A data-driven real-time defect prediction system and method for unmanned aerial vehicle flight control systems

Abstract

The application relates to the technical field of unmanned aerial vehicle control, in particular to a data-driven real-time defect prediction system and method for an unmanned aerial vehicle flight control system, and the application proposes the following scheme: the system comprises a data acquisition module, a sensor verification module, a defect prediction module, a credibility analysis module and a flight control optimization module. The sensor verification module identifies sensor data anomalies and distinguifies fault types through data consistency, timing anomaly detection and environment correlation analysis. The defect prediction module analyzes the verified data by using machine learning to predict potential faults. The credibility analysis module verifies the prediction results and generates a credibility score. The flight control optimization module selects an optimized control strategy according to the credibility score, adjusts flight parameters or executes safe operations such as returning and landing. The application improves the fault identification and prediction capability of the unmanned aerial vehicle in a complex environment, and enhances the flight safety and system reliability.

Description

Technical Field

[0001] This invention relates to the field of unmanned aerial vehicle (UAV) control technology, and in particular to a data-driven real-time defect prediction system and method for UAV flight control systems. Background Technology

[0002] Flight Control System (FCS) is one of the core technologies for enabling autonomous flight, mission execution, and safety assurance of unmanned aerial vehicles (UAVs). During flight, UAVs are susceptible to various malfunctions due to complex environmental factors and hardware limitations, which can lead to flight interruptions or even crashes. Therefore, how to predict potential malfunctions in the UAV flight control system in real time and improve its safety and stability has become an urgent problem to be solved in the field of UAV technology.

[0003] For example, Chinese patent CN116954197A addresses the problem that existing technologies cannot fully extract the mapping relationship between parameters and faults, leading to inaccurate defect prediction. First, it uses a whale optimization algorithm to generate a large amount of sample data, ensuring high coverage of flight control system defects. Second, it extracts relevant parameters through parameter combination extraction methods. Third, it uses a feature extraction method based on convolutional operations to extract the mapping relationship between relevant parameters and faults. Finally, it uses a convolutional neural network to establish a model of flight control sensor monitoring parameters and flight control system defects. Predicting flight control system defects using this model achieves very high prediction accuracy.

[0004] The above patents have the problems mentioned in the background: the current technology lacks an effective real-time verification mechanism for sensor data, making it difficult to quickly identify sensor faults and abnormal data, thereby affecting the accuracy of the system's judgment of flight status. In order to solve the above problems, this application designs a data-driven real-time defect prediction system and method for UAV flight control system. Summary of the Invention

[0005] The technical problem this invention aims to solve is to address the shortcomings of existing technologies by providing a data-driven real-time defect prediction system and method for UAV flight control systems. The system includes a data acquisition module, a sensor verification module, a defect prediction module, a reliability analysis module, and a flight control optimization module. The sensor verification module identifies sensor data anomalies and distinguishes fault types through data consistency, time-series anomaly detection, and environmental correlation analysis. The defect prediction module uses machine learning to analyze the verified data and predict potential faults. The reliability analysis module verifies the prediction results and generates a reliability score. The flight control optimization module selects an optimized control strategy based on the reliability score, adjusting flight parameters or performing safety operations such as return-to-home or landing. This invention improves the fault identification and prediction capabilities of UAVs in complex environments, enhancing flight safety and system reliability.

[0006] To achieve the above objectives, the present invention provides the following technical solution:

[0007] A data-driven real-time defect prediction system for UAV flight control system, the system comprising a data acquisition module, a sensor verification module, a defect prediction module, a reliability analysis module, and a flight control optimization module;

[0008] The sensor verification module is equipped with a sensor fault detection strategy, which is used to detect the output of the data acquisition module and determine whether the sensor is faulty.

[0009] The defect prediction module is configured with a data processing strategy, which is used to process verified sensor data, build a defect prediction model through machine learning algorithms, and predict potential failures of the UAV in future flights based on the defect prediction model.

[0010] The credibility analysis module is used to evaluate the credibility of the prediction results output by the defect prediction module and generate a credibility score.

[0011] The sensor verification module is used to identify anomalies from sensor data and distinguish whether there is a sensor malfunction, specifically including:

[0012] The data consistency check unit is used to perform consistency checks on sensor data. By analyzing the data change trends and characteristics over different time periods, it identifies whether there are any abnormal signals.

[0013] The timing anomaly detection unit is used to analyze the timing patterns of sensor data with abnormal signals and determine whether the sensor data conforms to the normal timing pattern.

[0014] The environmental correlation unit is used to perform correlation analysis between sensor data that does not conform to the normal timing pattern and external environmental data to determine whether the abnormal sensor data is related to environmental factors. If they are not related, it indicates that the sensor is faulty.

[0015] The sensor fault detection strategy includes data integrity check logic, time series stability analysis logic, and environmental correlation analysis logic.

[0016] The data integrity check logic is configured within the data consistency check unit, the time series stability analysis logic is configured within the time series anomaly detection unit, and the environment correlation analysis logic is configured within the environment correlation unit.

[0017] The time series stability analysis logic is used to detect the stability of data through a time series model and determine whether the data conforms to a normal time series pattern. Specifically, it includes:

[0018] A hybrid time series model is constructed, and the sensor data is processed through the hybrid time series model to output short-term anomaly scores and trend anomaly scores;

[0019] Based on the flight status of the drone, a weighting coefficient is set, and the short-term anomaly score and the trend anomaly score are weighted and calculated to generate a comprehensive anomaly score.

[0020] The comprehensive anomaly score is compared with a risk threshold. If it is greater than or equal to the risk threshold, the sensor corresponding to the sensor data is marked as a faulty sensor. If it is less than the risk threshold, the sensor corresponding to the sensor data is marked as a normal sensor.

[0021] The hybrid time series model includes:

[0022] The short-term detection layer is used to merge sensor data into a linear sequence. It performs autoregressive prediction on the linear sequence within the window through a sliding window to obtain the next data point within the prediction window. By calculating the residual between the predicted value and the actual value, it obtains the short-term anomaly score within each window.

