Self-adaptive unmanned aerial vehicle early fault early warning method

Abstract

The invention provides a self-adaptive unmanned aerial vehicle early fault early warning method, and mainly aims to effectively distinguish the normal state and the potential fault of an unmanned aerial vehicle, reduce false alarm and missing alarm, and monitor the health condition of the unmanned aerial vehicle in real time. According to the method, the multi-modal operation parameters of the unmanned aerial vehicle are collected, dimensionality reduction and cleaning are carried out on the multi-modal operation parameters, and the health state feature matrix of the flight state of the unmanned aerial vehicle is obtained; a hyperspherical model is constructed based on the health state feature matrix, training is completed, and the radius of the hyperspherical model after training is completed serves as an early warning threshold value; and early warning of early faults of the unmanned aerial vehicle is realized. According to the method, the boundary construction and fault detection capability of the support vector data model is optimized, the decision boundary can be dynamically adjusted according to the data distribution by introducing a sliding window and a Bayesian parameter optimization mechanism, and the accuracy and robustness of early fault early warning are further improved.

Description

Technical Field

[0001] The present invention belongs to the field of unmanned aerial vehicle (UAV) fault warning, and particularly relates to an adaptive early fault warning method for UAVs. Background Art

[0002] UAVs are widely used in fields such as agricultural monitoring, logistics distribution, environmental protection, and disaster relief. They can perform tasks in dangerous or remote areas, improving operation accuracy while greatly reducing labor costs. The working principle of a UAV is to generate lift through rotating propellers, enabling it to hover and fly directionally, and perform tasks such as shooting, mapping, and transportation while hovering and flying directionally. If a UAV fails, it may lead to task interruption and data loss, and in severe cases, cause damage to the UAV or even mission failure. Therefore, it is crucial to monitor the health of UAVs and perform condition-based maintenance. If repair measures can be taken in the initial stage of a fault, it will effectively reduce downtime, prevent further damage to the UAV, and thus ensure flight safety, improve work efficiency, and reduce long-term operating costs.

[0003] Traditional early fault warning of UAVs usually relies on regular manual inspections and experience-based judgments, mainly by monitoring physical parameters such as vibration, temperature, and sound to identify potential faults. The advantage of this method is simplicity and low cost. However, its disadvantages are also very obvious. It relies on manual operation and is difficult to detect tiny and early fault signals in a timely manner; it cannot monitor the flight state of the UAV in real-time and continuously, and is prone to missing key fault symptoms. In addition, experience-based judgments are often subjective, which may lead to misjudgments or missed judgments and cannot guarantee efficient fault warning. The development of machine learning technology provides a new idea for early fault warning of UAVs. Compared with traditional methods, machine learning does not need to rely on manually set rules or thresholds, can identify subtle changes and implicit patterns in data, extract potential abnormal features from complex signals, and detect potential faults earlier, improving the accuracy and sensitivity of early fault warning. However, existing machine learning-based early fault warning methods for UAVs have problems of false alarms and missed alarms. In order to effectively distinguish the normal state and potential fault state of UAVs, reduce false alarms and missed alarms, and monitor the health status of UAVs in real-time, the present invention provides an adaptive early fault warning method for UAVs to achieve more accurate and reliable early weak fault warning. Summary of the Invention

[0004] To overcome the deficiencies of the prior art, the present invention provides an adaptive early fault warning method for unmanned aerial vehicles (UAVs). The present invention can capture subtle abnormal signals of UAVs in a strong noise environment and adapt to different working conditions and health states by adaptively adjusting the decision boundary. The present invention combines the advantages of an enhanced denoising autoencoder network and an adaptive Bayesian support vector data description model, can effectively distinguish normal states and potential fault states, reduce false alarms and missed alarms, monitor the health status of UAVs, and thus achieve more accurate and reliable early weak fault warning.

