CN114757086A - Multi-rotor unmanned aerial vehicle real-time remaining service life prediction method and system - Google Patents

Multi-rotor unmanned aerial vehicle real-time remaining service life prediction method and system Download PDF

Info

Publication number
CN114757086A
CN114757086A CN202111561949.5A CN202111561949A CN114757086A CN 114757086 A CN114757086 A CN 114757086A CN 202111561949 A CN202111561949 A CN 202111561949A CN 114757086 A CN114757086 A CN 114757086A
Authority
CN
China
Prior art keywords
level fusion
transformer model
aerial vehicle
unmanned aerial
preset
Prior art date
Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
Pending
Application number
CN202111561949.5A
Other languages
Chinese (zh)
Inventor
宋佳
艾绍洁
苏江城
赵凯
蔡国飙
Current Assignee (The listed assignees may be inaccurate. Google has not performed a legal analysis and makes no representation or warranty as to the accuracy of the list.)
Beihang University
Original Assignee
Beihang University
Priority date (The priority date is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the date listed.)
Filing date
Publication date
Application filed by Beihang University filed Critical Beihang University
Priority to CN202111561949.5A priority Critical patent/CN114757086A/en
Publication of CN114757086A publication Critical patent/CN114757086A/en
Pending legal-status Critical Current

Links

Images

Classifications

    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06F—ELECTRIC DIGITAL DATA PROCESSING
    • G06F30/00—Computer-aided design [CAD]
    • G06F30/20—Design optimisation, verification or simulation
    • G06F30/27—Design optimisation, verification or simulation using machine learning, e.g. artificial intelligence, neural networks, support vector machines [SVM] or training a model
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06F—ELECTRIC DIGITAL DATA PROCESSING
    • G06F18/00—Pattern recognition
    • G06F18/20—Analysing
    • G06F18/21—Design or setup of recognition systems or techniques; Extraction of features in feature space; Blind source separation
    • G06F18/214—Generating training patterns; Bootstrap methods, e.g. bagging or boosting
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06F—ELECTRIC DIGITAL DATA PROCESSING
    • G06F18/00—Pattern recognition
    • G06F18/20—Analysing
    • G06F18/25—Fusion techniques
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06F—ELECTRIC DIGITAL DATA PROCESSING
    • G06F30/00—Computer-aided design [CAD]
    • G06F30/10—Geometric CAD
    • G06F30/15—Vehicle, aircraft or watercraft design
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
    • G06N3/00—Computing arrangements based on biological models
    • G06N3/02—Neural networks
    • G06N3/04—Architecture, e.g. interconnection topology
    • G06N3/045—Combinations of networks
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
    • G06N3/00—Computing arrangements based on biological models
    • G06N3/02—Neural networks
    • G06N3/08—Learning methods
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06F—ELECTRIC DIGITAL DATA PROCESSING
    • G06F2119/00—Details relating to the type or aim of the analysis or the optimisation
    • G06F2119/04—Ageing analysis or optimisation against ageing

Abstract

The invention provides a method and a system for predicting the real-time residual service life of a multi-rotor unmanned aerial vehicle, which comprises the following steps: constructing a multi-level fusion Transformer model based on multi-sensor data and target external factors of a multi-rotor unmanned aerial vehicle; the multi-level fusion transform model comprises a multi-scale feature extraction layer and a decoder attention layer embedded with an input tensor of a target external factor; training the multi-level fusion Transformer model based on a preset training data set to obtain a trained multi-level fusion Transformer model; and predicting the residual service life of the multi-rotor unmanned aerial vehicle by using the trained multi-level fusion Transformer model. The invention solves the technical problems that the prediction deviation of the residual service life of the existing data-driven unmanned aerial vehicle residual service life prediction method is large, the real-time performance is poor and the sequence prediction cannot be carried out.

