Unmanned aerial vehicle sensor fault monitoring method and system

Abstract

The invention discloses an unmanned aerial vehicle sensor fault monitoring method and system, and the system comprises a sensor module which is used for collecting an input characteristic signal of a measured object, and outputting a physical quantity signal having a certain relation with the measured object; the fault analysis module is used for monitoring the operation state information of each sensor in the sensor module and analyzing whether each sensor is abnormal, and the fault analysis module comprises a wavelet packet decomposition module, a feature extraction module, an SVM analysis module and a control module and is used for analyzing whether the sensor is abnormal or not after the sensor module is abnormal. Controlling the unmanned aerial vehicle to fly and return; an SVM algorithm analysis model based on wavelet packet analysis and an RBF kernel function is constructed and is used for unmanned aerial vehicle sensor fault analysis; the model combines the feature extraction capability of wavelet transform and the classification capability of SVM, energy features in various forms of fault states are associated with fault states, the association is stronger, and different fault states of the sensor can be effectively identified and analyzed.

Description

Technical Field

[0001] The present invention relates to the technical field of fault monitoring, and in particular to a method and system for monitoring faults of unmanned aerial vehicle sensors. Background Art

[0002] Data collection can be effectively performed by installing sensors on drones. However, during the process of data collection by drones, failures in the sensors themselves can cause changes in the data and affect the accuracy of the collected data, which can render the data collected by the drone useless or inaccurate information cannot be collected. If the failure of the data collection sensor is not discovered in time, it can easily lead to flight accidents of the drone, resulting in a waste of the drone's flight resources.

[0003] The patent application with application number CN201910609537.0 discloses a sensor-based monitoring system and method, which extracts the characteristic signal of the sensor for analysis to monitor whether the working condition of the sensor is abnormal. The specific method is to first classify the collected signal data into unbalanced data, and then use wavelet packet decomposition to analyze the working status of the sensor, and then send the result to the user terminal; it has a good effect on high-frequency transient characteristics, but the defect of this method is that its classification is difficult and the calculation complexity is high, which makes it difficult to monitor the data in real time. During the flight of the drone, in order to ensure flight safety, the monitoring system is limited by the deployment restrictions of the body and the need to monitor the sensor data in real time. Therefore, a system and method suitable for drone sensor fault monitoring is needed. Summary of the invention

[0004] In order to overcome the defects of the prior art, the technical problem to be solved by the present invention is to propose a drone sensor fault monitoring method and system to ensure that when a data acquisition sensor carried by the drone fails, causing the data collected by the drone to be useless or unable to be collected, the accuracy of the sensor collection information can be accurately identified, and the flight of the drone can be corrected or recalled in time to ensure the working efficiency and flight safety of the drone.

[0005] To achieve this object, the present invention adopts the following technical solutions:

[0006] The present invention provides a method for monitoring the fault of an unmanned aerial vehicle sensor, comprising the following steps: S00: collecting characteristic signals input by a sensor object of the unmanned aerial vehicle under test, and outputting a physical quantity signal having a definite relationship with the measured quantity, and converting the collected physical quantity signal into an electrical signal through signal processing during output; S10: extracting the characteristic values ​​of the signal output in step S00 by using a wavelet packet, and calculating the energy change before and after the signal; S20: analyzing the fault of the sensor by using an SVM according to the energy change before and after the signal; S30: after detecting that an abnormality occurs in the sensor of the unmanned aerial vehicle under test, controlling the unmanned aerial vehicle to correct the flight parameters and then fly or return.

[0007] A drone sensor fault monitoring system is provided to implement the above fault monitoring method, comprising the following modules: a sensor module, for collecting input characteristic signals of a measured object, and outputting physical quantity signals having a definite relationship with the measured object; a fault analysis module, for monitoring the operating status information of each sensor in the sensor module, and analyzing whether each sensor is abnormal, the fault analysis module comprising a wavelet packet decomposition module, a feature extraction module, and an SVM analysis module; a control module, for controlling the drone to fly and return after an abnormality occurs in the sensor module; during the operation of the system, the wavelet packet decomposition module performs wavelet decomposition on the input signal and calculates the relative energy characteristic distribution of each frequency band; the feature extraction module performs feature vector input into the SVM analysis module for the relative energy of each frequency band after wavelet packet decomposition; the SVM analysis module analyzes the input feature vector, determines the sensor status, and outputs the identification status to the control module.

