Unmanned aerial vehicle fault prediction method and system based on deep learning

Abstract

The invention relates to the technical field of unmanned aerial vehicle fault prediction, in particular to an unmanned aerial vehicle fault prediction method and system based on deep learning, and the method comprises the steps: collecting the multi-source heterogeneous sensor data of an unmanned aerial vehicle; performing preprocessing and feature engineering on the multi-source heterogeneous sensor data of the unmanned aerial vehicle to extract time sequence features representing the operation state of the unmanned aerial vehicle; training an unmanned aerial vehicle fault prediction model comprising a time sequence information processing unit and an attention mechanism based on the time sequence characteristics of the unmanned aerial vehicle operation state; and based on the trained unmanned aerial vehicle fault prediction model, processing the unmanned aerial vehicle operation state time sequence characteristics collected in real time, and outputting an unmanned aerial vehicle fault prediction result. According to the method, multi-source time sequence data are effectively fused through the deep learning model, high-precision real-time prediction of early and weak faults of the unmanned aerial vehicle is realized, feasibility of airborne deployment is ensured through a model lightweight technology, and safety and reliability of operation of the unmanned aerial vehicle are remarkably improved.

Description

Technical Field

[0001] The present invention relates to the technical field of drone fault prediction, and in particular to a drone fault prediction method and system based on deep learning. Background Art

[0002] In recent years, drones have demonstrated tremendous application potential and value in numerous fields, including military reconnaissance, geographic mapping, precision agriculture, power inspection, logistics, and even consumer entertainment. With the continuous expansion of application scenarios and the increasing complexity of missions, drone flight safety, operational reliability, and mission success have become crucial issues. However, currently used drone fault diagnosis and prediction methods still have some inherent limitations. For one thing, alarm mechanisms based on fixed thresholds or technologies that rely on simplified physical models for state judgment often struggle to capture early, subtle system anomalies. They are particularly inadequate for complex failure modes caused by the coupling of multiple factors, resulting in insufficient lead time for fault prediction and limited early warning capabilities. Furthermore, while approaches based on expert systems or rule bases can incorporate a certain amount of empirical knowledge, their construction is time-consuming and labor-intensive, and the rules themselves lack comprehensiveness and adaptability to unknown failure modes, limiting their generalization capabilities. Moreover, with the increase in the types and number of drone sensors, massive amounts of multi-source heterogeneous data are generated. Traditional signal processing and statistical analysis methods are inefficient in deeply mining the complex fault feature information contained in these high-dimensional, nonlinear, and time-series related data, making it difficult to fully utilize the value of the data. Summary of the Invention

[0003] The purpose of the present invention is to provide a UAV fault prediction method and system based on deep learning. By effectively integrating multi-source time series data through a deep learning model, high-precision real-time prediction of early and minor UAV faults is achieved, and the feasibility of airborne deployment is ensured through model lightweight technology, significantly improving the safety and reliability of UAV operation.

[0004] The present invention is achieved through the following technical solutions:

[0005] A UAV fault prediction method based on deep learning, the steps of the method include:

[0006] Collect multi-source heterogeneous sensor data from drones;

[0007] Preprocess and feature engineer the multi-source heterogeneous sensor data of drones to extract time series features that characterize the drone's operating status;

[0008] A drone fault prediction model consisting of a time series information processing unit and an attention mechanism is trained based on the time series features of the drone's operating status.

[0009] Based on the trained UAV fault prediction model, the real-time collected UAV operation status time series features are processed and the UAV fault prediction results are output.

[0010] Optionally, the collecting of multi-source heterogeneous sensor data from drones is specifically as follows:

[0011] Various types of sensors are deployed on the preset parts of the drone, including vibration sensors, attitude sensors, energy status sensors, position and altitude sensors;

[0012] Sampling the output signals of the plurality of types of sensors at a preset synchronous sampling frequency to obtain raw signal data;

[0013] The raw signal data is time-stamped and synchronized and aggregated via an airborne data bus to form multi-source heterogeneous sensor data of the UAV, which is represented as a raw state data set.

