CN113505445B - Real-time fault diagnosis method and system based on sequential random forest - Google Patents

Real-time fault diagnosis method and system based on sequential random forest Download PDF

Info

Publication number
CN113505445B
CN113505445B CN202111058850.3A CN202111058850A CN113505445B CN 113505445 B CN113505445 B CN 113505445B CN 202111058850 A CN202111058850 A CN 202111058850A CN 113505445 B CN113505445 B CN 113505445B
Authority
CN
China
Prior art keywords
fault
residual signal
diagnosed
time
random forest
Prior art date
Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
Active
Application number
CN202111058850.3A
Other languages
Chinese (zh)
Other versions
CN113505445A (en
Inventor
宋佳
艾绍洁
赵凯
苏江城
尚维泽
蔡国飙
Current Assignee (The listed assignees may be inaccurate. Google has not performed a legal analysis and makes no representation or warranty as to the accuracy of the list.)
Beihang University
Original Assignee
Beihang University
Priority date (The priority date is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the date listed.)
Filing date
Publication date
Application filed by Beihang University filed Critical Beihang University
Priority to CN202111058850.3A priority Critical patent/CN113505445B/en
Publication of CN113505445A publication Critical patent/CN113505445A/en
Application granted granted Critical
Publication of CN113505445B publication Critical patent/CN113505445B/en
Active legal-status Critical Current
Anticipated expiration legal-status Critical

Links

Images

Classifications

    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06F—ELECTRIC DIGITAL DATA PROCESSING
    • G06F30/00—Computer-aided design [CAD]
    • G06F30/10—Geometric CAD
    • G06F30/15—Vehicle, aircraft or watercraft design
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06F—ELECTRIC DIGITAL DATA PROCESSING
    • G06F18/00—Pattern recognition
    • G06F18/20—Analysing
    • G06F18/21—Design or setup of recognition systems or techniques; Extraction of features in feature space; Blind source separation
    • G06F18/214—Generating training patterns; Bootstrap methods, e.g. bagging or boosting
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06F—ELECTRIC DIGITAL DATA PROCESSING
    • G06F18/00—Pattern recognition
    • G06F18/20—Analysing
    • G06F18/24—Classification techniques
    • G06F18/243—Classification techniques relating to the number of classes
    • G06F18/24323—Tree-organised classifiers
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06F—ELECTRIC DIGITAL DATA PROCESSING
    • G06F30/00—Computer-aided design [CAD]
    • G06F30/20—Design optimisation, verification or simulation
    • G06F30/27—Design optimisation, verification or simulation using machine learning, e.g. artificial intelligence, neural networks, support vector machines [SVM] or training a model
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
    • G06N3/00—Computing arrangements based on biological models
    • G06N3/004—Artificial life, i.e. computing arrangements simulating life
    • G06N3/006—Artificial life, i.e. computing arrangements simulating life based on simulated virtual individual or collective life forms, e.g. social simulations or particle swarm optimisation [PSO]
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06F—ELECTRIC DIGITAL DATA PROCESSING
    • G06F2111/00—Details relating to CAD techniques
    • G06F2111/08—Probabilistic or stochastic CAD

Landscapes

  • Engineering & Computer Science (AREA)
  • Theoretical Computer Science (AREA)
  • Physics & Mathematics (AREA)
  • Evolutionary Computation (AREA)
  • General Physics & Mathematics (AREA)
  • Data Mining & Analysis (AREA)
  • General Engineering & Computer Science (AREA)
  • Geometry (AREA)
  • Artificial Intelligence (AREA)
  • Computer Vision & Pattern Recognition (AREA)
  • Life Sciences & Earth Sciences (AREA)
  • Evolutionary Biology (AREA)
  • Bioinformatics & Computational Biology (AREA)
  • Bioinformatics & Cheminformatics (AREA)
  • Computer Hardware Design (AREA)
  • Software Systems (AREA)
  • Mathematical Physics (AREA)
  • Aviation & Aerospace Engineering (AREA)
  • Molecular Biology (AREA)
  • Computing Systems (AREA)
  • Computational Linguistics (AREA)
  • Biophysics (AREA)
  • Automation & Control Theory (AREA)
  • General Health & Medical Sciences (AREA)
  • Biomedical Technology (AREA)
  • Computational Mathematics (AREA)
  • Mathematical Analysis (AREA)
  • Mathematical Optimization (AREA)
  • Pure & Applied Mathematics (AREA)
  • Health & Medical Sciences (AREA)
  • Medical Informatics (AREA)
  • Complex Calculations (AREA)

Abstract

The invention provides a real-time fault diagnosis method and a system based on a sequential random forest, which comprises the following steps: acquiring a residual signal to be detected of the nonlinear system to be diagnosed; extracting a time-frequency characteristic vector of a residual signal to be detected based on wavelet packet decomposition; substituting the time-frequency characteristic vector into a trained random forest fault separator, and carrying out sequential probability ratio inspection on a nonlinear system to be diagnosed to obtain target fault information; the target fault information comprises a target fault type and target fault occurrence time; and substituting the time-frequency characteristic vector and the target fault type into a trained regression random forest fault identifier to obtain fault size information corresponding to the target fault type. The invention solves the technical problems that the real-time fault diagnosis cannot be realized and the missed diagnosis rate and the misdiagnosis rate are high in the prior art.