[0005] An adaptive early fault warning method for unmanned aerial vehicles (UAVs) includes the following steps:

[0006] Step 1: Collect multi-modal flight parameters during the flight of the UAV, normalize the multi-modal flight parameters to obtain the normalized UAV flight parameters, and use the trained autoencoder network to perform dimensionality reduction and data cleaning on the normalized UAV flight parameters to obtain a health status feature matrix that effectively characterizes the flight state of the UAV;

[0007] The health status feature matrix eliminates the noise and outliers of the UAV flight parameters, reducing the influence of external environmental factors and measurement errors;

[0008] Step 2: Construct a hypersphere model based on the health status feature matrix and complete the training, and use the radius of the trained hypersphere model as the warning threshold;

[0009] Step 3: Use the trained hypersphere model to give an early weak fault warning for the UAV;

[0010] Process the multi-modal flight parameters of the UAV to be predicted according to the method of Step 1 to obtain the health status feature matrix of the UAV to be predicted, and then input the health status feature matrix of the UAV to be predicted into the trained hypersphere model. Calculate the distance between the health status feature matrix and the center of the hypersphere model. If the distance is less than or equal to the warning threshold, the UAV is considered to be in a normal state; if the distance is greater than the warning threshold, the UAV is considered to be in a potential fault state and a warning is triggered.

[0011] Furthermore, the step of collecting multi-modal flight parameters during the flight of the UAV and normalizing the multi-modal flight parameters to obtain the normalized UAV flight parameters is as follows:

[0012] Step S1: Collect multi-modal flight parameters during the flight of the UAV , , is the th multi-modal flight parameter of the UAV, is the number of multi-modal flight parameters;

[0013] Step S2: Process the multi-modal flight parameters using the standard deviation normalization method to obtain the normalized UAV flight parameters;

[0014] The normalized UAV flight parameters are:

[0015] , is the th normalized UAV flight parameter, , is the number of multi-modal flight parameters;

[0016] , is the mean of the th multi-modal flight parameter; the is the th standard deviation of the multi-modal flight parameter.

[0017] Furthermore, the autoencoder network is an enhanced denoising autoencoder network;

[0018] The enhanced denoising autoencoder network is a feedforward neural network FNN; the input of the enhanced denoising autoencoder network is the normalized UAV flight parameters and random noise ; the output of the enhanced denoising autoencoder network is the health state feature matrix ; random noise is a matrix; the size of random noise is the same as the dimension of the normalized UAV flight parameters; the elements of random noise are random numbers between 0 and 1;

[0019] The enhanced denoising autoencoder network includes an encoder and a decoder;

[0020] The encoder includes a batch normalization BN layer, a ReLU activation function layer, a fully connected layer, and a residual connection layer; the layers of the encoder are fully connected; the input of the encoder is the normalized UAV flight parameters and random noise ; the output of the encoder is the health state feature matrix ; where , , is the dimension of the features of the UAV health state feature matrix after dimensionality reduction, is the th column health state vector of the health state feature matrix;

[0021] The decoder It includes a residual connection layer, a fully connected layer, a BN layer, and a ReLU activation function layer; the layers of the decoder are fully connected; the input of the decoder is the output of the encoder , and the output of the decoder is the reconstructed and normalized UAV flight parameters ;

[0022] The loss function of the enhanced denoising autoencoder network is:

[0023]

[0024] wherein, represents the th input of the enhanced denoising autoencoder network, represents the th output of the enhanced denoising autoencoder network, is the total sample size of the UAV flight parameters;

[0025] The enhanced denoising autoencoder network updates the weights and biases of the network using the batch gradient descent method, optimizes the parameters, and minimizes the value of the loss function;

[0026] To improve the training efficiency of the enhanced denoising autoencoder network, an exponential decay strategy is used to dynamically adjust the learning rate of the enhanced denoising autoencoder network; the learning rate of the enhanced denoising autoencoder network in the th round of training is:

[0027] ;

[0028] wherein, is the initial learning rate, is the decay rate, is used to control the decay speed, takes a positive constant value;

[0029] The enhanced denoising autoencoder network is trained using the normalized UAV flight parameters, and when the specified number of training rounds is reached, it is considered that the training of the enhanced denoising autoencoder network is completed; the specified number of training rounds is 100 rounds.

