A physical-data hybrid fault diagnosis method and system for electric vertical takeoff and landing aircraft

Abstract

This invention relates to the field of unmanned aerial vehicle (UAV) control technology, and more particularly to a physical-data hybrid fault diagnosis method and system for electric vertical takeoff and landing (eVTOL) aircraft. First, suitable flight state variables are selected to construct an observer, generating a residual signal between the estimated rotor speed and the tracked value. Second, a random forest is introduced to achieve the localization and isolation of rotor faults. By integrating physical model data to construct a dataset and training a random forest classification model, a physical-data hybrid diagnostic model is established to capture the dynamic correlation between flight state and faults, achieving multi-fault diagnosis of eVTOL under closed-loop control.

Description

Technical Field

[0001] This invention relates to the field of unmanned aerial vehicle (UAV) control technology, and in particular to a physical-data hybrid fault diagnosis method and system for electric vertical take-off and landing (eVTOL) aircraft. Background Technology

[0002] With the advancement of global urbanization and population growth, the low-altitude economy is receiving increasing attention. As a core component of this economy, electric vertical takeoff and landing (eVTOL) aircraft play a crucial role in firefighting, logistics, and tourism. As a new type of transportation, the safety and reliability of eVTOLs directly impact their large-scale application. However, the complex sensors, actuators, and power systems of eVTOLs face multi-dimensional failure risks. Sensor failures (such as inertial measurement unit (IMU) deviations and airspeed sensor malfunctions) and actuator failures (such as rotor power attenuation and control surface jamming) can both lead to flight loss of control, especially in densely populated urban environments at low altitudes, where the consequences of such failures can be more severe. Although fault diagnosis technology for unmanned aerial vehicles (UAVs) has made some progress, the unique application scenarios and the complexity of the flight control systems of eVTOLs place higher safety demands on them. Therefore, tailoring fault diagnosis solutions for eVTOLs to support emergency decision-making is of great significance.

[0003] In practical applications, the problem of multi-fault diagnosis for electric vertical takeoff and landing (eVTOL) aircraft urgently needs to be solved. However, because eVTOLs employ closed-loop control systems, when multiple faults occur simultaneously, the system's dynamic compensation mechanism can spread the impact of these faults to other subsystems, leading to inter-fault coupling. For example, if a rotor motor in an eVTOL fails, causing a power reduction, the flight control system will reallocate the power output of each axis according to a power distribution algorithm. This will trigger power changes in other axis motors, creating the false impression that "other axis motors have also failed." Furthermore, when sensor signals from multiple faults are superimposed, their characteristic frequencies, amplitudes, or statistical characteristics may become similar, causing diagnostic methods based on fixed thresholds or one-dimensional features to fail. These issues all pose challenges to multi-fault diagnosis in eVTOLs.

[0004] Fault diagnosis methods for electric vertical takeoff and landing (eVTOL) aircraft can be categorized into model-based methods and data-driven methods. Model-based methods construct observers or filters based on the dynamic equations of the eVTOL system. By comparing the output of the designed normal-state observer with actual sensor signals to generate residual signals, and analyzing these residuals, actuator fault diagnosis is achieved. The advantages of this type of method are strong physical interpretability, direct reflection of the system's dynamic characteristics, and lower real-time requirements. However, feedback control actively adjusts healthy components to compensate for faults, which can mask early faults, significantly increasing the risk of model mismatch and failing to provide clear and direct anomaly information for fault detection. Furthermore, nonlinear characteristics and unmodeled dynamics reduce the accuracy of the observer, thus requiring robust design. Simultaneously, closed-loop control can lead to fault propagation, meaning a fault in one component may cause anomalies in multiple variables of other components. Traditional observers based on matching conditions often fail in closed-loop environments. In eVTOL, when an actuator fails, the controller adjusts the output of other healthy actuators through a feedback loop to maintain flight. This causes the healthy actuators to deviate from their normal operating points, creating a false impression of multiple faults where "healthy actuators behave abnormally due to compensation." Because complex systems are inherently difficult to model accurately, model-based fault diagnosis methods have low accuracy in diagnosing faults in complex systems. Furthermore, in multi-fault scenarios, the system complexity increases further, making it even more difficult for model-based methods to accurately isolate the fault source.

[0005] In recent years, data-driven methods have become a research hotspot in the field of unmanned aerial vehicle (UAV) fault detection. The core idea of ​​these methods is to identify faults by mining the statistical features of sensors and control signals. Without relying on physical models, data-driven methods can extract the mapping relationship between sensor data and fault types, and exhibit high accuracy in fault localization under multi-fault scenarios. Typical data-driven methods include Random Forest (RF), Convolutional Neural Network (CNN), Deep Forest, Dynamic Bayesian Network (DBN), and Support Vector Machine (SVM). To address the challenges of fault diagnosis in closed-loop systems, combining data-driven methods with dynamic feature extraction can automatically learn complex features in sensor data and accurately classify different fault scenarios, effectively improving fault isolation capabilities in closed-loop systems. However, data-driven classification methods are highly dependent on the quality and quantity of training data. Since eVTOLs are affected by various unknown perturbations, the data is uncertain, thus affecting classification accuracy and fault isolation performance. Furthermore, data-driven methods lack clear physical meaning, making the decision-making process of the fault diagnosis model difficult to interpret, which is a prominent problem in eVTOL fault diagnosis where safety requirements are extremely high. Summary of the Invention

[0006] To address the aforementioned technical problems, the present invention aims to provide a physical-data hybrid fault diagnosis method suitable for electric vertical takeoff and landing (eVTOL) aircraft. This method integrates an observer-based eVTOL physical model with a random forest (RF) algorithm for fault diagnosis. An ideal rotor speed is obtained from an unknown input observer, and this residual is normalized by combining it with the equivalent speed derived from control allocation. This residual, combined with the random forest algorithm, enables real-time detection, isolation, and severity assessment of multiple faults, thereby significantly improving diagnostic accuracy, interpretability, and system operational safety.

