Unmanned aerial vehicle power calculation method and system based on physical information neural network

Abstract

This invention discloses a method for calculating UAV power based on a physical information neural network. The method constructs a deep feature learning architecture including a bidirectional long short-term memory network (BiLSTM), which can bidirectionally encode the UAV's flight state from a temporal dimension, effectively solving the problem of large estimation errors in dynamic tasks by traditional models based on quasi-steady-state assumptions. Simultaneously, it innovatively proposes a structured physical embedding strategy, explicitly mapping and constraining the prior knowledge of the physical model to neurons in specific hidden layers of the network through a multi-objective composite loss function. This mechanism forces the internal representation of the "black box" neural network to follow the laws of aerodynamics, not only clarifying the physical meaning of the network output components but also ensuring that the model can still output physically logical predictions in flight areas not covered by training data. This greatly reduces model uncertainty and provides a reliable basis for safety-critical flight decisions.

Description

Technical Field

[0001] This invention relates to the field of unmanned aerial vehicle (UAV) power estimation technology, and in particular to a method and system for calculating UAV power based on a physical information neural network. Background Technology

[0002] With the increasingly widespread application of quadcopter drones in military, logistics, and agriculture, their limited endurance has become a key bottleneck restricting their mission range and operational performance. Therefore, achieving high-precision real-time power prediction has become a core issue in drone energy management and mission planning. Existing power estimation methods often rely on quasi-steady-state assumptions in traditional mechanism-driven models, making it difficult to fully couple the transient dynamic effects generated by drones during acceleration, deceleration, and complex maneuvers, thus limiting estimation accuracy.

[0003] To address dynamic operating conditions, while data-driven methods based on deep learning, thanks to their powerful nonlinear fitting capabilities and the ability to avoid building complex mechanistic models, are essentially "black box" models lacking physical interpretability. These models map data solely based on statistical regularities, and are prone to high uncertainty in flight conditions not covered by training data. Their predictions may even violate fundamental physical laws, making them unreliable in safety-critical decisions. Furthermore, existing hybrid modeling attempts often lack an effective mechanism to simultaneously capture complex temporal dependencies using neural networks (such as recurrent neural networks) while forcing the network's internal feature representations to adhere to a clear physical structure. This lack of physical constraint prevents the provision of power predictions for UAVs that are both highly accurate, robust, and physically logical. Therefore, there is an urgent need for a new method that can structurally embed prior physical knowledge into deep neural networks while simultaneously learning transient temporal features to improve the generalization ability and reliability of power estimation. Summary of the Invention

[0004] To address the technical problems of insufficient estimation accuracy and poor physical interpretability of existing UAV power estimation methods under dynamic conditions, this invention proposes a UAV power calculation method and system based on a physical information neural network.

[0005] The technical solution adopted in this invention is as follows:

[0006] A method for calculating the power of a UAV based on a physical information neural network, the method comprising the following steps:

[0007] Step 1: Construct a time-series flight status dataset consisting of several UAV flight status data and corresponding power data; wherein, the UAV flight status data includes horizontal flight speed and vertical flight data;

[0008] Step 2: Construct a deep neural network architecture that integrates temporal learning, wherein the deep neural network architecture includes cascaded fully connected layers and a bidirectional long short-term memory network;

[0009] Step 3: Based on the time-series flight state dataset constructed in Step 1, train the deep neural network architecture constructed in Step 2 to obtain the UAV power calculation model;

[0010] Step 4: Using the UAV power calculation model obtained in Step 3, calculate the real-time total power of the UAV under test.

[0011] Furthermore, the data constraint term defined by mean squared error is used as the loss function to train the deep neural network architecture constructed in step 2. The data constraint term is defined as follows:

[0012] ,

[0013] In the formula, This represents the total number of samples in the time-series flight state dataset. Representing the The total power prediction value corresponding to each sample; Representing the The total power measurement value corresponding to each sample.

[0014] Furthermore, a multi-objective composite loss function is used to train the deep neural network architecture formed in step 2. The multi-objective composite loss function is defined as follows:

[0015] ,

[0016] ,

[0017] ,

[0018] ,

[0019] In the formula, For multi-objective composite loss functions, For data constraints, For L2 regularization constraints, For smoothness constraint terms, , and They are respectively , and The weighting coefficients, This represents the total number of samples in the time-series flight state dataset. Representing the The total power prediction value corresponding to each sample Representing the The total power measurement value corresponding to each sample This represents any trainable parameter in a deep neural network; This represents the set of parameters consisting of all trainable parameters of the deep neural network, including weight parameters and bias parameters in the fully connected layer, bidirectional long short-term memory network, physical constraint layer, and output layer.

[0020] Furthermore, in step 2, the bidirectional long short-term memory network consists of forward LSTM units and backward LSTM units, wherein the forward LSTM units generate a forward hidden state sequence. The backward LSTM unit generates the backward hidden state sequence. The final output of the bidirectional long short-term memory network is , For time step Final output, ;

[0021] Each LSTM unit in the bidirectional long short-term memory network maintains its internal state through a gating mechanism, and the internal state is composed of hidden states. and cell state The common representation, whose update mechanism is as follows:

[0022]

[0023]

[0024]

[0025] In the formula, For time step The input vector, They are time steps The activation values ​​of the forget gate, input gate, and output gate; For time step Candidate cell state; and They are time steps The hidden state and the cellular state; All are learnable parameters. = ; This indicates element-wise multiplication.

