation control data of the current UAV.
[0061] The formula for calculating feedforward compensation control data is as follows:
[0062] in, For feedforward compensation control data, This is a characteristic of a reserve pool.
[0063] The output layer 130 can utilize the current reservoir characteristics to calculate and generate feedforward compensation control data in real time. The feedforward compensation control data is used to directly offset the aerodynamic interference experienced by the UAV in the current complex flight environment.
[0064] The computation process of output layer 130 is essentially a linear transformation, with its internal parameter being the output weight matrix. This internal parameter defines how to map the high-dimensional reservoir features to the low-dimensional control output space. In each control cycle, when the reservoir features are ready, output layer 130 performs a matrix-vector multiplication operation, multiplying the reservoir features by the output weight matrix to generate a specific, numerical form of feedforward compensation control data. This feedforward compensation control data is the prediction of the residual aerodynamic forces experienced by the UAV at the current moment and in the near future by the deep reservoir optical computing system, transmitted back to the UAV controller in real time as an electrical signal.
[0065] To ensure that the output layer 130 can continuously and accurately predict time-varying aerodynamic disturbances, its output weight matrix is not fixed but needs to be adaptively adjusted online according to changes in flight status and environment. This adjustment is performed in real time by the weight module of the output layer 130 according to a preset update function. The mathematical form of the update function comprehensively considers the accuracy of prediction, the generalization ability of the model, and numerical stability. Specifically, the update process requires the use of the reservoir characteristics and control data calculated at the current moment.
[0066] After generating nominal feedforward control data and feedback compensation control data in each control cycle, the UAV controller superimposes and fuses these two data sets to obtain the control data. In the update function of the weight module in output layer 130, the update function correlates the reservoir features with the control data through a recursive least squares calculation. Its purpose is to adjust the output weight matrix so that the feedforward compensation control data calculated by the weight module of output layer 130 based on the reservoir features can approximate as closely as possible the actual control quantity needed to compensate for unknown aerodynamic disturbances. The regularization coefficient in the update formula is used to prevent overfitting when data is limited or feature correlations are strong, ensuring that the solution of the output weight matrix is numerically stable; the identity matrix is used to guarantee the non-singularity of the matrix in the inversion operation, making the update process mathematically feasible at all times.
[0067] Therefore, the output layer 130 operates as a dynamic, closed-loop adaptive process. On one hand, it utilizes the reservoir characteristics to calculate and output the feedforward compensation in real time; on the other hand, it receives control data from the control closed loop as a supervisory signal, continuously fine-tuning its internal parameters through an update function. This design enables the deep reservoir optical computing system to learn and track the dynamic changes of aerodynamic disturbances online, thereby continuously optimizing the accuracy of its predictive compensation and ultimately helping the UAV achieve high-precision, stable flight in complex airflow environments.
[0068] Please see Figure 1 The control method also includes the following steps: S50, fusing the nominal feedforward control data, feedback compensation control data, and feedforward compensation control data to obtain the UAV's flight control data.
[0069] Specifically, nominal feedforward control data is pre-calculated based on the nominal rigid body dynamics model of the UAV. The main function of nominal feedforward control data is to provide a baseline control force to compensate for the gravity terms required to maintain flight, as well as the inertial forces and dynamic coupling terms required to drive the UAV along the desired operating state. In essence, nominal feedforward control data is an approximation of the control inputs required under ideal, undisturbed flight conditions.
[0070] Feedback compensation control data is generated based on the real-time error between the desired and actual operating states. Through the feedback control channel within the UAV controller, sensors continuously collect the UAV's actual position, velocity, attitude, and other state information, comparing it with the desired operating state to generate a tracking error. The nonlinear feedback controller within the UAV controller (e.g., a proportional-integral-derivative controller or an advanced variant thereof) calculates the feedback compensation control data according to this real-time error and a preset control law. The purpose of the feedback compensation control data is to suppress deviations caused by model uncertainties, initial state biases, and disturbances not completely canceled out by the feedforward channel.