[0007] To achieve the above objectives, the present invention adopts the following technical solution:

[0008] A physical-data hybrid fault diagnosis method applicable to electric vertical takeoff and landing (eVTOL) aircraft, comprising the following steps:

[0009] S1. Observer Construction and Ideal Rotor Speed ​​Estimation: A subset of states satisfying the stability design conditions of the unknown input observer is selected from the flight state variables. An unknown input observer is established based on eVTOL dynamics / kinematics to estimate the input control variables and obtain the ideal rotor speed for each rotor. ;

[0010] S2. Equivalent speed estimation and residual generation: Without configuring speed sensors, the equivalent speed of each rotor is calculated based on the monitored control signals and power distribution relationships. And obtain the relative speed residual;

[0011] S3. Feature Construction: Based on the three-axis kinematics / attitude, position, control surface deflection, equivalent rotation speed, and the relative rotation speed residual, combined with sliding time window statistics, a multi-dimensional feature vector containing mean, standard deviation, peak value, energy, and trend slope is formed.

[0012] S4. Random Forest Modeling and Optimization: Bootstrap sampling is performed on the training samples, and split candidates are randomly selected from the feature subset. A random forest classifier is trained to output the fault location and fault type, and the forest size is determined based on out-of-bag error and cross-validation.

[0013] S5. Online Diagnosis and Severity Assessment: In the online phase, real-time features are input into the trained model, fault location / type is obtained through majority voting, and the severity of the fault is aggregated and assessed based on the statistics of each leaf node.

[0014] Preferably, the subset of states satisfies the following observer design conditions: the system is in minimum phase and satisfies the observer matching condition rank(CD)=rank(D);

[0015] Linear systems with unknown inputs are typically described as shown in (1):

[0016] (1),

[0017] Where: x(t) is the state vector, u(t) is the known input vector, ω(t) is the unknown input vector, y(t) is the measurement output; A, B, C, D are the system matrices, and t is time;

[0018] Minimum phase condition:

[0019] (2),

[0020] Where: s is a complex frequency domain variable, I is an identity matrix of the same order as A, n is the system state dimension, and rank(.) represents the matrix rank;

[0021] Observer matching conditions:

[0022] rank(CD)=rank(D)=q(3)

[0023] Where q is the unknown input dimension.

[0024] Preferably, the control signal is ω=[dt,de,da,dr] T This indicates that the unknown input is considered an unknown input; the estimation of the unknown input is achieved by reconstructing the observer, and its mathematical description is shown in (4):

[0025] (4),

[0026] in, This represents the extended state vector, where dt, de, da, and dr are the throttle, elevator, aileron, and rudder commands, respectively.

[0027] To adapt to the form of a linear system, the reconstructed observation equation can be expressed as shown in (5):

[0028] (5);

[0029] in: To expand the output, To expand the measurement matrix;

[0030] Specify C1=[C CA] T F1=[0 CD] T as well as Then (5) is rewritten as (6):

[0031] (6);

[0032] Assume the following relationship exists: The extended linear system is then described as (7):

[0033] (7);

[0034] The Luenberger observer is selected to track the state vector. The observer is described as follows (8):

[0035] (8);

[0036] in: To extend the system matrix, G1 and H1 are the design matrices, and z(t) represents the internal state of the observer. For extended state estimation, L1 is the observer gain. The available extended measurement signal.

[0037] As a preferred option, the equivalent speed The calculations include:

[0038] a) Based on the estimated control signal, the rotor speed is estimated by combining the power distribution equation, as shown in equation (9).

[0039] M=FU(9)

[0040] Where M represents the controlled throttle and torque distribution components of each rotor, U represents the total throttle and total torque, and F represents the power distribution matrix.

[0041] b) Establish the relationship between the distribution quantity and the rotational speed using a third-order polynomial, as shown in equation (10). :

[0042] (10)

[0043] Among them, M i Let αi represent the controlled throttle distribution component of the i-th rotor, and α0, α1, α2, and α3 represent Mi. i The coefficient;

[0044] And / or, using the relative residual of rotational speed, as defined in (11):

[0045] (11),

[0046] in, For time t The residual speed at any given moment; and They represent time respectively t The equivalent rotor speed and the ideal rotor speed at any given time.

[0047] Preferably, the sliding time window width is a preset value W, and for each rotor signal s...j (t) Extract the following five statistical features:

[0048] mean ;

[0049] Standard deviation ;

[0050] peak ;

[0051] energy ;

[0052] Trend slope ;

[0053] Where: μ j and All are mean values ​​within the window. .