[0026] Furthermore, the method also includes:

[0027] Based on aerodynamic principles, a basic power model for drones is constructed:

[0028] ,

[0029] In the formula, For the UAV at a given velocity vector Total instantaneous power below, The horizontal flight speed of the drone, The vertical flight speed of the drone. This is the reference power for the drone when it is stationary and hovering. and These are the power increments compared to the hovering state, generated by horizontal and vertical motion, respectively.

[0030] The basic model of UAV power is used as the hard constraint rule of the deep neural network, and a physical constraint layer including three neuron nodes is added after the bidirectional long short-term memory network.

[0031] The three neuron nodes of the physical constraint layer The following mapping relationships are established with the three components in the basic power model of the UAV: ​​① Hovering reference mapping ② Horizontal dynamic mapping ③ Vertical dynamic mapping ;

[0032] After adding the physical constraint layer, the output layer is finally formed, resulting in the output of the deep neural network architecture:

[0033] ,

[0034] In the formula, For the output of a deep neural network architecture, They are respectively The activation value.

[0035] Furthermore, a multi-objective composite loss function is used to train the deep neural network architecture formed in step 2. The multi-objective composite loss function is defined as follows:

[0036] ,

[0037] ,

[0038] ,

[0039] ,

[0040] ,

[0041] ,

[0042] ,

[0043] ,

[0044] In the formula, For multi-objective composite loss functions, For physical constraint terms, For data constraints, For L2 regularization constraints, For smoothness constraint terms, , , and They are respectively , , and The weighting coefficients, , and These represent the horizontal motion loss term, the vertical motion loss term, and the hovering loss term, respectively. , and They are respectively , and The weighting coefficients, This means taking the mean over the training batches. This represents the total number of samples in the time-series flight state dataset. Representing the The total power prediction value corresponding to each sample Representing the The total power measurement value corresponding to each sample This represents any trainable parameter in a deep neural network; This represents the set of parameters consisting of all trainable parameters of the deep neural network, including weight parameters and bias parameters in the fully connected layer, bidirectional long short-term memory network, physical constraint layer, and output layer.

[0045] Furthermore, the method also includes:

[0046] Based on aerodynamic principles, a nonlinear power compensation term is superimposed on the basic power model of the UAV to compensate for transient dynamics, thus constructing a four-component power model for the UAV:

[0047] ,

[0048] In the formula, For the UAV at a given velocity vector Total instantaneous power below, The horizontal flight speed of the drone, The vertical flight speed of the drone. This is the reference power for the drone when it is stationary and hovering. and These are the power increments compared to the hovering state, generated by horizontal and vertical motion, respectively. This is the remainder term for nonlinear power compensation;

[0049] The four-component model of UAV power is used as a hard constraint rule for the deep neural network architecture, and a physical constraint layer including four neuron nodes is added after the bidirectional long short-term memory network.

[0050] The four neuron nodes of the physical constraint layer The following mapping relationships are established with the four components of the UAV power four-component model: ① Hovering reference mapping ② Horizontal dynamic mapping ③ Vertical dynamic mapping ④ Transient residual mapping ;

[0051] After adding the physical constraint layer, the output layer is finally formed, resulting in the output of the deep neural network architecture:

[0052] ,

[0053] In the formula, For the output of a deep neural network architecture, They are respectively The activation value.

[0054] Furthermore, a multi-objective composite loss function is used to train the deep neural network architecture formed in step 2. The multi-objective composite loss function is defined as follows:

[0055] ,

[0056] ,

[0057] ,

[0058] ,

[0059] ,

[0060] ,

[0061] ,

[0062] ,

[0063] ,

[0064] In the formula, For multi-objective composite loss functions, For physical constraint terms, For data constraints, For L2 regularization constraints, For smoothness constraint terms, , , and They are respectively , , and The weighting coefficients, , , and These represent the horizontal motion loss term, vertical motion loss term, hovering loss term, and additional loss term, respectively. , , and They are respectively , , and The weighting coefficients, This means taking the mean over the training batches. This represents the total number of samples in the time-series flight state dataset. Representing the The total power prediction value corresponding to each sample Representing the The total power measurement value corresponding to each sample This represents any trainable parameter in a deep neural network; This represents the set of parameters consisting of all trainable parameters of the deep neural network, including weight parameters and bias parameters in the fully connected layer, bidirectional long short-term memory network, physical constraint layer, and output layer.