Description

Real-time fault diagnosis method and system based on sequential random forest
Technical Field
The invention relates to the technical field of unmanned aerial vehicle fault diagnosis, in particular to a real-time fault diagnosis method and system based on sequential random forests.
Background
The existing methods for diagnosing faults through quantitative analysis are mainly classified into methods based on analytical models and methods based on data driving. The core idea of fault diagnosis based on the analytical model is to estimate the system variable observed value through the state and take the residual error generated between the system variable observed value and the system state variable true value as the diagnosis source and basis. The residual is typically obtained in two ways: the key point of the former is to design an observer and a filter with strong filtering and high dynamic response characteristics based on a state observer and based on a reference model, and the difficulty of the latter is to construct an accurate high-dimensional nonlinear system nominal model. The estimation precision of the state observer and the design precision of the reference model directly influence the accuracy of fault diagnosis to a certain extent.
In order to realize active fault-tolerant control of a system, fault diagnosis needs to realize two main aspects of functions: firstly, determining the type and severity of the fault, namely, having fault separation and identification capability; and secondly, a real-time system state monitoring mechanism is provided, and the occurrence of faults can be rapidly detected. The data driving method based on machine learning does not depend on a model, and the capability of the fault diagnosis system can be greatly improved by learning and extracting features from mass data to fit the nonlinear mapping relation between residual observation data and the system state. Random forests, which are typical methods for ensemble learning, have excellent overfitting resistance and noise insensitivity by introducing randomness. Compared with a neural network, the random forest has few hyper-parameters, does not need feature subset selection, can process high-dimensional data generated by a complex system, uses unbiased estimation for training and learning, and does not need normalized processing and extra verification set setting. The method is simple to implement and can realize rapid fault diagnosis. In addition, the random forest has the capability of fitting regression and isolating classification, wherein the random forest can realize the identification of the size of the fault, and the random forest can realize the classification of the fault mode. Although the random forest has the advantages, the dependence of state data on a space domain is only considered, the historical data relevance on a time domain is ignored, the diagnosis based on a single sample is limited by random disturbance such as sensor noise, real-time fault diagnosis cannot be realized, and the missed diagnosis rate and the misdiagnosis rate are high.
Disclosure of Invention
In view of the above, the present invention provides a real-time fault diagnosis method and system based on sequential random forest to solve the technical problems of the prior art that real-time fault diagnosis cannot be realized and the missed diagnosis rate and the misdiagnosis rate are high.
In a first aspect, an embodiment of the present invention provides a real-time fault diagnosis method based on a sequential random forest, including: acquiring a residual signal to be detected of the nonlinear system to be diagnosed; extracting a time-frequency characteristic vector of the residual signal to be detected based on wavelet packet decomposition; substituting the time-frequency characteristic vector into a trained random forest fault separator, and carrying out sequential probability ratio inspection on the nonlinear system to be diagnosed to obtain target fault information; the target fault information comprises a target fault type and target fault occurrence time; and substituting the time-frequency characteristic vector and the target fault type into a trained regression random forest fault identifier to obtain fault size information corresponding to the target fault type.
Further, acquiring a residual signal to be detected of the nonlinear system to be diagnosed, including: establishing a nominal model of the nonlinear system to be diagnosed, and acquiring a nominal system residual error signal of the nominal model; acquiring an observation residual signal of the nonlinear system to be diagnosed; and obtaining the residual signal to be detected based on the nominal system residual signal and the observation residual signal.
Further, the method further comprises: obtaining the residual signal to be measured based on the nominal system residual signal and the observation residual signal, including: determining the residual signal to be measured by the following formula: e.g. of the typel=es-g(en,tr,tc) (ii) a Wherein e islFor the residual signal to be measured, esFor the observed residual signal, enFor the nominal system residual signal, g (e)n,tr,tc) As a nominal residual injection function, tr,tcRespectively command signal change time instant and nominal residual injection duration.
Further, the method further comprises: acquiring a residual error sample training set of the nonlinear system to be diagnosed; the residual sample training set comprises residual sample data in a normal mode and residual sample data in a fault mode; and training a preset random forest fault separator based on the residual error sample training set, and optimizing the structural parameter combination of the preset random forest fault separator by using a sparrow optimization algorithm to obtain the trained random forest fault separator corresponding to the optimal structural parameter combination.
Further, the method further comprises: acquiring a fault type training set of the nonlinear system to be diagnosed; the fault type training set comprises residual sample data of the nonlinear system to be diagnosed under each fault type; and training a preset regression random forest fault identifier based on the fault type training set, and optimizing the structural parameter combination of the preset regression random forest fault identifier by using a sparrow optimization algorithm to obtain the trained regression random forest fault identifier corresponding to the optimal structural parameter combination.
Further, substituting the time-frequency characteristic vector into a random forest fault separator after training, and performing sequential probability ratio test on the nonlinear system to be diagnosed to obtain target fault information, wherein the steps comprise: substituting the time-frequency characteristic vector into a random forest fault separator after training, and obtaining the probability output of each fault type corresponding to the current state according to the forest voting result; calculating sequential probability ratio statistics for each fault type based on the probability outputs; and performing sequential probability ratio test on the nonlinear system to be diagnosed based on the sequential probability ratio statistic of each fault type to obtain target fault information.
