CN115469679A - Unmanned aerial vehicle flight state parameter prediction method and system - Google Patents

Abstract

The invention discloses a method and a system for predicting flight state parameters of an unmanned aerial vehicle, wherein an improved Transformer neural network model is built; acquiring flight state data of a typical flight state of the unmanned aerial vehicle; preprocessing the flight state data of the unmanned aerial vehicle; generating an unmanned aerial vehicle flight state data set by utilizing the preprocessed data, and then dividing the unmanned aerial vehicle flight state data set into a training set, a verification set and a test set; training the Transformer neural network model by using a training set, obtaining parameters of the Transformer neural network model by using a verification set, inputting data of H time sequences of t, t-1, … and t-H-1 in the test set into the trained Transformer neural network model, and predicting to obtain flight state parameters at the t +1 moment. The method can predict the flight state parameters of the unmanned aerial vehicle in real time and at high precision, lays a solid foundation for the precise control of the unmanned aerial vehicle, ensures the safe and reliable flight of the unmanned aerial vehicle, can execute different tasks of different industries, and expands the use scene of the unmanned aerial vehicle.

Description

A method and system for predicting flight state parameters of an unmanned aerial vehicle

technical field

The invention belongs to the technical field of unmanned aerial vehicles, and in particular relates to a method and system for predicting flight state parameters of an unmanned aerial vehicle.

Background technique

At present, drones have been widely used in many industries, such as aerial photography, surveying and mapping, inspection, air patrol and so on. Different industries require UAVs to complete different tasks, but the most fundamental requirement for UAVs is that UAVs can be precisely controlled to complete these tasks. To precisely control the drone, the controller of the drone needs to obtain real-time and accurate state parameters of the drone, such as the speed, acceleration, attitude angle, angular velocity, and angular acceleration of the drone.

Traditional methods of estimating and predicting flight state parameters are based on UAV dynamics models. The dynamic model is to calculate the acceleration information and angular acceleration information of the UAV based on Newton's second law and Euler's equation under certain assumptions and simplified conditions, and then obtain other flight state parameter information through integration, such as velocity, angular velocity and attitude angle etc. The dynamic modeling method makes some assumptions and simplifications for the actual flight of the UAV, and generally ignores the unsteady parameters and some uncertain states in the flight of the UAV, such as aerodynamic characteristics, so through the UAV dynamics The acceleration and angular acceleration information obtained by the model is not accurate enough, and there are certain errors. Other flight state parameters are calculated by integrating the acceleration and angular acceleration, and the integral operation will accumulate this error, so it is not accurate enough to obtain the flight state parameters of the UAV through the dynamic model.

In order to improve the accuracy of flight state parameter estimation, a joint model of dynamics model and convolutional neural network is currently used, in which the dynamics model predicts the deterministic part of the flight state, and the convolutional neural network predicts the uncertain part of the flight state; this Compared with the traditional simple dynamic modeling method, this method has greatly improved the accuracy of prediction results. However, since the flow field near the UAV and the state of the UAV change with time during the flight of the UAV, the flight state at the historical moment will affect the current flight state, while the traditional dynamic model and this A dynamics-convolutional neural network joint model does not fully consider this time-varying characteristic, and fails to make full use of the historical information of the UAV to predict the current flight state parameters.

In order to make full use of the historical information of the UAV to predict the current flight state parameters, the LSTM method is used alone or in combination with traditional dynamic modeling to estimate the flight state parameters of the UAV. However, due to the information bottleneck of the LSTM method, the input When the flight status data of a long time segment is too long, the previous information may be lost, so that this method can only use the flight status data of a shorter time segment, which will affect the prediction accuracy. At the same time, due to the use of a deep LSTM network, there will be gradient disappearance or explosion, the convergence of the network is poor, and the calculation is complex, the parallel ability is poor, the calculation takes a long time, and it is difficult to achieve real-time prediction.