[0065] Furthermore, the nonlinear power compensation remainder term is defined as:

[0066] ,

[0067] In the formula, Based on and The basis function row vector is formed by several nonlinear basis functions. , It is a coefficient column vector consisting of sparse coefficients corresponding to the nonlinear basis functions;

[0068] The coefficient column vector is obtained through a parameter identification method, specifically:

[0069] The linear subspace spanned by the three component vectors of the UAV power fundamental model is defined as... ,in, To correspond to the hovering reference power a constant vector, To correspond to the incremental power of horizontal motion The horizontal increment vector, To correspond to the incremental power of vertical motion The vertical increment vector;

[0070] Secondly, the target residual and All projected to orthogonal complementary space In the middle, the projected target residuals are obtained respectively. and the projected basis function row vectors ;in, For orthogonal projection operators, This is the measured total power vector;

[0071] Finally, for and Performing sparse regression, i.e., solving... This yields the coefficient column vector.

[0072] This invention also proposes a UAV power calculation system based on a physical information neural network that applies the method described above, the system comprising:

[0073] The data acquisition unit is used to collect UAV flight status data and construct a time-series flight status dataset consisting of several UAV flight status data and corresponding power data; wherein, the UAV flight status data includes horizontal flight speed and vertical flight data;

[0074] A network building unit is used to build a deep neural network architecture that integrates temporal learning, wherein the deep neural network architecture includes cascaded fully connected layers and a bidirectional long short-term memory network.

[0075] The network training unit is used to train the deep neural network architecture based on the time-series flight state dataset to obtain the UAV power calculation model;

[0076] The power calculation unit is used to calculate the real-time total power of the UAV under test using the obtained UAV power calculation model.

[0077] Beneficial effects:

[0078] 1. Significantly improves power prediction accuracy under dynamic flight conditions:

[0079] This invention constructs a deep feature learning architecture that includes a bidirectional long short-term memory network (BiLSTM), enabling bidirectional encoding of UAV flight states from a temporal perspective. Compared to traditional static models, this system can accurately capture the transient dynamic characteristics and long-term dependencies generated by UAVs during rapid acceleration, deceleration, and complex maneuvers, effectively solving the problem of large estimation errors in dynamic tasks by traditional models based on quasi-steady-state assumptions.

[0080] 2. Endow deep neural networks with physical interpretability and strong robustness:

[0081] This invention innovatively proposes a structured physical embedding strategy. Through a multi-objective composite loss function, it explicitly maps and constrains the prior knowledge of the physical model to neurons in specific hidden layers of the network. This mechanism forces the internal representation of the "black box" neural network to follow the laws of aerodynamics, not only clarifying the physical meaning of the network's output components but also ensuring that the model can still output physically logical predictions in flight regions not covered by training data. This significantly reduces model uncertainty and provides a reliable basis for safety-critical flight decisions. Attached Figure Description

[0082] Figure 1 This is a flowchart of the method of the present invention. Detailed Implementation

[0083] The present invention will be further described below with reference to the accompanying drawings and embodiments.

[0084] like Figure 1 As shown, the present invention proposes a method for calculating the power of a UAV based on a physical information neural network, which includes the following steps:

[0085] Step 1: Construct a time-series flight state dataset;

[0086] To provide high-quality, highly consistent input for subsequent neural network models, this step collects the UAV's state data and corresponding power data during flight. Data sources can integrate high-fidelity simulation datasets (such as AirSim simulation data) with real-world flight datasets to enhance the model's generalization ability.

[0087] The UAV flight status data mainly includes:

[0088] ① Horizontal flight speed :

[0089]

[0090] In the formula, Representing drones Axial velocity and Axial velocity, in units of .

[0091] ② Vertical flight speed :Right now Shaft speed, in units of The vertical upward direction is defined as positive.

[0092] The real-time power data is the total instantaneous power. This refers to the total power of the drone under a given flight condition, measured in watts (W).

[0093] Before being fed into the model for training, the collected raw data needs to undergo a series of preprocessing steps, including: format conversion, time synchronization (to ensure that the state data and power data are precisely aligned in time), data filtering (such as using a moving average filter to smooth high-frequency noise), outlier removal (such as removing extreme values), and sequence data construction (such as using a sliding window technique to construct time series samples), ultimately resulting in a high-quality dataset for model training.

[0094] Step 2: Construct a deep neural network architecture that integrates temporal learning;

[0095] The neural network architecture constructed in this invention aims to deeply integrate data-driven temporal feature learning with prior physical knowledge. The network architecture mainly includes a deep feature learner and a physical information fusion module.

[0096] Step 2.1: Constructing the Deep Feature Learner:

[0097] The deep feature learner is used to extract deep temporal features from the input flight state sequence. It is mainly composed of a fully connected layer (Dense Layer) and a bidirectional long short-term memory network (BiLSTM).

[0098] First, the fully connected layer is used to perform preliminary feature transformation on the input flight state data to achieve linear transformation and nonlinear mapping of features, thereby mapping the original input features to a high-dimensional feature space. For a given input vector... Its output The calculation formula is as follows:

[0099]

[0100] In the formula, It is a learnable weight matrix. It is a learnable bias vector. This represents a non-linear activation function. Through this layer, the model can map the original input features to a more expressive high-dimensional space.