[0023] The long-term detection layer is used to learn the historical sequence of sensor data through a bidirectional long short-term memory neural network, obtain long-term predicted values, calculate the deviation between the actual value of sensor data and the long-term predicted value, and gradually accumulate the deviation value to generate a trend anomaly score.

[0024] The self-learning update layer is used to adjust parameters based on the current environment and historical anomaly patterns.

[0025] The environmental correlation analysis logic is used to determine whether abnormal data is affected by the external environment, specifically including:

[0026] Principal component analysis is used to extract environmental features from external environmental data.

[0027] An environmental correlation model is constructed based on historical flight data and historical environmental parameters;

[0028] The environmental features are used as input parameters for the environmental association model, and the environmental association model is used to train the environmental features to output associated data.

[0029] Calculate the residual between the correlated data and the sensor data that does not conform to the normal timing pattern, perform fluctuation analysis on the residual, and determine whether it is within the preset environmental influence range. If the residual fluctuation is not within the range, it indicates that the sensor is faulty. If the residual fluctuation is within the range, it indicates that the abnormality of the sensor data is related to the environment and there is no fault.

[0030] The defect prediction module is used to analyze normal sensor data confirmed by the sensor verification module to predict potential faults of the UAV during flight, specifically including:

[0031] The data preprocessing unit is used to perform data cleaning, noise reduction, smoothing and normalization on the validated sensor data;

[0032] The feature extraction unit is used to extract feature variables from the preprocessed data and upgrade the extracted feature variables into high-dimensional feature vectors.

[0033] The defect prediction unit is used to build and train the defect prediction model, analyze the feature vectors generated by the feature extraction unit, and predict the failures of the UAV in future flights.

[0034] The data processing strategy includes feature extraction logic and defect prediction logic. The defect prediction logic is used to train the high-dimensional feature vector according to the defect prediction model and output the prediction result.

[0035] The feature extraction logic is configured within the feature extraction unit, and the defect prediction logic is configured within the defect prediction unit.

[0036] The feature extraction logic includes:

[0037] Sensor data is decomposed into short-time frequency signals, and the frequency components within each time period are calculated using Fourier transform to extract instantaneous change features.

[0038] The sensor signal is decomposed into multiple wavelet coefficient levels based on wavelet transform, and the multi-scale variation features of each wavelet coefficient level are extracted through wavelet coefficients.

[0039] The instantaneous change features and the multi-scale change features are aggregated to form a feature vector, and the feature vector is upgraded to a high-dimensional feature space through nonlinear mapping.

[0040] The defect prediction model includes:

[0041] Convolutional feature extraction layer is used to extract spatial and temporal features from high-dimensional feature vectors;

[0042] The time series pattern learning layer is used to learn time series patterns from high-dimensional feature vectors.

[0043] The convolutional pooling layer is used to convolve the spatial features and temporal features with the time series pattern using a 3×3 convolutional kernel, and obtain a 2×2 mapping feature through the ReLU activation function and max pooling operation;

[0044] The output layer is used to flatten the mapped features into a feature vector, perform multi-class classification on the feature vector through a multi-layer perceptron, generate a probability distribution for each potential fault type, generate a confidence score for each fault type, sort the faults according to the confidence scores, and take the fault type corresponding to the highest confidence score as the prediction result.

[0045] The credibility analysis module includes:

[0046] The result verification unit is used to verify the output results of the defect prediction module by comparing them with historical data, and output the verification results.

[0047] The anomaly stability analysis unit is used to assess the stability of anomalies in the prediction results and output the analysis results.

[0048] A credibility scoring unit is used to calculate the credibility score output by the defect prediction module based on the verification results and the analysis results.

[0049] The flight control optimization module is used to evaluate the reliability score and optimize the control of potential faults, specifically including:

[0050] The risk level assessment unit is used to assess the risk level of the credibility score, specifically including the severity of the fault and flight feasibility.

[0051] The control strategy decision unit is used to select the corresponding optimized control strategy based on the output of the risk level assessment unit.

[0052] The parameter adjustment unit is used to adjust the parameters of the flight control system according to the optimized control strategy output by the control strategy decision unit.

[0053] A data-driven method for real-time defect prediction in a UAV flight control system, comprising the following specific steps:

[0054] S1: Acquire sensor data from the various sensors of the drone;

[0055] S2: Verify the collected sensor data, identify abnormal data, and determine whether there is a sensor malfunction;

[0056] S3: Preprocess and extract features from sensor data that does not have sensor faults to generate high-dimensional feature vectors;

[0057] S4: Construct a defect prediction model, using the high-dimensional feature vector as the input parameter of the defect prediction model, and process the input parameter through the defect prediction model to output the prediction result;

[0058] S5: Evaluate the credibility of the prediction results and generate a credibility score;

[0059] S6: Optimize and adjust the flight control of the UAV based on the credibility score.

[0060] Compared with the prior art, the beneficial effects of the present invention are:

[0061] This invention enables the effective identification and prediction of potential faults during UAV flight, improving the accuracy of data anomaly detection and fault type differentiation. By utilizing machine learning algorithms to analyze sensor data, potential faults can be identified in advance, supporting timely preventative measures. Attached Figure Description

[0062] Other features, objects, and advantages of the invention will become more apparent from the following detailed description of non-limiting embodiments with reference to the accompanying drawings:

[0063] Figure 1 This is a flowchart illustrating a data-driven real-time defect prediction method for a UAV flight control system according to Embodiment 1 of the present invention.

[0064] Figure 2 This is a structural diagram of the hybrid timing model of Embodiment 1 of the present invention;

[0065] Figure 3 This is a structural diagram of the defect prediction model in Embodiment 1 of the present invention;

[0066] Figure 4 This is a block diagram of a data-driven real-time defect prediction system for a UAV flight control system according to Embodiment 2 of the present invention. Detailed Implementation

[0067] The technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings of the embodiments of the present invention. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments.

