UAV data fault detection method based on model joint computing

Abstract

The present invention provides a method for detecting drone data faults based on model-based joint computation, addressing the technical issues of poor fault detection accuracy and real-time performance in existing technologies. The method includes: 1) constructing a drone data fault detection system; 2) obtaining a training sample set and a test sample set; 3) constructing a long short-term memory (LSTM) network model; 4) iteratively training the LSTM network model; 5) constructing a data fault detection model at a fault detection center; and 6) obtaining drone data fault detection results. The linear self-convergent model and the long short-term memory network model in the data fault detection model of the present invention obtain weights in the final signal prediction value through a joint computation unit, reducing signal prediction errors and improving data fault detection accuracy. Furthermore, the linear model's compensation for signal value prediction accuracy reduces the structure of the long short-term memory network model, improving the real-time performance of drone data fault detection.

Description

Technical Field

[0001] The present invention belongs to the field of UAV flight safety technology, and specifically relates to a UAV data fault detection method based on model joint calculation, which can be used to detect sensor data faults during the real-time flight of the UAV and ensure the flight safety of the UAV. Background Art

[0002] The efficiency and convenience of drones make them widely used in people's daily production and life. Due to the lack of real-time control by pilots, the safety of drones is relatively low. Introducing a fault detection mechanism for drones is a feasible way to improve the flight safety of drones.

[0003] Because the real-time flight data recorded by a drone's sensors reflects its real-time flight status, this data exhibits temporal correlation. A data failure indicates that the drone's sensor signals are unreliable, violating established spatiotemporal correlations. Data failure detection methods utilize a time series model to simulate the spatiotemporal relationship between sensor input and output signals. The predicted output values ​​are then compared with the true values ​​to obtain residuals, which are then analyzed to determine if a data failure has occurred.

[0004] In their paper "Automatic Real-time Anomaly Detection for Autonomous Aerial Vehicles" (arXiv:1907.00511v1[eess.SY]1Jul2019), Azarakhsh Keipour et al. published a method for autonomous, real-time drone fault detection. This method uses a linear model to calculate the predicted value of state sensor signals, then obtains the residual between the predicted value and the true value, and uses a threshold to determine whether a fault has occurred. This method has a simple model structure and the advantages of low storage space usage, low time complexity, and fast computation speed. However, because it relies solely on a linear model to predict state sensor signals, it has poor universality for different state sensor signals. Furthermore, because linear models can only approximate the relationship between sensor input and output signals, the detection accuracy is low.

[0005] In their paper "Unmanned Aerial Vehicle Flight Data Anomaly Detection and Recovery Prediction Based on Spatio-Temporal Correlation" (IEEE TRANSACTIONS ON RELIABILITY, VOL. 71, NO. 1, MARCH 2022), Jie Zhong et al. published a method for detecting anomalies in drone flight data based on spatio-temporal correlation. This method uses an artificial neural network (ANN) and an LSTM network working in series. Each sensor data first enters the ANN network for correlation screening. The filtered high-correlation signal data then enters the LSTM network to calculate the predicted value of the target state sensor signal. By calculating the residual with the true value, the residual threshold is used to determine whether the target state sensor data at the current moment is normal. Due to the presence of the LSTM network, this method has a better fit for the temporal relationship of the state sensor signals and a lower fault detection latency. However, the training process of the complex neural network model is extremely cumbersome and prone to polarization phenomena such as underfitting or overfitting. In addition, the computation time for each data point is high, resulting in poor real-time fault detection.

[0006] The above-mentioned fault detection algorithms all rely on only one time series prediction model for signal prediction, and all have the contradiction between accuracy and real-time performance. Without the participation of a good combination algorithm, the simultaneous use of two different types of time series models for signal prediction cannot dynamically determine the weight of the prediction results of each model in the final result, that is, the problem of model weight imbalance. When a model has a large error, it will directly interfere with the prediction result, resulting in reduced detection accuracy, and even a situation where the detection efficiency is inferior to that of a single model. Summary of the Invention

[0007] To overcome the shortcomings of the aforementioned existing methods, the present invention aims to provide a method for drone data fault detection based on model-based joint computation. First, a drone data fault detection system is provided, comprising a fault detection center, state sensors, a drone flight controller, and a ground monitoring platform that communicate with the fault detection center, enabling real-time transmission and effective processing of fault information. Second, the data fault detection model deployed by the fault detection center includes a lightweight long-short-term memory (LSTM) model and a linear self-convergent (LAR) model operating in parallel, along with a joint computation unit connected in series with the outputs of the two models. This addresses the issue of unbalanced model weight distribution and enables the dynamic integration of predicted values ​​from different model signals.