Further, calculating sequential probability ratio statistics for each fault type based on the probability outputs, including: calculating a sequential probability ratio statistic for each fault type by:
Figure DEST_PATH_IMAGE001
(ii) a Wherein the content of the first and second substances,
Figure 614285DEST_PATH_IMAGE002
sequential probability ratio statistics of k fault types occurring at the time t of the nonlinear system to be diagnosed,
Figure DEST_PATH_IMAGE003
for the nonlinear system to be diagnosed at t-tsSequential probability ratio statistic, t, for the k-th fault type occurring at a timesFor the truncation time interval of the residual signal to be measured,Pk(Xt) Probability P of occurrence of k-th fault type at t moment of the nonlinear system to be diagnosed0(Xt) And the probability that the nonlinear system to be diagnosed is normal at the moment t is obtained.
In a second aspect, an embodiment of the present invention further provides a real-time fault diagnosis system based on a sequential random forest, including: the system comprises an acquisition module, an extraction module, a fault separation module and a fault identification module; the acquisition module is used for acquiring a residual signal to be detected of the nonlinear system to be diagnosed; the extraction module is used for extracting the time-frequency characteristic vector of the residual signal to be detected based on wavelet packet decomposition; the fault separation module is used for substituting the time-frequency characteristic vector into a trained random forest fault separator and carrying out sequential probability ratio inspection on the nonlinear system to be diagnosed to obtain target fault information; the target fault information comprises a target fault type and target fault occurrence time; and the fault identification module is used for substituting the time-frequency characteristic vector and the target fault type into a regression random forest fault identifier after training to obtain fault size information corresponding to the target fault type.
In a third aspect, an embodiment of the present invention further provides an electronic device, which includes a memory, a processor, and a computer program stored in the memory and executable on the processor, where the processor implements the steps of the method according to the first aspect when executing the computer program.
In a fourth aspect, the present invention further provides a computer-readable medium having non-volatile program code executable by a processor, where the program code causes the processor to execute the method according to the first aspect.
The invention provides a real-time fault diagnosis method and a real-time fault diagnosis system based on sequential random forests, which can realize that data is continuously increased until a fault is detected in the fault detection process by combining a random forest fault separator and sequential probability ratio detection, can realize real-time self-adaptive multi-sample fault diagnosis, and solves the technical problems that the real-time fault diagnosis cannot be realized and the missed diagnosis rate and the misdiagnosis rate are high in the prior art.
Drawings
In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the drawings used in the description of the embodiments or the prior art will be briefly described below, and it is obvious that the drawings in the following description are some embodiments of the present invention, and other drawings can be obtained by those skilled in the art without creative efforts.
Fig. 1 is a flowchart of a real-time fault diagnosis method based on a sequential random forest according to an embodiment of the present invention;
fig. 2 is a construction diagram of a four-rotor unmanned aerial vehicle system fault diagnosis residual observer provided by the embodiment of the invention;
fig. 3 is a schematic diagram of a real-time fault diagnosis system based on a sequential random forest according to an embodiment of the present invention;
fig. 4 is a schematic diagram of another real-time fault diagnosis system based on sequential random forests according to an embodiment of the present invention.
Detailed Description
The technical solutions of the present invention will be described clearly and completely with reference to the accompanying drawings, and it should be understood that the described embodiments are some, but not all embodiments of the present invention. All other embodiments, which can be derived by a person skilled in the art from the embodiments given herein without making any creative effort, shall fall within the protection scope of the present invention.
The first embodiment is as follows:
fig. 1 is a flowchart of a method for real-time fault diagnosis based on sequential random forests, which is applied to fault detection of a nonlinear system according to an embodiment of the present invention. As shown in fig. 1, the method specifically includes the following steps:
and step S102, acquiring a residual signal to be detected of the nonlinear system to be diagnosed.
And step S104, extracting the time-frequency characteristic vector of the residual signal to be detected based on wavelet packet decomposition.
And S106, substituting the time-frequency characteristic vector into the trained random forest fault separator, and carrying out sequential probability ratio inspection on the nonlinear system to be diagnosed to obtain target fault information. The target fault information comprises a target fault type and a target fault occurrence time.
And S108, substituting the time-frequency characteristic vector and the target fault type into the trained regression random forest fault identifier to obtain fault size information corresponding to the target fault type.
The invention provides a real-time fault diagnosis method based on sequential random forests, which can realize that data is continuously increased until a fault is detected in the fault detection process by combining a random forest fault separator and sequential probability ratio detection, can realize real-time self-adaptive multi-sample fault diagnosis, and solves the technical problems that the real-time fault diagnosis cannot be realized and the missed diagnosis rate and the misdiagnosis rate are high in the prior art.
Optionally, step S102 further includes the steps of:
and step S1021, establishing a nominal model of the nonlinear system to be diagnosed, and acquiring a nominal system residual signal of the nominal model.
Step S1022, an observation residual signal of the nonlinear system to be diagnosed is obtained.