[0068] Example 1:

[0069] Please see Figure 1 The present invention provides an embodiment of a data-driven real-time defect prediction method for a UAV flight control system, the specific steps of which are as follows:

[0070] S1: Acquire sensor data from the various sensors of the drone;

[0071] In this step, sensor data is acquired through the sensor. The sensor data includes:

[0072] Attitude data (such as pitch and roll angles) are used to monitor the current flight attitude of the UAV;

[0073] Velocity data (horizontal velocity, vertical velocity) is used to assess the drone's motion status;

[0074] Location data (such as GPS coordinates) is used to determine the geographical location of the drone;

[0075] Environmental data (such as temperature and humidity) are used to analyze the impact of the external environment on the operation of drones;

[0076] The frequency of sensor data acquisition is dynamically adjusted according to the drone's flight mode (such as high-speed flight or hovering) to ensure the real-time nature and effectiveness of the data.

[0077] S2: Verify the collected sensor data, identify abnormal data, and determine whether there is a sensor malfunction;

[0078] S3: Preprocess and extract features from sensor data that does not have sensor faults to generate high-dimensional feature vectors;

[0079] S4: Construct a defect prediction model, using the high-dimensional feature vector as the input parameter of the defect prediction model, and process the input parameter through the defect prediction model to output the prediction result;

[0080] S5: Evaluate the credibility of the prediction results and generate a credibility score;

[0081] S6: Optimize and adjust the flight control of the UAV based on the credibility score.

[0082] The specific steps of S2 are as follows:

[0083] S2.1: Analyze the continuity and consistency of sensor data. By comparing adjacent data points within the same time period, determine whether the data change trend is reasonable. For example, if the pitch angle suddenly fluctuates abnormally, mark the data as a potential anomaly.

[0084] Continuity and consistency analysis is performed on the data from each sensor to ensure that data changes conform to reasonable physical laws. This step examines the trend of data changes by comparing adjacent data points within the same time period. For example, during normal flight, the sensor output should be stable and have a certain continuity, especially attitude data such as pitch and roll angles.

[0085] S2.2: For data points marked as anomalous, further time series pattern analysis is performed to determine whether they conform to normal time series characteristics;

[0086] Time series analysis can capture the temporal dependence and pattern characteristics of sensor data, identifying data that deviates significantly from normal patterns at different time points. This method can effectively avoid misjudging short-term abnormal fluctuations as faults, while ensuring data consistency over time.

[0087] S2.3: Perform correlation analysis between sensor data that does not conform to the normal timing pattern and external environmental data to determine whether the abnormal sensor data is related to environmental factors. If they are not related, it indicates that the sensor is faulty.

[0088] Environmental correlation analysis can effectively distinguish between data anomalies caused by environmental changes and actual sensor malfunctions, thereby improving the accuracy of fault detection. For example, the data from a sensor may be too high in a high-temperature environment, but if this deviation is highly correlated with changes in temperature data, it can be determined that it is caused by environmental factors, thus avoiding misjudgment.

[0089] By constructing a hybrid time-series model, short-term and trend-based anomaly scores are comprehensively considered, and the weighting coefficients are dynamically adjusted according to the UAV's flight status to generate a comprehensive anomaly score. By comparing the comprehensive anomaly score with a risk threshold, normal fluctuations in sensor data and potential faults can be effectively distinguished, improving the sensitivity and accuracy of fault detection.

[0090] The specific steps of S2.2 are as follows:

[0091] S2.2.1: Construct a hybrid time series model, process the sensor data through the hybrid time series model, and output short-term anomaly scores and trend anomaly scores;

[0092] The short-term anomaly score is defined as the residual calculated within the sliding window. A larger residual between the actual and predicted values ​​results in a higher short-term anomaly score, indicating transient fluctuations. The trend anomaly score is calculated based on the long-term sequence prediction error of the Bi-LSTM, using accumulated residuals. An increase in the trend anomaly score indicates that the data deviates from the normal pattern over a longer period, making it suitable for detecting gradually occurring faults.

[0093] By decomposing sensor data into short-term and trend anomaly scores, different types of anomalies can be captured more accurately. For example, short-term anomaly scores can reflect transient sensor errors, while trend anomaly scores can reveal persistent deviations in the sensor's behavior. This approach can effectively improve the system's sensitivity and robustness, enabling earlier detection of potential faults.

[0094] S2.2.2: Set weighting coefficients based on the UAV's flight status, and perform a weighted calculation on the short-term anomaly score and the trend anomaly score to generate a comprehensive anomaly score;

[0095] The calculation of the comprehensive anomaly score needs to consider the current flight state of the drone to reasonably allocate the weights of short-term anomaly scores and trend anomaly scores. The fluctuation characteristics of sensor data may differ under different flight states of the drone. For example, short-term fluctuations may be more significant during high-speed flight; while long-term trend anomalies may be more indicative during hovering. The weight values ​​should be dynamically adjusted according to the current flight state of the drone (such as high-speed flight, low-speed flight, hovering, etc.).

[0096] By dynamically adjusting the weighting coefficients based on flight status, the system can reasonably assess anomaly risks under different flight modes. For example, during high-speed flight, short-term fluctuations increase, and the system can increase the weight of short-term anomaly scores to more sensitively detect transient anomalies; during stable hovering, trend anomaly scores have a higher weight, which can capture chronic fault signals earlier.

[0097] S2.2.3: Compare the comprehensive anomaly score with the risk threshold. If it is greater than or equal to the risk threshold, mark the sensor corresponding to the sensor data as a faulty sensor. If it is less than the risk threshold, mark the sensor corresponding to the sensor data as a normal sensor.

[0098] Risk thresholds are determined through analysis of historical data and statistical methods. Lower thresholds are set for critical sensors (such as attitude sensors), while relatively higher thresholds are set for non-critical sensors (such as environmental sensors). By setting risk thresholds, it can be ensured that only high-risk anomalies will trigger fault flags, avoiding false alarms caused by small fluctuations. This fault determination method helps to identify potential faults in a timely manner during UAV operation and to take timely flight control measures to ensure flight safety.

[0099] Please see Figure 2 This invention presents a hybrid time-series model structure diagram. The hybrid time-series model is used for multi-level time-series analysis of UAV sensor data to identify short-term fluctuation anomalies and long-term trend anomalies. The model includes a short-term detection layer, a long-term detection layer, and a self-learning update layer, improving the accuracy and sensitivity of fault detection through multi-layer processing. The hybrid time-series model includes:

[0100] The short-term detection layer is used to merge sensor data into a linear sequence. It performs autoregressive prediction on the linear sequence within the window through a sliding window to obtain the next data point within the prediction window. By calculating the residual between the predicted value and the actual value, it obtains the short-term anomaly score within each window.