[0054] As a preferred choice, a random forest (RF) consists of k base classifiers, expressed as equation (12):

[0055] (12)

[0056] Among them, T i (i=1,2,…,k) is the input feature set for each decision tree, h i (T i (i=1,2,…,k) represents the i-th classification regression tree (CART), which uses the Gini index to measure node impurity; when the sample is divided into C categories under feature t, the Gini value of feature t is expressed as equation (13):

[0057] (13),

[0058] Among them, P c This represents the proportion of the C-th class of samples in node t;

[0059] Each decision tree grows from the root node and continuously splits nodes based on specific sample subsets and feature subsets; the selection principle for features and split points is to minimize the Gini index, and the specific calculation method is shown in Equation (14):

[0060] (14)

[0061] Where a is the splitting threshold on feature t, Let n1 and n2 be the two subsets after splitting, and n be the number of samples in the corresponding subsets, and n be the total number of samples in the node.

[0062] As a preferred approach, a dual-validation mechanism is employed to optimize the hyperparameters of the random forest. By combining out-of-bag (OOB) error analysis and cross-validation, a balance is achieved between model complexity and generalization performance. The OOB error rate of the random forest can be obtained by comprehensively comparing the OOB validation results of all known labeled samples. The optimal number of decision trees, k... * Corresponding error rate e RF The first point where the convergence reaches the minimum value is represented by equation (15):

[0063] k * =min{argmine RF}(15)

[0064] Among them, e RF The OOB error rate represents the RF, and min{.} indicates taking the first k that reaches the global minimum and satisfies the convergence threshold;

[0065] K-fold cross-validation is used to maximize the average classification accuracy to prevent overfitting. The formula for calculating the average classification accuracy is shown in equation (16).

[0066] (16)

[0067] in, For the first k The number of correctly classified samples in the validation test. This represents the number of test samples for this test.

[0068] Furthermore, the present invention also provides an eVTOL multi-fault diagnosis system for implementing the method, comprising:

[0069] The sensing module is used to collect signals such as IMU, GPS and control surface / motor control commands, but does not include a rotor speed sensor;

[0070] The observer module is configured to calculate the ideal rotational speed based on the state subset and the UIO estimated input. ;

[0071] The equivalent speed and residual module is configured to calculate based on power / thrust distribution. and output ;

[0072] The feature engineering module is configured to generate multidimensional feature vectors that include sliding window statistics.

[0073] The classification and evaluation module is configured to run a random forest optimized with OOB and cross-validation, outputting the fault location / type and providing a severity score weighted by inverse variance.

[0074] Furthermore, the present invention also provides a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, causes the computer to implement the method described thereon.

[0075] Furthermore, the present invention also provides a computer program product, including a computer program or instructions that, when executed by a processor, implement the method.

[0076] This invention, by employing the aforementioned technical solution, significantly weakens the fault masking effect under closed-loop compensation through the dual construction of observer physical residuals and control allocation equivalent rotational speeds: Under conditions without rotational speed sensors, the equivalent rotational speed derived from power / thrust allocation forms a normalized residual with the ideal rotational speed given by UIO, actively eliminating disturbances under different operating conditions and loads, and improving the distinguishability and interpretability of abnormal deviations; Combining a multi-source feature space composed of five statistical features—three-axis motion, attitude, control variables, and sliding windows—with the collaborative optimization of out-of-bag error and cross-validation from random forests, it maintains robust isolation and hierarchical evaluation of concurrent faults in single / multi-rotor systems even when parameter perturbations and unmodeled dynamics exist; In the online phase, it reduces false alarms and quantifies severity through majority voting and inverse variance weighted aggregation, thereby achieving low-latency and high-accuracy diagnosis while simplifying hardware (RPM sensor-free). Simulation results show that the proposed scheme achieves a fault time identification accuracy of approximately 98% and a fault level identification accuracy of approximately 85% / 84% in single-fault and multi-fault scenarios, respectively, demonstrating stronger robustness, generalization ability and engineering application value compared to traditional pure model or pure data methods. Attached Figure Description

[0077] Figure 1 The workflow for the physical-data hybrid approach.

[0078] Figure 2 This describes the basic workflow of the observer-based method.

[0079] Figure 3 This describes the workflow of the RF algorithm.

[0080] Figure 4 Select a workflow for flight state variables.

[0081] Figure 5 This describes the growth process of a decision tree.

[0082] Figure 6 Optimize the out-of-bag (OOB) error curve.

[0083] Figure 7 This is the confusion matrix of the final model.

[0084] Figure 8 The equivalent rotational speed when rotor 1 experiences a misalignment fault.

[0085] Figure 9 The majority vote result for the severity of the fault when a 60% offset fault is injected into rotor 1.

[0086] Figure 10 The equivalent rotational speed when rotors 1 and 2 experience a misalignment fault. Detailed Implementation

[0087] The technical solutions in the embodiments of the present invention will be clearly and completely described below. Obviously, the described embodiments are only a part of the embodiments of the present invention, and not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those of ordinary skill in the art without creative effort are within the protection scope of the present invention.