Accurate prediction of UAV flight state parameters is the basis for precise control of UAVs, thereby ensuring safe and reliable flight of UAVs, and completing tasks in more complex scenarios, so real-time high-precision prediction of UAV flight State parameters are crucial for drones.

Contents of the invention

The technical problem to be solved by the present invention is to provide a method and system for predicting flight state parameters of UAVs in view of the deficiencies in the above-mentioned prior art, which are used to solve the problem of inaccurate and poor real-time prediction of UAV flight state parameters, which in turn affects the The control accuracy and control frequency of the UAV are technical issues.

The present invention adopts following technical scheme:

A method for predicting flight state parameters of an unmanned aerial vehicle, comprising the following steps:

S1. Build an improved Transformer neural network model;

S2. Obtain the flight state data of the typical flight state of the drone;

S3. Preprocessing the flight state data of the unmanned aerial vehicle acquired in step S2;

S4, utilize the preprocessed data of step S3 to generate the UAV flight state data set, and then divide the UAV flight state data set into a training set, a verification set and a test set;

S5, use the training set obtained in step S4 to train the Transformer neural network model obtained in step S1, use the verification set obtained in step S4 to obtain the parameters of the Transformer neural network model, and use the test set t obtained in step S4, t-1, …, t-H-1 A total of H time series data are input into the trained Transformer neural network model, and the flight status parameters at time t+1 are predicted.

Specifically, in step S1, the Transformer neural network model includes an input, an encoder, a decoder, and a fully-connected network layer. The time series data is passed through a linear mapping layer, and time stamp encoding is added as the input of the encoder and decoder, and the encoding Both the decoder and the decoder are stacked N times, the output of the encoder is used as the input of the cross-correlation attention mechanism layer in the decoder, and the output of the decoder passes through the fully connected network layer to output the prediction data.

Further, the input of the encoder and decoder is specifically:

Map the input H×m dimensional time series data to H×d dimension, H is the length of time series, m and d are the dimensions of flight state parameters, d≥m.

Further, the timestamp encoding is specifically:

P(k,2i)=sin(k*e ^-2i*ln100/d )

P(k,2i+1)＝cos(k*e ^{-(2i+1)*ln100/d} )

Among them, P(k, 2i) is the time code of the even-numbered dimension of the flight state data at time k, P(k, 2i+1) is the time code of the odd-numbered dimension of the flight state data at time k, and i represents the i-th dimension data , after the linear mapping layer, the data has a total of d dimensions.

Specifically, in step S2, typical flight states of the UAV include climb, cruise, hover and dive, and the flight state parameters include the time stamp, control signal, speed, acceleration and angular velocity data of the UAV.

Specifically, step S3 is specifically:

S301, using filtering processing to suppress the noise in the flight state data of the UAV acquired in step S2;

S302. Resample the data obtained in step S301, align time stamps of different sensor data, and generate N data in total of time series numbers [T ₁ , T ₂ , . . . , T _N ].

Further, in step S302, the flight state data T _{k at time k} is:

表示k时刻不同方向的加速度分量，[ω_xk,ω_yk,ω_zk]表示k时刻不同方向的角速度分量，

表示k时刻不同方向的角加速度分量，

表示k时刻姿态角。Among them, [τ _1k ,τ _2k ,τ _3k ,τ _4k ] represent the control input signals of roll, pitch, throttle and yaw channels at time k respectively, and [v _xk ,v _yk ,v _zk ] represent the control input signals of different directions at time k speed component,

Represents the acceleration components in different directions at time k, [ω _xk , ω _yk , ω _zk ] represents the angular velocity components in different directions at time k,

Indicates the angular acceleration components in different directions at time k,

Indicates the attitude angle at time k.

Specifically, in step S4, the N time series data obtained in step S3 are overlapped with time interval L and length H to obtain a data set D with length M, and then divided. The training set accounts for 60% of the total data. The validation set is 20% of the total data, and the test set is 20% of the total data.