[0101] Secondly, to capture the sequential dependencies of UAV flight status data over time, the output of the fully connected layer is input into the Bidirectional Long Short-Term Memory (BiLSTM) network. The BiLSTM consists of forward LSTM units and backward LSTM units. The forward LSTM unit starts from the beginning of the sequence. Until the end Process the input and generate a sequence of forward hidden states. The backward LSTM unit starts from the end of the sequence. Reverse processing to the beginning Generate a backward hidden state sequence .

[0102] at any time step The final output of the BiLSTM It is obtained by concatenating the forward hidden state and the backward hidden state at this moment, and the calculation formula is as follows:

[0103]

[0104] Each LSTM cell constituting the BiLSTM maintains its internal state through a gating mechanism, and the internal state consists of a hidden state. and cell state The common representation, whose update mechanism is as follows:

[0105]

[0106]

[0107]

[0108] In the formula, These are the activation values ​​for the three gates; These are learnable parameters; This represents element-wise multiplication. Through the bidirectional structure of BiLSTM, the final output is each feature vector in the sequence. It not only encodes information from past moments but also incorporates information from future moments, thus enabling a more comprehensive and accurate feature representation of the state at each point in time.

[0109] Step 2.2: Construction of the Physical Information Fusion Module:

[0110] To accurately estimate the power of multi-rotor UAVs in various flight missions, this invention establishes a hybrid power model framework combining physical principles and data-driven methods. Addressing the limitations of traditional power models based on aerodynamic principles, which typically rely on steady-state or quasi-steady-state flight assumptions and struggle to capture the additional nonlinear dynamic effects generated by UAVs during rapid acceleration, deceleration, and complex maneuvers, this invention decomposes the total power into a four-component model. The core of this model lies in the fundamental physical model... On top of this, a data-driven remainder term is superimposed to compensate for transient dynamics, thereby extending the applicability of the model from steady-state flight to all flight modes.

[0111] The mathematical expression for this four-component model framework is as follows:

[0112]

[0113] In the formula, The total instantaneous power of the drone at a given speed is expressed in watts (W). The horizontal flight speed of the drone is defined as The unit is meters per second (m / s); This refers to the vertical flight speed of the drone, i.e., the Z-axis speed, measured in meters per second (m / s), with the vertical upward direction defined as positive.

[0114] The first three components of this four-component model framework constitute the fundamental physical model. The reference power (W) for the drone in a stationary hovering state; and These are the power increments (W) compared to the hovering state, generated by horizontal and vertical motion, respectively. The fourth component... The nonlinear power compensation remainder term is used to characterize transient dynamic effects and multi-axis motion coupling effects that are not fully described by the fundamental physical model.

[0115] The nonlinear power compensation remainder term aims to identify and learn residual dynamics that the fundamental physical model fails to explain from the data. To this end, this invention constructs a nonlinear candidate basis function library consisting of high-order velocity polynomials and their cross-coupling terms, providing a mathematical foundation for describing energy dissipation in transient processes. By employing Lasso regression with cross-validation (LassoCV), the few terms that contribute most significantly to explaining the residuals are automatically identified from this large feature library, thereby constructing a concise and effective remainder term model. In the formula, Based on and Row vectors formed by several nonlinear basis functions. It is a column vector consisting of sparse coefficients corresponding to the nonlinear basis functions.

[0116] From horizontal flight speed and vertical flight speed It consists of constant terms, first-order terms, second-order terms, third-order terms, absolute value terms, and cross-coupling terms, and includes at least a constant term. First-order terms and Second-order terms , and Third-order terms , , and absolute value term and and cross-coupling terms and ; can be represented as: The remainder term is obtained by linearly combining the candidate nonlinear basis functions after sparse screening. This data-driven learned remainder term serves as a dynamic compensation for the fundamental physical model, thereby effectively extending the model's applicability from steady-state flight to general flight scenarios including acceleration, deceleration, and complex maneuvers.

[0117] This ensures that the remainder only learns information that the fundamental model cannot explain, without encroaching on the portion already explained by the three fundamental physics terms. To fundamentally address this issue, a sparse identification method based on orthogonal projection is employed:

[0118] First, the vector formed by the three components of the basic physical model on the sample sequence (i.e., the vector corresponding to the hovering reference power) is... constant vector , corresponding to the incremental power of horizontal motion vector and the power corresponding to the vertical motion increment vector The linear subspace spanned by is defined as .

[0119] To ensure complete decoupling of this remainder term from the fundamental physical components, a strict mathematical constraint is imposed: corresponding to the data-driven remainder term. The remainder vector Must be with subspace Orthogonal, that is This constraint is achieved through a projection operator. To achieve this, in which The subspace obtained through QR decomposition A set of orthonormal bases.

[0120] Before performing sparse regression, the target residual ( ) and All projected to orthogonal complementary space In the middle, the projected target residuals are obtained respectively. and the projected feature library In the formula, It is the identity matrix. This is the measured total power vector. This is the total power vector calculated from the fundamental physics model. The final sparse regression is applied to the projected variables. and The above is performed, i.e., the solution is obtained. .because All column vectors are located in the orthogonal complement space In, any linear combination of them (i.e., the final remainder) It must also be located in Therefore, this method mathematically guarantees that the learned remainder is completely decoupled from the three fundamental physics terms, and its value is a pure supplement to the predictive power of the fundamental model without recalculating or changing the contribution of the fundamental physics components.