[0071] The feedforward compensation control data is predicted and output in real time by the deep optical storage pool optical computing system. The deep optical storage pool optical computing system receives the actual operating state of the UAV as input, extracts high-dimensional nonlinear temporal features in the deep optical storage pool, and calculates the predicted results of the additional forces caused by complex aerodynamic disturbances (such as ground effect and eddy current interference) in the current flight environment from its output layer. This predicted result is the feedforward compensation control data, which serves as an active, learning-based compensation signal to proactively offset these nonlinear, time-varying aerodynamic disturbances, thereby reducing the adjustment burden on the feedback controller and significantly improving the ability to suppress rapidly changing disturbances.
[0072] The fusion process is mathematically represented as an additive synthesis. Specifically, the final generated total flight control data equals the sum of the nominal feedforward control data, the feedback compensation control data, and the feedforward compensation control data. This linear superposition is a typical "feedforward-feedback" composite control structure. The nominal feedforward control data provides the basic driving force, the feedback compensation control data handles general tracking errors and model mismatches, while the feedforward compensation control data provided by the deep-reservoir optical computing system specifically performs refined "active elimination" of complex dynamic disturbances. In this way, model prior knowledge, closed-loop feedback stability, and online adaptive learning capabilities for unknown disturbances are integrated.
[0073] The resulting fused flight control data corresponds in dimension and physical meaning to the UAV's control inputs (e.g., thrust or torque commands for each rotor). This flight control data is sent to the UAV's actuators, such as the electronic speed controller and brushless motors, driving the rotors to generate corresponding thrust and torque, thereby achieving precise, robust, and adaptive tracking of the desired operating state. The entire fusion process is completed at high speed within an embedded airborne computing system, ensuring real-time control and ultimately achieving stable attitude and trajectory control of the UAV in narrow, variable, and complex aerodynamically disturbed unknown environments.
[0074] As can be seen, compared with traditional compensation algorithms based on empirical models, this scheme, through data-driven and optical high-dimensional nonlinear mapping, can more accurately learn and predict unsteady, time-varying residual aerodynamic forces caused by ground effects, eddy current coupling, etc., thus effectively overcoming the problem of limited accuracy and adaptability of empirical models. Secondly, compared with methods based on complex digital neural networks such as multilayer perceptrons, temporal convolutional networks, and recurrent neural networks, the core computing unit of this scheme is an optical reservoir that only needs to train a linear readout layer. This fundamentally avoids the stability problems of gradient vanishing or exploding, while reducing the training time to the sub-millisecond level. Furthermore, by utilizing the high bandwidth and low latency characteristics of optical computing, it meets the stringent requirements of airborne platforms for real-time performance and low power consumption, solving the problems of traditional machine learning methods being complex, time-consuming to train, and having high computational power consumption, which are unfavorable for real-time deployment.
[0075] Please see Figure 3 The present invention also discloses a UAV control device based on reservoir calculation, and the above control method can be applied to the control device. The control device may include an operation data acquisition module 210, a feedforward data generation module 220, a feedback data generation module 230, a compensation data generation module 240, and a control data fusion module 250.
[0076] The system comprises the following modules: Operational data acquisition module 210 acquires the expected and actual operating states of the UAV in real time. Feedforward data generation module 220 inputs the expected and actual operating states of the current UAV into a preset rigid body dynamics nominal model to obtain nominal feedforward control data. Feedback data generation module 230 performs feedback control calculations based on the deviation between the expected and actual operating states of the current UAV to obtain feedback compensation control data. Compensation data generation module 240 inputs the actual operating state of the current UAV into a deep-reservoir optical computing system, which predicts the aerodynamic interference experienced by the UAV in the current environment in real time and outputs corresponding feedforward compensation control data. Control data fusion module 250 fuses the nominal feedforward control data, feedback compensation control data, and feedforward compensation control data to obtain the UAV's flight control data.