Specifically, in step S5, the Transformer neural network model is trained using the Adma optimization algorithm to minimize the loss function LOSS, and the loss function LOSS is specifically:

为k时刻的预测值，n为训练集样本个数。Among them, y _k is the real value at time k,

is the predicted value at time k, and n is the number of samples in the training set.

In a second aspect, an embodiment of the present invention provides a system for predicting flight state parameters of an unmanned aerial vehicle, including:

Build modules and build an improved Transformer neural network model;

The parameter module is used to obtain the flight state parameters of the typical flight state of the UAV;

The preprocessing module preprocesses the flight status data of the UAV acquired by the parameter module;

The division module uses the data preprocessed by the preprocessing module to generate the UAV flight state data set, and then divides the UAV flight state data set into a training set, a verification set and a test set;

For the prediction module, use the training set obtained by the division module to train the Transformer neural network model obtained by the construction module, use the verification set obtained by the division module to obtain the parameters of the Transformer neural network model, and divide the test set obtained by the division module into t, t-1 ,…, t-H-1, a total of H time series data input the trained Transformer neural network model, and predict the flight status parameters at time t+1.

Compared with the prior art, the present invention has at least the following beneficial effects:

The present invention is a method for predicting the flight state parameters of an unmanned aerial vehicle. By improving the Transformer neural network model and using the flight data of the typical flight state of the unmanned aerial vehicle to train the model, the flight state parameters of the unmanned aerial vehicle can be predicted in real time and with high precision. The precise control of drones lays a solid foundation to ensure safe and reliable flight of drones. At the same time, it can perform different tasks in different industries and expand the use scenarios of drones.

Furthermore, the improved Transformer neural network model, which proposes time stamp encoding and unified encoder-decoder input, has better parallel performance and higher prediction accuracy than the most original model, so it can realize UAV state parameters Real-time high-precision forecasting.

Furthermore, the model maps the input H×m dimensional time series data to H×d dimension, which maps the input data to a higher dimension, which is more conducive to learning to represent the complex relationship between UAV state parameters.

Furthermore, since the attention mechanism in the Transformer neural network model does not consider the relationship between input variables, and there is a strong time sequence between flight data, in order to better represent this time sequence relationship, the present invention proposes a Timestamp encoding method.

Further, the typical flight states of UAVs include climbing, cruising, circling and diving. The present invention collects these typical flight state data for improving the training of the Transformer neural network model, so that the model can learn the Typical flight status, which greatly expands the scope of application of the present invention.

Furthermore, filtering the collected flight data can eliminate the noise signal in the data; since the sampling frequency of the data collected by different sensors on the UAV is different, resampling the collected data can align the different sensor data. timestamp.

Further, the preprocessed flight data includes flight state parameters in all dimensions. Using the flight status parameters of all dimensions as the input of the network to train the model can enable the model to learn a global model, which in turn can improve the prediction accuracy of the model.

Further, overlapping sampling can expand the training data and avoid over-fitting of the model.

Furthermore, the Adma optimization algorithm can make the model convergence more stable. It can be understood that, for the beneficial effects of the second aspect above, reference may be made to the relevant description in the first aspect above, and details are not repeated here.

In summary, the present invention can predict the flight state parameters of UAVs in real time and with high precision, lay a solid foundation for precise control of UAVs, ensure safe and reliable flight of UAVs, and at the same time perform different tasks in different industries and expand the range of drones. Human-machine usage scenarios.

The technical solutions of the present invention will be described in further detail below with reference to the accompanying drawings and embodiments.

Description of drawings

Fig. 1 is a schematic flow chart of the present invention;

Fig. 2 is a schematic diagram of the improved Transformer neural network model of the present invention.

detailed description

The following will clearly and completely describe the technical solutions in the embodiments of the present invention with reference to the accompanying drawings in the embodiments of the present invention. Obviously, the described embodiments are some of the embodiments of the present invention, but not all of them. Based on the embodiments of the present invention, all other embodiments obtained by persons of ordinary skill in the art without making creative efforts belong to the protection scope of the present invention.