[0101] The size of the sliding window dynamically changes according to the drone's flight state (such as high-speed flight, low-speed cruise, hovering, etc.). During high-speed flight, the window width can be appropriately reduced to improve sensitivity to instantaneous fluctuations; during hovering, the window can be appropriately enlarged to reduce interference from short-term fluctuations.

[0102] By employing an adaptive sliding window and autoregressive prediction, the short-term detection layer can effectively capture transient anomalies in sensor data, especially during high-speed flight, enabling rapid response to short-term fluctuations in the data. This method avoids the errors caused by a fixed window size in different flight modes, thus improving detection sensitivity.

[0103] The long-term detection layer is used to learn the historical sequence of sensor data through a bidirectional long short-term memory neural network, obtain long-term predicted values, calculate the deviation between the actual value of sensor data and the long-term predicted value, and gradually accumulate the deviation value to generate a trend anomaly score, which is suitable for identifying slowly developing faults.

[0104] Bidirectional LSTM networks can simultaneously consider the dependencies between time series data, thus capturing long-term trends more accurately during prediction. A bidirectional long short-term memory neural network model is trained using historical data to generate data predictions for future moments.

[0105] By calculating the cumulative deviation between actual data and the predictions of a bidirectional LSTM network, a trend anomaly score is generated, which is suitable for detecting progressive faults. For example, a temperature sensor may show a continuous, slow increase in temperature.

[0106] The self-learning update layer dynamically adjusts parameters based on the current environment and historical anomaly patterns, ensuring the accuracy of the hybrid time series model under different environmental conditions. This layer guarantees the robustness of the model by adaptively adjusting key parameters such as the sliding window size and the bidirectional LSTM network model parameters.

[0107] The system automatically updates key parameters of the model based on external environmental data. For example, under high humidity conditions, the system will appropriately relax the threshold for trend anomaly scores to reduce the impact of humidity on the data.

[0108] Based on current environmental data (such as temperature and humidity) and historical anomaly patterns from sensors, the parameters of the hybrid time-series model are automatically adjusted. If more false alarms or missed alarms are detected under new environmental conditions, the threshold or sliding window size is automatically adjusted. The self-learning update layer enables adaptive adjustment of model parameters, ensuring the hybrid time-series model remains efficient under different environmental and flight conditions. By introducing a historical anomaly pattern library, the self-learning update layer can better cope with complex and changing environmental conditions, ensuring the accuracy of fault detection.

[0109] The specific steps of S2.3 are as follows:

[0110] S2.3.1: Extract environmental features from external environmental data using principal component analysis;

[0111] To simplify environmental data and extract key influencing factors, Principal Component Analysis (PCA) was used for dimensionality reduction. Environmental data typically includes multiple dimensions such as temperature, humidity, air pressure, wind speed, and light intensity. PCA can transform this data into a few principal components, which serve as environmental features input for subsequent analysis. PCA extracts these environmental features, reducing the multi-dimensional environmental data to a small number of key factors, thus simplifying the data processing. This dimensionality reduction not only reduces computational complexity but also preserves the main trends in environmental data, ensuring the reliability of subsequent analysis.

[0112] S2.3.2: Construct an environmental correlation model based on historical flight data and historical environmental parameters;

[0113] Environmental correlation models are used to analyze the correlation between environmental parameters and sensor data to determine the impact of environmental changes on sensor data. Based on historical flight data and historical environmental parameters, an environmental correlation model is constructed using a neural network model to capture the patterns of how environmental parameters affect sensor data.

[0114] The environmental correlation model uses a multilayer perceptron network, taking historical environmental data and corresponding sensor data as input and output. The model is trained by minimizing the prediction error to capture the impact of environmental parameters on sensor data.

[0115] By constructing an environmental correlation model, a quantitative relationship between environmental characteristics and sensor data can be established, enabling the prediction of the impact of environmental factors on sensor data during actual flight. This correlation provides a benchmark for subsequent residual analysis, improving the accuracy of environmental impact identification.

[0116] S2.3.3: The extracted environmental features are used as input to the environmental association model. The association model is then used to predict the expected values ​​of sensor data under the current environmental conditions, generating associated data. This associated data represents the expected performance of sensor data under normal conditions in the current environment.

[0117] By generating correlated data in real time, expected values ​​of sensor data under the current environment can be provided, serving as a benchmark for subsequent residual analysis. This step ensures reasonable predictions of sensor data under specific environmental conditions, thereby helping to distinguish between environmental influences and sensor malfunctions.

[0118] S2.3.4: Calculate the residual between the associated data and the sensor data that does not conform to the normal timing pattern, perform fluctuation analysis on the residual, and determine whether it is within the preset environmental influence range. If the residual fluctuation is not within the range, it indicates that the sensor is faulty. If the residual fluctuation is within the range, it indicates that the abnormality of the sensor data is related to the environment and there is no fault.

[0119] The specific steps of S3 are as follows:

[0120] S3.1: Perform data cleaning, noise reduction, smoothing, and normalization on the validated sensor data;

[0121] Sensor data may contain noise, missing values, and unstable fluctuations, which can affect subsequent feature extraction. By cleaning, denoising, smoothing, and normalizing the data, clean, stable, and standardized sensor data can be obtained.

[0122] Detect and remove outliers and missing values. Outliers can be removed using statistical methods (such as the three-standard-deviation rule); missing values ​​can be repaired using interpolation or data imputation techniques.

[0123] High-frequency noise can be filtered out using low-pass filters or moving average methods. For complex noise scenarios, adaptive noise suppression algorithms are applied to enhance signal stability.

[0124] Data is smoothed using a sliding window to reduce short-term fluctuations. The size of the sliding window is adaptively adjusted based on the drone's current flight status to ensure smoothing across different conditions.

[0125] Scaling the data to a uniform range ensures that data from various sensors are on the same scale, facilitating unified processing of subsequent features;

[0126] Data cleaning and preprocessing ensured the quality and consistency of sensor data, reduced the impact of noise on feature extraction, and provided a reliable data foundation for subsequent steps.