[0030] Furthermore, the steps of constructing a hypersphere model based on the health state feature matrix, completing the training, and using the radius of the trained hypersphere model as the warning threshold are:

[0031] Step S4-1: Preprocess the health state feature matrix using the sliding window algorithm to obtain local features ;

[0032] The sliding window algorithm for the health state feature matrix The processing process is as follows:

[0033] Set a sliding window on the health status feature matrix, and the size of the sliding window is where The range of is , is the total sample size of the UAV flight parameters; the step size of the sliding window is greater than 1;

[0034] Move the sliding window and extract the local features within each sliding window; the local features of the th sliding window are represented as where ; is the size of the sliding window, is the total sample size of the UAV flight parameters;

[0035] Divide the local features into a training set and a test set ;

[0036] Step S4-2: Use the training set to establish a hypersphere model;

[0037] The steps to establish a hypersphere model are as follows:

[0038] Step S4-2-1, the center of the hypersphere model is , the radius of the hypersphere model is , the input of the hypersphere model is the training , determine the objective function of and is:

[0039] (2)

[0040] where, is the slack variable; is the weight parameter, and the weight parameter controls the weight of the penalty term;

[0041] Step S4-2-2: Use the training set to train the hypersphere model. When the number of training times reaches the preset maximum number of iterations, the hypersphere model is trained and saved at this time, and the radius of the hypersphere model is used as the warning threshold.

[0042] Furthermore, the steps to optimize the slack variable and weight parameter of the hypersphere model are:

[0043] Use the training set to train the hypersphere model, and the training set After each piece of data in the input hyper-sphere model, the Bayesian optimization algorithm is used to optimize the weight parameters and the slack variables to obtain the optimized weight parameters and the optimized slack variables . The optimized weight parameters are used to update the weight parameters of the hyper-sphere model, and the optimized slack variables are used to update the slack variables of the hyper-sphere model; then continue to input the next piece of data in the training set until all the data in the training set are traversed; the Bayesian optimization algorithm is used to optimize the hyper-sphere model, improving the accuracy of the hyper-sphere model prediction.

[0044] The beneficial effects of the present invention are as follows:

[0045] 1. The present invention adopts an adaptive Bayesian support vector data description method to realize the early warning of weak faults of unmanned aerial vehicles. This method can effectively distinguish the normal state and the potential fault state, and the learning can characterize the important features of the monitoring parameters of unmanned aerial vehicles during strong noise flight, and early warning of weak faults during the flight of unmanned aerial vehicles.

[0046] 2. The present invention adopts a data dimensionality reduction and data cleaning technology based on an enhanced denoising autoencoder network, which is applicable to complex multi-modal monitoring data, noise data and abnormal data generated during the flight of unmanned aerial vehicles, and can adaptively extract key features representing the health state of unmanned aerial vehicles from the above data, and eliminate redundant and irrelevant information.

[0047] 3. The present invention uses an adaptive Bayesian support vector data description model to determine the decision boundary, optimizing the boundary construction and fault detection capabilities of the support vector data model. By introducing a sliding window and a Bayesian parameter optimization mechanism, the decision boundary can be dynamically adjusted according to the data distribution instead of being fixed, further improving the accuracy and robustness of early fault warning. Description of the Drawings

[0048] Figure 1 is a schematic diagram of an adaptive early fault warning method for unmanned aerial vehicles in this application;

[0049] Figure 2 is a schematic diagram of the enhanced denoising autoencoder network in this application;

[0050] Figure 3 is a schematic diagram of the hyper-sphere model in this application;

[0051] Figure 4 is a schematic diagram of the early weak fault warning process for unmanned aerial vehicles in this application;

[0052] Figure 5It is a schematic diagram of the features obtained by the enhanced denoising autoencoder in this application;

[0053] Figure 6 It is a schematic diagram of the early weak fault warning result of the drone in this application. Specific implementation manners

[0054] The multi-sensor data of the drone, as an important indicator for monitoring the flight state of the drone, covers parameters such as vibration, temperature, eddy current, and rotational speed during the flight of the drone. These parameters are of great significance in monitoring the flight state of the drone, evaluating the health status, and predicting potential faults. By deeply analyzing the monitored parameters, potential fault modes during the flight of the drone can be revealed, and key features related to faults can be identified, thus providing a scientific basis for fault warning. Based on the warning results, a reasonable predictive maintenance plan is formulated to help develop a more intelligent maintenance decision-making system, which can not only effectively reduce the unplanned downtime, but also reduce the maintenance cost and ensure the stable operation of the drone.