In the description of the present invention, it should be understood that the terms "comprising" and "comprising" indicate the presence of described features, integers, steps, operations, elements and/or components, but do not exclude one or more other features, Presence or addition of wholes, steps, operations, elements, components and/or collections thereof.

It should also be understood that the terminology used in the description of the present invention is for the purpose of describing particular embodiments only and is not intended to be limiting of the present invention. As used in this specification and the appended claims, the singular forms "a", "an" and "the" are intended to include plural referents unless the context clearly dictates otherwise.

It should also be further understood that the term "and/or" used in the description of the present invention and the appended claims refers to any combination and all possible combinations of one or more of the associated listed items, and includes these combinations , for example, A and/or B, may mean: A exists alone, A and B exist simultaneously, and B exists alone. In addition, the character "/" in this article generally indicates that the contextual objects are an "or" relationship.

It should be understood that although the terms first, second, third, etc. may be used in the embodiments of the present invention to describe preset ranges, etc., these preset ranges should not be limited to these terms. These terms are only used to distinguish preset ranges from one another. For example, without departing from the scope of the embodiments of the present invention, the first preset range may also be called the second preset range, and similarly, the second preset range may also be called the first preset range.

Depending on the context, the word "if" as used herein may be interpreted as "at" or "when" or "in response to determining" or "in response to detecting". Similarly, depending on the context, the phrases "if determined" or "if detected (the stated condition or event)" could be interpreted as "when determined" or "in response to the determination" or "when detected (the stated condition or event) )" or "in response to detection of (a stated condition or event)".

Various structural schematic diagrams according to the disclosed embodiments of the present invention are shown in the accompanying drawings. The figures are not drawn to scale, with certain details exaggerated and possibly omitted for clarity of presentation. The shapes of various regions and layers shown in the figure and their relative sizes and positional relationships are only exemplary, and may deviate due to manufacturing tolerances or technical limitations in practice, and those skilled in the art may Regions/layers with different shapes, sizes, and relative positions can be additionally designed as needed.

The invention provides a method for predicting the flight state parameters of the UAV, which makes full use of the historical flight state information of the UAV, avoids the use of dynamic model methods based on assumptions and simplification, and thus achieves high-precision prediction; adopts the improved The Transformer neural network model has very good parallel computing capabilities, so that this method can meet the requirements of real-time prediction; it has laid a solid foundation for the safe and reliable flight of UAVs and the execution of complex tasks in complex environments, thus greatly expanding the unmanned machine usage scenarios.

Please refer to Fig. 1, a kind of unmanned aerial vehicle flight state parameter prediction method of the present invention comprises the following steps:

S1. Build an improved Transformer neural network model for predicting flight state parameters;

Please refer to Figure 2. The Transformer neural network model adopts the time encoding method and the linear mapping layer, and at the same time unifies the input of the encoder and decoder; the input time series data passes through the linear mapping layer to encode and increase the dimension of the data to describe the time series Complex relationships between data are then encoded with timestamps as input to encoders and decoders. Timestamp coding can effectively describe the sequential relationship between time series data.

The Transformer neural network model includes input, encoder, decoder and fully connected network layers,

The encoder consists of a multi-head self-attention layer, a residual normalization layer, a feedforward network layer and a normalization layer. The encoder can be repeated N times to better describe the complex relationship between time series data.

The input of the decoder is the same as that of the encoder. The input data passes through the multi-head self-attention and residual normalization layer. The multi-head mutual attention layer fuses the output data of the encoder and the input data of the decoder, and then passes through the residual normalization layer. , the feedforward network layer and the residual normalization layer get the output of the decoder. The feedforward network layer consists of two linear network layers with an activation function in between. The decoder also repeats N times to better describe the complex relationship between time series data. The output of the decoder goes through fully connected layers, and finally the predicted flight state parameters are obtained.