[0088] The method proposed in this invention is a physics-data hybrid fault diagnosis method applicable to electric vertical takeoff and landing (eVTOL) aircraft. It integrates an observer-based eVTOL physical model with a random forest (RF) algorithm for fault diagnosis. The workflow of the proposed method is as follows: Figure 1 As shown: The physical model based on the unknown observer generates residual data of the eVTOL rotor speed and merges it with the original data in the eVTOL physical model to form the original dataset; Features are extracted from the original dataset to obtain the feature dataset used to build the random forest model; Each decision tree independently processes the feature dataset to obtain the classification result, and the classification results of all decision trees are integrated through a majority voting mechanism to finally obtain the fault diagnosis result.

[0089] Because eVTOL's flight control system uses closed-loop control, when a rotor malfunctions, the power compensation mechanism redistributes the throttle of each rotor to maintain stable flight. This results in no significant changes to the aircraft's flight state variables (such as linear velocity and attitude angles), but the power redistribution causes the rotor speed to deviate from the normal range. Therefore, rotor speed is a crucial indicator for eVTOL fault diagnosis. Figure 2 As shown, an observer is constructed using the dynamic equations of eVTOL, and the ideal rotor speed (IRS) obtained from the observer is used... This indicates that, since a speed sensor is not equipped, the actual rotor speed cannot be directly measured, but the actual control signal can be monitored through sensors. Therefore, the actual rotor speed is replaced by the equivalent speed (ERS). (This is expressed as an expression), the equivalent speed is calculated by combining the control signal monitored by eVTOL with the power distribution equation. The relative residual of the motor speed is extracted by the difference between the ideal speed and the actual speed.

[0090] like Figure 3As shown, the Random Forest (RF) algorithm extracts features from the original dataset to obtain a feature dataset for training a fault diagnosis classification model. Building a Random Forest model involves four key steps: feature data extraction, sample set generation, decision tree construction and ensemble, and parameter optimization. The trained Random Forest model processes real-time data generated by the eVTOL model to perform fault diagnosis in real-time. Model construction based on unknown input observers...

[0091] The control signals for eVTOL need to be estimated based on the flight state, and can be regarded as the unknown input of a linear system derived from the dynamic equations and kinematic equations. A typical linear system with unknown inputs is usually described as shown in (1):

[0092] (1),

[0093] To ensure the existence of an asymptotically stable observer for tracking unknown inputs, the linear system must satisfy the minimum-phase system condition and the observer matching condition, which can be described as follows (2) and (3):

[0094] 1) Minimum phase condition:

[0095] (2),

[0096] 2) Observer matching conditions:

[0097] (3),

[0098] Here, n and q represent the number of states and the number of unknown inputs in the linear system, respectively. It's important to note that not all combinations of flight states satisfy both conditions. Therefore, before constructing the unknown input observer, a subset of flight states that meet these conditions needs to be selected. The workflow for selecting flight state variables is as follows: Figure 4 As shown. Based on the kinematics and dynamics principles of electric vertical takeoff and landing (eVTOL) aircraft, the selected candidate parameters include linear velocity V. E =[u,v,w] T Linear displacement X E =[x e ,y e ,h e ] T Euler angular velocity ω in the body coordinate system B =[p,q,r] T Euler angles in the body coordinate system η = [ϕ, θ, ψ] T .

[0099] Based on the selection results of a subset of flight state variables, an unknown input observer is constructed for control signal estimation and rotor speed tracking. The control signal is represented by ω=[dt,de,da,dr]. T This indicates that it is considered an unknown input. The core idea of ​​unknown input estimation is to treat the unknown input as part of the extended state and estimate the unknown input by reconstructing the observer. Its mathematical description is shown in (4):

[0100] (4),

[0101] in This represents the extended state vector.

[0102] To adapt to the form of a linear system, the reconstructed observation equation can be expressed as shown in (5);

[0103] (5);

[0104] Specify C1=[C CA] T F1=[0 CD] T as well as Then (5) can be rewritten as (6):

[0105] (6);

[0106] Assume the following relationship exists: Then the extended linear system can be described as follows (7):

[0107] (7);

[0108] This invention uses a Luenberger observer to track the state vector, and therefore the observer can be described as follows (8):

[0109] (8).

[0110] Based on the estimated control signal, the rotor speed is estimated by combining the power distribution equation, as shown in equation (9):

[0111] M=FU(9)

[0112] Where M represents the controlled throttle and torque distribution components of each rotor, U represents the total throttle and total torque (as input to the eVTOL linear system), and F represents the power distribution matrix. The rotor speed is obtained based on the relationship between throttle and torque. This invention uses a third-order polynomial relationship to describe this correlation, as shown in equation (10):

[0113] (10)

[0114] Among them, M i Indicates the first i The controlled throttle distribution components of each rotor, where α0, α1, α2, and α3 represent M i The coefficients. The ideal rotor speed is calculated from the estimated throttle and torque, while the equivalent rotor speed is calculated by monitoring the control signals.

[0115] Based on the ideal rotor speed calculated above, the difference between the ideal rotor speed and the equivalent rotor speed can be obtained. To reduce the uncertainty of the absolute rotor speed under different flight conditions, this invention uses the relative residual of the rotor speed, which is defined as shown in (11):

[0116] (11),

[0117] in, For time t The rotational speed residual at any given moment. and They represent time respectively t The equivalent rotor speed and the ideal rotor speed at any given time.