[0008] The beneficial effects of the present invention are:

[0009] The present invention constructs an SVM algorithm analysis model based on wavelet packet analysis and RBF kernel function for drone sensor fault analysis. In the drone sensor scenario, for the six common fault types of drone sensors (periodic faults, drift, deviation, short circuit, open circuit, etc.), feature vectors are constructed through energy distribution differences, which reflects the scenario adaptability; this model combines the feature extraction capability of wavelet transform and the classification capability of SVM, and associates the energy characteristics of various fault states with the fault state, and its correlation is stronger, which can effectively identify and analyze different fault states of the sensor. During the drone flight phase, the fault analysis module will continuously collect the energy signal of the data acquisition sensor carried by the drone in real time and perform feature extraction, and then perform sensor fault judgment through the SVM calculation decision model; at the same time, the conservation of energy is verified by Parseval's theorem, and the feature vector is constructed based on multi-band energy distribution, which enhances the physical significance of fault characterization. BRIEF DESCRIPTION OF THE DRAWINGS

[0010] Figure 1It is a schematic diagram of wavelet packet decomposition in a method for monitoring faults of a drone sensor provided in a specific embodiment of the present invention;

[0011] Figure 2 It is the energy characteristic value distribution diagram under normal state;

[0012] Figure 3 It is the energy characteristic value distribution diagram under the deviation fault state;

[0013] Figure 4 It is the energy characteristic value distribution diagram under the circuit breaker fault state;

[0014] Figure 5 It is the energy characteristic value distribution diagram under the short-circuit fault state;

[0015] Figure 6 It is the energy characteristic value distribution diagram under periodic fault state;

[0016] Figure 7 It is the energy characteristic value distribution diagram under drift fault state;

[0017] Figure 8 It is a schematic flow chart of the steps of a method for monitoring a sensor fault of an unmanned aerial vehicle provided in a specific embodiment of the present invention. DETAILED DESCRIPTION

[0018] The technical solution of the present invention is further described below with reference to the accompanying drawings and through specific implementation methods.

[0019] like Figure 8 As shown, the present invention provides a method for monitoring sensor faults of unmanned aerial vehicles, comprising the following steps: S00: collecting characteristic signals input by the sensor object of the unmanned aerial vehicle under test, and outputting physical quantity signals having a definite relationship with the measured quantity, and converting the collected physical quantity signals into electrical signals through signal processing when outputting; S10: extracting the eigenvalues ​​of the signal output in step S00 using wavelet packets, and calculating the energy change before and after the signal; S20: analyzing the fault of the sensor using SVM according to the energy change before and after the signal; S30: after detecting that the sensor of the unmanned aerial vehicle under test is abnormal, controlling the unmanned aerial vehicle to correct the flight parameters and then fly or return.

[0020] A drone sensor fault monitoring system is provided to implement the above fault monitoring method, comprising the following modules: a sensor module, for collecting input characteristic signals of a measured object, and outputting physical quantity signals having a definite relationship with the measured object; a fault analysis module, for monitoring the operating status information of each sensor in the sensor module, and analyzing whether each sensor is abnormal, the fault analysis module comprising a wavelet packet decomposition module, a feature extraction module, and an SVM analysis module; a control module, for controlling the drone to fly and return after an abnormality occurs in the sensor module; during the operation of the system, the wavelet packet decomposition module performs wavelet decomposition on the input signal and calculates the relative energy characteristic distribution of each frequency band; the feature extraction module performs feature vector input into the SVM analysis module for the relative energy of each frequency band after wavelet packet decomposition; the SVM analysis module analyzes the input feature vector, determines the sensor status, and outputs the identification status to the control module.

[0021] In summary, this scheme mainly converts the measured sensor physical quantity signal into an electrical signal (the electrical signal can be enhanced or amplified when necessary), and then processes the collected signal in the following order: first extract the eigenvalue through wavelet decomposition, then calculate the energy of each frequency band to construct a eigenvector, and then analyze the eigenvector using SVM, and finally determine the working status of the sensor.