Timestamp coding keeps the chronological order of time series data; the linear mapping layer maps the input time series to a higher dimension, and increases the dimension of the data, so as to better represent the complex relationship between time series data;

The decoder uses the same input as the encoder, avoids the use of autoregressive for sequential prediction, reduces cumulative errors, and avoids sequential prediction can also improve the parallel efficiency of the algorithm and speed up the prediction time.

After the time series data is increased by the linear mapping layer, it is added to the time-coded data as the input of the encoder and the decoder. The encoder and the decoder of the present invention all use the encoder and the decoder of the original Transformer neural network model, and the decoder The output goes through the fully connected layer to get the final output result.

Time encoding is used as follows:

P(k,2i)=sin(k*e ^-2i*ln100/d )

P(k,2i+1)＝cos(k*e ^{-(2i+1)*ln100/d} )

Among them, k represents the timestamp data, i represents the i-th dimension data, d represents the dimension of the data after the linear mapping layer, and d=512.

S2. Obtain the flight state parameters of the typical flight state of the UAV;

The typical flight states of UAVs include climbing, cruising, hovering, and diving; the time stamp, control signal, speed, acceleration, and angular velocity data of the UAV are collected by different sensors, and then sent to the LOG of the flight control system. The module records, reads the historical flight status data recorded by the LOG module, and then integrates the angular velocity to obtain the attitude angle data, and differentiates the angular velocity to obtain the angular acceleration data.

For each different flight state, 25mins of flight data are collected, a total of 100mins of data.

S3. Preprocessing the flight state data of the unmanned aerial vehicle acquired in step S2;

S301. Using filtering processing to suppress noise in the flight status data of the drone;

Filtering includes hardware filtering and software filtering, and the filtering algorithm is not limited to Kalman filtering.

S302. Since the sampling frequencies of different sensors are different, the time intervals for collecting data are also different. It is necessary to resample the data obtained in step S301 to align the time stamps of different sensor data, and then generate a time series number [T ₁ , T ₂ ,…,T _N ] a total of N data.

The flight state data T _{k at time k} is:

表示k时刻不同方向的加速度分量，[ω_xk,ω_yk,ω_zk]表示k时刻不同方向的角速度分量，

表示k时刻不同方向的角加速度分量，

表示k时刻姿态角。Among them, [τ _1k ,τ _2k ,τ _3k ,τ _4k ] represent the control input signals of roll, pitch, throttle and yaw channels at time k respectively, and [v _xk ,v _yk ,v _zk ] represent the control input signals of different directions at time k speed component,

Represents the acceleration components in different directions at time k, [ω _xk , ω _yk , ω _zk ] represents the angular velocity components in different directions at time k,

Indicates the angular acceleration components in different directions at time k,

Indicates the attitude angle at time k.

S4, utilize the preprocessed data of step S3 to generate the UAV flight state data set, and then divide the UAV flight state data set into a training set, a verification set and a test set;

The data set generated in S4 is the N time series data obtained by preprocessing in step S3, the time interval is L, the length is H overlap sampling, and the data set D with length M is obtained, and then according to, the training set accounts for the total data 60% of the total data, the verification set accounts for 20% of the total data, and the test set accounts for 20% of the total data. Random sampling is used to obtain the training set, verification set and test set.

Dataset D is:

D＝{[T ₁ ,T ₂ ,…,T _H ],[T _1+L ,T _2+L ,…,T _H+L ],…,[T _N-H+1 ,T _{N-H+ 2} ,...,T _N ]}.

S5, use the training set obtained in step S4 to train the Transformer neural network model obtained in step S1, use the verification set obtained in step S4 to obtain the parameters of the Transformer neural network model, use the test set t obtained in step S4, t-1, ..., t-H-1 total H time series data, use the trained model to predict the flight status parameters at time t+1.

S501. Input and output of Transformer neural network model

The input of the Transformer neural network model is the data 1s before time t, and the time interval of the time series data after preprocessing is 0.01s. A total of 100 time series data are input, each time series data T _k is 19 dimensions, and the input data is A 100-by-19-dimensional vector.