[0008] To achieve the above object, the technical solution adopted by the present invention includes the following steps:

[0009] (1) Build a drone data fault detection system:

[0010] Construct a UAV data fault detection system that includes a monitoring platform set up on the ground, a fault detection center on the UAV, a flight controller, and I status sensors. The fault detection center can communicate with the flight controller and the monitoring platform, and I ≥ 1;

[0011] (2) Obtain training sample set and test sample set:

[0012] (2a) Select the state sensor data of I at N consecutive time points during a UAV’s fault-free flight from the database At the same time, the fault detection center obtains I state sensor data at K consecutive time points during the real-time flight of the UAV in, represents the i-th state sensor data at N consecutive time points, u i (n) and y i (n) represents the input value and output value of the i-th state sensor data at the n-th moment, N>1000; represents the i-th state sensor data at K consecutive time points, u i (k) and y i (k) represents the input signal value and output signal value of the i-th state sensor data at the k-th moment, 500<K<N;

[0013] (2b) Through a sliding window of length D Divide into (ND) and (KD) data segments respectively, and The input value and historical output value of each data segment are used as training samples, and the current output value is used as the training label, which consists of (ND) training samples. and (ND) training labels The training sample set and The input value and historical output value of each data segment are used as test samples, and the current output value is used as the test label, which consists of (KD) test samples. and (KD) test labels The test sample set, where 10≤D<20;

[0014] (3) Constructing a long short-term memory network LSTM model:

[0015] Construct a long short-term memory network LSTM model F′ consisting of I parallel arranged sub-networks corresponding to the number of state sensors lstm, the i-th sub-network is f′ i , f′ i It is composed of two LSTM layers and a fully connected layer connected in sequence;

[0016] (4) Long short-term memory network LSTM model F′ lstm Perform iterative training:

[0017] (4a) Initialize the number of iterations to t, the maximum number of iterations to T, and iterate each sub-network f′ in the tth iteration i The weight parameter And let t = 1, where, Indicates rounding down, S indicates the amount of training data for each iteration, 5<S<20;

[0018] (4b) S training samples will be randomly selected from the training sample set without replacement as F′ lstm Input of each sub-network f′ i After extracting features from each training sample, perform sequence prediction on the extracted features to obtain S prediction vectors in represents the set of real numbers;

[0019] (4c) The root mean square error loss function is used, and each training label Y(s) and its corresponding prediction vector Calculate each subnetwork f′ i The loss value is then used to calculate the network weight parameters using the gradient descent algorithm. The gradient of Make updates;

[0020] (4d) Determine whether t≥T holds. If so, obtain the trained LSTM model F lstm , otherwise, let t = t + 1 and execute step (4b);

[0021] (5) Build a data fault detection model in the fault detection center:

[0022] The fault detection center is constructed by parallel arrangement of trained LSTM models F lstm and a linear self-convergent LAR model, and a data fault detection model comprising a joint computing unit, a fault detection unit, and a fault recovery unit cascaded with the output ends of the two models, wherein the fault detection unit comprises a sudden fault detection subunit and a progressive fault detection subunit arranged in parallel;

[0023] (6) Obtaining drone data fault detection results:

[0024] (6a) The fault detection center initializes the data fault detection model and the historical fault state variable isFailure, setting isFailure = 0, k = D + 1;