[0014] Optionally, preprocess the UAV's multi-source heterogeneous sensor data, which specifically includes:

[0015] Perform time alignment processing on the data stream in the original state data set, and use interpolation algorithm to fill in the data missing points to obtain a time-aligned data set;

[0016] Applying a filtering algorithm to the signals of the specified type of sensors in the time-aligned dataset to obtain a filtered dataset;

[0017] Abnormal data points in the filtered data set are detected and processed based on a statistical threshold to obtain a preprocessed data set.

[0018] Optionally, feature engineering can be performed on the multi-source heterogeneous sensor data of the drone, specifically:

[0019] Applying a time sliding window with preset parameters to the preprocessed data set to perform segmentation processing to obtain window data segments;

[0020] In each window data segment, the time domain statistical characteristics of the data signal are calculated;

[0021] Performing frequency domain transformation processing on the data signal in each window data segment and extracting the transformed spectrum features;

[0022] The time domain statistical features are combined with the frequency spectrum features to form an initial feature vector of the UAV.

[0023] Optionally, feature engineering is performed on the multi-source heterogeneous sensor data of the drone, including:

[0024] Performing time-frequency domain decomposition processing on the data signal within each window data segment and extracting the energy distribution characteristics after decomposition;

[0025] Adding the energy distribution feature to the initial feature vector of the UAV to form a second feature vector of the UAV;

[0026] The second eigenvector of the UAV is normalized to obtain the third eigenvector of the UAV, and a preset oversampling algorithm is applied to perform sample balancing on the third eigenvector of the UAV to obtain the time series characteristics of the UAV's operating status.

[0027] Optionally, the UAV fault prediction model specifically includes:

[0028] The input layer is used to receive the time series features of the UAV’s operating status;

[0029] Bidirectional long short-term memory network layer, used to process the time series characteristics of the UAV's operating status and output the time series processing results;

[0030] The attention mechanism layer receives the time series processing results and outputs the weighted fusion feature representation;

[0031] Dropout layer processes the feature representation after weighted fusion;

[0032] The output layer consists of a fully connected layer and a Softmax activation function, which is used to receive the processing results of the Dropout layer and output the drone fault prediction results.

[0033] Optionally, a drone fault prediction model including a time series information processing unit and an attention mechanism is trained based on the time series features of the drone's operating status. Specifically:

[0034] The time series features of the UAV's operating status are used as a dataset and divided into a training subset and a test subset;

[0035] The UAV fault prediction model is trained using the training subset to obtain an initial UAV fault prediction result;

[0036] The loss value between the test subset and the initial prediction result of the drone fault is calculated, and the drone fault prediction model is iteratively updated through the loss value until the maximum number of iterations is reached, completing the training of the drone fault prediction model.

[0037] Optionally, the loss value between the test subset and the initial prediction result of the drone failure is calculated as follows:

[0038]

[0039] Among them, L is the total loss, y ic Is the true category of sample i c, p icTo predict the probability that sample i belongs to category c, λ is the preset regularization strength coefficient and w is the weight.

[0040] Optionally, the output drone fault prediction result also includes early warning logic, which is specifically:

[0041] Receive the drone fault prediction results output by the drone fault prediction model, including the real-time predicted probability value of the drone fault type f;

[0042] Set the sliding time window T and the initial probability threshold within the window;

[0043] For each UAV fault type f, within the current time window T, calculate the number of time steps N that meet the real-time prediction probability value greater than the initial threshold of the probability within the window;

[0044] Set a minimum duration step threshold. When the time step is greater than or equal to the minimum duration step threshold, calculate the average probability of the time moments that meet the condition in the window as the confidence score.

[0045] Setting a confidence warning threshold, and comparing the calculated confidence score with the confidence warning threshold;

[0046] When the confidence score exceeds the confidence warning threshold, a warning trigger signal corresponding to the fault type f is generated;

[0047] According to the warning trigger signal, the operations of outputting warning information and recording warning events are executed.

[0048] The UAV fault prediction system based on deep learning includes:

[0049] The acquisition unit collects multi-source heterogeneous sensor data from drones;

[0050] The feature extraction unit preprocesses and performs feature engineering on the multi-source heterogeneous sensor data of the UAV to extract the time series features that characterize the UAV's operating status;

[0051] The model training unit trains a UAV fault prediction model based on the time series features of the UAV's operating status, which includes a time series information processing unit and an attention mechanism.