Description

Multi-rotor unmanned aerial vehicle real-time remaining service life prediction method and system
Technical Field
The invention relates to the technical field of machine learning, in particular to a method and a system for predicting the real-time remaining service life of a multi-rotor unmanned aerial vehicle.
Background
Many rotor unmanned aerial vehicle possess ability, payload that VTOL and hover in the air and carry ability, autonomous flight or remote control flight ability, the flight mode is nimble changeable. The special advantages enable the aircraft to be widely applied to civil fields such as agricultural transportation, and meanwhile, the aircraft can switch various flight modes randomly in low-altitude, narrow and severe space environments, and has great military application potential. In the face of changeable battlefield situations, the high-mobility multi-rotor unmanned aerial vehicle needs to overcome the difficulties of complex disturbance, physical limitation, flight constraint and the like, and ensures the quick response capability in various working conditions. In order to maximize the performance mission capability of a multi-rotor drone, the remaining useful life of the flight is a necessary measure for the planning of the mission and the evaluation of the remaining flight capability. Flight durability is directly related to the total mass of the unmanned aerial vehicle, so a lithium polymer (Li-Po) battery with high energy density is usually used as a power supply to drive a high-mobility multi-rotor, and prediction of the Remaining service Life (RUL) of the Li-Po battery attracts the attention of experts in various fields, and becomes a research hotspot in the field of unmanned aerial vehicle fault Prediction and Health Management (PHM).
The RUL of the Li-Po battery is a variable that cannot be directly observed and measured and must be obtained through indirect partial correlation measurement, so that high uncertainty exists in estimation and prediction of the RUL battery, and the flight plan is highly conservative and is not beneficial to the utilization of the hitting capacity of the high-mobility unmanned aerial vehicle. The driving ability of the battery has temperature dependence and electrical load dependence, and RUL is related to battery voltage, discharge current, load, temperature, and flight regime. The existing RUL prediction technology is mainly divided into two types, namely a statistical modeling method and a machine learning method. The former, such as a Bayes estimation method, can fully use prior information, but is highly dependent on a pre-estimated model based on a system degradation mechanism, the battery discharge process is difficult to accurately model the probability distribution through a solvable mathematical model, and the high coupling and the high nonlinearity of the high maneuvering unmanned aerial vehicle further increase the inaccuracy of the model. The latter includes classical machine learning methods and deep learning methods.
Classic machine learning methods such as support vector machines and random forests enable data to train classifiers with optimal performance through advanced data processing technologies and strong and brisk algorithm technologies, but the data feature extraction process is often very tedious and time-consuming due to high precision requirements, and is not beneficial to real-time tasks such as online planning. In the field of deep learning methods, researchers often use Recurrent Neural Networks (RNNs) as a basic framework for sequence data processing, and develop leading edge horizontal baseline methods such as Gated Recurrent Units (GRUs) and long-term memory networks (LSTM). However, these methods are based on sequential processing sequences and cannot perform sequence-to-sequence prediction.
Disclosure of Invention
In view of this, the present invention aims to provide a method and a system for predicting the remaining service life of a multi-rotor unmanned aerial vehicle in real time, so as to solve the technical problems that the prediction deviation of the remaining service life of the existing data-driven unmanned aerial vehicle is large, the real-time performance is poor, and the sequence prediction cannot be performed in the existing method for predicting the remaining service life of the data-driven unmanned aerial vehicle.
In a first aspect, an embodiment of the present invention provides a method for predicting a real-time remaining service life of a multi-rotor unmanned aerial vehicle, including: constructing a multi-level fusion Transformer model based on multi-sensor data and target external factors of a multi-rotor unmanned aerial vehicle; the observation encoder of the multilevel fusion Transformer model comprises a multi-scale feature extraction layer; the multi-scale feature extraction layer is used for performing convolution calculation on the multi-sensor data based on a preset mining operator to obtain a time feature vector; the predictive decoder of the multi-level fusion Transformer model comprises a decoder attention layer embedded with input tensors of the external factors of the target; the target external factors include a load weight, a flight speed, and an ambient temperature of the multi-rotor drone; training the multi-level fusion Transformer model based on a preset training data set to obtain a trained multi-level fusion Transformer model; and predicting the residual service life of the multi-rotor unmanned aerial vehicle by using the trained multi-level fusion Transformer model.
Further, based on multi-sensor data and target external factors of the multi-rotor unmanned aerial vehicle, a multi-level fusion Transformer model is constructed, and the method comprises the following steps: adding the multi-scale feature extraction layer into an observation encoder with a preset Transformer model to construct the observation encoder with the multi-level fusion Transformer model; the multi-scale feature extraction layer is further used for performing one-dimensional convolution calculation on the multi-sensor data along a time dimension through the preset mining operator to obtain a time feature vector; and adding a decoder attention layer embedded with the input tensor of the target external factor into the predictive decoder of the preset Transformer model to construct the multi-level fused predictive decoder of the Transformer model.
Further, the preset mining operator is M (·; Θ); Θ: determining a mining scale as (k, p), wherein k represents a kernel size and p represents a fill size; d preset excavation operators are simultaneously set in each scale; d is the dimension of the multi-sensor data.
Further, the preset training data set includes: a prediction dataset and an observation dataset; training the multi-level fusion Transformer model based on a preset training data set, comprising: training the multi-level fusion Transformer model by using a target loss function based on the preset training data set; the objective loss function includes:
Figure BDA0003418228430000031
Figure BDA0003418228430000032
Figure BDA0003418228430000033
the output is predicted for the network in order to,
Figure BDA0003418228430000034
for the label value of the remaining lifetime, PairwiseDistance () is the pixel-level ohmic distance.
In a second aspect, an embodiment of the present invention further provides a system for predicting a remaining service life of a multi-rotor drone in real time, including: the device comprises a construction module, a training module and a prediction module; the building module is used for building a multi-level fusion Transformer model based on multi-sensor data and target external factors of the multi-rotor unmanned aerial vehicle; the observation encoder of the multilevel fusion Transformer model comprises a multi-scale feature extraction layer; the multi-scale feature extraction layer is used for performing convolution calculation on the multi-sensor data based on a preset mining operator to obtain a time feature vector; the predictive decoder of the multi-level fusion Transformer model comprises a decoder attention layer embedded with input tensors of the external factors of the target; the target external factors include a load weight, a flight speed, and an ambient temperature of the multi-rotor drone; the training module is used for training the multi-level fusion Transformer model based on a preset training data set to obtain a trained multi-level fusion Transformer model; and the prediction module is used for predicting the residual service life of the multi-rotor unmanned aerial vehicle by utilizing the trained multi-level fusion Transformer model.