[0025] (6b) The fault detection center takes the k-time sample X(k) and the label, i.e., the true value of the output signal Y(k), as input to the data fault detection model, the LSTM model F lstm Through X(k), calculate the output signal LSTM model prediction vector As output, at the same time, the LAR model constructs the data matrix Φ(k) through X(k), and calculates the output signal LAR model prediction vector through Φ(k) and the k-1 moment parameter matrix R(k-1) As output, and through Φ(k), and Y(k) iteratively calculate R(k), and then the joint calculation unit is and Y(k) extract the joint calculation matrix set By jointly calculating the matrix The eigenvector corresponding to the maximum eigenvalue is given by Construct the weight matrix W(k), and then pass Y(k) and W(k) calculate the output signal joint prediction vector and the joint residual vector Finally, the fault detection unit normalizes E(k) to obtain the normalized vector The sudden fault detection subunit generates a sudden fault sensor number set J1={j||z j (k)|>3.5}, and at the same time, the progressive fault detection subunit calculates the likelihood vector through Z(k) Generate progressive fault sensor number set J2 = {j|lh j >lh0}, lh0 is the progressive failure threshold;

[0026] (6c) The fault detection center calculates the fault sensor number set J = J1 ∪ J2 and determines whether J is an empty set. If so, set isFailure = 0 and execute step (6f). Otherwise, execute step (6d).

[0027] (6d) The fault detection center determines whether isFailure=0. If so, it sets isFailure=1 and executes step (6e). Otherwise, the fault detection center sends a forced landing command to the flight controller, and the UAV ends its flight.

[0028] (6e) The fault detection center sends a fault warning message to the monitoring platform. At the same time, the fault recovery unit of the data fault detection model generates a correction instruction, which is then sent to the flight controller for execution.

[0029] (6f) The fault detection center determines whether k≥K holds. If so, it obtains (KD) data fault detection results. Otherwise, it sets k=k+1 and executes step (6b).

[0030] Compared with the prior art, the present invention has the following advantages:

[0031] First, the present invention uses a linear self-convergent LAR model and a lightweight long-short-term memory (LSTM) network model working in parallel to calculate sensor signal predictions. Because the linear model compensates for the accuracy of the predicted signal values, the LSTM network model optimizes the model structure, sacrificing some accuracy in exchange for fewer parameters. The optimized LSTM network model consists of only two LSTM layers and one fully connected layer, significantly reducing the computational complexity during model runtime. This not only reduces the model's memory usage, but also improves model training efficiency and reduces computational time.

[0032] Second, the present invention combines the prediction values ​​of the two models through a joint calculation unit to obtain a joint prediction value and a joint residual value. The joint calculation unit constructs a joint calculation matrix based on the residuals of the prediction values ​​of each model signal, and extracts the weight of each model prediction value in the joint prediction value through the eigenvector corresponding to the largest eigenvalue of the matrix. The joint calculation matrix quantifies the relative relationship between the residuals of the prediction values ​​of each model and dynamically calculates the proportion of each model prediction value in the joint prediction value, effectively solving the problem of unbalanced weight distribution of different models, improving the prediction accuracy of sensor signals, and reducing fault detection delay.

[0033] Third, the fault detection unit of the present invention uses a structure in which a sudden fault detection subunit and a gradual fault detection subunit operate in parallel to determine abnormal values ​​of the joint residual value. The sudden fault detection subunit mainly detects large-scale sudden changes in sensor signal values ​​over a short period of time, while the gradual fault detection subunit mainly detects small-scale drift anomalies in sensor signal values ​​over a long period of time. This parallel fault detection unit structure can handle data failures caused by various conditions during the actual flight of the drone, effectively reducing the rate of missed fault detection and greatly improving the safety of drone flight. BRIEF DESCRIPTION OF THE DRAWINGS

[0034] Figure 1 Flowchart for the implementation of the present invention;

[0035] Figure 2 This is a comparison diagram of the root mean square error simulation of the pitch signal predicted by the present invention and the prior art;

[0036] Figure 3 This is a simulation comparison chart of the average single-point detection time consumption of the present invention and the prior art. DETAILED DESCRIPTION

[0037] In order to make the purpose, technical solutions and advantages of this application clearer, the present invention is further described clearly and completely below in conjunction with specific embodiments.