[0055] Since the drone monitoring parameters usually have high-dimensionality, non-linearity, and time-series characteristics, and are affected by external environmental factors, simple machine learning models and statistical models are difficult to capture the complex correlation relationships between the parameters. The correlation between these parameters may not be reflected by explicit expressions, and traditional methods often show limitations when facing these complex data. In addition, the drone flight parameters also face problems such as noise, outliers, and insufficient fault samples. Noise and outliers may result from sensor failures, external environmental changes, or measurement errors. These factors may mask the true fault signals, making data analysis more difficult and further exacerbating the difficulty of early weak fault warning. At the same time, the high reliability requirement of drone flight makes fault samples scarce. The lack of sufficient normal and fault samples leads to poor generalization ability of model training and increases the error rate of early fault identification. Therefore, how to improve the accuracy and reliability of early fault warning in a complex, noisy, and unbalanced data environment has become an important challenge faced by the current drone fault warning field.

[0056] The following will elaborate on the specific implementation details of the present invention in detail with reference to the accompanying drawings, aiming to more deeply reveal its design concept, the technical problems to be solved, the core features of the technical solution, and the technical benefits generated. However, it must be clear that the elaboration of these implementation cases is only for illustrative purposes and is not intended to limit the specific scope of the present invention.

[0057] An adaptive early fault warning method for drones includes the following steps:

[0058] Step S1: Collect multi-modal flight parameters during the flight of the drone , , is the th multimodal flight parameter of the UAV, is the number of multimodal flight parameters;

[0059] Step S2: Process the multimodal flight parameters using the standard deviation normalization method to obtain the normalized UAV flight parameters;

[0060] The normalized UAV flight parameters are:

[0061] , is the th normalized UAV flight parameter, , is the number of multimodal flight parameters;

[0062] , is the mean of the th multimodal flight parameter; the th is the th standard deviation of the multimodal flight parameter;

[0063] The normalization process eliminates the dimensional difference, adjusts the values of the multimodal flight parameters to a unified scale or interval; prevents the influence of the dimensional difference of different flight parameters on the results of the subsequent feature extraction model, and at the same time accelerates the convergence speed of the feature extraction model;

[0064] Step S3: Use an autoencoder network to perform dimensionality reduction and data cleaning on the normalized UAV flight parameters to obtain a health state feature matrix that effectively represents the UAV flight state;

[0065] In this step, the autoencoder network used is an enhanced denoising autoencoder network; the enhanced denoising autoencoder network is as Figure 2 shown; the schematic diagram of the health state feature matrix obtained by the enhanced denoising autoencoder is as Figure 5 shown;

[0066] Take the normalized UAV flight parameters and random noise as the input of the enhanced denoising autoencoder network, perform dimensionality reduction and data cleaning on the normalized UAV flight parameters, suppress the noise of the normalized UAV flight parameters, and obtain a health state feature matrix that effectively represents the UAV flight state to enhance the sensitivity to weak faults;

[0067] The enhanced denoising autoencoder network mentioned above is a feedforward neural network FNN; the input of the enhanced denoising autoencoder network is the normalized UAV flight parameters and random noise ; The output of the enhanced denoising autoencoder network is the healthy state feature matrix ; Random noise is a matrix; Random noise has the same size as the dimension of the normalized UAV flight parameters; Random noise The elements of are random numbers between 0 and 1;

[0068] The enhanced denoising autoencoder network includes an encoder and a decoder;