Further, the multi-level fusion Transformer model comprises an observation encoder; the observation encoder includes: the multi-head self-attention layer, the residual connecting layer, the feedforward full-connecting layer, the residual connecting layer and the multi-scale feature extraction layer; and the multi-scale feature extraction layer is further used for performing one-dimensional convolution calculation on the multi-sensor data along a time dimension through the preset mining operator to obtain a time feature vector.
Further, the multi-level fusion Transformer model further comprises a prediction decoder; the predictive decoder includes: a multi-headed self-attention layer, a residual connection layer, a feed-forward full connection layer, a residual connection layer, a coder decoder attention layer and a decoder attention layer embedding the input tensor of the external factors of the target.
Further, the preset training data set includes: a prediction dataset and an observation dataset; the training module is further configured to: training the multi-level fusion Transformer model by using a target loss function based on the preset training data set; the objective loss function includes:
Figure BDA0003418228430000041
Figure BDA0003418228430000042
Figure BDA0003418228430000043
the output is predicted for the network in order to,
Figure BDA0003418228430000044
for the label value of the remaining lifetime, PairwiseDistance () is the pixel-level ohmic distance.
In a third aspect, an embodiment of the present invention further provides an electronic device, which includes a memory, a processor, and a computer program stored in the memory and executable on the processor, where the processor implements the steps of the method according to the first aspect when executing the computer program.
In a fourth aspect, the present invention further provides a computer-readable medium having non-volatile program code executable by a processor, where the program code causes the processor to execute the method according to the first aspect.
The embodiment of the invention provides a method and a system for predicting the real-time residual service life of a multi-rotor unmanned aerial vehicle, which consider the influence of the input saturation of the multi-rotor unmanned aerial vehicle in the edge state on the residual flight time, utilize a Transformer model network to perform multi-sensor data characteristic layer fusion to weaken the dependence on a single discharge voltage signal, and greatly increase the prediction precision and speed under various flight maneuvers; the Transformer model processes input data in parallel through an attention mechanism, so that the processing speed is increased, the time sequence characteristic representation is enhanced, and partial missing observation data can be processed; meanwhile, on the basis of a transform model, a multi-scale processing mechanism is added to perform processing representation according to fine time characteristics, and a multi-level embedded layer is further added to add external influence factors to the transform model network so as to solve feasibility assessment of a predetermined flight plan, so that the technical problems that the prediction deviation of the residual service life of the existing data-driven unmanned aerial vehicle is large, the real-time performance is poor and the sequence prediction cannot be performed in the existing data-driven unmanned aerial vehicle residual service life prediction method are solved.
Drawings
In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the drawings used in the description of the embodiments or the prior art will be briefly described below, and it is obvious that the drawings in the following description are some embodiments of the present invention, and other drawings can be obtained by those skilled in the art without creative efforts.
Fig. 1 is a flowchart of a method for predicting a real-time remaining service life of a multi-rotor unmanned aerial vehicle according to an embodiment of the present invention;
FIG. 2 is a comparison graph of a single remaining life prediction result provided by an embodiment of the present invention;
fig. 3 is a schematic diagram of a sequence remaining service life prediction result of a multi-level fusion fransformer model according to an embodiment of the present invention;
fig. 4 is a schematic diagram of a real-time remaining service life prediction system for a multi-rotor drone according to an embodiment of the present invention;
fig. 5 is a schematic structural diagram of a multi-level fusion transform model according to an embodiment of the present invention.
Detailed Description
The technical solutions of the present invention will be described clearly and completely with reference to the accompanying drawings, and it should be understood that the described embodiments are some, but not all embodiments of the present invention. All other embodiments, which can be obtained by a person skilled in the art without making any creative effort based on the embodiments in the present invention, belong to the protection scope of the present invention.
The first embodiment is as follows:
fig. 1 is a flowchart of a method for predicting real-time remaining service life of a multi-rotor drone according to an embodiment of the present invention. As shown in fig. 1, the method specifically includes the following steps:
and S102, constructing a multi-level fusion Transformer model based on multi-sensor data of the multi-rotor unmanned aerial vehicle and target external factors.
The observation encoder with the multilevel fusion Transformer model comprises a multiscale feature extraction layer; the multi-scale feature extraction layer is used for performing convolution calculation on the multi-sensor data based on a preset mining operator to obtain a time feature vector.
The predictive decoder of the multi-level fusion Transformer model comprises a decoder attention layer embedded with the input tensor of the external factors of the target; the target external factors include the load weight, flight speed, and ambient temperature of the multi-rotor drone.
And step S104, training the multi-level fusion Transformer model based on a preset training data set to obtain the trained multi-level fusion Transformer model.
And S106, predicting the residual service life of the multi-rotor unmanned aerial vehicle by using the trained multi-level fusion Transformer model.
The embodiment of the invention provides a method for predicting the real-time remaining service life of a multi-rotor unmanned aerial vehicle, which considers the influence of the input saturation of the multi-rotor unmanned aerial vehicle in the edge state on the remaining flight time, utilizes a Transformer model network to perform multi-sensor data characteristic layer fusion to weaken the dependence on a single discharge voltage signal, and greatly increases the prediction precision and speed under various flight maneuver conditions; the Transformer model processes input data in parallel through an attention mechanism, so that the processing speed is increased, the time sequence characteristic representation is enhanced, and partial missing observation data can be processed; meanwhile, on the basis of a transform model, a multi-scale processing mechanism is added to perform processing representation according to fine time characteristics, and a multi-level embedded layer is further added to add external influence factors to the transform model network so as to solve feasibility assessment of a predetermined flight plan, so that the technical problems that the prediction deviation of the residual service life of the existing data-driven unmanned aerial vehicle is large, the real-time performance is poor and the sequence prediction cannot be performed in the existing data-driven unmanned aerial vehicle residual service life prediction method are solved.