[0038] Reference Figure 1 , the present invention comprises the following steps:

[0039] (1) Build a drone data fault detection system:

[0040] A UAV data fault detection system is constructed, which includes a monitoring platform set up on the ground, a fault detection center, a flight controller, and I state sensors carried on the UAV. The fault detection center can communicate with the flight controller and the monitoring platform, where I ≥ 1. In this example, I = 3, where i = 1 represents the roll state sensor, i = 2 represents the pitch state sensor, and i = 3 represents the vertical roll state sensor.

[0041] (2) Obtain training sample set and test sample set:

[0042] (2a) The training sample data for this example comes from the AirLab laboratory, which is the roll, pitch, and yaw state sensor data of the Carbon Z T-28 model UAV at N consecutive time points during a trouble-free flight. At the same time, the fault detection center obtains the roll, pitch, and vertical roll state sensor data of K consecutive time points during the real-time flight of the UAV. in, represents the i-th state sensor data at N consecutive time points, u i (n) and y i (n) represents the input value and output value of the i-th state sensor data at the n-th moment, N>1000; represents the i-th state sensor data at K consecutive time points, u i (k) and y i (k) represents the input signal value and output signal value of the i-th state sensor data at the k-th moment, 500<K<N. In this example, N=2531, K=1508;

[0043] (2b) Through a sliding window of length D Divide into (2531-D) and (1508-D) data segments respectively, and The input value of each data segment is used as a training sample, and the output value is used as a training label, which consists of (2531-D) training samples. and (2531-D) training labels The training sample set and The input value of each data segment is used as the test sample, and the output value is used as the test label, which consists of (1508-D) test samples. and (1508-D) test tags The test sample set, where 10≤D<25, in this example, D=19, in the training sample set, In the test sample set,

[0044] (3) Constructing a long short-term memory network LSTM model:

[0045] Construct a long short-term memory network LSTM model F′ consisting of three parallel sub-networks corresponding to the number of state sensors lstm , the i-th sub-network is f′ i , f′ i It is composed of two LSTM layers and a fully connected layer connected in sequence. The number of hidden nodes in the first LSTM layer is 64, and the number of hidden nodes in the second LSTM layer is 32.

[0046] (4) Long Short-Term Memory Network LSTM Model F′ lstm Perform iterative training:

[0047] (4a) Initialize the number of iterations to t, the maximum number of iterations to T, and iterate each sub-network f′ in the tth iteration i The weight parameter And let t = 1, where, Indicates rounding down, S indicates the amount of training data for each iteration, 5<S<20. In this example, T=50, S=9;

[0048] (4b) Randomly select 9 training samples from the training sample set without replacement as F′ lstm Input of each sub-network f′ i After extracting features from each training sample, sequence prediction is performed on the extracted features to obtain 9 prediction vectors in represents the set of real numbers;

[0049] (4c) The root mean square error loss function is used, and each training label Y(s) and its corresponding prediction vector Calculate each subnetwork f′ i The loss value is then used to calculate the network weight parameters using the gradient descent algorithm. The gradient of Make updates;

[0050] (4d) Determine whether t≥50 is true. If so, obtain the trained LSTM model F lstm , otherwise, let t = t + 1 and execute step (4b);

[0051] (5) Build a data fault detection model in the fault detection center:

[0052] The fault detection center is constructed by parallel arrangement of trained LSTM models F lstm and a linear self-convergent LAR model, and a data fault detection model comprising a joint computing unit, a fault detection unit, and a fault recovery unit cascaded with the output ends of the two models, wherein the fault detection unit comprises a sudden fault detection subunit and a progressive fault detection subunit arranged in parallel;

[0053] (6) Obtaining drone data fault detection results:

[0054] (6a) The fault detection center initializes the data fault detection model and the historical fault state variable isFailure, setting isFailure = 0, k = 10;

[0055] The historical fault status variable records historical fault status information. A value of 0 indicates no fault, and a value of 1 indicates a fault affecting the current moment. When the fault detection center detects a fault, it uses the value of isFailure to determine whether the fault is a first-time occurrence or a persistent condition. For first-time occurrences, the fault recovery unit will attempt to repair the fault. For persistent conditions, the fault cannot be recovered through data methods, and an emergency landing will be performed to avoid a crash.