[0052] The prediction unit processes the real-time collected UAV operation status time series features based on the trained UAV fault prediction model and outputs the UAV fault prediction results.

[0053] The technical solution of the present invention has at least the following advantages and beneficial effects:

[0054] This paper leverages deep learning models to analyze multi-source time series data, significantly improving the accuracy and lead time of early-stage, subtle UAV fault prediction. By lightweighting the model and optimizing its deployment, the prediction system ensures efficient, real-time operation on the airborne platform, effectively enhancing the operational safety, reliability, and mission success rate of the UAV. BRIEF DESCRIPTION OF THE DRAWINGS

[0055] Figure 1 A schematic diagram of the flow chart of the UAV fault prediction method based on deep learning provided by the present invention;

[0056] Figure 2 Schematic diagram of the principle of the deep learning-based UAV fault prediction system provided by the present invention. DETAILED DESCRIPTION

[0057] To make the objectives, technical solutions, and advantages of the present invention more apparent, the following will provide a clear and complete description of the technical solutions of the present invention in conjunction with the accompanying drawings. It should be understood that the description is only a portion of the present invention, not all of it. The components of the present invention generally described and illustrated in the drawings herein may be arranged and designed in a variety of different configurations.

[0058] like Figure 1 As shown, the present invention provides one embodiment: a UAV fault prediction method based on deep learning, the method comprising the following steps:

[0059] Collect multi-source heterogeneous sensor data from drones;

[0060] Preprocess and feature engineer the multi-source heterogeneous sensor data of drones to extract time series features that characterize the drone's operating status;

[0061] A drone fault prediction model consisting of a time series information processing unit and an attention mechanism is trained based on the time series features of the drone's operating status.

[0062] Based on the trained UAV fault prediction model, the real-time collected UAV operation status time series features are processed and the UAV fault prediction results are output.

[0063] In implementation, this embodiment deploys various types of sensors in key locations within the drone's core components, such as the motor, battery management system, flight control unit, and fuselage structure. These sensors include an inertial measurement unit (IMU) containing an accelerometer and gyroscope, a GPS module, a barometer, current / voltage sensors, and a vibration sensor. All sensors acquire data at a preset synchronous sampling frequency of 200Hz, with timestamps aligned and aggregated via the onboard data bus to ensure a comprehensive and time-synchronized stream of raw operational status data. The raw data undergoes filtering to remove high-frequency noise, interpolation to fill missing values ​​due to signal loss, and the application of a statistical threshold-based 3σ criterion to identify and address outliers. By setting a sliding time window (as can be understood in this embodiment, which has a time window length of 1 second and a 50% overlap), the continuous data is segmented. Within each window, time series features characterizing the drone's operational status are calculated and extracted. These include time-domain statistical features such as mean, variance, and kurtosis, and frequency-domain features such as spectral peaks and energy distribution obtained through fast Fourier transforms. These features are then combined into a high-dimensional feature vector and normalized. Based on the time series features of the drone's operating status, a drone fault prediction model is trained, consisting of a time series information processing unit and an attention mechanism. The processed feature dataset is divided into training and test sets. The training set is used to iteratively update the model parameters through a backpropagation algorithm and optimizer. The attention mechanism helps the model focus on the most discriminative time steps or feature dimensions. During actual flight missions, the trained drone fault prediction model is used to process the drone's operating status time series features, collected in real time and generated through the same preprocessing and feature engineering process, and outputs the drone fault prediction results in real time. The model rapidly infers the input real-time feature sequence and outputs a probability vector indicating the likelihood of a fault type such as "motor failure," "battery anomaly," or "sensor failure" occurring at the current moment, providing a basis for subsequent warnings and decision-making.

[0064] Specifically, the acquisition of multi-source heterogeneous sensor data from drones is as follows:

[0065] Various types of sensors are deployed on the preset parts of the drone, including vibration sensors, attitude sensors, energy status sensors, position and altitude sensors;

[0066] Sampling the output signals of the plurality of types of sensors at a preset synchronous sampling frequency to obtain raw signal data;

[0067] The raw signal data is time-stamped and synchronized and aggregated via an airborne data bus to form multi-source heterogeneous sensor data of the UAV, which is represented as a raw state data set.