[0118] Furthermore, the method proposed in this invention employs a physical-data hybrid diagnostic architecture. This method constructs a high-precision classification model through offline training and combines it with real-time feature analysis to achieve fault diagnosis. The construction steps of the Random Forest (RF) classification model are as follows:

[0119] 1. Feature Extraction: Feature extraction is a crucial prerequisite and is highly correlated with the performance of Random Forest (RF). The original dataset is processed through feature extraction to form a feature dataset. The original dataset includes three-axis motion parameters (linear velocity, angular velocity, attitude angle), position parameters, control surface deflections (throttle, elevator, aileron, rudder), equivalent motor speed data, and motor speed residual data. To further enhance feature representation capabilities, a sliding window statistical analysis method is used to calculate five types of time-domain features for each motor speed signal within a time window. The constructed 129-dimensional feature space covers three-axis motion parameters, position parameters, control surface deflections, equivalent motor speed, motor speed residuals, and five types of window statistical features for all motors.

[0120] The following lists five types of time-domain features calculated for motor speed signals: c1 represents the mean, reflecting the operating baseline; c2 represents the standard deviation, characterizing the intensity of fluctuations; c3 represents the peak value, used to detect abnormal pulses; c4 represents the energy, quantifying the impact of faults; and c5 represents the trend slope, capturing performance degradation. Features extracted from the same data segment constitute training and testing samples, while the corresponding fault types are set as labels. The feature variables are as follows:

[0121] mean Standard deviation Peak value ; energy Trend slope ; where s j (t) represents the first element in the window. j The speed sequence of each motor, For the mean within the window This is the midpoint of time.

[0122] Sample and Feature Subset Generation: The significant improvement in classification accuracy of random forests stems from the ensemble of decision trees. Training subsets are generated by randomly selecting training samples and subsets of features from those samples. For random sample selection, each subset is generated using Bootstrap sampling. Bootstrap sampling draws an equal number of samples from the original training set. The unselected samples constitute the out-of-bag (OOB) dataset, used for cross-validation. When randomly selecting features from the training samples, m = ⌊(M) is randomly drawn from the complete feature set. 1⁄2 Each set of features is used to form a feature subset paired with a subset of samples. Here, M represents the total number of features in the complete dataset and feature sets. This aims to reduce the correlation between decision trees and increase their diversity.

[0123] 2. Construction and Ensemble of Decision Trees: As a tree-structured ensemble learning method, Random Forest (RF) consists of k base classifiers, which can be expressed as Equation (12):

[0124] (12)

[0125] Among them, T i (i=1,2,…,k) is the input feature set for each decision tree, h i (T i (i=1,2,…,k) represents the i-th classification regression tree (CART), which uses the Gini index to measure node impurity. Since the Gini value reflects the classification impurity of the sample set, when the sample is divided into C categories under feature t, the Gini value of feature t can be expressed as (13):

[0126] (13),

[0127] Among them, P c Indicates the first node in node t C The proportion of samples of each class.

[0128] The modeling process of random forest is as follows: Figure 5 As shown, this illustrates the process of growing all decision trees. Each decision tree grows from the root node and continuously splits nodes based on specific sample subsets and feature subsets.

[0129] In this invention, the selection principle for features and split points is to minimize the Gini index, and the specific calculation method is shown in (14):

[0130] (14)

[0131] Where a is the splitting threshold on feature t, Let n1 and n2 be the two subsets after splitting, and n be the number of samples in the corresponding subsets, and n be the total number of samples in the node.

[0132] The decision tree stops growing when all samples in each leaf node belong to the same category. Finally, the random forest is an ensemble of all the constructed decision trees, and its classification result is obtained by majority voting on the results generated by each decision tree.

[0133] 3. Parameter Optimization: In this invention, the system employs a dual-validation mechanism to optimize the hyperparameters of the random forest. The number of decision trees is a key parameter in random forests. By combining out-of-bag (OOB) error analysis and cross-validation, a balance is achieved between model complexity and generalization performance. The OOB error rate of the random forest can be obtained by comprehensively comparing the OOB validation results of all known labeled samples. The optimal number of decision trees, k... * Corresponding error rate e RF The first point that converges to the minimum value can be expressed as equation (15):

[0134] k * =min{argmine RF}(15)

[0135] Among them, e RF This represents the OOB error rate of RF.

[0136] Furthermore, the generalization ability of the model was evaluated using five-fold cross-validation. The training set was divided into five subsets, and the parameter combination that maximized the average classification accuracy of the cross-validation was selected. The formula for calculating the average classification accuracy is shown in equation (16) below:

[0137] (16)

[0138] in, For the first k The number of correctly classified samples in the validation.

[0139] After determining the optimal parameters for fault diagnosis, a physical-data hybrid fault diagnosis model based on an observer model and random forest (RF) can be constructed. Through offline design and training, the established fault diagnosis model can be applied to online fault diagnosis, achieving real-time fault detection and isolation. The specific steps of fault diagnosis are as follows:

[0140] 1) Select a suitable subset of the flight states of an electric vertical takeoff and landing (eVTOL) aircraft as the state vector to satisfy the minimum phase condition and observer matching condition required to construct a stable unknown input observer.

[0141] 2) Under normal operating conditions, the observer estimates the input control signal and calculates the ideal rotor speed based on the power distribution equation, thereby obtaining the rotor speed residual signal.