[0069] The encoder includes a batch normalization BN layer, a ReLU activation function layer, a fully connected layer, and a residual connection layer; The layers of the encoder are fully connected; The input of the encoder is the normalized UAV flight parameters and random noise ; The output of the encoder is the healthy state feature matrix ; Wherein , , is the dimension of the features of the reduced-dimension UAV healthy state feature matrix, is the th column healthy state vector of the healthy state feature matrix;

[0070] The decoder includes a residual connection layer, a fully connected layer, a BN layer, and a ReLU activation function layer; The layers of the decoder are fully connected; The input of the decoder is the output of the encoder , and the output of the decoder is the reconstructed normalized UAV flight parameters ;

[0071] The loss function of the enhanced denoising autoencoder network is:

[0072]

[0073] Wherein, represents the th input of the enhanced denoising autoencoder network, represents the th output of the enhanced denoising autoencoder network, is the total sample size of the UAV flight parameters;

[0074] The enhanced denoising autoencoder network uses the batch gradient descent method to update the weights and biases of the network, optimize the parameters, and minimize the value of the loss function;

[0075] To improve the training efficiency of the enhanced denoising autoencoder network, an exponential decay strategy is used to dynamically adjust the learning rate of the enhanced denoising autoencoder network; The enhanced denoising autoencoder network is trained for the Learning rate of the round is as follows:

[0076] ;

[0077] Among them, is the initial learning rate, is the attenuation rate, which is used to control the attenuation speed, and the value of

[0078] is a positive constant value;

[0079] The enhanced denoising autoencoder network is trained using the normalized UAV flight parameters, and when the specified number of training rounds is reached, it is considered that the training of the enhanced denoising autoencoder network is completed; the specified number of training rounds is 100 rounds;

[0080] Step S4: Construct a hypersphere model based on the health status feature matrix;

[0081] Step S4-1: Preprocess the health status feature matrix using the sliding window algorithm to obtain local features ;

[0082] The process of the sliding window algorithm for processing the health status feature matrix is as follows:

[0083] Set a sliding window on the health status feature matrix, and the size of the sliding window is where ranges from , and

[0084] is the total sample size of the UAV flight parameters; the step size of the sliding window is greater than 1; Move the sliding window to extract the local features within each sliding window; the local features of the th sliding window are denoted as , where is the size of the sliding window, is the total sample size of the UAV flight parameters;

[0085] Divide the local features into a training set and a test set ;

[0086] Training set Used to train the proposed Bayesian support vector data description model, test set It is used to test the performance of the trained Bayesian support vector data description model and realize early warning of weak faults of UAVs;

[0087] Step S4-2: Using the training set Build a hypersphere model;

[0088] The steps to build a hypersphere model are:

[0089] Step S4-2-1, the center of the hypersphere model is , the radius of the hypersphere model is , the input of the hypersphere model is training ,Sure and The objective function is:

[0090] (2)

[0091] in, is the slack variable; is the weight parameter, which controls the weight of the penalty term. The meaning of formula (2) is to satisfy Under the condition of The minimum value of

[0092] Using the training set Train the hypersphere model, training set After each data in the hypersphere model is input, the weight parameters are optimized using the Bayesian optimization algorithm. and slack variables Optimize and obtain the optimized weight parameters and the optimized slack variables , using the optimized weight parameters Update the weight parameters of the hypersphere model, using the optimized slack variables Update the relaxation variables of the hypersphere model; then continue to input the next data in the training set until all the data in the training set are traversed;

[0093] When the number of training times reaches the preset maximum number of iterations, the training of the hypersphere model is completed, the hypersphere model is saved, and the radius of the hypersphere model is used as the warning threshold;

[0094] A hypersphere model is established so that normal data is contained in the hypersphere as much as possible, while potential fault data is outside the hypersphere. The schematic diagram is shown in the figure. Figure 3 As shown in Figure 2, the distribution of normal state data is effectively described by optimizing the radius and center position of the hypersphere, and points that deviate from the normal state data distribution are detected as faults.

[0095] Step S5: Calculate the distance to the center of the hypersphere model for the test set ; ;

[0096]

[0097] If the distance is less than or equal to the warning threshold, i.e., , it is considered to be in a normal state; otherwise, it is considered to be in a potential fault state and a warning is triggered.