[0068] During implementation, to comprehensively and accurately obtain the operating status information of the drone during flight, this embodiment installs a vibration sensor on the motor housing or bracket of the power system to monitor abnormal vibrations. An attitude sensor is integrated near the center of gravity of the fuselage or on the flight control unit. In this embodiment, the attitude sensor is specifically an inertial measurement unit (IMU), which includes a gyroscope and accelerometer to capture attitude dynamics such as roll, pitch, and yaw. An energy status sensor is installed at the battery interface or power management module. The energy status sensor includes voltage and current sensors to monitor the health of the power source. A GPS receiver and barometer are used as position and altitude sensors to determine the spatial trajectory and altitude of the drone. After the sensors are deployed, the output signals of various types of sensors are collected at a preset synchronous sampling frequency to ensure that the relevant dynamic information can be captured without distortion, thereby obtaining a series of raw signal data representing the changes of various physical quantities over time. The raw signal data is then time-stamped and synchronized and aggregated via the onboard data bus to form a structured drone multi-source heterogeneous sensor data set.

[0069] Furthermore, the multi-source heterogeneous sensor data of the UAV is preprocessed, which specifically includes:

[0070] Perform time alignment processing on the data stream in the original state data set, and use interpolation algorithm to fill in the data missing points to obtain a time-aligned data set;

[0071] Applying a filtering algorithm to the signals of the specified type of sensors in the time-aligned dataset to obtain a filtered dataset;

[0072] Abnormal data points in the filtered data set are detected and processed based on a statistical threshold to obtain a preprocessed data set.

[0073] Specifically, after acquiring multi-source heterogeneous sensor data from a drone, represented as a raw state dataset, this embodiment performs time alignment on the data streams in the raw state dataset to ensure data consistency and quality. Because different sensors may have slight startup delays, varying internal sampling clock accuracies, or differences in data transmission paths, the raw data may not be perfectly aligned on the time axis. Therefore, a unified time base is selected, and all data points are aligned to the closest common timestamp grid. The timestamps of each data stream are adjusted to ensure consistency at any point in time. During this process, an interpolation algorithm is used to fill in missing data points to ensure the continuity of the data sequence, thereby obtaining a time-aligned dataset. To mitigate the effects of sensor noise or environmental interference on signal quality, especially for sensor signals sensitive to high-frequency noise, such as acceleration and angular velocity signals from an IMU, filtering algorithms are applied to the signals of the specified sensor types in the time-aligned dataset. Common filtering methods include Kalman filtering, Butterworth low-pass filtering or moving average filtering. After processing, a filtered data set is obtained. The 3σ criterion is used to calculate the mean and standard deviation of the data segment. Data points that exceed the preset threshold range are identified as outliers. Depending on the specific situation, they are removed or replaced with the mean or median of the neighboring points to obtain the preprocessed data set.

[0074] In this embodiment, feature engineering is performed on the multi-source heterogeneous sensor data of the drone, specifically:

[0075] Applying a time sliding window with preset parameters to the preprocessed data set to perform segmentation processing to obtain window data segments;

[0076] In each window data segment, the time domain statistical characteristics of the data signal are calculated;

[0077] Performing frequency domain transformation processing on the data signal in each window data segment and extracting the transformed spectrum features;

[0078] The time domain statistical features are combined with the frequency spectrum features to form an initial feature vector of the UAV.

[0079] Also includes:

[0080] Performing time-frequency domain decomposition processing on the data signal within each window data segment and extracting the energy distribution characteristics after decomposition;

[0081] Adding the energy distribution feature to the initial feature vector of the UAV to form a second feature vector of the UAV;

[0082] The second eigenvector of the UAV is normalized to obtain the third eigenvector of the UAV, and a preset oversampling algorithm is applied to perform sample balancing on the third eigenvector of the UAV to obtain the time series characteristics of the UAV's operating status.