[0142] 3) Based on triaxial motion data, position data, control surface deflection data, equivalent motor speed data, motor speed residual data, and five types of window statistical feature data for all motors, a 129-dimensional feature dataset was constructed. This feature dataset was used to train the random forest model.

[0143] 4) Generate and construct the decision tree. In a decision tree, each new feature sample and its corresponding fault label (fault type and fault level) are placed into a leaf node along the path from the root node to the leaf node using a node splitting strategy. Subsequently, the probability distribution of the fault type and the statistics of the fault level are stored in the leaf nodes to generate the final decision tree. The optimal parameters of the model are determined through testing and validation, and the finally trained Random Forest (RF) model is saved.

[0144] 5) Decision trees are used for fault identification, and the severity of the fault is calculated using the inverse variance weighted average method. A fault diagnosis conclusion is obtained by majority voting on the classification results of all decision trees.

[0145] The effectiveness of the proposed method is verified through experiments. Experimental data comes from a simulation platform of an 8-axis, 16-rotor electric vertical takeoff and landing (eVTOL) aircraft, which allows for the setting of fault types and their levels. It should be noted that each rotor is equipped with two motors; therefore, the letters "U" and "L" are abbreviations for "upper" and "lower," respectively, used to distinguish the motors above and below the rotor. For example, 1-U represents the motor above rotor 1, and 1-L represents the motor below rotor 1. The eVTOL is equipped with sensors including an inertial measurement unit (IMU), a global positioning system (GPS), and a control signal acquisition unit, capable of acquiring signals such as linear velocity, position, Euler angles, angular rate, throttle opening, and throttle torque. Next, single-fault and multi-fault experiments will be conducted to verify the effectiveness of the proposed method.

[0146] 1. Validation of the Random Forest Model

[0147] During the modeling process, raw data of the physical model under normal operating conditions and raw data corresponding to each fault type were collected. After feature extraction, the raw dataset was transformed into a feature dataset with a feature space dimension of 129, including timestamps, three-dimensional linear velocities (u, v, w), angular velocities (p, q, r), attitude angles (ϕ, θ, ψ), positions (x, y, z), and control surface deflections (d). t ,d e ,d a ,d r The equivalent rotational speeds of the upper rotor (omega1U~omega8U) and the lower rotor (omega1L~omega8L), the rotational speed residuals of the upper and lower rotors, and the five types of window statistical characteristics of each rotor (the sliding window is set to 30).

[0148] Based on the collected fault samples, determine the number of decision trees, k. * —This is defined as the minimum number of decision trees required to achieve the lowest error rate. First, an initial random forest (RF) model is trained, with the maximum number of decision trees set to maxTrees=300. During training, the out-of-bag (OOB) error for each decision tree is calculated, and a curve is plotted to show the trend of OOB error as the number of decision trees increases. By analyzing the relative rate of change of OOB error, the first starting point is determined when the relative rate of change for 10 consecutive trees is less than 1%. This starting point indicates that beyond this number, adding more decision trees can no longer significantly reduce the error. The number of decision trees at this point is defined as k. * As the number of decision trees increases, the out-of-bag (OOB) error rate gradually decreases, and at k... * The minimum value is reached when the value is 100. The OOB error optimization curve is shown below. Figure 6 As shown. According to equation (15), the optimal number of decision trees is determined to be k. * =100. To avoid overfitting due to excessive tree depth, the minimum number of samples per leaf node is set to 5 through empirical optimization. The optimal number of features for each feature subset is determined by m = ⌊(M). 1⁄2 ⌋=10, which means that when a node splits, the optimal split point is selected from 10 randomly chosen features.

[0149] Five-fold cross-validation was used to divide the training and test sets, with each test set containing approximately 20% of the samples and the training set comprising 80%. The cumulative confusion matrix was used to analyze class confusion. After determining the optimal parameters through cross-validation, the proposed model was trained using the training set. The confusion matrix of the trained model is shown below. Figure 7 As shown. Experimental results show that faults marked f1 to f6 can all be correctly isolated.

[0150] 2. Effectiveness verification under single rotor failure

[0151] Based on the validity verification of the random forest model, the feasibility of the diagnostic method under closed-loop control was verified through a single-rotor fault experiment. In this experiment, at a simulation time of 150s, a drift fault with a 60% power loss was injected into the No. 1 rotor of the simulated electric vertical takeoff and landing (eVTOL) aircraft. Figure 8 The experimental results show that when rotor 1 experienced a misalignment fault at 150s, the equivalent rotor speed deviated. The fault occurred at the moment when the motor speed changed abruptly.

[0152] After injecting a fault into rotor 1 at different time points, the time of fault occurrence was detected. As shown in Table 2, the experimental results show that the method proposed in this invention achieves a fault time identification accuracy of 98%.

[0153] The effectiveness of fault severity identification was verified by injecting offset faults with power losses ranging from 30% to 80% into rotor 1. During the offline training phase, training samples included fault level labels. Each node (including leaf nodes) of the decision tree stored a three-dimensional information matrix: fault type, fault location, and fault level. The specific numerical value of the fault level was calculated using an inverse variance weighted average, and the final result was determined through a voting mechanism. Taking the injection of an offset fault with a power loss of 60% into rotor 1 as an example, the voting results of the decision tree for the fault level are as follows: Figure 9 As shown in Table 3, the experimental results demonstrate that the proposed method achieves a fault level identification accuracy of 85%.