[0077] For specific limitations regarding the control device, please refer to the limitations of the control method above, which will not be repeated here. Each module in the aforementioned control device can be implemented entirely or partially through software, hardware, or a combination thereof. These modules can be embedded in hardware or independently of the memory in the electronic device, or stored in software form in the memory of the electronic device, so that the memory can call and execute the operations corresponding to each module.
[0078] Please see Figure 4The electronic device 300 may include a memory 310, a processor 320 and a bus, and may also include a computer program stored in the memory 310 and executable on the processor 320, such as a program for a drone control method based on a reservoir calculation.
[0079] Furthermore, the memory 310 includes at least one type of readable storage medium, including flash memory, portable hard drive, multimedia card, card-type memory (e.g., SD or DX memory), magnetic memory, magnetic disk, optical disk, etc. In some embodiments, the memory 310 can be an internal storage unit of the electronic device 300, such as the portable hard drive of the electronic device 300. In other embodiments, the memory 310 can also be an external storage device of the electronic device 300, such as a plug-in portable hard drive, smart media card (SMC), secure digital (SD) card, flash card, etc., equipped on the electronic device 300. Furthermore, the memory 310 can also include both internal and external storage units of the electronic device 300. The memory 310 can be used not only to store application software and various types of data installed on the electronic device 300, such as code for a drone control method based on a reservoir calculation, but also to temporarily store data that has been output or will be output.
[0080] Specifically, in some embodiments, the processor 320 may be composed of integrated circuits, such as a single packaged integrated circuit or multiple integrated circuits packaged with the same or different functions, including combinations of one or more central processing units (CPUs), microprocessors, digital processing chips, graphics processors, and various control chips. The processor 320 is the control unit of the electronic device 300, connecting various components of the entire electronic device 300 through various interfaces and lines. It executes programs or modules stored in the memory 310 (e.g., programs for drone control methods based on reservoir calculations) and calls data stored in the memory 310 to perform various functions of the electronic device 300 and process data.
[0081] The processor 320 executes the operating system of the electronic device 300 and various installed applications. The processor 320 executes the applications to implement the steps in the above-described UAV control method based on reservoir computing. The computer program can be divided into one or more modules, one or more of which are stored in the memory 310 and executed by the processor 320 to complete this application. The one or more modules can be a series of computer program instruction segments capable of performing specific functions, which describe the execution process of the computer program in the electronic device 300. For example, the computer program can be divided into a running data acquisition module 210, a feedforward data generation module 220, a feedback data generation module 230, a compensation data generation module 240, a control data fusion module 250, etc.
[0082] The above embodiments are merely illustrative of the principles and effects of the present invention and are not intended to limit the invention. Any person skilled in the art can modify or alter the above embodiments without departing from the spirit and scope of the present invention. Therefore, all equivalent modifications or alterations made by those skilled in the art without departing from the spirit and technical concept disclosed in the present invention should still be covered by the claims of the present invention.
Claims
1. A method for controlling unmanned aerial vehicles (UAVs) based on reservoir calculation, characterized in that, include: Real-time acquisition of the expected and actual operating status of the drone; Based on the expected and actual operating states of the UAV, nominal feedforward control data are calculated using a pre-set rigid body dynamics nominal model. Based on the deviation between the expected and actual operating states of the UAV, feedback control calculations are performed to obtain feedback compensation control data. The actual operating status of the current UAV is input into the deep reservoir optical computing system, which then predicts the aerodynamic interference experienced by the UAV in the current environment in real time and outputs the corresponding feedforward compensation control data. The nominal feedforward control data, the feedback compensation control data, and the feedforward compensation control data are fused to obtain the flight control data of the UAV.
2. The UAV control method based on reservoir calculation according to claim 1, characterized in that, The nominal feedforward control data is calculated using a rigid body dynamics nominal model: in, For nominal feedforward control data, This represents the current actual location of the drone. For the current expected speed of drones, To accelerate the expected performance of current drones, For the quality matrix, The matrix of Coriolis force and centrifugal force. This is the gravity vector.