The output T _t+1 of the Transformer neural network model is the flight state parameters at time t+1, with a total of 15 dimensions, and the output data format is a 1×15-dimensional vector.

The output T _t+1 of the Transformer neural network model is specifically:

表示t+1时刻不同方向的加速度分量，[ω_x,t+1,ω_y,t+1,ω_z,t+1]表示t+1时刻不同方向的角速度分量，

表示t+1时刻不同方向的角加速度分量，

表示t+1时刻的姿态角。Among them, [v _x,t+1 ,v _y,t+1 ,v _z,t+1 ] represent the velocity components in different directions at time t+1,

Indicates the acceleration components in different directions at time t+1, [ω _x,t+1 ,ω _y,t+1 ,ω _z,t+1 ] indicates the angular velocity components in different directions at time t+1,

Indicates the angular acceleration components in different directions at time t+1,

Indicates the attitude angle at time t+1.

S502, parameter initialization

Initialize the parameters of the entire network, and select the optimal value of the hyperparameters of the network according to a large number of experiments on the verification set; among them, the number of layers of the encoder and decoder is 8 layers, the dimension of the linear mapping layer is 128, and the number of heads of the multi-head attention layer It is 8, and the dimension of the fully connected layer is 1024.

S503, model training and testing

The activation function in the network selects the Relu function, and when predicting the flight state parameters of the drone, the mean square error (Mean square error, MSE) is selected as the loss function.

为k时刻的预测值，n为训练集样本个数。Among them, y _k is the real value at time k,

is the predicted value at time k, and n is the number of samples in the training set.

Using the Adma optimizer, the learning rate uses an adaptive adjustment strategy to minimize the loss function, and the training set is used to complete the training of the model. The selection of hyperparameters is done using the validation set.

S504. Prediction accuracy evaluation

Root mean square error (Root mean square, RSE) is used as the prediction accuracy evaluation index, as follows:

On the test set, use the data at time t and before, and use the trained Transformer neural network model to predict the flight state parameters at time t+1.

The predicted flight state parameters include velocities in three directions, accelerations in three directions, angular velocities in three directions, angular accelerations in three directions and three attitude angles, totaling 15 dimensions.

In yet another embodiment of the present invention, a UAV flight state parameter prediction system is provided, which can be used to implement the above-mentioned UAV flight state parameter prediction method. Specifically, the UAV flight state parameter prediction system includes building module, parameter module, preprocessing module, partition module, and prediction module.

Among them, build modules and build an improved Transformer neural network model;

The parameter module is used to obtain the flight state data of the typical flight state of the UAV;

The preprocessing module preprocesses the flight status data of the UAV acquired by the parameter module;

The division module uses the data preprocessed by the preprocessing module to generate the UAV flight state data set, and then divides the UAV flight state data set into a training set, a verification set and a test set;

For the prediction module, use the training set obtained by the division module to train the Transformer neural network model obtained by the construction module, use the verification set obtained by the division module to obtain the parameters of the Transformer neural network model, and divide the test set obtained by the division module into t, t-1 ,…, t-H-1, a total of H time series data input the trained Transformer neural network model, and predict the flight status parameters at time t+1.

In order to make the purpose, technical solutions and advantages of the embodiments of the present invention clearer, the technical solutions in the embodiments of the present invention will be clearly and completely described below in conjunction with the drawings in the embodiments of the present invention. Obviously, the described embodiments It is a part of embodiments of the present invention, but not all embodiments. The components of the embodiments of the invention generally described and illustrated in the drawings herein may be arranged and designed in a variety of different configurations. Accordingly, the following detailed description of the embodiments of the invention provided in the accompanying drawings is not intended to limit the scope of the claimed invention, but merely represents selected embodiments of the invention. Based on the embodiments of the present invention, all other embodiments obtained by persons of ordinary skill in the art without creative efforts fall within the protection scope of the present invention.