[0098] The early weak fault warning process of the UAV based on the hypersphere model is as shown in Figure 4 : First, input multi-modal flight parameter data and perform standard deviation normalization on the data. Then, use an enhanced denoising autoencoder network for data dimensionality reduction and data cleaning. When the network reaches the maximum training iteration number, save the health state feature matrix output by the encoder module, and perform a sliding time window process on the health state feature matrix to obtain local features. Next, divide the above local features into training set local features and test set local features. Input the training set local features into the support vector data description model, and use the Bayesian optimization algorithm to optimize the weight parameters and relaxation variables of the Bayesian support vector data description model. After training, obtain the trained hypersphere model. Finally, input the test set local features into the trained hypersphere model to output the early weak warning result.

[0099] Figure 6 shows the early weak fault warning results obtained using the hypersphere model. It can be seen that the UAV goes from the healthy stage (#1 second to #20000 seconds) to the degradation stage (#20001 seconds to #29000 seconds), and then to the final fault occurrence (#30000 seconds). This solution effectively reduces the influence of noise through a feature extraction model based on an enhanced denoising autoencoder and obtains key information characterizing the health state of the device. Then, using the hypersphere model, it can keenly capture early weak abnormal signals deviating from the normal mode. Even in the case of limited training data, this solution has an accurate approximate boundary. New samples outside the decision boundary of the model are marked as the degradation state, and the distance from the normal state is characterized to identify early weak faults, ensuring a high warning accuracy rate and providing a reliable basis for UAV health monitoring. Through the warning results, maintenance personnel can more intuitively obtain potential fault information of the UAV and formulate reasonable maintenance decision measures to ensure the reliability and safety of the UAV.

Claims

1. An adaptive early fault warning method for unmanned aerial vehicles, characterized in that, It includes the following steps: Step 1: Collect multi-modal flight parameters during the flight of the UAV, normalize the multi-modal flight parameters to obtain the normalized UAV flight parameters, and use the trained autoencoder network to reduce the dimension and clean the data of the normalized UAV flight parameters to obtain a health state feature matrix that effectively characterizes the UAV flight state; Step 2: Construct a hypersphere model based on the health state feature matrix, complete the training, and use the radius of the trained hypersphere model as the warning threshold; Step 3: Use the trained hypersphere model to give an early warning of the incipient weak faults of the UAV; Process the multi-modal flight parameters of the UAV to be predicted according to the method in Step 1 to obtain the health state feature matrix of the UAV to be predicted, and then input the health state feature matrix of the UAV to be predicted into the trained hypersphere model. Calculate the distance between the health state feature matrix and the center of the hypersphere model. If the distance is less than or equal to the warning threshold, the UAV is considered to be in a normal state; if the distance is greater than the warning threshold, the UAV is considered to be in a potential fault state and a warning is triggered.

2. The adaptive early fault warning method for unmanned aerial vehicles according to claim 1, characterized in that, The step of collecting the multi-modal flight parameters during the flight of the UAV and normalizing the multi-modal flight parameters to obtain the normalized UAV flight parameters is as follows: Step S1: Collect multi-modal flight parameters during the flight of the drone , , is the th multi-modal flight parameter of the drone, is the number of multi-modal flight parameters; Step S2: Process the multi-modal flight parameters using the standard deviation normalization method to obtain the normalized UAV flight parameters; Normalized drone flight parameters are as follows: , is the th normalized flight parameter of the drone, , is the number of multimodal flight parameters; , is the mean of the -th multimodal flight parameter; the is the -th standard deviation of the multimodal flight parameter.