And step S1023, obtaining a residual signal to be measured based on the nominal system residual signal and the observation residual signal. Specifically, the residual signal to be measured is determined by the following equation: e.g. of the typel=es-g(en,tr,tc) (ii) a Wherein e islFor the residual signal to be measured, esTo observe the residual signal, enFor a nominal system residual signal, g (e)n,tr,tc) As a nominal residual injection function, tr,tcRespectively command signal change time instant and nominal residual injection duration.
In the following, the embodiment of the present invention takes a quad-rotor unmanned aerial vehicle system as an example of a nonlinear system to be diagnosed, and describes a specific process for acquiring a residual signal to be detected. It should be noted that the quad-rotor drone system in the embodiment of the present invention satisfies the following conditions:
(1) the four-rotor unmanned aerial vehicle body is a rigid body, and deformation and quality change do not occur in the flight process;
(2) the mass distribution of the quad-rotor unmanned aerial vehicle is uniform, and the gravity center is positioned at the geometric center of the unmanned aerial vehicle body;
(3) the quad-rotor unmanned aerial vehicle flies at low altitude and in short distance, so that the surface of the earth is seen as a plane, and the gravity acceleration is unchanged in the flying process;
(4) the influence of earth rotation and revolution on the motion of the quad-rotor unmanned aerial vehicle is ignored, and the ground coordinate system is an inertial coordinate system.
Selecting a coordinate system of the body, origin ObAt the centroid of quadrotor unmanned plane, xbAxis directed to the head, ybAxis pointing to the right of the fuselage, zbPointing to the lower part of the fuselage. Selecting a ground coordinate system, origin OeAt any point on the ground, xeThe axis being in a horizontal plane and pointing in a certain direction, yeThe axis being perpendicular to the ground and directed towards the centre of the earth, zeThe axial direction is determined by the right hand rule.
Defining Euler angle to represent flight attitude of quad-rotor unmanned aerial vehicle: the roll angle phi is the angle of the body coordinate system rotating around the x-axis (roll right is a positive direction), the pitch angle theta is the angle of the body coordinate system rotating around the y-axis (upward along the nose direction is a positive direction), and the yaw angle psi is the angle of the body coordinate system rotating around the z-axis (yaw right is a positive direction).
A mathematical model of the quad-rotor unmanned aerial vehicle, which is composed of a translation model and a rotation model, is constructed as follows:
Figure 219489DEST_PATH_IMAGE004
(1)
in the formula (I), the compound is shown in the specification,x,y,zis the three-axis direction of the ground coordinate system,J x ,J y ,J z is a three-axis moment of inertia,j r is the rotary inertia of the rotor, U2The rolling moment, U, being the rotation of the body about its x-axis3To wind the machine bodyyPitching moment of shaft rotation, U4Yawing moment, omega, for quad-rotor unmanned aerial vehicles1,ω2,ω3,ω4At four rotary-wing speeds, U1As a total tension, Dx,Dy,DzThe air resistance coefficients of the three-axis direction are respectively, m is the mass of the quad-rotor unmanned aerial vehicle, and g is the gravity acceleration.
To model the failure of a sensor accelerometer measuring pitch angle, equation (1) is simplified to a longitudinal channel model without loss of generality:
Figure 170127DEST_PATH_IMAGE005
(2)
a common sensor fault model is established as follows:
constant value deviation fault:
Figure 114950DEST_PATH_IMAGE006
；
and (3) jamming failure:
Figure 467434DEST_PATH_IMAGE007
；
gain variation failure:
Figure 409982DEST_PATH_IMAGE008
；
outlier data failure:
Figure 660966DEST_PATH_IMAGE009
。
sensor faults that the solution is capable of diagnosing here include, but are not limited to, the above 4 faults.
In order to better realize the attitude control of the quad-rotor unmanned aerial vehicle, a Linear Active Disturbance Rejection Control (LADRC) controller is adopted, and a pitch angle residual signal is obtained from a linear extended observer (LESO) of the controller. Based on vertical passageway (2) formula design third-order LESO of four rotor unmanned aerial vehicle:
Figure 81583DEST_PATH_IMAGE010
(3)
where e is the error between observed quantity z1 and controlled object output y, y is the output of the controller, and β1，β2，β3Is an observer adjustable parameter. Fig. 2 is a construction diagram of a four-rotor unmanned aerial vehicle system fault diagnosis residual observer provided by the embodiment of the invention.
Taking the formula (1) as a nominal model of the quad-rotor unmanned aerial vehicle system, inputting an observation residual signal of the quad-rotor unmanned aerial vehicle system and a residual signal of the nominal model into a simulation unmanned aerial vehicle system model LESO, and obtaining a final residual signal to be detected:
el=es-g(en,tr,tc)(4)
in the formula, el,es,enRespectively a residual signal to be measured, an observation residual signal and a nominal system residual signal. g (e, t)1,t2)=e,t1<t<t2Is a nominal residual injection function, where tr,tcRespectively, the command signal change time instant and the nominal residual injection duration, the latter being determined empirically from the residual behavior time resulting from the control change.
And then, carrying out 3-layer wavelet packet decomposition on the residual signal to be detected, reconstructing the decomposed signals of 8 frequency bands, and calculating 9 statistical parameters including waveform indexes, margin indexes, pulse indexes, peak indexes, absolute mean values, mean square errors, skewness, kurtosis and square root amplitudes of all reconstructed signals, thereby obtaining the extracted 72-dimensional characteristic vector of the residual signal sample.
Optionally, the method provided by the embodiment of the present invention further includes a training and optimizing process for the preset random forest fault separator, where the process is an offline training part. Specifically, the method comprises the following steps:
acquiring a residual error sample training set of a nonlinear system to be diagnosed; the residual sample training set comprises residual sample data in a normal mode and residual sample data in a fault mode;
training a preset random forest fault separator based on a residual sample training set, and optimizing a structural parameter combination of the preset random forest fault separator by using a sparrow optimization algorithm to obtain a trained random forest fault separator corresponding to the optimal structural parameter combination.