[0083] During implementation, this embodiment applies a time sliding window with preset parameters to the preprocessed data set, segmenting it into a series of fixed-length window segments. Within each window segment, the data signal's time-domain statistical features, including mean, variance, kurtosis, peak-to-peak value, and RMS value, are calculated to capture the signal's overall amplitude, fluctuations, and morphology. Frequency domain transformation is also performed on the data signal within each window segment, extracting spectral features such as dominant frequency, energy in each frequency band, and spectral entropy to reveal the signal's frequency domain distribution. The time-domain statistical features are combined with the spectral features to form the initial UAV feature vector. To further capture the non-stationary nature of the signal, which varies simultaneously in both time and frequency, this embodiment also performs time-frequency decomposition on the data signal within each window segment, decomposing the signal into multiple frequency bands. The energy distribution characteristics of the decomposed data are then extracted, and the energy value or energy contribution of each wavelet packet node is calculated. The energy distribution features are added to the initial drone feature vector, and through vector concatenation, a higher-dimensional, more informative second drone feature vector is formed. Because the dimensions and numerical ranges of different features vary significantly, to eliminate the impact of these differences on subsequent model training, this embodiment performs normalization on the drone's second feature vector, ensuring that all features have zero mean and unit variance, resulting in the drone's third feature vector. It is understood that this embodiment also applies a preset oversampling algorithm, such as SMOTE, to perform sample balancing on the drone's third feature vector, ensuring a relatively balanced number of samples across different categories.

[0084] In the specific application of this embodiment, the UAV fault prediction model specifically includes:

[0085] The input layer is used to receive the time series features of the UAV’s operating status;

[0086] Bidirectional long short-term memory network layer, used to process the time series characteristics of the UAV's operating status and output the time series processing results;

[0087] The attention mechanism layer receives the time series processing results and outputs the weighted fusion feature representation;

[0088] Dropout layer processes the feature representation after weighted fusion;

[0089] The output layer consists of a fully connected layer and a Softmax activation function, which is used to receive the processing results of the Dropout layer and output the drone fault prediction results.

[0090] Optionally, a drone fault prediction model including a time series information processing unit and an attention mechanism is trained based on the time series features of the drone's operating status. Specifically:

[0091] The time series features of the UAV's operating status are used as a dataset and divided into a training subset and a test subset;

[0092] The UAV fault prediction model is trained using the training subset to obtain an initial UAV fault prediction result;

[0093] The loss value between the test subset and the initial prediction result of the drone fault is calculated, and the drone fault prediction model is iteratively updated through the loss value until the maximum number of iterations is reached, completing the training of the drone fault prediction model.

[0094] During implementation, this embodiment combines the resulting time-series feature dataset of the drone's operating status into a complete dataset and divides it into a training subset and a test subset according to a predetermined ratio. The initialized drone fault prediction model is trained using the training subset. During training, the model receives training samples, performs forward propagation to obtain a predicted output, and then calculates the loss between the model's predicted output and the true label of the training sample. This loss is then used to calculate the gradient using the backpropagation algorithm. The weights and bias parameters within the drone fault prediction model are then iteratively updated using the Adam optimizer. This iterative update process continues until the maximum number of iterations is met, completing the training of the drone fault prediction model.

[0095] Specifically, the loss between the model's predicted output and the true label of the training sample is calculated as follows:

[0096]

[0097] Among them, L is the total loss, y ic Is the true category of sample i c, p ic To predict the probability that sample i belongs to category c, λ is the preset regularization strength coefficient and w is the weight.

[0098] In further applications, the output of drone fault prediction results also includes early warning logic, which is specifically:

[0099] Receive the drone fault prediction results output by the drone fault prediction model, including the real-time predicted probability value of the drone fault type f;

[0100] Set the sliding time window T and the initial probability threshold within the window;

[0101] For each UAV fault type f, within the current time window T, calculate the number of time steps N that meet the real-time prediction probability value greater than the initial threshold of the probability within the window;

[0102] Set a minimum duration step threshold. When the time step is greater than or equal to the minimum duration step threshold, calculate the average probability of the time moments that meet the condition in the window as the confidence score.

[0103] Setting a confidence warning threshold, and comparing the calculated confidence score with the confidence warning threshold;

[0104] When the confidence score exceeds the confidence warning threshold, a warning trigger signal corresponding to the fault type f is generated;

[0105] According to the warning trigger signal, the operations of outputting warning information and recording warning events are executed.