[0127] S3.2: By using Fourier transform, the time-domain sensor data is converted into a frequency-domain signal to analyze the frequency components within each time period. Short-time frequency signals can reflect the instantaneous changes in sensor data and capture dynamic changes across different frequency bands;

[0128] The sensor data is divided into several overlapping time periods, and a Fourier transform is performed within each time period to obtain instantaneous frequency characteristics. The overlapping portion helps to smooth frequency changes. By analyzing the amplitude spectrum after the Fourier transform, the main frequency components and their corresponding amplitudes are identified to extract the instantaneous change characteristics of the sensor data. The window size and overlap rate of the short-time Fourier transform are dynamically adjusted according to the UAV's flight status to adapt to the frequency characteristics of the data. By calculating the energy proportion of each frequency component, the main frequency components within each time period are extracted to reflect the instantaneous dynamic changes of the sensor data.

[0129] Analysis of short-time frequency signals helps to capture rapid changes and instantaneous anomalies in sensor data, improves sensitivity to sudden failures, and enables real-time detection of instantaneous fluctuations in sensor data during UAV flight.

[0130] S3.3: Decompose the sensor signal into multiple wavelet coefficient levels based on wavelet transform, and extract the multi-scale variation features of each wavelet coefficient level through wavelet coefficients;

[0131] Wavelet transform is a multi-scale analysis method that decomposes a signal into different resolution levels, thereby extracting feature information at different scales. By performing wavelet decomposition on sensor signals, details and approximation coefficients at each level can be obtained, which can be used to analyze the multi-scale variation characteristics of the data.

[0132] Multi-level wavelet decomposition is performed on sensor data to generate wavelet coefficients at each level. Each level represents the signal characteristics at different frequency bands. Statistical analysis is performed on the wavelet coefficients at each level to extract variation features at various scales, such as mean, standard deviation, and energy. Wavelet transform provides decomposition of data in both the time and frequency domains, enabling the capture of data variation characteristics at different scales. Multi-scale feature analysis is particularly sensitive to progressive failure modes and can identify progressive failures during long-duration UAV flights.

[0133] S3.4: The instantaneous and multi-scale features extracted through Fourier transform and wavelet transform are aggregated to form a multi-dimensional feature vector. Then, the feature vector is upgraded to a higher-dimensional feature space through nonlinear mapping to enhance the discriminative power of the features.

[0134] Please see Figure 3The present invention provides a defect prediction model structure diagram. This defect prediction model aims to predict potential fault types of unmanned aerial vehicles (UAVs) based on high-dimensional feature vectors extracted from sensor data. The model includes a convolutional feature extraction layer, a temporal pattern learning layer, a convolutional pooling layer, and an output layer. The defect prediction model includes:

[0135] A convolutional feature extraction layer is used to extract spatial and temporal features from a high-dimensional feature vector. The convolutional feature extraction layer includes a spatial feature extraction sub-layer and a temporal feature extraction sub-layer.

[0136] In multi-sensor systems of unmanned aerial vehicles (UAVs), the spatial relationships between sensor data can contain important fault information. For example, abnormal fluctuations in attitude sensor data can affect velocity or acceleration data. Therefore, by extracting spatial features, the model can capture the interaction information of multi-dimensional data, which helps to more comprehensively identify potential system faults. This design enhances the model's robustness in handling complex data patterns.

[0137] The spatial feature extraction sublayer is used to identify the spatial structure features of the data from the high-dimensional feature vector. Spatial features describe the correlation between sensor data in multiple dimensions and can reflect the interrelationships between different sensor data. For example, the attitude sensor and accelerometer data of a UAV may be spatially correlated; by extracting spatial features, the mutual influence between sensor data can be better understood.

[0138] First, the correlation matrix between each sensor is calculated to measure the linear relationship between the outputs of different sensors. This method can reveal the direct relationship between data in various dimensions. Then, through feature cross-referencing, spatial features are constructed based on the correlation matrix, thereby capturing the complex relationships in multi-dimensional data.

[0139] Spatial features reflect the dependencies between different sensor data, especially in UAV flight control, where data from different sensors often influence each other. Examples include the correlation between velocity and acceleration, and the relationship between attitude and altitude. By extracting spatial features, models can identify the interactions and dependencies between different data dimensions, thereby gaining a more accurate understanding of the current flight state and providing more comprehensive information for fault detection.

[0140] In drone operations, fault signals often exhibit temporal correlation. A temporal feature extraction sublayer can identify temporal patterns in sensor data. For example, an abnormal increase in acceleration data may be a precursor to a fault. The temporal feature extraction sublayer is designed to capture short-term changes and trends in sensor data, enabling the model to detect both sudden and gradual faults early, thereby improving prediction accuracy.

[0141] The time feature extraction sublayer is used to identify the time-dependent characteristics of sensor data. Time features mainly describe the changing patterns of sensor data over time, helping the model understand the fluctuations and trends of data in the short term. For example, sensor data may experience abnormal fluctuations in a short period of time, and these fluctuations have a time correlation. This dynamic feature can be captured by time feature extraction. The rate of change of data is calculated by first-order or second-order difference to capture the increasing or decreasing trend of data over time.

[0142] Temporal features help models understand the dynamic changes in sensor data. For example, during a drone's flight, sensor data may exhibit specific trends or periodic changes over time. Temporal feature extraction enables models to capture these patterns of change, thus allowing for more sensitive detection of potential fault signs.

[0143] The temporal pattern learning layer is used to learn time series patterns from high-dimensional feature vectors;

[0144] The time-series pattern learning layer is primarily responsible for learning long-term time-series patterns from sensor data. While the short-term time features obtained through the time feature extraction sublayer only reflect changes in sensor data at small time scales, the time-series pattern learning layer further models the long-term dependencies and trends of the data. This layer uses a recurrent neural network to capture the time-series features of the sensor data.

[0145] The time-series pattern learning layer helps the model capture the long-term dependencies of sensor data, which is particularly important for identifying progressive failures. For example, a temperature sensor failure might manifest as a gradual increase in data over a long period, rather than a sudden fluctuation. By learning time-series patterns, the model can identify progressive failure trends and provide early warnings.