[0154] 3. Effectiveness verification under multi-rotor failure

[0155] To verify the feasibility of multi-fault diagnosis under closed-loop control, a 60% power loss offset fault was simultaneously injected into rotors 1 and 2 during a simulation time of 150 seconds. For example... Figure 10 As shown, similar to the single-rotor fault experiment, the rotor speed changes drastically when a fault is injected.

[0156] As shown in Table 3, the accuracy of fault detection time reaches 98%, indicating that the proposed method is effective in fault detection under multi-rotor fault scenarios. Simultaneously, similar to the single-rotor fault experiment, the fault level identification effect was verified by injecting offset faults with power losses ranging from 30% to 80% into rotors 1 and 2. The experimental results presented in Table 4 show that the fault level identification accuracy of the proposed method reaches 84%.

[0157] Table 1 shows the failure time detection results under a single failure of rotor No. 1.

[0158]

[0159] Table 2 shows the failure severity detection results under a single fault in rotor No. 1.

[0160]

[0161] Table 3 shows the failure time detection results for rotors 1 and 2 under multiple faults.

[0162]

[0163] Table 4 shows the failure severity detection results for multiple faults in rotors 1 and 2.

[0164]

[0165] In summary, this invention proposes a physical-data hybrid fault diagnosis method that integrates an unknown input observer model and random forest (RF), solving the problem that single-model methods struggle to identify complex faults and fault masking under closed-loop control. The proposed method comprises the following five steps: 1) selecting appropriate flight state variables to construct an unknown input observer; 2) generating rotor speed residual data through the unknown input observer; 3) extracting features to generate a sample dataset; 4) constructing a decision tree and training the model; and 5) obtaining fault diagnosis results through a voting mechanism. The effectiveness of the proposed method is verified through single-fault and multi-fault experiments.

[0166] The foregoing description of embodiments of the present invention, through which those skilled in the art are able to implement or use the present invention, will be readily apparent to those skilled in the art. Various modifications to these embodiments will be readily apparent to those skilled in the art. The general principles defined in the present invention may be implemented in other embodiments without departing from the spirit or scope of the invention. Therefore, the present invention is not to be limited to the embodiments shown herein, but is to be accorded the widest scope consistent with the principles and novelty disclosed herein.

[0167] Those skilled in the art will understand that embodiments of this application can be provided as methods, systems, or computer program products. Therefore, this application can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, this application can take the form of a computer program product embodied on one or more computer-usable storage media (including but not limited to disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0168] This application is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of this application. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, generate instructions for implementing the flowchart... Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0169] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes.

[0170] These computer program instructions may also be loaded onto a computer or other programmable data processing equipment to cause a series of operational steps to be performed on the computer or other programmable equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.

[0171] In a typical configuration, a computing device includes one or more processors (CPU), input / output interfaces, network interfaces, and memory.

[0172] Memory may include non-persistent memory in computer-readable media, such as random access memory (RAM) and / or non-volatile memory, such as read-only memory (ROM) or flash RAM. Memory is an example of computer-readable media.

[0173] Computer-readable media includes both permanent and non-permanent, removable and non-removable media that can store information using any method or technology. Information can be computer-readable instructions, data structures, modules of programs, or other data. Examples of computer storage media include, but are not limited to, phase-change memory (PRAM), static random access memory (SRAM), dynamic random access memory (DRAM), other types of random access memory (RAM), read-only memory (ROM), electrically erasable programmable read-only memory (EEPROM), flash memory or other memory technologies, CD-ROM, digital versatile optical disc (DVD) or other optical storage, magnetic tape, magnetic disk storage or other magnetic storage devices, or any other non-transferable medium that can be used to store information accessible by a computing device. As defined herein, computer-readable media does not include transient computer-readable media, such as modulated data signals and carrier waves.

Claims

1. A physical-data hybrid fault diagnosis method for an electric vertical take-off and landing aircraft (eVTOL), characterized in that, The method comprises the following steps: S1, observer construction and ideal speed estimation: select a subset of states from the flight state variables that meet the stable design conditions of the unknown input observer, establish an unknown input observer based on the dynamics and kinematics of the eVTOL, estimate the input control quantity, and obtain the ideal rotor speed of each rotor ; S2, equivalent rotation speed estimation and residual generation: based on the monitored control signals and power distribution relationship, the equivalent rotation speed of each rotor is calculated without configuring a rotation speed sensor and a relative rotation speed residual is obtained; S3, feature construction: based on three-axis kinematics / pose, position, control surface deflection, equivalent rotation speed, and the relative rotation speed residual, combined with sliding time window statistics, a multi-dimensional feature vector containing mean, standard deviation, peak value, energy, and trend slope is formed; S4, random forest modeling and optimization: the training samples are sampled by the bootstrap method and random splitting candidates are selected in the feature subset, a random forest classifier is trained to output fault location and fault type, and the forest size is determined based on out-of-bag error and cross-validation; S5, online diagnosis and severity evaluation: in the online stage, the trained model is input with real-time features, fault location and type are obtained by majority voting, and fault severity is aggregated and evaluated according to the statistics of each tree leaf node.