In the embodiment of the invention, the multi-sensor data of the multi-rotor unmanned aerial vehicle can be obtained in a mode of obtaining experimental data, and the multi-sensor data of the multi-rotor unmanned aerial vehicle can also be obtained in a mode of simulation.
Preferably, the embodiment of the invention acquires multi-sensor data of the multi-rotor unmanned aerial vehicle in a simulation mode. Specifically, a simulation model is established based on a high-precision four-rotor power mechanism mathematical model considering disturbance and a lithium ion battery discharge state equation considering temperature influence, and simulation model parameters are obtained through fitting of real flight data of the four-rotor unmanned aerial vehicle in specific flight speed, flight mode and flight environment temperature.
Establishing a propeller aerodynamic model (comprising propeller thrust T and propeller moment M):
Figure BDA0003418228430000071
in the formula, N is the rotating speed of the propeller,
Figure BDA0003418228430000072
is the air density, where h is the flying height, TtIs ambient temperature, CTIs the coefficient of thrust of the propeller, CMIs the torque coefficient of the propeller, DpIs the blade diameter.
Establishing a motor model:
Figure BDA0003418228430000081
in the formula of UmIs motor equivalent voltage, ImIs motor equivalent current, M is motor load torque, N is motor rotation speed (i.e. propeller rotation speed), thetamIs the basic parameter of the motor.
Establishing an electronic speed regulator model:
Figure BDA0003418228430000082
in the formula, thetae，ΘbBasic parameters of the electric regulation and basic parameters of the battery are respectively, sigma is an input throttle of the electric regulation, and UeFor electrically regulating the input voltage, IeThe current is input for electrical regulation.
The state transition equation for establishing the battery model is as follows:
Figure BDA0003418228430000083
in the formula, RintIs the internal resistance of the battery, SOC is the discharge state, EcritFor the initial total cell energy, P is the power consumed, k is the sampling time, and ω is the process noise. The measurement equation is established as follows:
V(k)＝voc(k)-i(k)·R(k)(k)int
in the formula, vocIs the open circuit voltage, i is the discharge current, and η is the measurement noise.
The power consumption is mainly determined by the flight mode and the flight speed of the unmanned aerial vehicle, and a four-rotor energy consumption model is established in a hovering state:
Figure BDA0003418228430000084
in a climbing state:
Figure BDA0003418228430000085
in a descending state:
Figure BDA0003418228430000091
in a flat flight state:
Figure BDA0003418228430000092
wherein W is the total weight of the unmanned aerial vehicle, AtIs the total area of the rotor, Vc、Vd、VhorThe speed in various flight modes. Etac、ηd、ηhorIs an efficiency factor v in various flight modeshorAn induced velocity in flat flight, avFor the angle of attack, their trend of change with flight speed can be obtained by actual flight linear fitting.
Based on the state space equation, a high-precision battery discharge failure model is obtained by applying a particle filter algorithm. On the basis of the battery voltage obtained by filtering estimation, the influence of temperature on battery discharge is further considered:
V＝(voc-θ1(i-istd))·exp(θ2·sign(Tt-Tstd)·||Tt-Tstd||2)
in the formula, theta1，θ2As a parameter of the temperature model, istd，TstdDischarge current and temperature at baseline.
The method comprises the following steps of operating a constructed high-precision simulation model to obtain multi-sensor data, and specifically comprises the following steps: battery voltage, discharge current, motor throttle, and state of discharge (SOC), etc., where SOC is indirectly derived from the discharge current integration. Each time the initial SOC is simulated, the initial SOC is randomly reduced to a small degree so as to simulate the initial power consumption, and the simulation is stopped when the operation is performed to meet the condition that the initial SOC is temporarily reduced to one of the following two conditions (namely when the remaining service life of the unmanned aerial vehicle is 0): (1) the battery voltage drops to a cut-off voltage; (2) the throttle exceeds the maximum throttle limited by the minimum attitude control capability when the unmanned aerial vehicle suspends. The total weight (no-load weight + load weight), the ambient temperature and the flying speed of the quad-rotor unmanned aerial vehicle are set values of an automatic control program so as to simulate a preset remote control value and weather forecast data in actual flight and serve as external factor signals of future sudden changes to be input into a prediction network.
After obtaining the simulation data, it is necessary to perform a pre-processAnd constructing a target measuring label. The RUL curve of the quad-rotor unmanned aerial vehicle changes along with different simulation conditions, influences are very slight and are ignored in the single flight process due to system degradation, and RUL data are generated by a linear descending assignment method in a self-adaptive mode and serve as prediction labels: rul (t) ═ min ({ t | { σ |)hover(t)≥σmax}∪{t|U(t)≤Uthr}}). Wherein sigmahoverFor hovering throttle, σmaxMaximum throttle upper limit (input saturation value), UthrIs the cut-off voltage.
Formally, for the ith flight, at TsldRespectively constructing an observation data set consisting of multiple sensor signals by intercepting the acquired data at intervals
Figure BDA0003418228430000101
Figure BDA0003418228430000102
Prediction data set composed of increment of RUL tag value and initial value of RUL prediction start time
Figure BDA0003418228430000103
External input data set composed of external factor signals
Figure BDA0003418228430000104
Figure BDA0003418228430000105
Each group of data sets is of a size
Figure BDA0003418228430000106
The raw data, after being normalized by z-score, is divided into a training set, a validation set and a test set according to a certain proportion.
Optionally, step S102 further includes:
adding a multi-scale feature extraction layer in an observation encoder with a preset Transformer model to construct an observation encoder with a multi-level fusion Transformer model; and the multi-scale feature extraction layer is also used for performing one-dimensional convolution calculation on the multi-sensor data along the time dimension through a preset mining operator to obtain a time feature vector.
Optionally, presetting a mining operator as M (·.;) and (theta); Θ: determining a mining metric as (k, p), wherein k represents a kernel size and p represents a fill size; d preset excavation operators are simultaneously set in each scale; d is the dimension of the multi-sensor data.
And adding a decoder attention layer embedded with the input tensor of the target external factors into a predictive decoder of a preset Transformer model to construct a multi-level predictive decoder fused with the Transformer model.
In an embodiment of the invention, the input data is acquired by linear projection
Figure BDA0003418228430000107
Original input data are mapped to a high-dimensional feature space to realize distributed expression, and multi-level fusion TF (namely a multi-level fusion Transformer model) is convenient to process input features:
Figure BDA0003418228430000111
wherein f islinearIs a linear projection layer, WxIs a matrix of coefficients. Correspondingly, the output of the ith flight at the time t is a D-dimensional vector
Figure BDA0003418228430000112
It will be re-projected back to the prediction spatial dimensions by the above inverse transformation, thus enabling RUL prediction data embedding.
TF realizes the feature layer fusion of multi-sensor data by taking attention to the correlation between data at different moments and different data features into account through an attention mechanism, but a parallel computing mechanism of TF is not sensitive to sequence features. To solve the above problem, a "position coding" operation is applied to encode each historical and future time instant, stamping each data to be embedded with a corresponding time stamp
Figure BDA0003418228430000113
Figure BDA0003418228430000114
Wherein the timestamp is defined as:
Figure BDA0003418228430000115