[0146] The convolutional pooling layer is used to convolve the spatial features and temporal features with the time series pattern using a 3×3 convolutional kernel, and to obtain a 2×2 mapped feature by using the ReLU activation function and max pooling operation, the max pooling operation including a 2×2 pooling kernel.

[0147] The output layer is used to flatten the mapped features into a compact feature vector, perform multi-class classification on the feature vector through a multi-layer perceptron, generate a probability distribution for each potential fault type, generate a confidence score for each fault type, sort the faults according to the confidence scores, and take the fault type corresponding to the highest confidence score as the prediction result.

[0148] Features obtained from the temporal pattern learning layer are classified to predict the probability distribution of each potential fault, and the faults are ranked according to confidence scores. The output layer uses a multilayer perceptron to map the extracted features to each fault type and outputs the probability of each fault type. This mapping of features to specific fault types provides the UAV system with clear fault prediction information, along with the probability and confidence score for each fault. This not only helps in the accurate identification of fault types but also provides reliable decision support for subsequent flight control.

[0149] The specific steps of S5 are as follows:

[0150] S5.1: Verify the prediction results by comparing them with historical data and output the verification results;

[0151] Using a pattern recognition-based method, the current sensor data is compared with fault patterns in historical data based on K-nearest neighbor similarity matching to evaluate their similarity. If the similarity is greater than or equal to a similarity threshold, the prediction result has higher confidence, and the verification result is output as true; if the similarity is less than the similarity threshold, the prediction result has lower confidence, and the verification result is output as false, with a value of 0.

[0152] The accuracy of the current prediction is verified by comparing it with historical failure modes and data. Historical data includes sensor data recorded under similar flight conditions and known failure modes. This verification helps to determine whether the current prediction is consistent with similar past failure modes, thereby assessing the reliability of the prediction.

[0153] The design intent of historical data comparison is to utilize existing failure modes and empirical data to verify the rationality of current predictions. This method can quickly detect anomalies and determine whether they conform to past failure characteristics, thus providing more evidence to support the prediction results.

[0154] S5.2: Perform a stability assessment on anomalies in the prediction results and output the analysis results;

[0155] The abnormal signals are accumulated, and their intensity is calculated over time. If the abnormal signal accumulates continuously over a period of time, it may be a real fault; if the intensity fluctuates unstablely, it may be transient noise. Linear regression is used to analyze the overall trend of the abnormal signal. If the abnormal signal shows a gradual increasing trend, it indicates that the anomaly has both cumulative and stable characteristics.

[0156] Stability assessment is used to analyze whether anomalous signals in the prediction results are persistent, thereby determining the reliability of the anomalous signals. In UAV sensor data, real fault signals often have a certain persistence and cumulative effect, while transient anomalies caused by noise usually do not persist. Stability assessment can effectively distinguish between these two situations;

[0157] Stability assessments effectively filter out transient, fluctuating anomalies, ensuring the model focuses only on genuine, persistent anomalies. This approach is particularly important in unmanned aerial vehicle (UAV) systems, as many faults exhibit gradual progression rather than instantaneous fluctuations. Stability analysis enhances the model's ability to identify persistent fault signals.

[0158] S5.3: Based on the verification results and stability assessment results from the first two steps, a comprehensive score is generated for the prediction results to produce a credibility score. The credibility score is a standardized value that reflects the reliability of the prediction results. A higher score indicates that the model has a higher level of confidence in the prediction results, while a low score means that the prediction results may contain more uncertainty.

[0159] Credibility scoring provides a quantitative standard for the reliability of prediction results, helping to categorize different prediction outcomes. A high credibility score indicates that the prediction result is reliable and suitable for automated decision-making; while a low credibility score suggests that more careful handling is required. By combining the verification results from the first two steps and stability analysis, credibility scoring can effectively distinguish between high and low credibility prediction results, providing a reliable basis for subsequent automated decision-making and ensuring the safety and stability of the UAV system.

[0160] By assessing the risk level of the credibility score, judging the severity of the fault and the feasibility of flight, and selecting the appropriate optimized control strategy, the safety and stability of the UAV under different fault scenarios are ensured. The specific steps of S6 are as follows:

[0161] S6.1: After analyzing the credibility scores, the failure risk is classified according to the scores. The risk level assessment includes two key dimensions: failure severity and flight feasibility.

[0162] Fault severity is used to assess the degree of threat a predicted fault poses to the drone system, and is categorized into three levels: low, medium, and high. For example, abnormal battery power is considered high severity, while slight fluctuations in sensor power might be considered low severity.

[0163] Flight feasibility is used to assess whether safe flight can continue in the event of the malfunction, or whether an immediate return or landing is necessary. Flight feasibility is also classified into three levels: high (flight is possible), medium (recommendation to reduce speed or adjust), and low (immediate return or landing required).

[0164] The credibility score and failure prediction results are mapped to a risk level matrix, with different risk levels set for different score ranges. For example, a high score corresponds to low risk, and a low score corresponds to high risk.

[0165] S6.2: Select the corresponding optimized control strategy based on the risk level assessment results;

[0166] The risk level assessment results include low risk, medium risk, and high risk, and the optimized control strategies include:

[0167] In low-risk situations, the drone can continue to perform its current flight mission and monitor for abnormal situations without requiring major adjustments to flight parameters. Specific measures include the drone continuing to fly along its original flight path without making significant adjustments, increasing the monitoring frequency of sensors that show abnormalities to prevent the abnormalities from worsening, sending low-risk warnings to the flight control center for ground monitoring personnel to notice, and slightly reducing the sensor data sampling frequency to reduce system resource consumption while ensuring normal flight.

[0168] In medium-risk situations, drones need to adjust their flight parameters to reduce potential risks. They can continue to perform missions but require more cautious flight. Specific measures include reducing flight speed, minimizing rapid acceleration or sharp turns to avoid exacerbating malfunctions; adjusting flight paths within mission-permissible limits to avoid complex or dangerous areas such as mountainous regions and densely built-up areas; moderately reducing the power consumption of non-critical components, such as weakening or suspending monitoring by non-core sensors to reduce battery consumption and extend flight time; increasing the frequency of uploading abnormal data to feed real-time status information back to the ground control center so that ground personnel can track and assess the situation and prepare emergency plans for return or landing based on further risk development, maintaining an appropriate level of readiness for return.