2. The method of claim 1, wherein, The state subset satisfies the following observer design conditions: the system is minimum phase, and the observer matching condition rank(CD)=rank(D) is satisfied; The unknown input linear system is described as shown in formula (1): （1）， Wherein: x(t) is a state vector, u(t) is a known input vector, ω(t) is an unknown input vector, y(t) is a measurement output; A, B, C, and D are system matrices, and t is time; The minimum phase condition is: （2）， Wherein: s is a complex frequency variable, I is a unit matrix of the same order as A, n is the state dimension of the system, and rank(.) represents the rank of the matrix; The observer matching condition is: rank(CD)=rank(D)=q (3), Wherein: q is the dimension of the unknown input.

3. The method of claim 1, wherein, Control signals with ω = [dt, de, da, dr] T denotes the unknown input, which is estimated by the reconstruction observer, whose mathematical description is shown in equation (4): （4）， wherein, denotes the extended state vector, dt, de, da, dr are the throttle, elevator, aileron, rudder commands, respectively; To adapt to the form of the linear system, the reconstructed observation equation is shown in formula (5): （5）， wherein: to extend the output, is an extended measurement matrix; C1 = [C CA] T F1 = [0 CD] T and Equation (5) is rewritten as Equation (6): （6）； Assuming such a relationship exists: then the extended linear system is described as equation (7): （7）； The Luenberger observer is selected to track the state vector, and the observer description is shown in formula (8): （8）， where: is the extended system matrix, G1, H1 are the design matrices, z(t) is the internal state of the observer, is the extended state estimate, L1 is the observer gain, is the available extended measurement signal.

4. The method of claim 1, wherein, Equivalent rotational speed The calculation of the equivalent rotational speed comprises: a) Based on the estimated control signal, the rotor speed is estimated by combining the power distribution equation, as shown in formula (9) M=FU (9), Wherein, M represents the distribution components of the controlled throttle and torque of each rotor, U represents the total throttle and total torque, and F represents the power distribution matrix; b) establishing a relationship between the dispensed amount and the rotational speed with a third order polynomial, as in equation (10) : （10）， where M i represents the controlled throttle distribution component of the ith rotor, and a0, a1, a2, and a3 represent coefficients of M i . And / or, the relative residual of the rotation speed is used, which is defined as shown in formula (11): （11）， wherein, is the rotational speed residual at time t; and respectively represent the equivalent rotor speed and the ideal rotor speed at time t.

5. The method of claim 1, wherein, The sliding time window width is a preset value W, and the rotor signal s j (t) Extract the following five types of statistical features: mean value ; Standard deviation ; peak ; energy ; Trend slope ; where: μ j with are the mean values within the window, .

6. The method of claim 1, wherein, The random forest RF is composed of k base classifiers, as shown in formula (12): （12）， where T i (i=1,2,…,k) is the input feature set of each decision tree, h i (T i )(i=1,2,…,k) represents the i-th classification and regression tree, which uses the Gini index to measure the node impurity; when the sample is divided into C categories under the feature t, the Gini value of the feature t is represented as formula (13): （13）, wherein P c represents the proportion of the Cth class samples in the node t; Each decision tree grows from the root node and continuously splits nodes according to a specific sample subset and feature subset; the selection principle of features and splitting points is to minimize the Gini index, and the specific calculation method is shown in formula (14): （14）， where a is a split threshold on feature t, are the two subsets after splitting, n1, n2 are the corresponding sample numbers, and n is the total sample number of the node.

7. The method of claim 6, wherein, The double verification mechanism is used to optimize the hyperparameters of the random forest, and the balance between the model complexity and the generalization performance is realized by combining the out-of-bag OOB error analysis and cross-validation. The out-of-bag error rate of the random forest can be obtained by comparing the out-of-bag verification results of all known label samples. The optimal number of decision trees k * The corresponding error rate e RF The first point converging to the minimum value is expressed as formula (15): k * = min{argmin RF} (15), where e RF represents the OOB error rate of the RF, min{.} denotes taking the first k that reaches the global minimum and satisfies the convergence threshold; And supplemented by K-fold cross-validation to maximize the average classification accuracy to prevent overfitting, the calculation formula of the average classification accuracy is shown in formula (16); （16）， wherein, is the number of samples classified correctly in the kth fold validation, is the number of test samples in the fold.

8. An eVTOL multi-fault diagnostic system for implementing the method of any one of claims 1-7, characterized by, It comprises: A sensing module for collecting IMU, GPS, and control surface / motor control instruction signals, and does not include a rotor speed sensor; an observer module configured to estimate inputs to the UIO based on the state subset and compute an ideal speed ; An equivalent rotational speed and residual error module is configured to calculate from the power / thrust allocation and output ; A feature engineering module configured to generate a multi-dimensional feature vector containing sliding window statistics; a classification and evaluation module configured to run a random forest optimized by OOB and cross-validation, output fault location / type, and give inverse variance weighted severity.

9. A computer readable storage medium having stored thereon a computer program or instructions, characterized in that, The computer program or instructions are executed by the processor to implement the method of any one of claims 1-7.

10. A computer program product comprising computer programs or instructions, characterized in that, The computer program or instructions, when executed by a processor, implement the method of any one of claims 1-7.

Images