3. An early fault warning method for an adaptive unmanned aerial vehicle according to claim 1, characterized in that The step of constructing a hypersphere model based on the health state feature matrix, completing the training, and using the radius of the trained hypersphere model as the warning threshold is as follows: Step S4-1: Preprocess the health status feature matrix using a sliding window algorithm to obtain local features ; The process of the sliding window algorithm for processing the health status feature matrix is as follows: Set a sliding window on the health status feature matrix, and the size of the sliding window is , where ranges from , is the total sample size of the UAV flight parameters; the step size of the sliding window is greater than 1; Move the sliding window and extract the local features within each sliding window; the local features of the th sliding window are represented as , where ; is the size of the sliding window, and is the total sample size of the UAV flight parameters; Divide the local features into a training set and a test set ; Step S4-2: Use the training set Establish a hypersphere model; The step of establishing the hypersphere model is as follows: Step S4-2-1: The center of the hypersphere model is , the radius of the hypersphere model is , the input of the hypersphere model is the training , determine and 's objective function is: (2) wherein, is a slack variable; is a weight parameter, and the weight parameter controls the weight of the penalty term; Step S4-2-2: Use the training set Train the hypersphere model. When the number of training times reaches the preset maximum number of iterations, the hypersphere model at this time is trained and saved. The radius of the hypersphere model is used as the warning threshold.

4. An adaptive early fault warning method for an unmanned aerial vehicle according to claim 3, characterized in that, The step of optimizing the slack variable and weight parameter of the hypersphere model is as follows: Using the training set Train the hypersphere model. After each data in the training set is input into the hypersphere model, use the Bayesian optimization algorithm to optimize the weight parameters and the slack variables to obtain the optimized weight parameters and the optimized slack variables . Use the optimized weight parameters to update the weight parameters of the hypersphere model, and use the optimized slack variables to update the slack variables of the hypersphere model; then continue to input the next data in the training set until all data in the training set are traversed; using the Bayesian optimization algorithm to optimize the hypersphere model improves the prediction accuracy of the hypersphere model.

5. An early fault warning method for an adaptive unmanned aerial vehicle according to claim 1, characterized in that, The autoencoder network is an enhanced denoising autoencoder network; The enhanced denoising autoencoder network is a feedforward neural network FNN; the input of the enhanced denoising autoencoder network is the normalized UAV flight parameters and random noise ; the output of the enhanced denoising autoencoder network is the health state feature matrix ; the random noise is a matrix; the random noise has the same size as the dimension of the normalized UAV flight parameters; the elements of the random noise are random numbers between 0 and 1; The enhanced denoising autoencoder network includes an encoder and a decoder; The encoder includes a batch normalization (BN) layer, a ReLU activation function layer, a fully connected layer, and a residual connection layer; all layers of the encoder are fully connected; the input of the encoder is the normalized flight parameters of the drone and random noise ; the output of the encoder is a health state feature matrix ; where , , is the dimension of the features of the health state feature matrix of the drone after dimensionality reduction, is the th column health state vector of the health state feature matrix; The decoder includes a residual connection layer, a fully connected layer, a BN layer, and a ReLU activation function layer; all layers of the decoder are fully connected; the input of the decoder is the output of the encoder and the output of the decoder is the reconstructed and normalized UAV flight parameters ; The enhanced denoising autoencoder network uses the batch gradient descent method to update the weights and biases of the network, optimize the parameters, and minimize the value of the loss function; To improve the training efficiency of the enhanced denoising autoencoder network, an exponential decay strategy is used to dynamically adjust the learning rate of the enhanced denoising autoencoder network; the learning rate for the round of training of the enhanced denoising autoencoder network is as follows: ； Among them, is the initial learning rate, is the decay rate, used to control the decay speed, and its value is a positive constant value; Use the normalized UAV flight parameters to train the enhanced denoising autoencoder network. When the specified number of training rounds is reached, it is considered that the enhanced denoising autoencoder network is trained. The specified number of training rounds is 100 rounds.

6. The adaptive early fault warning method for unmanned aerial vehicles according to claim 5, wherein, The loss function of the enhanced denoising autoencoder network is as follows: ； Among them, represents the th input of the enhanced denoising autoencoder network, represents the th output of the enhanced denoising autoencoder network, is the total sample size of the UAV flight parameters.

7. A computer-readable storage medium storing a computer program therein; characterized in that, When the computer program is executed by the processor, it implements an adaptive UAV early fault warning method according to any one of claims 1-6.

Images