[0169] In high-risk situations, the drone prioritizes safety and immediately initiates return-to-home or emergency landing measures to ensure that the malfunction does not affect flight safety. Specific measures include: if return-to-home is feasible, immediately activating automatic return-to-home mode. The drone will choose the shortest path to return, reduce flight speed to extend battery life, and minimize unnecessary operations to conserve power; if return-to-home is not feasible or the risk continues to escalate, it will select a safe area to land immediately. Before landing, all non-critical components will be shut down or weakened to ensure power stability during landing. During return-to-home or landing, monitoring of unnecessary sensors and operation of non-critical equipment will be stopped, concentrating energy on critical components (such as navigation and control systems) to ensure a smooth return-to-home or landing. High-priority emergency information will be sent to the ground control center in real time to notify ground personnel to monitor or prepare for rescue. During return-to-home or landing, continuous self-checks will be performed to determine if the malfunction is worsening and dynamically adjust flight parameters to adapt to the malfunction situation.

[0170] S6.3: Adjust the parameters of the flight control system according to the optimized control strategy.

[0171] Example 2:

[0172] Please see Figure 4 The present invention provides an embodiment: a data-driven real-time defect prediction system for UAV flight control system, the system comprising a data acquisition module, a sensor verification module, a defect prediction module, a reliability analysis module, and a flight control optimization module;

[0173] The sensor verification module is equipped with a sensor fault detection strategy, which is used to detect the output of the data acquisition module and determine whether the sensor is faulty.

[0174] The defect prediction module is configured with a data processing strategy, which is used to process verified sensor data, build a defect prediction model through machine learning algorithms, and predict potential failures of the UAV in future flights based on the defect prediction model.

[0175] The credibility analysis module is used to evaluate the credibility of the prediction results output by the defect prediction module and generate a credibility score.

[0176] The sensor verification module is used to identify anomalies from sensor data and distinguish whether there is a sensor malfunction, specifically including:

[0177] The data consistency check unit is used to perform consistency checks on sensor data. By analyzing the data change trends and characteristics over different time periods, it identifies whether there are any abnormal signals.

[0178] The timing anomaly detection unit is used to analyze the timing patterns of sensor data with abnormal signals and determine whether the sensor data conforms to the normal timing pattern.

[0179] The environmental correlation unit is used to perform correlation analysis between sensor data that does not conform to the normal timing pattern and external environmental data to determine whether the abnormal sensor data is related to environmental factors. If they are not related, it indicates that the sensor is faulty.

[0180] The sensor fault detection strategy includes data integrity check logic, time series stability analysis logic, and environmental correlation analysis logic.

[0181] The data integrity check logic is configured within the data consistency check unit, the time series stability analysis logic is configured within the time series anomaly detection unit, and the environment correlation analysis logic is configured within the environment correlation unit.

[0182] The defect prediction module is used to analyze normal sensor data confirmed by the sensor verification module to predict potential faults of the UAV during flight, specifically including:

[0183] The data preprocessing unit is used to perform data cleaning, noise reduction, smoothing and normalization on the validated sensor data;

[0184] The feature extraction unit is used to extract feature variables from the preprocessed data and upgrade the extracted feature variables into high-dimensional feature vectors.

[0185] The defect prediction unit is used to build and train the defect prediction model, analyze the feature vectors generated by the feature extraction unit, and predict the failures of the UAV in future flights.

[0186] The data processing strategy includes feature extraction logic and defect prediction logic. The defect prediction logic is used to train the high-dimensional feature vector according to the defect prediction model and output the prediction result.

[0187] The feature extraction logic is configured within the feature extraction unit, and the defect prediction logic is configured within the defect prediction unit.

[0188] The credibility analysis module includes:

[0189] The result verification unit is used to verify the output results of the defect prediction module by comparing them with historical data, and output the verification results.

[0190] The anomaly stability analysis unit is used to assess the stability of anomalies in the prediction results and output the analysis results.

[0191] The credibility scoring unit is used to calculate the credibility score output by the defect prediction module based on the verification results and the analysis results.

[0192] The flight control optimization module is used to evaluate the reliability score and optimize the control of potential faults, specifically including:

[0193] The risk level assessment unit is used to assess the risk level of the credibility score, specifically including the severity of the fault and flight feasibility.

[0194] The control strategy decision unit is used to select the corresponding optimized control strategy based on the output of the risk level assessment unit.

[0195] The parameter adjustment unit is used to adjust the parameters of the flight control system according to the optimized control strategy output by the control strategy decision unit.

[0196] Although embodiments of the present invention have been shown and described above, it is understood that the above embodiments are exemplary and should not be construed as limiting the present invention. Those skilled in the art can make changes, modifications, substitutions and variations to the above embodiments within the scope of the present invention.