[0022] In the process of using wavelet decomposition and calculating the energy of each frequency band to construct the feature vector:

[0023] In step S10, the wavelet packet analysis is performed to extract the eigenvalues ​​by the following steps:

[0024] S11: Perform wavelet decomposition on the sensor signal to obtain the decomposition coefficient and energy of each frequency band;

[0025] The energy calculation formula for each frequency band is:

[0026]

[0027] Where: E j is the energy in the frequency band; x jk Represents the amplitude of the discrete points of the signal in each frequency band, i.e., the decomposition coefficient;

[0028] S12: extract these energies as feature vectors;

[0029] So, let the output signal f(t) be, according to the wavelet analysis recursive formula:

[0030]

[0031] In the formula, α is the scale factor; τ is the translation factor; t is the time, is the complex conjugate of the signal, WT f (a,τ) are the transformation coefficients;

[0032] In order to reduce the amount of calculation during wavelet decomposition transformation, first let Get the discrete wavelet transform:

[0033]

[0034] Among them, j is the scale index; k is the translation index; WT f (j, k) Wavelet packet decomposition to extract energy features;

[0035] After discrete wavelet transform, the coefficients at each scale represent the energy distribution of the signal at that scale. By squaring these coefficients and summing them, we can get the energy of the signal in each frequency band. The energy characteristics can reflect the intensity changes of the signal in different frequency bands. By using these energy values ​​as feature vectors for subsequent sensor fault analysis, the energy of the output signal f(t) in the time domain is X(t), and the expression of X(t) is:

[0036]

[0037] in: is a constant;

[0038] In order to improve the accuracy and stability of the subsequent SVM classification analysis, the extracted feature vectors need to be normalized, that is, according to the energy integral equation of Parseval's energy conservation theorem:

[0039]

[0040] Among them: j is the scale index, k is the translation index, c j,k is a wavelet transform system;

[0041] Thus, in S12, the energy distribution of each frequency band of the sensor output signal under six states, including normal state, periodic fault state, drift fault state, deviation fault state, short circuit fault state, and open circuit fault state, is analyzed by wavelet, and then the feature vectors under different states are extracted, and the frequency band energy values ​​are constructed into a feature vector according to the node order.

[0042] Furthermore, in order to use SVM to analyze the faults of sensors, it is necessary to build an SVM decision function model to accurately determine the fault type. This model collects the feature vectors of the signal data of the sensor in six states, including normal state, periodic fault state, drift fault state, deviation fault state, short circuit fault state, and open circuit fault state, before the UAV sensor is mounted on the UAV, and then inputs the feature vectors into the support vector machine SVM for classification training.

[0043] First, let {x i ,y i} are two separable samples, namely, {x i ,y i} is a training sample, y i is the class label of the sample, x i is the sample feature vector (i.e., T i ); Assuming that there is a hyperplane that can completely separate these two types of samples, the classification hyperplane in n-dimensional Euclidean space can be expressed as:

[0044] f(x)=<w,x> +b

[0045] in:<w,x> is the inner product of two vectors; w is the normal vector of the hyperplane; x is the new sample to be classified and analyzed; b is the displacement;

[0046] Therefore, the data information can be divided into the optimization problem of linear separable classification and the optimization problem of nonlinear separable classification:

[0047] (1) Optimization of linearly separable classification. When the data is linearly separable, finding the optimal hyperplane is transformed into a quadratic programming problem as shown in the following formula:

[0048] Minimize:

[0049] Satisfy: y i ( <w,x i >+b)≥1(i=1,2,…,N);

[0050] Introducing the Lagrange operator a i ≥0(i=1,2,…,N) to obtain the dual form of the optimization problem,

[0051] maximize:

[0052] To satisfy:

[0053] in,<w,x> is the inner product of two vectors; w is the normal vector of the hyperplane; b is the displacement;

[0054] (2) Optimization of nonlinear separable classification. The basic idea is to map the nonlinear sample data to another feature space through a certain function so that it can be linearly separable in this space, that is, to use the kernel function (select radial basis kernel function) for processing. In the feature space, the optimal classification surface can be obtained according to the linear SVM training algorithm to realize the classification of sample data. That is, when processing nonlinear data, the data is first linearized: x→ψ(x). Then the expression of the optimization problem is:

[0055] maximize:

[0056] To satisfy:

[0057] Furthermore, we use the kernel function K(x,x i ) to replace the inner product in the feature space <ψ(x i ),ψ(x j )>, so in step S20, the final expression of the decision function of SVM is:

[0058]

[0059] In the formula, x is the new sample to be classified and analyzed; {x i ,y i} is a training sample, y i is the class label of the sample, xi is the sample feature vector; l is the number of training samples, is the Lagrange multiplier learned through training, indicating the importance of each support vector; b * is the bias term, which determines the location of the decision boundary; K(x,x i ) is the kernel function, which is a "vote" contributed by each training sample to the classification of the new sample to be analyzed. The weight of the "vote" is determined by the Lagrange multiplier and the class label yi of the sample, and is determined by the kernel function K(x,x i ) represents the similarity between the new sample and the training sample. The higher the similarity, the greater the "vote". The final result is the sum of the "votes" plus the bias term b. * Together they determine the category of the new sample, thus determining the fault state of the sensor to be tested.

[0060] In this way, a SVM algorithm analysis model based on wavelet packet analysis and RBF kernel function is formed for UAV sensor fault analysis. i It can reflect the characteristic changes of sensor signals in different states, that is, SVM is based on these characteristic vectors T i After training, the sensor can be distinguished in different states, and the feature vectors T in different states of the six states mentioned in this case (normal state, periodic fault state, drift fault state, deviation fault state, short circuit fault state, open circuit fault state) are i Different from this, during the flight phase of the drone, the energy signals of the data acquisition sensors carried by the drone will be continuously collected and features will be extracted, and then the sensor faults will be judged through the SVM calculation decision model.

[0061] In summary, in actual use, for new sensor signals, wavelet packet analysis is also performed through the above steps to extract new feature vectors T, and then it is input into the trained SVM model. SVM determines which category the feature vector belongs to based on the decision function model. The decision function of SVM is based on the feature vector T for classification, and the form of the decision function is:

[0062]

[0063] Among them, T is the new sample to be classified and analyzed; T i is the feature vector of the training sample, y i is the class label of the sample (+1 and -1), l is the number of training samples, is the Lagrange multiplier learned through training, indicating the importance of each support vector; b * is the bias term, which determines the location of the decision boundary; K(T,T i ) is the kernel function, which is used to calculate the similarity between feature vectors.

[0064] In this way, the fault type can be analyzed and determined through the decision function, specifically:

[0065] When f(T)>0, the new sample is classified into the category and T in the decision function i The categories are the same;

[0066] When f(T) < 0, the new sample is classified into the category and T in the decision function i The categories are different;

[0067] When f(T) = 0, the new sample is on the classification decision boundary and usually requires further processing.

[0068] In this way, after the fault analysis module judges and analyzes the fault of the drone sensor, the result will be directly output to the control module. When the control module recognizes that the sensor has a periodic fault state, a drift fault state, a deviation fault state, a short circuit fault state, or an open circuit fault state, the control module will send an alarm to the background terminal and prompt whether to make corrections or whether the drone should return. The background terminal can issue corresponding instructions. After receiving the instructions, the drone will perform corresponding operations on the drone through the control module.

[0069] The present invention is described by preferred embodiments, and those skilled in the art will appreciate that various changes or equivalent substitutions may be made to these features and embodiments without departing from the spirit and scope of the present invention. The present invention is not limited to the specific embodiments disclosed herein, and other embodiments falling within the claims of this application are within the scope of protection of the present invention.

Claims

1. A method for monitoring failure of a drone sensor, characterized in that: The following steps are involved: S00: Collect characteristic signals input by the sensor object of the measured UAV, and output physical quantity signals that have a definite relationship with the measured quantity. When outputting, the collected physical quantity signals are converted into electrical signals through signal processing; S10: extracting the eigenvalues ​​of the signal output in step S00 using wavelet packets, and calculating the energy change of the signal before and after; S20: Analyze the sensor failure using SVM according to the change of signal energy before and after; S30: After detecting that the sensor of the drone under test is abnormal, the drone is controlled to calibrate the flight parameters and then fly or return.