Specifically, the sampling time of residual sample data obtained in each of k +1 modes (including a normal mode and k fault modes) is TsAnd intercept to obtain [ t ]f-tw,tf+tw]All length t insidewOf a sample of (1), intercept interval tsAnd obtaining characteristic data samples through characteristic extraction, respectively labeling the data samples, and generating a fault residual error sliding sample training set.
And setting a search space of a structural parameter combination (the number of decision trees contained in the random forest and the number of characteristic variables used for the binary tree in the nodes) of a preset random forest model, and designating population scale and maximum genetic algebra. And setting a population step structure comprising the proportion of a leader, a follower and a cautioner, and giving a danger threshold. And randomly generating an initial population in a search space, and finishing the initialization of the improved sparrow structure parameter optimization algorithm.
All the parameter combinations in the set search space range of the parameter combinations are regarded as a sparrow, the positions of individuals in the population are updated by using a sparrow algorithm based on the mixed performance fitness function, and the step updating is realized on the population. The hybrid performance fitness function is defined as follows:
fit f (X)=ξ f1, oob f (X)+ξ f2, T f (X)(5)
in the formula (I), the compound is shown in the specification,Xin order to combine the variables for the parameters,T(X) Is composed ofXTraining time for classifying the training set obtained in a by the constructed random forest classifier oob: (X) Out-of-bag error, ξ, for corresponding random forest classificationiThe coefficient of interest is the mixing performance.
And carrying out Rivie flight variation on specific population individuals. And selecting the sparrow individuals with a certain quantity and returned to participate in the tournament each time by utilizing tournament selection, selecting the individual with the minimum fitness to carry out Levy flight variation, and repeating the selection step until the number of the selected individuals is equal to the total number of population individuals.
And updating and selecting the individual position, and replacing the original position with the position with smaller fitness after the flight of the Levy. The position updating formula is as follows:
Figure 554152DEST_PATH_IMAGE011
in the formula (I), the compound is shown in the specification,
Figure 729919DEST_PATH_IMAGE012
for point-to-point multiplication, Levy (d) is the flight variation factor of Levie, and d is the dimension of the structural parameter to be optimized.
Figure 655150DEST_PATH_IMAGE013
Is the position of the ith individual of the variation in the it generation.
And after the maximum genetic algebra is reached, exiting the iterative computation, selecting the individual with the minimum fitness as the optimal individual, and constructing a classified random forest according to the corresponding optimal structural parameter combination to finish off-line training.
Optionally, step S106 further includes the following steps:
step S1061, substituting the time-frequency characteristic vector into a trained random forest fault separator, and obtaining the probability output of each fault type corresponding to the current state according to the forest voting result;
step S1062, calculating sequential probability ratio statistics for each fault type based on the probability output; specifically, the sequential probability ratio statistic for each fault type is calculated by the following equation:
Figure 879457DEST_PATH_IMAGE014
wherein the content of the first and second substances,
Figure 754004DEST_PATH_IMAGE015
sequential probability ratio statistics of the k fault type of the nonlinear system to be diagnosed at the time t,
Figure 38355DEST_PATH_IMAGE016
for the non-linear system to be diagnosed at t-tsSequential probability ratio statistic, t, for the k-th fault type occurring at a timesFor the truncation of the residual signal to be measured, Pk(Xt) Probability, P, of occurrence of k-th fault type at time t for the nonlinear system to be diagnosed0(Xt) The probability that the nonlinear system to be diagnosed is normal at the moment t is obtained;
and step S1063, based on the sequential probability ratio statistic of each fault type, performing sequential probability ratio test on the nonlinear system to be diagnosed to obtain target fault information.
The current time t to t-twInputting residual sampling samples of all the preorders into a trained classification regression forest, and obtaining the probability output of each fault corresponding to the current state according to the forest voting result:
Figure 513198DEST_PATH_IMAGE017
(6)
in the formula, k is a fault mode label;
Figure 541197DEST_PATH_IMAGE018
when the voting result of the binary tree at the time t is a fault k, the voting result is 1, otherwise, the voting result is 0;
Figure 722780DEST_PATH_IMAGE019
a classification strategy based on m randomly selected characteristic variables for the ith binary tree; n is the number of decision trees contained in the random forest, and m is the number of characteristic variables used for the binary tree in the node; i is the cumulative count factor.
Taking the fault mode as a discrete random variable X = {0,1,2, …, k }, and satisfying h in the formula (6) at the time tn,m(Xt) Probability of (2)And (4) distribution. Calculating sequential probability ratio statistics at t moment aiming at kth fault:
Figure 725502DEST_PATH_IMAGE020
(7)
in the formula, Pk(X) represents the probability of the system experiencing a class k fault, P0(X) represents the probability that the system is normal.
Iterative computation of sequential probability ratio statistic LR according to equation (7)tCarrying out sequential inspection:
(1) if it is
Figure 625325DEST_PATH_IMAGE021
And stopping checking and judging that the system is in a normal mode. In the formula, PFNRFor missed diagnosis rate, PFPRThe misdiagnosis rate is indicated.
(2) If it is
Figure 191435DEST_PATH_IMAGE022
And stopping checking and judging that the system is in the kth fault mode.
(3) If it is
Figure 555421DEST_PATH_IMAGE023
And continuously waiting for the sample at the next sampling moment to be tested, and judging that the system is in the state of the previous sampling moment.
In the embodiment of the invention, because the sequential probability ratio test can only test the fault occurrence or non-occurrence characteristic, the sequential test decision rule is improved to be applied to the fault test of multiple fault types, and the improvement is as follows:
(1) simultaneous computation of sequential probability ratio statistics for k failure modes
Figure 915995DEST_PATH_IMAGE024
And independently performing the iterative calculation and inspection steps, and judging the result
Figure 303114DEST_PATH_IMAGE025
Then the independent test is continued.
(2) If the judgment result has and only has 1 type fault
Figure 483035DEST_PATH_IMAGE026
If yes, judging that the system has the kth fault; if the judgment result has 2 or more types of faults