[0121] Built using a data-driven approach, its basic principle is as follows: First, based on the measured total power... The residual is constructed from the difference between the power predicted by the fundamental physical model and the power predicted by the fundamental physical model. Then based on horizontal flight speed and vertical flight speed Constructing nonlinear basis function row vectors And obtain the coefficient vectors corresponding to each basis function through parameter identification. Thus, the nonlinear power compensation remainder term is obtained. By introducing this remainder term, the basic physical model can be dynamically compensated while maintaining its clear structure, thereby improving the accuracy and applicability of power estimation in acceleration, deceleration, and complex maneuvering scenarios.

[0122] Step 3: Establish a structured physical constraint mapping mechanism;

[0123] To overcome the "black box" nature of deep neural networks and give their internal reasoning processes a clear physical meaning, this step, based on the "physical information fusion module" constructed in step 2, establishes a rigorous structured physical constraint mapping mechanism. This mechanism explicitly maps abstract feature activation values ​​to specific components of the UAV power physical model.

[0124] Step 3.1: Introduce a physical decomposition model as a priori mapping benchmark.

[0125] First, a four-component power model for UAVs based on aerodynamic principles is introduced as the theoretical benchmark for mapping. This benchmark defines the total power as a linear superposition of four physical components. This step uses the physical driving model as a hard constraint rule for the network topology, forcing the penultimate layer of the network to remain isomorphic to this physical structure.

[0126] Step 3.2: Functional assignment and mapping of specific neurons

[0127] In the physical information fusion module, four specific neuron nodes... To lock the functionality, establish the following one-to-one mapping relationship:

[0128] Hover reference mapping ( ): Specifies the first neuron Specifically used to characterize hovering power. Its activation value is used in subsequent training. The node is constrained to approximate the hovering power value, thereby enabling the node to learn the static energy consumption characteristics of the drone.

[0129] Horizontal dynamic mapping ( ): Specifies the second neuron Specifically designed to characterize the power increment generated by horizontal motion. Its activation value... Mapped to horizontal velocity The function is responsible for capturing the nonlinear relationship between induced power and parasitic power as a function of velocity.

[0130] Vertical dynamic mapping ( ): Specify the third neuron Specifically designed to characterize the power increment generated by vertical motion. Its activation value... Mapped to vertical velocity The function is responsible for capturing the dynamic characteristics of thrust work and gravitational potential energy changes.

[0131] Transient residual mapping ( ): Specify the fourth neuron It is specifically used to characterize transient residuals that cannot be explained by fundamental physical models. Its activation value... It is mapped to a data-driven remainder term used to compensate for acceleration, deceleration, and multi-axis coupling effects.

[0132] Step 3.3: Physical aggregation during forward propagation

[0133] Based on the above mapping mechanism, the aggregation logic of the network output layer is defined. Unlike the weighted summation of traditional fully connected layers, this step forces the computational logic of the output layer to strictly follow the physical summation relationship of energy conservation, that is, the output layer only performs a simple summation operation:

[0134]

[0135] Through this structured mapping mechanism, the outputs of the intermediate layers of a neural network are no longer uninterpretable numerical vectors, but are endowed with clearly defined physical state variables. This allows for precise control of the network's learning trajectory during subsequent training by applying an independent physical loss function to each specific neuron.

[0136] Step 4: Construct a multi-objective composite loss function that includes physical priors;

[0137] To collaboratively optimize the prediction accuracy, physical consistency, and generalization ability of the model, this embodiment constructs a multi-objective composite loss function. Total Loss Function It consists of a linear combination of four weighted terms: physical constraint term Data constraints L2 regularization constraint term With smoothness constraint term . These represent the weights of each loss term. The final expression is shown below:

[0138]

[0139] Step 4.1 Construction of physical constraint terms:

[0140] The physical constraints This is the core of the invention, aiming to encode the physical decomposition model as prior knowledge into the network. This involves applying physical laws to four specific neurons in the physical information fusion module described in step 2 (i.e., the penultimate layer of the network), as expressed below:

[0141]

[0142] In the formula, , , , These represent the horizontal motion loss function value, the vertical motion loss function value, the hovering loss function value, and the additional loss function value, respectively. , , , These represent the weight coefficients for each term. Each sub-loss term aims to minimize the difference between the activation value and gradient of the neuron in that layer and its theoretical behavior according to the corresponding physical formula. The specific expression is as follows:

[0143]

[0144]

[0145]

[0146]

[0147] In the formula, This means calculating the mean over the training batch. , These are the theoretical calculation values ​​of each power component previously derived. These are the activation values ​​of four neurons in the physical constraint layer, which are constrained to learn the dynamics of horizontal, vertical, hovering, and nonlinear power compensation terms, respectively. Through this sophisticated loss function design, rich prior physical knowledge is effectively injected into the network's internal representation learning.