[0056] (6b) The fault detection center inputs the k-time sample X(k) and the label, i.e., the true value of the output signal Y(k), into the data fault detection model:

[0057] LSTM model F lstm Through X(k), calculate the output signal LSTM model prediction vector As output;

[0058] At the same time, the LAR model constructs the data matrix Φ(k) through X(k), and calculates the output signal LAR model prediction vector through Φ(k) and the k-1 moment parameter matrix R(k-1) As output, and through Φ(k), And Y(k) iteratively calculate R(k), the specific calculation formula is:

[0059]

[0060]

[0061]

[0062]

[0063]

[0064]

[0065] R(k)=[δ 1 (k)…δ 3 (k)],

[0066] Among them, the data vector of the i-th state sensor at time k is From X(k) Stretching to a one-dimensional vector, I3 is a three-dimensional unit column vector, Represents the LAR model residual value of the i-th state sensor at time k, Δ i (k) represents the change matrix of the i-th state sensor at time k, (·) T represents the transpose operation, represents the convergence matrix of the i-th state sensor at time k, δ i (k) represents the parameter column vector of the i-th state sensor at time k, is a (18×18) matrix of all 1s, ||δ i (k)||0 represents the vector δ i (k) is the zero norm. In particular, for k = 10, R (9) is an all-zero matrix. is the identity matrix, δ i (9) is the zero vector;

[0067] Then, the joint computing unit passes and Y(k) extract the joint calculation matrix set By jointly calculating the matrix The eigenvector corresponding to the maximum eigenvalue is given by Construct the weight matrix W(k), and then pass Y(k) and W(k) calculate the output signal joint prediction vector and the joint residual vector The specific calculation formula is:

[0068]

[0069]

[0070]

[0071]

[0072]

[0073]

[0074] Among them, λ i for The maximum eigenvalue of i (k) Corresponding to λ i The characteristic vector of sum(v i (k)) is ν i (k) The cumulative sum of all elements, ω i (k) is the weight vector of state sensor i at time k. The joint calculation matrix eliminates the dimension of the residual value of the state sensor of each model through division, obtains the relative relationship of the accuracy of each model after quantization at time k, and maps the relative relationship of quantization accuracy to the weight vector through the eigenvector of the matrix to complete the calculation of the joint prediction value and the joint residual value;

[0075] Finally, the fault detection unit normalizes E(k) to obtain the normalized vector Each z i The calculation formula for (k) is:

[0076]

[0077]

[0078]

[0079] in, and σ i (k) represents the mean and standard deviation of the joint residual of the state sensor i at time k; the sudden fault detection subunit generates the sudden fault sensor number set J1 = {j||z j (k)|>3.5}, the purpose is to detect whether the UAV has a short-term large-scale mutation data failure. At the same time, the progressive fault detection subunit calculates the likelihood vector through Z(k) The purpose is to detect whether the drone has a long-term small-amplitude drift data fault. i The calculation formula for (k) is:

[0080]

[0081] Regenerate progressive fault sensor number set J2 = {jlh j >lh0}, lh0 is the progressive failure threshold, and the specific calculation formula is:

[0082]

[0083] Where ln(·) represents the natural logarithm, β represents the false alarm rate, and γ represents the missed detection rate. In this example, β = 0.002, γ = 0.002;

[0084] (6c) The fault detection center calculates the fault sensor number set J = J1 ∪ J2 and determines whether J is an empty set. If so, set isFailure = 0 and execute step (6f). Otherwise, execute step (6d).

[0085] (6d) The fault detection center determines whether isFailure=0. If so, it sets isFailure=1 and executes step (6e). Otherwise, the fault detection center sends a forced landing command to the flight controller, and the UAV ends its flight.

[0086] (6e) The fault detection center sends a fault warning message to the monitoring platform. At the same time, the fault recovery unit of the data fault detection model generates a correction instruction, which is then sent to the flight controller for execution.