Figure 373630DEST_PATH_IMAGE027
Then it is determined that the system is maximized
Figure 170685DEST_PATH_IMAGE028
Corresponding class k fault.
(3) Class k failures continue to be checked independently, other classes of failures suspend checking, and
Figure 841838DEST_PATH_IMAGE029
。
(4) if the k-th type fault is judged
Figure 15330DEST_PATH_IMAGE030
And (5) returning to the step (1) to continue the calculation of all fault modes after judging that the fault performance is finished.
In order to eliminate the fault occurrence detection delay caused by the accumulation of negative values below a negative threshold and the fault performance detection delay caused by the accumulation of positive values above a positive threshold, the embodiment of the invention also corrects the iterative calculation of the sequential probability ratio statistic:
Figure 26012DEST_PATH_IMAGE031
in the formula (I), the compound is shown in the specification,
Figure 807017DEST_PATH_IMAGE032
representing the increment of the sequential probability ratio statistic at time t.
Optionally, the method provided in the embodiment of the present invention further includes a training and optimizing process for the pre-set regression random forest fault identifier, where the process is an offline training part. Specifically, the method comprises the following steps:
acquiring a fault type training set of a nonlinear system to be diagnosed; the fault type training set comprises residual sample data of the nonlinear system to be diagnosed under each fault type;
and training the preset regression random forest fault identifier based on the fault type training set, and optimizing the structural parameter combination of the preset regression random forest fault identifier by using a sparrow optimization algorithm to obtain the trained regression random forest fault identifier corresponding to the optimal structural parameter combination.
Specifically, k regression random forest identifiers are respectively established for k different faults, the identifiers are independent in pairs, and the establishing method is the same. The offline portion of the fault identification comprises the steps of:
carrying out equal difference value on the size of the kth fault, and sampling residual sample data obtained under different fault sizes for TsSimply discrete sampling and clipping to obtain [ t ]f-tw+ts,tf+ts]And (3) obtaining characteristic data samples through characteristic extraction, respectively labeling the data samples, wherein the labels are values of the size of the fault, and generating a fault residual error sample training set.
And setting a search space of random forest structure parameters (the number of decision trees contained in the random forest and the number of characteristic variables used for the binary tree in the nodes), and specifying the population size and the maximum genetic algebra. And setting a population step structure comprising the proportion of a leader, a follower and a cautioner, and giving a danger threshold. And randomly generating an initial population in a search space, and finishing the initialization of the improved sparrow structure parameter optimization algorithm.
All parameter combinations in the set parameter search space range are regarded as a sparrow, the positions of individuals in the population are updated by using a sparrow algorithm based on the mixed performance fitness function, and the step updating is realized on the population. The hybrid performance fitness function is defined as follows:
Figure 903149DEST_PATH_IMAGE033
and carrying out Rivie flight variation on specific population individuals. And selecting the sparrow individuals with a certain quantity and returned to participate in the tournament each time by utilizing tournament selection, selecting the individual with the minimum fitness to carry out Levy flight variation, and repeating the selection step until the number of the selected individuals is equal to the total number of population individuals.
And updating and selecting the individual position, and replacing the original position with the position with smaller fitness after the flight of the Levy. The position updating formula is as follows:
Figure 614753DEST_PATH_IMAGE034
and after the maximum genetic algebra is reached, exiting the iterative computation, selecting the individual with the minimum fitness as the optimal individual, and constructing a regression random forest according to the corresponding optimal structure parameter combination to finish off-line training.
The online part of the fault identification comprises the following steps:
after the fault mode is obtained by fault separation, selecting a regression random forest identifier of the corresponding mode, and identifying the current time from t to t-twAnd inputting residual sampling samples of all the preamble moments into a trained regression random forest identifier, and averaging recognition results of each decision tree of the random forest to obtain recognition results.
It can be known from the above description that the embodiment of the present invention provides a real-time fault diagnosis method based on a sequential random forest, which treats fault types as random variables, takes the proportion of voting results of each decision tree constituting the random forest to the total voting number as the probability of different fault mode types, and calculates the likelihood probability ratio of the current time of the random variable based on the distribution probability, where the probability ratio depends on the distribution condition of the random variables at all times in the preamble. The built sequential random forest algorithm can continuously increase data in the fault detection process until the fault is detected, and can realize real-time self-adaptive multi-sample fault diagnosis. Meanwhile, in the process of training a diagnostic model, a random forest hyper-parameter is optimized by using an improved sparrow algorithm, a Levin variation factor is introduced, a mixed performance fitness function is determined based on the error outside the bag and the training time to realize optimization of a random forest structure, high precision can be maintained, and meanwhile, the fault diagnosis data processing time is further reduced to shorten the diagnosis time delay.
Compared with the prior art, the real-time fault diagnosis method based on the sequential random forest provided by the embodiment of the invention has the following technical effects:
1. the method has a faster fault diagnosis speed, and the fault diagnosis time delay is only one ts duration in the embodiment.
2. The method can realize accurate fault diagnosis, can realize the recognition of tiny faults in the noise level and even the noise level, needs less training data, and has stronger robustness and generalization capability.
3. The fault diagnosis during the change of the command signal can be realized, and the method is suitable for fault diagnosis of a normal state variable system.