2. The method for monitoring a drone sensor failure according to claim 1, characterized in that: In step S10, the wavelet packet analysis is performed to extract the eigenvalues ​​by the following steps: S11: Perform wavelet decomposition on the sensor signal to obtain the decomposition coefficient and energy of each frequency band; The energy calculation formula for each frequency band is: Where: E j is the energy in the frequency band; x jk Represents the amplitude of the discrete points of the signal in each frequency band, i.e., the decomposition coefficient; S12: Extract these energies as feature vectors.

3. The method for monitoring a drone sensor failure according to claim 2, characterized in that: Assume the output signal f(t), according to the wavelet analysis recursive formula: In the formula, α is the scale factor; τ is the translation factor; t is the time, is the complex conjugate of the signal, WT f (a,τ) are the transformation coefficients; make Get the discrete wavelet transform: Among them, j is the scale index; k is the translation index; WT f (j, k) Wavelet packet decomposition to extract energy features; The energy of the output signal f(t) in the time domain is X(t), and the expression of X(t) is: in: is a constant; The energy integral equation according to Parseval's energy conservation theorem is: Among them: j is the scale index, k is the translation index, c j,k It is a wavelet transform system.

4. The method for monitoring a drone sensor failure according to claim 3, characterized in that: In S12, the energy distribution of each frequency band of the sensor output signal under six states, including normal state, periodic fault state, drift fault state, deviation fault state, short circuit fault state, and open circuit fault state, is analyzed by wavelet, and then the feature vectors under different states are extracted, and the frequency band energy values ​​are constructed into a feature vector according to the node order.

5. The method for monitoring a drone sensor failure according to claim 4, characterized in that: In step S20, the decision function of SVM is expressed as: Where x is the new sample vector to be classified; {x i ,y i } is a training sample, y i is the class label of the sample, x i is the sample feature vector; l is the number of training samples, is the Lagrange multiplier learned through training, indicating the importance of each support vector; b * is the bias term, which determines the location of the decision boundary; K(x,T i ) is the kernel function.

6. The method for monitoring failure of a drone sensor according to claim 5, characterized in that: {x i ,y i } are two separable samples. Assuming there is a hyperplane that can completely separate these two types of samples, the classification hyperplane in n-dimensional Euclidean space can be expressed as: f(x)=<w,x> +b in:<w,x> is the inner product of two vectors; w is the normal vector of the hyperplane; b is the displacement.

7. The method for monitoring a drone sensor failure according to claim 6, characterized in that: When the data is linearly separable, finding the optimal hyperplane is transformed into a quadratic programming problem as shown in the following formula: Minimize: Satisfy: y i ( <w,x i >+b)≥1(i=1,2,…,N); Introducing the Lagrange operator a i ≥0(i=1,2,…,N) to obtain the dual form of the optimization problem, maximize: To satisfy: in,<w,x> is the inner product of two vectors; w is the normal vector of the hyperplane; b is the displacement.

8. The method for monitoring failure of a drone sensor according to claim 6, characterized in that: When the data is nonlinearly separable, first linearize the data: x→ψ(x), then the expression of the optimization problem is: maximize: To satisfy:

9. A drone sensor fault monitoring system, used to implement the drone sensor fault monitoring method according to any one of claims 1 to 8, characterized in that: Includes the following modules: The sensor module is used to collect the input characteristic signal of the measured object and output a physical quantity signal that has a definite relationship with the measured quantity; A fault analysis module, used to monitor the operating status information of each sensor in the sensor module and analyze whether each sensor is abnormal, the fault analysis module includes a wavelet packet decomposition module, a feature extraction module, and a SVM analysis module; The control module is used to control the UAV to fly and return after an abnormality occurs in the sensor module.

10. The drone sensor fault monitoring system according to claim 9, characterized in that: The wavelet packet decomposition module performs wavelet decomposition on the input signal and calculates the relative energy characteristic distribution of each frequency band; the feature extraction module inputs the feature vector of the relative energy of each frequency band after wavelet packet decomposition into the SVM analysis module; the SVM analysis module analyzes the input feature vector, determines the sensor status, and outputs the identification status to the control module.

Images