[0087] (6f) The fault detection center determines whether k≥1508 is established. If so, 1489 data fault detection results are obtained. Otherwise, k=k+1 is set and step (6b) is executed.

[0088] The effects of the present invention will be further explained below in conjunction with simulation experiments.

[0089] 1. Simulation experiment conditions:

[0090] The hardware platform of the simulation experiment of the present invention is: the processor is Intel(R) Core(TM) i7-9750H CPU, and the main frequency is 2.60GHz.

[0091] The software platform for the simulation experiment of the present invention is: a VMware virtual machine with 4GB memory, Ubuntu 21.04 operating system, python 3.9.5, and keras 2.6.0.

[0092] 2. Simulation experiment data:

[0093] The simulation experiment data of the present invention comes from the real UAV flight data of the AirLab laboratory, published by Azarakhsh Keipour et al. at http: / / theairlab.org / alfa-dataset / . It includes 47 independent flight sequences, covering 9 types of flight data, including no faults and 8 types of faults, including engine, left aileron, right aileron, left and right ailerons, elevator, left rudder, right rudder, and aileron-rudder faults.

[0094] 3. Simulation experiment content:

[0095] The simulation experiment of the present invention uses the present invention and the existing technology to perform fault detection on 11 groups of UAV flight data with data failure periods.

[0096] In the simulation experiment, the existing technology refers to:

[0097] Azarakhsh Keipour et al. published an autonomous real-time anomaly detection method for autonomous aerial vehicles (ARX) in “Automatic Real-time Anomaly Detection for Autonomous Aerial Vehicles” (arXiv:1907.00511v1[eess.SY]1Jul 2019).

[0098] Jie Zhong et al. published a method for detecting anomalies in UAV flight data based on spatio-temporal correlation in "Unmanned Aerial Vehicle Flight Data Anomaly Detection and Recovery Prediction Based on Spatio-Temporal Correlation" (IEEE TRANSACTIONS ON RELIABILITY, VOL. 71, NO. 1, MARCH 2022), referred to as the STC-LSTM method.

[0099] Simulation 1: The root mean square error of the pitch signal predicted by the present invention, ARX method and STC-LSTM method in the fault-free period is compared and simulated. The results are as follows: Figure 2 shown.

[0100] Simulation 2: A comparative simulation of the fault detection delays of the present invention, the ARX method, and the STC-LSTM method is performed, and the results are shown in Table 1. Fault detection delay refers to the time delay from the occurrence of a fault to its actual detection.

[0101] Simulation 3: The average calculation time of each data point of the present invention, ARX method and STC-LSTM method is compared and simulated. The results are as follows: Figure 3 shown.

[0102] 4. Simulation results:

[0103] The results of simulation 1 are as follows Figure 2As shown in the figure, the horizontal axis represents the comparison experiment group number, and the vertical axis represents the root mean square error (RMS) of the predicted pitch sensor signal during trouble-free flight. The white graph represents the present invention, while the gray and black graphs represent the ARX method and the STC-LSTM method, respectively. Across 11 comparison experiments, the RMS error of the present invention for predicting pitch sensor signals during trouble-free periods was significantly lower than that of the prior art. The mean RMS error of each method across all 11 experimental groups showed that the ARX method and the STC-LSTM method achieved 3.17 degrees and 5.76 degrees, respectively, while the mean RMS error of the present invention's method was 1.16 degrees. Compared to the other two methods, the present invention improved pitch signal prediction accuracy by 63.4% and 79.9%, respectively.

[0104] The results of Simulation 2 are shown in Table 1. A "-" indicates that this method failed to detect a data failure in the reorganized fault detection experiment. Across 11 comparative experiments, the average fault detection delays for the ARX method and the STC-LSTM method were 6.52s and 3.55s, respectively, while the average fault detection delay for the proposed method was 1.51s, representing reductions of 76.8% and 57.5%, respectively, compared to the existing methods. Furthermore, the fault detection rates for the ARX method and the STC-LSTM method were 45.4% and 27.3%, respectively, while the proposed method achieved a 72.7% fault detection rate, a significant improvement over existing technologies.