4. And a fault diagnosis model based on sequential random forests is completely and autonomously constructed, so that the subjectivity and the cost of artificial parameter adjustment are reduced.
Example two:
fig. 3 is a schematic diagram of a real-time fault diagnosis system based on a sequential random forest according to an embodiment of the present invention. As shown in fig. 3, the system includes: the system comprises an acquisition module 10, an extraction module 20, a fault separation module 30 and a fault identification module 40.
Specifically, the obtaining module 10 is configured to obtain a residual signal to be detected of the nonlinear system to be diagnosed.
Optionally, the obtaining module 10 is further configured to: establishing a nominal model of the nonlinear system to be diagnosed, and acquiring a nominal system residual error signal of the nominal model; acquiring an observation residual signal of a nonlinear system to be diagnosed; and obtaining a residual signal to be measured based on the nominal system residual signal and the observation residual signal.
Specifically, the residual signal to be measured is determined by the following equation: e.g. of the typel=es-g(en,tr,tc) (ii) a Wherein e islFor the residual signal to be measured, esTo observe the residual signal, enIs a nominal systemSystematic residual signal, g (e)n,tr,tc) As a nominal residual injection function, tr,tcRespectively command signal change time instant and nominal residual injection duration.
And the extraction module 20 is configured to extract a time-frequency feature vector of the residual signal to be detected based on wavelet packet decomposition.
The fault separation module 30 is used for substituting the time-frequency characteristic vector into the trained random forest fault separator, and performing sequential probability ratio test on the nonlinear system to be diagnosed to obtain target fault information; the target failure information includes a target failure type and a target failure occurrence time.
Optionally, the fault separation module 30 is further configured to substitute the time-frequency feature vector into the trained random forest fault separator, and obtain probability output of each fault type corresponding to the current state according to a forest voting result; calculating sequential probability ratio statistics for each fault type based on the probability outputs; and performing sequential probability ratio test on the nonlinear system to be diagnosed based on the sequential probability ratio statistic of each fault type to obtain target fault information.
Specifically, the fault separation module 30 is further configured to calculate the sequential probability ratio statistic for each fault type by the following equation:
Figure 542258DEST_PATH_IMAGE035
wherein the content of the first and second substances,
Figure 681115DEST_PATH_IMAGE036
sequential probability ratio statistics of the k fault type of the nonlinear system to be diagnosed at the time t,
Figure 264543DEST_PATH_IMAGE037
for the non-linear system to be diagnosed at t-tsSequential probability ratio statistic, t, for the k-th fault type occurring at a timesFor the truncation of the residual signal to be measured, Pk(Xt) To diagnose non-linearitiesProbability of occurrence of k-th fault type, P, in system at time t0(Xt) The probability that the nonlinear system to be diagnosed is normal at the moment t is obtained.
And the fault identification module 40 is used for substituting the time-frequency characteristic vector and the target fault type into the trained regression random forest fault identifier to obtain fault size information corresponding to the target fault type.
The invention provides a real-time fault diagnosis system based on sequential random forests, which can realize that data is continuously increased until a fault is detected in the fault detection process by combining a random forest fault separator and sequential probability ratio detection, can realize real-time self-adaptive multi-sample fault diagnosis, and solves the technical problems that the real-time fault diagnosis cannot be realized and the missed diagnosis rate and the misdiagnosis rate are high in the prior art.
Optionally, fig. 4 is a schematic diagram of another real-time fault diagnosis system based on a sequential random forest according to an embodiment of the present invention. As shown in fig. 4, the system further includes: a first training module 50 and a second training module 60.
Specifically, the first training module 50 is configured to obtain a residual sample training set of the nonlinear system to be diagnosed; the residual sample training set comprises residual sample data in a normal mode and residual sample data in a fault mode; training a preset random forest fault separator based on a residual sample training set, and optimizing a structural parameter combination of the preset random forest fault separator by using a sparrow optimization algorithm to obtain a trained random forest fault separator corresponding to the optimal structural parameter combination.
A second training module 60, configured to obtain a fault type training set of the nonlinear system to be diagnosed; the fault type training set comprises residual sample data of the nonlinear system to be diagnosed under each fault type; and training the preset regression random forest fault identifier based on the fault type training set, and optimizing the structural parameter combination of the preset regression random forest fault identifier by using a sparrow optimization algorithm to obtain the trained regression random forest fault identifier corresponding to the optimal structural parameter combination.
The embodiment of the present invention further provides an electronic device, which includes a memory, a processor, and a computer program stored in the memory and capable of running on the processor, and when the processor executes the computer program, the steps of the method in the first embodiment are implemented.
The embodiment of the invention also provides a computer readable medium with a non-volatile program code executable by a processor, wherein the program code causes the processor to execute the method in the first embodiment.
Finally, it should be noted that: the above embodiments are only used to illustrate the technical solution of the present invention, and not to limit the same; while the invention has been described in detail and with reference to the foregoing embodiments, it will be understood by those skilled in the art that: the technical solutions described in the foregoing embodiments may still be modified, or some or all of the technical features may be equivalently replaced; and the modifications or the substitutions do not make the essence of the corresponding technical solutions depart from the scope of the technical solutions of the embodiments of the present invention.