[0106] After the drone fault prediction model is trained and deployed in actual flight missions, this embodiment also includes an early warning process. First, the drone fault prediction results output by the drone fault prediction model are received. The drone fault prediction results represent the real-time predicted probability of drone fault type f, such as motor 1 overheating, battery voltage too low, or GPS signal loss. To smooth instantaneous fluctuations and confirm the persistence of the fault state, a time window-based judgment mechanism is introduced. Specifically, a 5-second sliding time window T and a preliminary threshold for probability within the window are set, here 0.7. For each potential drone fault type f, the number of time steps N within the current most recent time window T that meet the real-time predicted probability value greater than the preliminary threshold for probability within the window is calculated in real time. To further eliminate short-term, non-persistent probability spikes, a minimum duration step threshold N_min is set (e.g., corresponding to 60% of the length of window T). Only when the calculated number of time steps is greater than or equal to this minimum duration step threshold is the fault type f considered to have exhibited significant and persistent abnormal signs within window T. At this time, calculate the average probability of all moments in the window where the real-time predicted probability values ​​that meet the conditions are greater than the preliminary probability threshold in the window, and use this average value as the confidence score for the occurrence of the fault. Set a confidence warning threshold, which is 0.85 here, and compare the calculated confidence score with the confidence warning threshold. When the confidence score exceeds the confidence warning threshold, it is determined that the fault type f is very likely to occur or has occurred, and generate a warning trigger signal corresponding to the fault type f. According to the warning trigger signal, the warning information is automatically output. It can be understood that the warning information includes displaying a warning through the ground station interface, issuing a sound alarm, or sending a notification to the operator, and storing key information such as the fault type, occurrence time, and confidence score in the flight log, so that the operator can take timely countermeasures and use it for subsequent analysis.

[0107] like Figure 2As shown, the present invention also provides another embodiment: a UAV fault prediction system based on deep learning, comprising:

[0108] The acquisition unit collects multi-source heterogeneous sensor data from drones;

[0109] The feature extraction unit preprocesses and performs feature engineering on the multi-source heterogeneous sensor data of the UAV to extract the time series features that characterize the UAV's operating status;

[0110] The model training unit trains a UAV fault prediction model based on the time series features of the UAV's operating status, which includes a time series information processing unit and an attention mechanism.

[0111] The prediction unit processes the real-time collected UAV operation status time series features based on the trained UAV fault prediction model and outputs the UAV fault prediction results.

[0112] The above is merely a preferred embodiment of the present invention and is not intended to limit the present invention. Those skilled in the art will readily appreciate that the present invention is susceptible to various modifications and variations. Any modifications, equivalent substitutions, or improvements made within the spirit and principles of the present invention shall be included within the scope of protection of the present invention.

Claims

1. A UAV fault prediction method based on deep learning, characterized by: The steps of the method include: Collect multi-source heterogeneous sensor data from drones; Preprocess and feature engineer the multi-source heterogeneous sensor data of drones to extract time series features that characterize the drone's operating status; A drone fault prediction model consisting of a time series information processing unit and an attention mechanism is trained based on the time series features of the drone's operating status. Based on the trained UAV fault prediction model, the real-time collected UAV operation status time series features are processed and the UAV fault prediction results are output.

2. The UAV fault prediction method based on deep learning according to claim 1 is characterized in that: The acquisition of multi-source heterogeneous sensor data from drones is specifically as follows: Various types of sensors are deployed on the preset parts of the drone, including vibration sensors, attitude sensors, energy status sensors, position and altitude sensors; Sampling the output signals of the plurality of types of sensors at a preset synchronous sampling frequency to obtain raw signal data; The raw signal data is time-stamped and synchronized and aggregated via an airborne data bus to form multi-source heterogeneous sensor data of the UAV, which is represented as a raw state data set.

3. The UAV fault prediction method based on deep learning according to claim 2, characterized in that: Preprocessing of multi-source heterogeneous sensor data from drones includes: Perform time alignment processing on the data stream in the original state data set, and use interpolation algorithm to fill in the data missing points to obtain a time-aligned data set; Applying a filtering algorithm to the signals of the specified type of sensors in the time-aligned dataset to obtain a filtered dataset; Abnormal data points in the filtered data set are detected and processed based on a statistical threshold to obtain a preprocessed data set.