[0105]

[0106] Table 1

[0107] The results of simulation three are as follows Figure 3 As shown, the horizontal axis represents the comparative experimental group number, the vertical axis represents the average single-point calculation time of the experimental group, the triangular solid line represents the present invention, the square long dashed line represents the ARX method, the circle short dashed line represents the STC-LSTM method, and the light gray solid line is the average line of the average single-point calculation time of each method in 11 comparative experiments. In the 11 comparative experiments, the average line values ​​of the average single-point calculation time of the ARX method and the STC-LSTM method were 0.041ms and 0.514ms, respectively, while the average line value of the single-point calculation time of the present invention method was 0.235ms. Because the ARX method is based on a linear model, its calculation time is relatively short. Compared with the STC-LSTM method, which relies on a complex neural network model, the average single-point calculation time of the present invention is reduced by 54.3% overall.

[0108] The above simulation analysis demonstrates that the present invention improves the accuracy of drone status sensor signal prediction and data fault detection efficiency, as well as reduces computational time compared to existing technologies that rely on complex neural network models.

[0109] Parts of the present invention that are not described in detail belong to common knowledge among those skilled in the art.

[0110] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Obviously, for professionals in this field, after understanding the content and principles of the present invention, they may make various modifications and changes in form and details without departing from the principles and structure of the present invention. However, these modifications and changes based on the ideas of the present invention are still within the scope of protection of the claims of the present invention.

Claims

1. A method for detecting UAV data faults based on model joint computing, characterized in that: The following steps are involved: (1) Build a drone data fault detection system: Construct a UAV data fault detection system that includes a monitoring platform set up on the ground, a fault detection center on the UAV, a flight controller, and I status sensors. The fault detection center can communicate with the flight controller and the monitoring platform, and I ≥ 1; (2) Obtain training sample set and test sample set: (2a) Select the state sensor data of I at N consecutive time points during a UAV’s fault-free flight from the database At the same time, the fault detection center obtains I state sensor data at K consecutive time points during the real-time flight of the UAV in, represents the i-th state sensor data at N consecutive time points, u i (n) and y i (n) represents the input value and output value of the i-th state sensor data at the n-th moment, N>1000; represents the i-th state sensor data at K consecutive time points, u i (k) and y i (k) represents the input signal value and output signal value of the i-th state sensor data at the k-th moment, 500<K<N; (2b) Through a sliding window of length D Divide into (ND) and (KD) data segments respectively, and The input value and historical output value of each data segment are used as training samples, and the current output value is used as the training label, which consists of (ND) training samples. and (ND) training labels The training sample set and The input value and historical output value of each data segment are used as test samples, and the current output value is used as the test label, which consists of (KD) test samples. and (KD) test labels The test sample set, where 10≤D<25; (3) Constructing a long short-term memory network LSTM model: Construct a long short-term memory network LSTM model F′ consisting of I parallel arranged sub-networks corresponding to the number of state sensors lstm , the i-th sub-network is f′ i , f′ i It is composed of two LSTM layers and a fully connected layer connected in sequence; (4) Long short-term memory network LSTM model F′ lstm Perform iterative training: (4a) Initialize the number of iterations to t, the maximum number of iterations to T, and iterate each sub-network f′ in the tth iteration i The weight parameter And let t = 1, where, Indicates rounding down, S indicates the amount of training data for each iteration, 5<S<20; (4b) S training samples will be randomly selected from the training sample set without replacement as F′ lstm Input of each sub-network f′ i After extracting features from each training sample, perform sequence prediction on the extracted features to obtain S prediction vectors in represents the set of real numbers; (4c) The root mean square error loss function is used, and each training label Y(s) and its corresponding prediction vector Calculate each subnetwork f′ i The loss value is then used to calculate the network weight parameters using the gradient descent algorithm. The gradient of Make updates; (4d) Determine whether t≥T holds. If so, obtain the trained LSTM model F lstm , otherwise, let t = t + 1 and execute step (4b); (5) Build a data fault detection model in the fault detection center: The fault detection center is constructed by parallel arrangement of trained LSTM models F lstm and a linear self-convergent LAR model, and a data fault detection model comprising a joint computing unit, a fault detection unit, and a fault recovery unit cascaded with the output ends of the two models, wherein the fault detection unit comprises a sudden fault detection subunit and a progressive fault detection subunit arranged in parallel; (6) Obtaining drone data fault detection results: (6a) The fault detection center initializes the data fault detection model and the historical fault state variable isFailure, setting isFailure = 0, k = D + 1; (6b) The fault detection center takes the k-time sample X(k) and the label, i.e., the true value of the output signal Y(k), as input to the data fault detection model, the LSTM model F lstm Through X(k), calculate the output signal LSTM model prediction vector As output, at the same time, the LAR model constructs the data matrix Φ(k) through X(k), and calculates the output signal LAR model prediction vector through Φ(k) and the k-1 moment parameter matrix R(k-1) As output, and through Φ(k), and Y(k) iteratively calculate R(k), and then the joint calculation unit is and Y(k) extract the joint calculation matrix set By jointly calculating the matrix The eigenvector corresponding to the maximum eigenvalue is given by Construct the weight matrix W(k), and then pass Y(k) and W(k) calculate the output signal joint prediction vector and the joint residual vector Finally, the fault detection unit normalizes E(k) to obtain the normalized vector The sudden fault detection subunit generates a sudden fault sensor number set J1={j||z j (k)|>3.5}, and at the same time, the progressive fault detection subunit calculates the likelihood vector through Z(k) Generate progressive fault sensor number set J2 = {j|lh j >lh0}, lh0 is the progressive failure threshold; (6c) The fault detection center calculates the fault sensor number set J = J1 ∪ J2 and determines whether J is an empty set. If so, set isFailure = 0 and execute step (6f). Otherwise, execute step (6d). (6d) The fault detection center determines whether isFailure=0. If so, it sets isFailure=1 and executes step (6e). Otherwise, the fault detection center sends a forced landing command to the flight controller, and the UAV ends its flight. (6e) The fault detection center sends a fault warning message to the monitoring platform. At the same time, the fault recovery unit of the data fault detection model generates a correction instruction, which is then sent to the flight controller for execution. (6f) The fault detection center determines whether k≥K holds. If so, it obtains (KD) data fault detection results. Otherwise, it sets k=k+1 and executes step (6b).