[0148] Step 4.2: Constructing Data Constraints:

[0149] The data constraints Data-driven, this approach aims to quantify the difference between the model's predicted total power and the actual measured power to ensure the model's empirical accuracy. This embodiment defines this term using mean squared error (MSE):

[0150]

[0151] In the formula, This represents the total number of samples in the training batch. The representative model is for the first The total power prediction value calculated from each input sample; and That is the actual power measurement value corresponding to the sample.

[0152] Step 5: Training and real-time estimation of the network model;

[0153] During the model training phase, the UAV flight state time-series dataset constructed in step 1 is input into the neural network. The backpropagation algorithm is used to iteratively update the network weights to minimize the multi-objective composite loss function constructed in step 4. In this process, data constraints drive the network to fit the true power measurements, while physical constraints, through a gradient penalty mechanism, force four specific neurons in the physical information fusion module to learn the physical evolution laws of hovering, horizontal, vertical, and additional terms, respectively. This training mechanism ensures that when the network converges, it can not only accurately capture transient temporal features using BiLSTM but also maintain a high degree of consistency between its internal representation and aerodynamic principles, thereby achieving a dual optimization of physical interpretability and data fitting accuracy.

[0154] After training, the system enters the real-time estimation phase, inputting the real-time flight status data of the UAV under test (or preset data in the mission planning) into the trained network model. The data stream passes through a deep feature learner to extract dynamic features, and is then parsed into specific physical components in the physical information fusion module. Finally, a high-precision real-time total power estimate is obtained at the output layer through simple linear summation. This estimation result, due to the incorporation of physical priors, possesses extremely high robustness and can be directly and reliably applied to practical tasks such as online energy management, remaining flight time prediction, and energy-optimal path planning for UAVs.

[0155] Example:

[0156] This embodiment uses a quadcopter drone performing an actual flight mission as an example to illustrate the specific application process of the power estimation method proposed in this invention.

[0157] 1. System Configuration and Initialization

[0158] Before using this method, it is necessary to first determine the hyperparameter configuration of the network model and calibrate the physical parameters of the UAV.

[0159] (1) Model hyperparameter settings

[0160] Based on the Bayesian optimization results, the hyperparameters of the neural network are configured as follows to ensure the model is in the optimal learning state:

[0161] Based on the Bayesian optimization results, the specific topology and hyperparameter configuration of the physical information neural network constructed in this embodiment are as follows to ensure that the model is in the optimal learning state:

[0162] • Input layer: Receives time-series data on the UAV's flight status (including horizontal flight speed) and vertical flight speed ).

[0163] • Feedforward mapping layer (fully connected layer): Contains two cascaded fully connected layers. The first layer has 84 neurons and the second layer has 48 neurons. It is used to perform preliminary high-dimensional nonlinear mapping on the original input features.

[0164] • Temporal Feature Extraction Layer (BiLSTM Layer): The number of hidden units is set to 148 to capture complex temporal dependencies during transient flight processes bidirectionally. To prevent overfitting, a Dropout mechanism is introduced into the network, with a dropout probability set to 0.28.

[0165] • Physical constraint layer: Contains 4 specific neuron nodes that receive the output of the BiLSTM layer and are explicitly mapped to hover power, horizontal power increment, vertical power increment, and nonlinear power compensation remainder, respectively.

[0166] • Output layer: Contains 1 neuron, which performs a linear summation on the 4 activation values ​​of the physical constraint layer and outputs the total instantaneous power prediction value.

[0167] • Global optimization parameters: The model is trained using a multi-objective composite loss function, with the smoothness constraint weight set to 0.0003 and the L2 regularization constraint weight set to 0.0008; the model optimizer learning rate is set to 0.0012.

[0168] (2) Physical model parameter calibration

[0169] For a specific model of quadcopter UAV to be tested, the empirical constants in its physical decomposition model are calibrated in one go through basic hovering and uniform flight experiments.

[0170] 2. Specific estimation execution process

[0171] Step 1: Real-time data acquisition and preprocessing

[0172] During the mission, the onboard flight control system of the drone collects flight status data in real time at a frequency of 10Hz.

[0173] Scenario assumption: During flight time At that time, the drone was in the acceleration and ascent phase.

[0174] Data collected: Airborne sensors read the current horizontal flight speed. Vertical ascent speed .

[0175] Sequence construction: Using a sliding window (window size = 10), the current time step is... The state data from the past 9 time points are combined into a time series vector, which serves as the input to the model.

[0176] Step Two: Model Inference and Physics Integration

[0177] Input the constructed sequence data into the trained network model:

[0178] Temporal feature extraction: The data first passes through a BiLSTM layer, and the network automatically extracts the dynamic temporal features of the drone during the current acceleration process.

[0179] Physical constraint analysis: Data flows through the physical information fusion module (second to last layer), where four specific neurons output estimated values ​​for each component based on physical laws.

[0180] Hovering component neuron output: (Maintain stability).

[0181] Horizontal component neuron output: (Reflects overcoming horizontal resistance).