Claims

1. A data-driven real-time defect prediction system for unmanned aerial vehicle (UAV) flight control systems, characterized in that, The system includes a data acquisition module, a sensor verification module, a defect prediction module, a reliability analysis module, and a flight control optimization module. The sensor verification module is equipped with a sensor fault detection strategy, which is used to detect the output of the data acquisition module and determine whether the sensor is faulty. The defect prediction module is equipped with a data processing strategy, which is used to process verified sensor data, build a defect prediction model through machine learning algorithms, and predict potential failures of the UAV in future flights based on the defect prediction model. The credibility analysis module is used to evaluate the credibility of the prediction results output by the defect prediction module and generate a credibility score. The sensor verification module is used to identify anomalies from sensor data and distinguish whether there is a sensor malfunction, specifically including: The data consistency check unit is used to perform consistency checks on sensor data. By analyzing the data change trends and characteristics over different time periods, it identifies whether there are any abnormal signals. The timing anomaly detection unit is used to analyze the timing patterns of sensor data with abnormal signals and determine whether the sensor data conforms to the normal timing pattern. The environmental correlation unit is used to perform correlation analysis between sensor data that does not conform to the normal timing pattern and external environmental data to determine whether the abnormal sensor data is related to environmental factors. If they are not related, it indicates that the sensor is faulty. The sensor fault detection strategy includes data integrity check logic, time series stability analysis logic, and environmental correlation analysis logic. The data integrity check logic is configured within the data consistency check unit, the time series stability analysis logic is configured within the time series anomaly detection unit, and the environmental correlation analysis logic is configured within the environmental correlation unit. The time series stability analysis logic is used to detect the stability of data through a time series model and determine whether the data conforms to a normal time series pattern. Specifically, it includes: A hybrid time series model is constructed, and the sensor data is processed through the hybrid time series model to output short-term anomaly scores and trend anomaly scores; Based on the flight status of the drone, a weighting coefficient is set, and the short-term anomaly score and the trend anomaly score are weighted and calculated to generate a comprehensive anomaly score. The comprehensive anomaly score is compared with the risk threshold. If it is greater than or equal to the risk threshold, the sensor corresponding to the sensor data is marked as a faulty sensor. If it is less than the risk threshold, the sensor corresponding to the sensor data is marked as a normal sensor. The hybrid time series model includes: The short-term detection layer is used to merge sensor data into a linear sequence. It performs autoregressive prediction on the linear sequence within the window through a sliding window to obtain the next data point within the prediction window. By calculating the residual between the predicted value and the actual value, it obtains the short-term anomaly score within each window. The long-term detection layer is used to learn the historical sequence of sensor data through a bidirectional long short-term memory neural network, obtain long-term predicted values, calculate the deviation between the actual value of sensor data and the long-term predicted value, and gradually accumulate the deviation value to generate a trend anomaly score. A self-learning update layer is used to adjust parameters based on the current environment and historical anomaly patterns; The environmental correlation analysis logic is used to determine whether abnormal data is affected by the external environment, specifically including: Principal component analysis is used to extract environmental features from external environmental data. An environmental correlation model is constructed based on historical flight data and historical environmental parameters; The environmental features are used as input parameters for the environmental association model, and the environmental association model is used to train the environmental features to output associated data. Calculate the residual between the correlated data and the sensor data that does not conform to the normal timing pattern, perform fluctuation analysis on the residual, and determine whether it is within the preset environmental influence range. If the residual fluctuation is not within the range, it indicates that the sensor is faulty. If the residual fluctuation is within the range, it indicates that the abnormality of the sensor data is related to the environment and there is no fault.

2. The data-driven real-time defect prediction system for UAV flight control system according to claim 1, characterized in that, The defect prediction module is used to analyze normal sensor data confirmed by the sensor verification module to predict potential faults of the UAV during flight, specifically including: The data preprocessing unit is used to perform data cleaning, noise reduction, smoothing and normalization on the validated sensor data; The feature extraction unit is used to extract feature variables from the preprocessed data and upgrade the extracted feature variables into high-dimensional feature vectors. The defect prediction unit is used to build and train the defect prediction model, analyze the feature vectors generated by the feature extraction unit, and predict the failures of the UAV in future flights. The data processing strategy includes feature extraction logic and defect prediction logic. The defect prediction logic is used to train the high-dimensional feature vector according to the defect prediction model and output the prediction result. The feature extraction logic is configured within the feature extraction unit, and the defect prediction logic is configured within the defect prediction unit.

3. The data-driven real-time defect prediction system for UAV flight control system according to claim 2, characterized in that, The feature extraction logic includes: Sensor data is decomposed into short-time frequency signals, and the frequency components within each time period are calculated using Fourier transform to extract instantaneous change features. The sensor signal is decomposed into multiple wavelet coefficient levels based on wavelet transform, and the multi-scale variation features of each wavelet coefficient level are extracted through wavelet coefficients. The instantaneous change features and the multi-scale change features are aggregated to form a feature vector, and the feature vector is upgraded to a high-dimensional feature space through nonlinear mapping.

4. The data-driven real-time defect prediction system for UAV flight control system according to claim 2, characterized in that, The defect prediction model includes: Convolutional feature extraction layer is used to extract spatial and temporal features from high-dimensional feature vectors; The temporal pattern learning layer is used to learn time series patterns from high-dimensional feature vectors; The convolutional pooling layer is used to convolve the spatial features and temporal features with the time series pattern using a 3×3 convolutional kernel, and obtain a 2×2 mapping feature through the ReLU activation function and max pooling operation; The output layer is used to flatten the mapped features into a feature vector, perform multi-class classification on the feature vector through a multi-layer perceptron, generate a probability distribution for each potential fault type, generate a confidence score for each fault type, sort the faults according to the confidence scores, and take the fault type corresponding to the highest confidence score as the prediction result.

5. The data-driven real-time defect prediction system for UAV flight control system according to claim 1, characterized in that, The credibility analysis module includes: The result verification unit is used to verify the output results of the defect prediction module by comparing them with historical data, and output the verification results. The anomaly stability analysis unit is used to assess the stability of anomalies in the prediction results and output the analysis results. A credibility scoring unit is used to calculate the credibility score output by the defect prediction module based on the verification results and the analysis results. The flight control optimization module is used to evaluate the reliability score and optimize the control of potential faults, specifically including: The risk level assessment unit is used to assess the risk level of the credibility score, specifically including the severity of the fault and flight feasibility. The control strategy decision unit is used to select the corresponding optimized control strategy based on the output of the risk level assessment unit. The parameter adjustment unit is used to adjust the parameters of the flight control system according to the optimized control strategy output by the control strategy decision unit.

6. A data-driven real-time defect prediction method for a UAV flight control system, implemented based on a data-driven real-time defect prediction system for a UAV flight control system as described in any one of claims 1-5, characterized in that, The specific steps of the method are as follows: S1: Acquire sensor data from the various sensors of the drone; S2: Verify the collected sensor data, identify abnormal data, and determine whether there is a sensor malfunction; S3: Preprocess and extract features from sensor data that does not have sensor faults to generate high-dimensional feature vectors; S4: Construct a defect prediction model, using the high-dimensional feature vector as the input parameter of the defect prediction model, and process the input parameter through the defect prediction model to output the prediction result; S5: Evaluate the credibility of the prediction results and generate a credibility score; S6: Optimize and adjust the flight control of the UAV based on the credibility score.

Images