2. The method according to claim 1, wherein: The output signal LAR model prediction vector described in step (6b) And the calculation formulas of the parameter matrix R(k) at time k are: R(k)=[δ 1 (k) … d I (k)], Among them, the data vector From X(k) Stretch out, I I is the I-dimensional unit column vector, Represents the LAR model residual value of the i-th state sensor at time k, Δ i (k) represents the change matrix of the i-th state sensor at time k, (·) T represents the transpose operation, represents the convergence matrix of the i-th state sensor at time k, δ i (k) represents the parameter column vector of the i-th state sensor at time k, is a (2D×2D) matrix of all 1s, ||δ i (k)||0 represents the vector δ i (k) is the zero norm, in particular, R(D) is an all-zero matrix, is the identity matrix, δ i (D) is the zero vector.

3. The method according to claim 1, wherein: The joint calculation matrix described in step (6b) And the output signal joint prediction vector The calculation formulas for the joint residual vector E(k) are: Among them, λ i for The maximum eigenvalue of i (k) Corresponding to λ i The characteristic vector of sum(v i (k)) is ν i (k) The cumulative sum of all elements, ω i (k) is the weight vector of the state sensor i at time k.

4. The method according to claim 1, wherein: The normalized vector described in step (6b) Each z i The calculation formula for (k) is: in, and σ i (k) represents the mean and standard deviation of the joint residual of state sensor i at time k, respectively.

5. The method according to claim 1, wherein: The likelihood value lh of the state sensor i at time k in step (6b) i The calculation formulas for (k) and the progressive failure threshold lh0 are: Where ln(·) represents the natural logarithm, β represents the false alarm rate, and γ represents the missed detection rate.

Images