[0182] Vertical component neuron output: (Reflects overcoming gravity and vertical resistance).

[0183] Additional neuron output: (Nonlinear errors such as aerodynamic disturbances that are automatically compensated by the model).

[0184] Step 3: Obtain the final estimation result

[0185] The network's output layer sums these four components to obtain the final real-time power estimate:

[0186]

[0187] Step 4: Application of Results

[0188] Mission planning: Based on the acquired real-time power estimate (484.8 W) and the current remaining battery energy, the onboard computer predicts and verifies the global energy feasibility for all subsequent mission waypoints.

[0189] Decision support: If the predicted total energy consumption to complete the global task exceeds the system safety boundary, or the current instantaneous power approaches the hardware output limit, the system will autonomously issue trajectory adjustment instructions (such as reducing the target climb rate) to ensure the safe return of the UAV.

[0190] Through the above steps, the present invention can transform abstract flight data into accurate and interpretable power indicators, directly serving the actual operation management of UAVs.

[0191] Those skilled in the art will understand that, for the sake of convenience and brevity, the specific working process of the modules described above can be referred to the corresponding process in the aforementioned method implementation, and will not be repeated here.

[0192] The modules described as separate components may or may not be physically separate. Similarly, the components shown as modules may or may not be physical modules; they may be located in one place or distributed across multiple network modules. Some or all of the modules can be selected to achieve the purpose of this embodiment, depending on actual needs.

[0193] Furthermore, the functional modules in the various embodiments of this application can be integrated into one processing module, or each module can exist physically separately, or two or more modules can be integrated into one module. The integrated module can be implemented in hardware or in a combination of hardware and software functional modules.

[0194] The integrated modules implemented as software functional modules described above can be stored in a computer-readable storage medium. These software functional modules, stored in a storage medium, include several instructions to cause a computer system (which may be a personal computer, server, or network system, etc.) or processor to execute some steps of the methods described in the various embodiments of this application. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0195] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of this application, and are not intended to limit them. Although this application has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of this application.

Claims

1. A method for calculating the power of a UAV based on a physical information neural network, characterized in that, The method includes the following steps: Step 1: Construct a time-series flight status dataset consisting of several UAV flight status data and corresponding power data; wherein, the UAV flight status data includes horizontal flight speed and vertical flight data; Step 2: Construct a deep neural network architecture that integrates temporal learning, wherein the deep neural network architecture includes cascaded fully connected layers and a bidirectional long short-term memory network; Step 3: Based on the time-series flight state dataset constructed in Step 1, train the deep neural network architecture constructed in Step 2 to obtain the UAV power calculation model; Step 4: Using the UAV power calculation model obtained in Step 3, calculate the real-time total power of the UAV under test.

2. The method as described in claim 1, characterized in that, The data constraint term defined by mean squared error is used as the loss function to train the deep neural network architecture constructed in step 2. The data constraint term is defined as follows: ， In the formula, This represents the total number of samples in the time-series flight state dataset. Representing the The total power prediction value corresponding to each sample; Representing the The total power measurement value corresponding to each sample.

3. The method as described in claim 2, characterized in that, The deep neural network architecture formed in step 2 is trained using a multi-objective composite loss function, which is defined as follows: ， ， ， ， In the formula, For multi-objective composite loss functions, For data constraints, For L2 regularization constraints, For smoothness constraint terms, , and They are respectively , and The weighting coefficients, This represents the total number of samples in the time-series flight state dataset. Representing the The total power prediction value corresponding to each sample Representing the The total power measurement value corresponding to each sample This represents any trainable parameter in a deep neural network; This represents the set of parameters consisting of all trainable parameters of the deep neural network, including weight parameters and bias parameters in the fully connected layer, bidirectional long short-term memory network, physical constraint layer, and output layer.

4. The method as described in claim 1, characterized in that, In step 2, the bidirectional long short-term memory network consists of forward LSTM units and backward LSTM units, wherein the forward LSTM units generate a forward hidden state sequence. The backward LSTM unit generates the backward hidden state sequence. The final output of the bidirectional long short-term memory network is , For time steps Final output, ; Each LSTM unit in the bidirectional long short-term memory network maintains its internal state through a gating mechanism, and the internal state is composed of hidden states. and cell state The common representation, whose update mechanism is as follows: ， ， ， In the formula, For time steps The input vector, They are time steps The activation values ​​of the forget gate, input gate, and output gate; For time steps Candidate cell state; and They are time steps The hidden state and the cellular state; All are learnable parameters. = ; This indicates element-wise multiplication.

5. The method as described in claim 1, characterized in that, The method further includes: Based on aerodynamic principles, a basic power model for unmanned aerial vehicles (UAVs) is constructed: ， In the formula, For the UAV at a given velocity vector Total instantaneous power below, The horizontal flight speed of the drone, The vertical flight speed of the drone. This is the reference power for the drone when it is stationary and hovering. and These are the power increments compared to the hovering state, generated by horizontal and vertical motion, respectively. The basic model of UAV power is used as the hard constraint rule of the deep neural network, and a physical constraint layer including three neuron nodes is added after the bidirectional long short-term memory network. The three neuron nodes of the physical constraint layer The following mapping relationships are established with the three components in the basic power model of the UAV: ​​① Hovering reference mapping ② Horizontal dynamic mapping ③ Vertical dynamic mapping ; After adding the physical constraint layer, the output layer is finally formed, resulting in the output of the deep neural network architecture: ， In the formula, For the output of a deep neural network architecture, They are respectively The activation value.