4. The UAV fault prediction method based on deep learning according to claim 3 is characterized in that: Feature engineering is performed on multi-source heterogeneous sensor data from drones, specifically: Applying a time sliding window with preset parameters to the preprocessed data set to perform segmentation processing to obtain window data segments; In each window data segment, the time domain statistical characteristics of the data signal are calculated; Performing frequency domain transformation processing on the data signal in each window data segment and extracting the transformed spectrum features; The time domain statistical features are combined with the frequency spectrum features to form an initial feature vector of the UAV.

5. The UAV fault prediction method based on deep learning according to claim 4 is characterized in that: Feature engineering for multi-source heterogeneous sensor data from drones, including: Performing time-frequency domain decomposition processing on the data signal within each window data segment and extracting the energy distribution characteristics after decomposition; Adding the energy distribution feature to the initial feature vector of the UAV to form a second feature vector of the UAV; The second eigenvector of the UAV is normalized to obtain the third eigenvector of the UAV, and a preset oversampling algorithm is applied to perform sample balancing on the third eigenvector of the UAV to obtain the time series characteristics of the UAV's operating status.

6. The UAV fault prediction method based on deep learning according to claim 5 is characterized in that: The UAV fault prediction model specifically includes: The input layer is used to receive the time series features of the UAV’s operating status; Bidirectional long short-term memory network layer, used to process the time series characteristics of the UAV's operating status and output the time series processing results; The attention mechanism layer receives the time series processing results and outputs the weighted fusion feature representation; Dropout layer processes the feature representation after weighted fusion; The output layer consists of a fully connected layer and a Softmax activation function, which is used to receive the processing results of the Dropout layer and output the drone fault prediction results.

7. The UAV fault prediction method based on deep learning according to claim 6, characterized in that: Based on the time series features of the UAV's operating status, a UAV fault prediction model including a time series information processing unit and an attention mechanism is trained. Specifically: The time series features of the UAV's operating status are used as a dataset and divided into a training subset and a test subset; The UAV fault prediction model is trained using the training subset to obtain an initial UAV fault prediction result; The loss value between the test subset and the initial prediction result of the drone fault is calculated, and the drone fault prediction model is iteratively updated through the loss value until the maximum number of iterations is reached, completing the training of the drone fault prediction model.

8. The UAV fault prediction method based on deep learning according to claim 7 is characterized in that: The loss value between the calculated test subset and the initial prediction result of the UAV failure is specifically: Among them, L is the total loss, y ic Is the true category of sample i c, p ic To predict the probability that sample i belongs to category c, λ is the preset regularization strength coefficient and w is the weight.

9. The UAV fault prediction method based on deep learning according to claim 8, characterized in that: The output of the UAV fault prediction result also includes early warning logic, which is specifically: Receive the drone fault prediction results output by the drone fault prediction model, including the real-time predicted probability value of the drone fault type f; Set the sliding time window T and the initial probability threshold within the window; For each UAV fault type f, within the current time window T, calculate the number of time steps N that meet the real-time prediction probability value greater than the initial threshold of the probability within the window; Set a minimum duration step threshold. When the time step is greater than or equal to the minimum duration step threshold, calculate the average probability of the time moments that meet the condition in the window as the confidence score. Setting a confidence warning threshold, and comparing the calculated confidence score with the confidence warning threshold; When the confidence score exceeds the confidence warning threshold, a warning trigger signal corresponding to the fault type f is generated; According to the warning trigger signal, the operations of outputting warning information and recording warning events are executed.

10. UAV fault prediction system based on deep learning, characterized by: include: The acquisition unit collects multi-source heterogeneous sensor data from drones; The feature extraction unit preprocesses and performs feature engineering on the multi-source heterogeneous sensor data of the UAV to extract the time series features that characterize the UAV's operating status; The model training unit trains a UAV fault prediction model based on the time series features of the UAV's operating status, which includes a time series information processing unit and an attention mechanism. The prediction unit processes the real-time collected UAV operation status time series features based on the trained UAV fault prediction model and outputs the UAV fault prediction results.

Images