ensuring its uniqueness within 10000 time steps.
When a high-mobility four-rotor (namely, the multi-rotor unmanned aerial vehicle in the embodiment of the invention) is loaded with a limit weight load or is in a limit temperature environment, the input of the high-mobility four-rotor unmanned aerial vehicle is saturated or the driving voltage is reduced to a cut-off voltage level, so that the high-mobility four-rotor unmanned aerial vehicle is out of control and is likely to crash in the takeoff process. In order to shorten the prediction blank period at the beginning of takeoff, the TF network is expected to achieve the expected prediction effect by using as little observation data as possible. Under the condition, the time information of observation data is limited, and the embodiment of the invention is inspired by the thought of multi-granularity scanning in deep forests, constructs a multi-scale feature extraction mechanism, and is combined with a TF encoder to refine a time channel and deeply mine the time sequence information embedded in multi-sensor data.
Specifically, the multi-level fusion Transformer model obtains the embedded multi-sensor input
Figure BDA0003418228430000121
The 1d convolution calculation is performed on the input tensor along the time dimension via the mining operator M (·; Θ). Wherein, Θ: the mining metric is determined as (k, p), where k denotes the kernel size and p denotes the fill size. For specific s scales, the mining operator respectively processes the input tensor on each scale to obtain the time dimension of
Figure BDA0003418228430000122
And D mining operators are simultaneously set in each scale to ensure the characteristic dimensionality of the multi-sensor. After the time characteristics are refined in a multi-scale mode, selecting the elite characteristics as a final processing result:
Figure BDA0003418228430000123
it should be noted that the above equation limits the correspondence between k and p:
Figure BDA0003418228430000124
in the multilevel fusion transform model constructed in the embodiment of the invention, except for the embedded multi-scale feature extraction layer, an observation encoder and a prediction decoder are both composed of a plurality of basic layers with attention mechanisms, and each basic layer has three components: a multi-headed self-attention layer, a feed-forward fully-connected layer, and two remaining connected layers.
Specifically, the multi-head self-attention layer is realized by h self-attention modules in parallel computing, and for j self-attention modules, trainable hyper-parameter query vectors
Figure BDA0003418228430000125
Keypoint vector
Figure BDA0003418228430000126
Figure BDA0003418228430000127
Sum value vector
Figure BDA0003418228430000128
Respectively composed of a query matrix WQ and a key point matrix WKAnd the value matrix WVThe weight calculation mechanism of attention is determined and formed together:
Figure BDA0003418228430000129
in the formula (d)kAfter each self-attention module calculation is completed, parallel attention calculation is applied to realize information integration from different characterization subspacesCombining:
Figure BDA00034182284300001210
in the formula, WATo be an attention matrix, Concat () is a tensor concatenation. The feedforward full-connection module consists of linear transformation and a ReLU activation function, and acts on each observation time step by the same weight:
Figure BDA00034182284300001211
wherein the content of the first and second substances,
Figure BDA0003418228430000131
and
Figure BDA0003418228430000132
is a matrix of coefficients.
From the lithium ion battery discharge failure state space equation and end-of-life conditions, RUL is also affected by a number of external factors, including load weight, flight speed and ambient temperature. These factors all appear as sudden signs and the symptom-free nature results in unpredictable effects on remaining useful life and failure to match the demand for predicting high maneuver flight capability over a future period of time.
In order to enable the TF network to take into account the strong correlation between RUL time sequences and mutated extrinsic factors, embodiments of the present invention enable embedded fusion of extrinsic factors by adding an extrinsic-decoder attention layer at the decoder stage.
The objective of the observation encoder is to create a time-series representation for the multi-sensor embedded signal, giving memory to the TF network, while its trained keypoint vector KencSum value vector VencWill be shared to the decoder level. External factors can be regarded as features similar to observed values of multiple sensors and are only in different time sequence spaces, future predictability enables the external factors to directly act on a prediction time period, and therefore processing of data of the multiple sensors is achievedInput tensor gamma applied to external factorsextTo obtain
Figure BDA0003418228430000133
Figure BDA0003418228430000134
To prevent the predictive information from changing the attention of the TF network to historical observations, the feature coupling update is chosen to be done at the decoder embedding stage instead of the encoder-decoder attention stage:
Figure BDA0003418228430000135
the learned potential characteristics are fed back to the TF network through the multi-stage encoder, so that the importance of external factors and the related attention of the network to the TF network are enhanced.
Optionally, in an embodiment of the present invention, the presetting of the training data set includes: a prediction dataset and an observation dataset; training a multi-level fusion Transformer model based on a preset training data set, comprising:
training a multi-level fusion Transformer model by using a target loss function based on a preset training data set;
the objective loss function includes:
Figure BDA0003418228430000141
Figure BDA0003418228430000142
the output is predicted for the network and,
Figure BDA0003418228430000143
for the tag value of remaining lifetime, PairwiseDistance () is the pixel level ohmic distance.
In the embodiment of the invention, the residual service life prediction algorithm of the high-mobility unmanned aerial vehicle learns the nonlinear change rule of the RUL along with the discharge failure of the lithium battery, external factors and multi-sensor data through an offline training multi-level fusion Transformer model, and the trained model is directly applied to an online prediction stage, so that the accurate and real-time residual service life prediction of the high-mobility unmanned aerial vehicle in various complex flight processes is realized. In the off-line training process, the back propagation algorithm utilizes an Adam optimizer to minimize errors to achieve non-linear fit.
For the purpose of clearly describing the objects and technical solutions of the present invention, the present invention will be further described in detail with reference to the following simulation examples.
The application example carries out parameter fitting and setting on the simulation model established in the embodiment of the invention through a normal flight experiment and a battery temperature experiment box experiment of the quad-rotor unmanned aerial vehicle in a multi-flight mode, a variable flight speed and a variable load weight. The wheelbase of the four-rotor unmanned aerial vehicle frame is 550mm, the unloaded weight is 1.357kg, and the normal flight behavior at 30 ℃ and 120m is taken as a baseline condition; the battery is a 5100mAh 8C lithium polymer battery, the internal resistance of the battery is 27m omega, and the cut-off voltage is 10.3V; an Air Gear 350KV920 model MOTOR of T-MOTOR is adopted, and 2 pairs of T-MOTOR T9545-B propellers and 4T-MOTOR AIR 20A electric MOTORs are arranged. The high-mobility four-rotor simulation system utilizes PID to realize automatic control, and considers the ground effect, gust disturbance and sensor noise when being close to the ground. The load weight range of the unmanned aerial vehicle is within 0.5kg, the flight space is set to 1000m multiplied by 200m, and the horizontal flight speed is Vhor，x∈[-8，8]m/s，Vhor，y∈[-8，8]m/s, maximum descent velocity Vd，max2m/s, maximum climbing speed Vc，maxThe maximum throttle is 0.95 at 4.5 m/s. The temperature variation range is T epsilon [0, 60]DEG C. And acquiring 850 times of flight data of four flight modes under the conditions of different external influence factors through a simulation experiment with the sampling time of 1 s. Based on the real data and the actual demand, the prediction algorithm needs to predict the future RUL of 48s from the observed multi-sensor data of 32s, so 850 times of flight data are intercepted with a sliding time window of length 80s and a sliding distance of 80s to form a raw data set. The multisensor observation input dimension is 32 × 7, the external factor input dimension is 48 × 5, and the RUL prediction input and output dimensions are 48 × 2.