6. The method as described in claim 5, characterized in that, The deep neural network architecture formed in step 2 is trained using a multi-objective composite loss function, which is defined as follows: ， ， ， ， ， ， ， ， In the formula, For multi-objective composite loss functions, For physical constraints, For data constraints, For L2 regularization constraints, For smoothness constraint terms, , , and They are respectively , , and The weighting coefficients, , and These represent the horizontal motion loss term, the vertical motion loss term, and the hovering loss term, respectively. , and They are respectively , and The weighting coefficients, This means taking the mean over the training batches. This represents the total number of samples in the time-series flight state dataset. Representing the The total power prediction value corresponding to each sample Representing the The total power measurement value corresponding to each sample This represents any trainable parameter in a deep neural network; This represents the set of parameters consisting of all trainable parameters of the deep neural network, including weight parameters and bias parameters in the fully connected layer, bidirectional long short-term memory network, physical constraint layer, and output layer.

7. The method as described in claim 5, characterized in that, The method further includes: Based on aerodynamic principles, a nonlinear power compensation term is superimposed on the basic power model of the UAV to compensate for transient dynamics, thus constructing a four-component power model for the UAV: ， In the formula, For the UAV at a given velocity vector Total instantaneous power below, The horizontal flight speed of the drone, The vertical flight speed of the drone. This is the reference power for the drone when it is stationary and hovering. and These are the power increments compared to the hovering state, generated by horizontal and vertical motion, respectively. This is the remainder term for nonlinear power compensation; The four-component model of UAV power is used as a hard constraint rule for the deep neural network architecture, and a physical constraint layer including four neuron nodes is added after the bidirectional long short-term memory network. The four neuron nodes of the physical constraint layer The following mapping relationships are established with the four components of the UAV power four-component model: ① Hovering reference mapping ② Horizontal dynamic mapping ③ Vertical dynamic mapping ④ Transient residual mapping ; After adding the physical constraint layer, the output layer is finally formed, resulting in the output of the deep neural network architecture: ， In the formula, For the output of a deep neural network architecture, They are respectively The activation value.

8. The method as described in claim 7, characterized in that, The deep neural network architecture formed in step 2 is trained using a multi-objective composite loss function, which is defined as follows: ， ， ， ， ， ， ， ， ， In the formula, For multi-objective composite loss functions, For physical constraints, For data constraints, For L2 regularization constraints, For smoothness constraint terms, , , and They are respectively , , and The weighting coefficients, , , and These represent the horizontal motion loss term, vertical motion loss term, hovering loss term, and additional loss term, respectively. , , and They are respectively , , and The weighting coefficients, This means taking the mean over the training batches. This represents the total number of samples in the time-series flight state dataset. Representing the The total power prediction value corresponding to each sample Representing the The total power measurement value corresponding to each sample This represents any trainable parameter in a deep neural network; This represents the set of parameters consisting of all trainable parameters of the deep neural network, including weight parameters and bias parameters in the fully connected layer, bidirectional long short-term memory network, physical constraint layer, and output layer.

9. The method as described in claim 7, characterized in that, The nonlinear power compensation remainder term is defined as follows: ， In the formula, Based on and The basis function row vector is formed by several nonlinear basis functions. , It is a coefficient column vector consisting of sparse coefficients corresponding to the nonlinear basis functions; The coefficient column vector is obtained through a parameter identification method, specifically: The linear subspace spanned by the three component vectors of the UAV power fundamental model is defined as... ,in, To correspond to the hovering reference power a constant vector, To correspond to the incremental power of horizontal motion The horizontal increment vector, To correspond to the incremental power of vertical motion The vertical increment vector; Secondly, the target residual and All projected to orthogonal complementary space In the middle, the projected target residuals are obtained respectively. and the projected basis function row vectors ;in, For orthogonal projection operators, This is the measured total power vector; Finally, for and Performing sparse regression, i.e., solving... This yields the coefficient column vector.

10. A UAV power calculation system based on a physical information neural network using the method described in any one of claims 1 to 9, characterized in that, The system includes: The data acquisition unit is used to collect UAV flight status data and construct a time-series flight status dataset consisting of several UAV flight status data and corresponding power data; wherein, the UAV flight status data includes horizontal flight speed and vertical flight data; A network building unit is used to build a deep neural network architecture that integrates temporal learning, wherein the deep neural network architecture includes cascaded fully connected layers and a bidirectional long short-term memory network. The network training unit is used to train the deep neural network architecture based on the time-series flight state dataset to obtain the UAV power calculation model; The power calculation unit is used to calculate the real-time total power of the UAV under test using the obtained UAV power calculation model.

Images