3. The UAV control method based on reservoir calculation according to claim 1, characterized in that, The step involves performing feedback control calculations based on the deviation between the expected and actual operating states of the current UAV, to obtain feedback compensation control data, including: Calculate the speed deviation based on the expected speed of the current UAV in its expected operating state and the actual speed in its actual operating state; The integrals of the speed deviation and the historical speed deviation are independently adjusted and then fused to calculate the feedback compensation control data.
4. The UAV control method based on reservoir calculation according to claim 3, characterized in that, The formula for calculating the feedback compensation control data is as follows: in, The actual speed of the current drone, For the current expected speed of drones, , These are the proportional weight matrix and the integral weight matrix, respectively.
5. The UAV control method based on reservoir calculation according to claim 1, characterized in that, The process involves inputting the current actual operating status of the UAV into the deep reservoir optical computing system, which then predicts in real time the aerodynamic disturbances experienced by the UAV in the current environment and outputs corresponding feedforward compensation control data, including: Through the input layer of the deep reservoir optical computing system, the actual operating status of the current UAV is photoelectrically converted and modulated to form multi-channel modulated optical signals. Through the reservoir layer of the deep reservoir optical computing system, high-dimensional nonlinear temporal features are extracted from the multi-channel modulated optical signals to generate reservoir features. The output layer of the deep reservoir optical computing system calculates and outputs feedforward compensation control data corresponding to the aerodynamic interference experienced by the UAV in the current environment, based on its internal parameters and the characteristics of the reservoir.
6. The UAV control method based on reservoir calculation according to claim 5, characterized in that, The reservoir layer comprises multiple cascaded reservoirs. The reservoir layer of the deep reservoir optical computing system performs high-dimensional nonlinear temporal feature extraction on the multi-path modulated optical signals to generate reservoir features, including: For each reserve pool: Based on the received current input optical signal and historical input optical signal, the current time series feature is obtained through high-dimensional nonlinear time series feature extraction; wherein, the input of the first reservoir is a multi-channel modulated optical signal, and the input of each subsequent reservoir is a modulated optical signal obtained by processing the time series feature output by the previous reservoir through an optical injection locking method. The time-series features of all the reserve pools are concatenated to generate the reserve pool features.
7. The UAV control method based on reservoir calculation according to claim 5, characterized in that, The internal parameters of the output layer of the deep reservoir optical computing system are updated in real time according to a preset update function, the calculation formula of which is as follows: in, These are the internal parameters of the output layer. As a characteristic of the reserve pool, The regularization coefficient is . It is the identity matrix. This refers to the control data obtained by fusing the nominal feedforward control data and the feedback compensation control data of the current UAV.
8. A drone control device based on reservoir calculation, characterized in that, include: The data acquisition module is used to acquire the expected and actual operating status of the drone in real time. The feedforward data generation module is used to calculate nominal feedforward control data based on the expected and actual operating states of the current UAV and through a preset rigid body dynamics nominal model. The feedback data generation module is used to perform feedback control calculations based on the deviation between the expected operating state and the actual operating state of the current UAV, and obtain feedback compensation control data. The compensation data generation module is used to input the actual operating status of the current UAV into the deep reserve pool optical computing system, which then predicts the aerodynamic interference experienced by the UAV in the current environment in real time and outputs the corresponding feedforward compensation control data. The control data fusion module is used to fuse the nominal feedforward control data, the feedback compensation control data, and the feedforward compensation control data to obtain the flight control data of the UAV.
9. An electronic device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that, When the processor executes the computer program, it implements the steps of the unmanned aerial vehicle control method based on reservoir calculation as described in any one of claims 1 to 7.
10. A computer-readable storage medium storing a computer program, characterized in that, When the computer program is executed by the processor, it implements the steps of the unmanned aerial vehicle control method based on reservoir calculation as described in any one of claims 1 to 7.