Since the output of the dynamic model is about the acceleration in three directions and the angular acceleration in three directions, the comparison of the root mean square errors of the predictions of different prediction models is shown in Table 1:

Table 1 Comparison of root mean square errors of predictions of different prediction models

By table 1, the root mean square error of dynamic model, dynamic model-convolutional neural network hybrid model, LSTM model and three direction angular velocities of the present invention and angular acceleration is compared, as seen from table 1, the error of the present invention minimum.

Simultaneously, the present invention compares the time that different models predict one time, as shown in Table 2:

Table 2 Comparison of forecasting time of different forecasting models

Table 2 compares dynamics model, dynamics model-convolutional neural network hybrid model, LSTM model and the prediction time of the present invention, as can be seen from the table, the prediction time of the present invention is 9ms, reaches the real-time requirement of unmanned aerial vehicle (The control frequency of the general controller is 100Hz).

In summary, the present invention provides a method and system for predicting flight state parameters of UAVs, which can realize high-precision real-time prediction of flight state data.

Those skilled in the art should understand that the embodiments of the present application may be provided as methods, systems, or computer program products. Accordingly, the present application may take the form of an entirely hardware embodiment, an entirely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, the present application may take the form of a computer program product embodied on one or more computer-usable storage media (including but not limited to disk storage, CD-ROM, optical storage, etc.) having computer-usable program code embodied therein.

The present application is described with reference to flowcharts and/or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of the present application. It should be understood that each procedure and/or block in the flowchart and/or block diagram, and a combination of procedures and/or blocks in the flowchart and/or block diagram can be realized by computer program instructions. These computer program instructions may be provided to a general purpose computer, special purpose computer, embedded processor, or processor of other programmable data processing equipment to produce a machine such that the instructions executed by the processor of the computer or other programmable data processing equipment produce a An apparatus for realizing the functions specified in one or more procedures of the flowchart and/or one or more blocks of the block diagram.

These computer program instructions may also be stored in a computer-readable memory capable of directing a computer or other programmable data processing apparatus to operate in a specific manner, such that the instructions stored in the computer-readable memory produce an article of manufacture comprising instruction means, the instructions The device realizes the function specified in one or more procedures of the flowchart and/or one or more blocks of the block diagram.

These computer program instructions can also be loaded onto a computer or other programmable data processing device, causing a series of operational steps to be performed on the computer or other programmable device to produce a computer-implemented process, thereby The instructions provide steps for implementing the functions specified in the flow diagram procedure or procedures and/or block diagram procedures or blocks.

The above content is only to illustrate the technical ideas of the present invention, and cannot limit the protection scope of the present invention. Any changes made on the basis of the technical solutions according to the technical ideas proposed in the present invention shall fall within the scope of the claims of the present invention. within the scope of protection.

Claims (10)

1. An unmanned aerial vehicle flight state parameter prediction method is characterized by comprising the following steps:

s1, building an improved Transformer neural network model;

s2, acquiring flight state data of a typical flight state of the unmanned aerial vehicle;

s3, preprocessing the flight state data of the unmanned aerial vehicle acquired in the step S2;

s4, generating an unmanned aerial vehicle flight state data set by using the data preprocessed in the step S3, and then dividing the unmanned aerial vehicle flight state data set into a training set, a verification set and a test set;

and S5, training the Transformer neural network model obtained in the step S1 by using the training set obtained in the step S4, obtaining parameters of the Transformer neural network model by using the verification set obtained in the step S4, inputting data of H time sequences of the t, t-1, … and t-H-1 in the test set obtained in the step S4 into the trained Transformer neural network model, and predicting to obtain flight state parameters at the time of t + 1.

2. The method of claim 1, wherein in step S1, the transform neural network model includes an input, an encoder, a decoder, and a fully-connected network layer, the time-series data is passed through the linear mapping layer, and then the time-stamp coding is used as the input of the encoder and the decoder, the encoder and the decoder are stacked repeatedly N times, the output of the encoder is used as the input of a cross-correlation attention mechanism layer in the decoder, and the output of the decoder is passed through the fully-connected network layer to output the prediction data.