The parameter of a multi-level fusion transform model which can realize high-precision and real-time prediction requirements is set to be D (512), multi-scale mining factors are selected to be k (1), 3 and 5 scales, corresponding filling size p (0, 1 and 2), an encoder and a decoder are composed of 6 basic layers, and each multi-head self-attention module is provided with 8 attention heads. During model training, the batch size is set to 80, the maximum iteration times is 30, and the sizes of a training set and a test set are set to be 3: 1: 1. adam optimizer parameter setting to beta1＝0.9，β2＝0.98，ε＝10-9. The experimental result shows that the proposed multi-level fusion Transformer model can quickly predict the residual service life with high precision, the average prediction time is 0.017029s, the error of the training set converges to 0.1296, the average absolute deviation (MAD) of the verification set converges to 24.20, and the MAD of the test set converges to 24.68. The prediction method has better effect than LSTM and Transformer in the residual service life prediction at the next moment (single point prediction) and the residual service life prediction in a future period (sequence prediction).
And further using a multilevel fusion Transformer to predict the residual flight life of the four-rotor aircraft in high maneuvering flight at a certain time, wherein the prediction error is 58.4723, which is lower than 61.3311 of the Transformer, and the change trend of the residual service life can be more reasonably predicted under various flight states. The specific prediction results are shown in fig. 2 and 3. Fig. 2 is a comparison diagram of a single residual service life prediction result provided according to an embodiment of the present invention, and fig. 3 is a schematic diagram of a sequence residual service life prediction result of a multi-level fusion fransformer model provided according to an embodiment of the present invention.
Example two:
fig. 4 is a schematic diagram of a real-time remaining service life prediction system for a multi-rotor drone, according to an embodiment of the invention. As shown in fig. 4, the system includes: a building module 10, a training module 20 and a prediction module 30.
Specifically, the building module 10 is used for building a multi-level fusion transform model based on multi-sensor data and target external factors of the multi-rotor unmanned aerial vehicle.
The observation encoder with the multilevel fusion Transformer model comprises a multiscale feature extraction layer; the multi-scale feature extraction layer is used for performing convolution calculation on multi-sensor data based on a preset mining operator to obtain a time feature vector; the predictive decoder of the multi-level fusion Transformer model comprises a decoder attention layer embedded with the input tensor of the external factors of the target; the target external factors include the load weight, flight speed, and ambient temperature of the multi-rotor drone.
The training module 20 is configured to train the multi-level fusion Transformer model based on a preset training data set to obtain the trained multi-level fusion Transformer model.
And the prediction module 30 is used for predicting the residual service life of the multi-rotor unmanned aerial vehicle by using the trained multi-level fusion Transformer model.
Optionally, the presetting of the training data set comprises: a prediction dataset and an observation dataset; a training module further to:
training a multi-level fusion Transformer model by using a target loss function based on a preset training data set;
the objective loss function includes:
Figure BDA0003418228430000161
Figure BDA0003418228430000162
the output is predicted for the network in order to,
Figure BDA0003418228430000163
for the tag value of remaining lifetime, PairwiseDistance () is the pixel level ohmic distance.
The embodiment of the invention provides a multi-rotor unmanned aerial vehicle real-time remaining service life prediction system, which considers the influence of input saturation of a multi-rotor unmanned aerial vehicle in an edge state on remaining flight time, utilizes a Transformer model network to perform multi-sensor data characteristic layer fusion to weaken the dependence on a single discharge voltage signal, and greatly increases the prediction precision and speed under various flight maneuver conditions; the Transformer model processes input data in parallel through an attention mechanism, so that the processing speed is increased, the time sequence characteristic representation is enhanced, and partial missing observation data can be processed; meanwhile, on the basis of a transform model, a multi-scale processing mechanism is added to perform processing representation according to fine time characteristics, and a multi-level embedded layer is further added to add external influence factors to the transform model network so as to solve feasibility assessment of a predetermined flight plan, so that the technical problems that the prediction deviation of the residual service life of the existing data-driven unmanned aerial vehicle is large, the real-time performance is poor and the sequence prediction cannot be performed in the existing data-driven unmanned aerial vehicle residual service life prediction method are solved.
Optionally, fig. 5 is a schematic structural diagram of a multi-level fusion transform model according to an embodiment of the present invention. As shown in FIG. 5, the multi-level fusion Transformer model includes an observation encoder; the observation encoder includes: the multi-head self-attention layer, the residual connecting layer, the feedforward full-connecting layer, the residual connecting layer and the multi-scale feature extraction layer;
and the multi-scale feature extraction layer is also used for performing one-dimensional convolution calculation on the multi-sensor data along the time dimension through a preset mining operator to obtain a time feature vector.
Specifically, as shown in fig. 5, the multi-level fusion Transformer model further includes a predictive decoder; the predictive decoder includes: the multi-head self-attention layer, the residual connection layer, the feedforward full connection layer, the residual connection layer, the encoder decoder attention layer and the decoder attention layer of the input tensor embedded in the external factors of the target.
The embodiment of the present invention further provides an electronic device, which includes a memory, a processor, and a computer program stored in the memory and capable of running on the processor, and when the processor executes the computer program, the steps of the method in the first embodiment are implemented.
The embodiment of the invention also provides a computer readable medium with a non-volatile program code executable by a processor, wherein the program code causes the processor to execute the method in the first embodiment.
Finally, it should be noted that: the above embodiments are only used to illustrate the technical solution of the present invention, and not to limit the same; while the invention has been described in detail and with reference to the foregoing embodiments, it will be understood by those skilled in the art that: the technical solutions described in the foregoing embodiments may still be modified, or some or all of the technical features may be equivalently replaced; and the modifications or the substitutions do not make the essence of the corresponding technical solutions depart from the scope of the technical solutions of the embodiments of the present invention.

Claims (10)

1. The utility model provides a real-time remaining service life prediction method of many rotor unmanned aerial vehicle which characterized in that includes:
constructing a multi-level fusion Transformer model based on multi-sensor data and target external factors of a multi-rotor unmanned aerial vehicle; the observation encoder of the multilevel fusion Transformer model comprises a multi-scale feature extraction layer; the multi-scale feature extraction layer is used for performing convolution calculation on the multi-sensor data based on a preset mining operator to obtain a time feature vector; the predictive decoder of the multi-level fusion transform model comprises a decoder attention layer embedded with the input tensor of the target external factors; the target external factors include a load weight, a flight speed, and an ambient temperature of the multi-rotor drone;
training the multi-level fusion Transformer model based on a preset training data set to obtain a trained multi-level fusion Transformer model;
and predicting the residual service life of the multi-rotor unmanned aerial vehicle by using the trained multi-level fusion Transformer model.
2. The method of claim 1, wherein constructing a multi-level fusion transform model based on multi-sensor data and target external factors for a multi-rotor drone comprises:
adding the multi-scale feature extraction layer into an observation encoder with a preset Transformer model to construct the observation encoder with the multi-level fusion Transformer model;
the multi-scale feature extraction layer is further used for performing one-dimensional convolution calculation on the multi-sensor data along a time dimension through the preset mining operator to obtain a time feature vector;
and adding a decoder attention layer embedded with the input tensor of the target external factors into the predictive decoder of the preset Transformer model to construct the multi-level fused predictive decoder of the Transformer model.
3. The method of claim 2, wherein the predetermined mining operator is M (·; determining a mining scale by (k, p), wherein k represents a kernel size, and p represents a filling size; d preset excavation operators are simultaneously set in each scale; d is the dimension of the multi-sensor data.
4. The method of claim 1, wherein the preset training data set comprises: a prediction dataset and an observation dataset; training the multi-level fusion Transformer model based on a preset training data set, comprising:
training the multi-level fusion Transformer model by using a target loss function based on the preset training data set;
the objective loss function includes:
Figure FDA0003418228420000021
Figure FDA0003418228420000022
the output is predicted for the network and,
Figure FDA0003418228420000023
for the label value of the remaining lifetime, PairwiseDistance () is the pixel-level ohmic distance.
5. The utility model provides a real-time remaining service life prediction system of many rotor unmanned aerial vehicle which characterized in that includes: the system comprises a construction module, a training module and a prediction module; wherein, the first and the second end of the pipe are connected with each other,
the building module is used for building a multi-level fusion transform model based on multi-sensor data and target external factors of the multi-rotor unmanned aerial vehicle; the observation encoder of the multilevel fusion Transformer model comprises a multi-scale feature extraction layer; the multi-scale feature extraction layer is used for performing convolution calculation on the multi-sensor data based on a preset mining operator to obtain a time feature vector; the predictive decoder of the multi-level fusion transform model comprises a decoder attention layer embedded with the input tensor of the target external factors; the target external factors include a load weight, a flight speed, and an ambient temperature of the multi-rotor drone;
the training module is used for training the multi-level fusion Transformer model based on a preset training data set to obtain a trained multi-level fusion Transformer model;
and the prediction module is used for predicting the residual service life of the multi-rotor unmanned aerial vehicle by utilizing the trained multi-level fusion Transformer model.
6. The system of claim 5, wherein the multi-level fusion fransformer model comprises an observation encoder; the observation encoder includes: a multi-head self-attention layer, a residual connecting layer, a feedforward full-connecting layer, a residual connecting layer and the multi-scale feature extraction layer;
and the multi-scale feature extraction layer is further used for performing one-dimensional convolution calculation on the multi-sensor data along a time dimension through the preset mining operator to obtain a time feature vector.
7. The system of claim 5, wherein the multi-level fusion fransformer model further comprises a predictive decoder; the predictive decoder includes: a multi-headed self-attention layer, a residual connection layer, a feed-forward full connection layer, a residual connection layer, a coder decoder attention layer and a decoder attention layer embedding the input tensor of the external factors of the target.
8. The system of claim 5, wherein the preset training data set comprises: a prediction dataset and an observation dataset; the training module is further configured to:
training the multi-level fusion Transformer model by using a target loss function based on the preset training data set;
the objective loss function includes:
Figure FDA0003418228420000031
Figure FDA0003418228420000032
the output is predicted for the network in order to,
Figure FDA0003418228420000033
for the label value of the remaining lifetime, PairwiseDistance () is the pixel-level ohmic distance.
9. An electronic device comprising a memory, a processor and a computer program stored on the memory and executable on the processor, wherein the steps of the method of any of the preceding claims 1 to 4 are implemented when the computer program is executed by the processor.
10. A computer-readable medium having non-volatile program code executable by a processor, wherein the program code causes the processor to perform the method of any of claims 1-4.
CN202111561949.5A 2021-12-17 2021-12-17 Multi-rotor unmanned aerial vehicle real-time remaining service life prediction method and system Pending CN114757086A (en)