3. The method for predicting the flight parameters of an unmanned aerial vehicle according to claim 2, wherein the inputs of the encoder and the decoder are specifically:

and mapping the input H multiplied by m dimensional time sequence data to H multiplied by d dimension, wherein H is the length of the time sequence, m and d are the dimensions of flight state parameters, and d is more than or equal to m.

4. The unmanned aerial vehicle flight state parameter prediction method according to claim 2, wherein the timestamp coding specifically comprises:

P(k,2i)＝sin(k*e ^-2i*ln100/d )

P(k,2i+1)＝cos(k*e ^{-(2i+1)*ln100/d} )

p (k, 2 i) is time coding of an even bit dimension of the flight state data at the moment k, P (k, 2i + 1) is time coding of an odd bit dimension of the flight state data at the moment k, i represents ith dimension data, and the data after passing through the linear mapping layer have d dimensions.

5. The method of claim 1, wherein in step S2, the typical flight status of the drone includes climb, cruise, hover and dive, and the flight status parameters include timestamp, control signal, speed, acceleration and angular velocity data of the drone.

6. The unmanned aerial vehicle flight state parameter prediction method according to claim 1, wherein step S3 specifically is:

s301, suppressing noise in the unmanned aerial vehicle flight state data acquired in the step S2 by adopting filtering processing;

s302, resampling the data obtained in the step S301, aligning the time stamps of different sensor data, and generating a time series number [ T ₁ ,T ₂ ,…,T _N ]For a total of N data.

7. The unmanned aerial vehicle flight state parameter prediction method according to claim 6, wherein in step S302, flight state data T at time k is obtained _k Comprises the following steps:

wherein [ tau ] _1k ,τ _2k ,τ _3k ,τ _4k ]Control input signals representing the roll, pitch, throttle and yaw channels at time k, respectively, [ v ] _xk ,v _yk ,v _zk ]Indicating the time of kThe velocity components in the different directions of the wave,

represents the acceleration components in different directions at time k, [ omega ] _xk ,ω _yk ,ω _zk ]Representing angular velocity components in different directions at time k,

representing angular acceleration components in different directions at time k,

representing the attitude angle at time k.

8. The unmanned aerial vehicle flight state parameter prediction method according to claim 1, wherein in step S4, the N time series data obtained in step S3 are subjected to overlapping sampling of time interval L and length H to obtain a data set D with length M, and then the data set D is divided into a training set accounting for 60% of the total data, a verification set accounting for 20% of the total data, and a test set accounting for 20% of the total data.

9. The unmanned aerial vehicle flight state parameter prediction method of claim 1, wherein in step S5, an Adma optimization algorithm minimization of LOSS function LOSS is used to train the Transformer neural network model, and the LOSS function LOSS is specifically:

wherein, y _k Is the true value for the time instant k,

is the predicted value at the moment k, and n is the number of training set samples.

10. An unmanned aerial vehicle flight status parameter prediction system, characterized in that includes:

the building module is used for building an improved Transformer neural network model;

the parameter module is used for acquiring flight state data of a typical flight state of the unmanned aerial vehicle;

the preprocessing module is used for preprocessing the flight state data of the unmanned aerial vehicle acquired by the parameter module;

the division module is used for generating an unmanned aerial vehicle flight state data set by using data preprocessed by the preprocessing module, and then dividing the unmanned aerial vehicle flight state data set into a training set, a verification set and a test set;

and the prediction module is used for training the Transformer neural network model obtained by the building module by using the training set obtained by the dividing module, obtaining parameters of the Transformer neural network model by using the verification set obtained by the dividing module, inputting the data of the t, t-1, … and t-H-1 time sequences in the test set obtained by the dividing module into the trained Transformer neural network model, and predicting to obtain flight state parameters at the t +1 moment.

Images