Priority Applications (1)

Application Number Priority Date Filing Date Title
CN202111561949.5A CN114757086A (en) 2021-12-17 2021-12-17 Multi-rotor unmanned aerial vehicle real-time remaining service life prediction method and system

Applications Claiming Priority (1)

Application Number Priority Date Filing Date Title
CN202111561949.5A CN114757086A (en) 2021-12-17 2021-12-17 Multi-rotor unmanned aerial vehicle real-time remaining service life prediction method and system

Publications (1)

Publication Number Publication Date
CN114757086A true CN114757086A (en) 2022-07-15

Family

ID=82325485

Family Applications (1)

Application Number Title Priority Date Filing Date
CN202111561949.5A Pending CN114757086A (en) 2021-12-17 2021-12-17 Multi-rotor unmanned aerial vehicle real-time remaining service life prediction method and system

Country Status (1)

Country Link
CN (1) CN114757086A (en)

Cited By (1)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
WO2024016586A1 (en) * 2022-07-18 2024-01-25 中国电信股份有限公司 Machine room temperature control method and apparatus, and electronic device and storage medium

Cited By (1)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
WO2024016586A1 (en) * 2022-07-18 2024-01-25 中国电信股份有限公司 Machine room temperature control method and apparatus, and electronic device and storage medium

Similar Documents

Publication Publication Date Title
CN109034376B (en) L STM-based unmanned aerial vehicle flight state prediction method and system
US11823562B2 (en) Unsupervised anomaly detection for autonomous vehicles
CN111310965A (en) Aircraft track prediction method based on LSTM network
CN107092987B (en) Method for predicting autonomous landing wind speed of small and medium-sized unmanned aerial vehicles
Chuentawat et al. The comparison of PM2. 5 forecasting methods in the form of multivariate and univariate time series based on support vector machine and genetic algorithm
CN110083168A (en) Small-sized depopulated helicopter based on enhancing study determines high control method
CN107065897A (en) Three Degree Of Freedom helicopter explicit model forecast Control Algorithm
CN112784920A (en) Cloud-side-end-coordinated dual-anti-domain self-adaptive fault diagnosis method for rotating part
Sindhwani et al. Unsupervised anomaly detection for self-flying delivery drones
CN114757086A (en) Multi-rotor unmanned aerial vehicle real-time remaining service life prediction method and system
Goecks Human-in-the-loop methods for data-driven and reinforcement learning systems
Tang et al. A power transfer model-based method for lithium-ion battery discharge time prediction of electric rotatory-wing UAV
Gu et al. A coordinate descent method for multidisciplinary design optimization of electric-powered winged uavs
CN113609762A (en) Electric cooling and heating load combined prediction method and system based on GRU-MTL
Li et al. Method for predicting failure rate of airborne equipment based on optimal combination model
CN114384931B (en) Multi-target optimal control method and equipment for unmanned aerial vehicle based on strategy gradient
CN111580389A (en) Three-degree-of-freedom helicopter explicit model prediction control method based on deep learning
CN114580705B (en) Method for predicting residual life of avionics product
Yu et al. Longitudinal wind field prediction based on DDPG
CN114444763A (en) Wind power prediction method based on AFSA-GNN
Qian et al. An improved aircraft landing distance prediction model based on particle swarm optimization—Extreme learning machine method
Zhang et al. A novel particle swarm neural network model to optimize aircraft fuel consumption
Martinez-Gonzalez et al. ROAM-ML: A reduced order aerodynamic module augmented with neural network digital surrogates
Zhao et al. Active Learning Gaussian Process Model Predictive Control Method for Quadcopter
Li Hybrid Optimization for High Aspect Ratio Wings with Convolutional Neural Networks and Squirrel Optimization Algorithm

Legal Events

Date Code Title Description
PB01 Publication
PB01 Publication
SE01 Entry into force of request for substantive examination
SE01 Entry into force of request for substantive examination