Claims (6)

1. A real-time fault diagnosis method based on sequential random forests is characterized by comprising the following steps:
acquiring a residual signal to be detected of the nonlinear system to be diagnosed;
extracting a time-frequency characteristic vector of the residual signal to be detected based on wavelet packet decomposition;
substituting the time-frequency characteristic vector into a trained random forest fault separator, and carrying out sequential probability ratio inspection on the nonlinear system to be diagnosed to obtain target fault information; the target fault information comprises a target fault type and target fault occurrence time;
substituting the time-frequency characteristic vector and the target fault type into a trained regression random forest fault identifier to obtain fault size information corresponding to the target fault type;
substituting the time-frequency characteristic vector into a random forest fault separator after training, and carrying out sequential probability ratio inspection on the nonlinear system to be diagnosed to obtain target fault information, wherein the steps comprise:
substituting the time-frequency characteristic vector into a random forest fault separator after training, and obtaining the probability output of each fault type corresponding to the current state according to the forest voting result;
calculating sequential probability ratio statistics for each fault type based on the probability outputs;
performing sequential probability ratio test on the nonlinear system to be diagnosed based on the sequential probability ratio statistic of each fault type to obtain target fault information;
obtaining a residual signal to be detected of a nonlinear system to be diagnosed, comprising the following steps:
establishing a nominal model of the nonlinear system to be diagnosed, and acquiring a nominal system residual error signal of the nominal model;
acquiring an observation residual signal of the nonlinear system to be diagnosed;
obtaining the residual signal to be detected based on the nominal system residual signal and the observation residual signal;
the method further comprises the following steps: obtaining the residual signal to be measured based on the nominal system residual signal and the observation residual signal, including:
determining the residual signal to be measured by the following formula: e.g. of the typel=es-g(en,tr,tc) (ii) a Wherein e islFor the residual signal to be measured, esFor the observed residual signal, enFor the nominal system residual signal, g (e)n,tr,tc) As a nominal residual injection function, tr,tcRespectively changing time of the command signal and nominal residual injection duration;
the method further comprises the following steps:
acquiring a residual error sample training set of the nonlinear system to be diagnosed; the residual sample training set comprises residual sample data in a normal mode and residual sample data in a fault mode;
and training a preset random forest fault separator based on the residual error sample training set, and optimizing the structural parameter combination of the preset random forest fault separator by using a sparrow optimization algorithm to obtain the trained random forest fault separator corresponding to the optimal structural parameter combination.
2. The method of claim 1, further comprising:
acquiring a fault type training set of the nonlinear system to be diagnosed; the fault type training set comprises residual sample data of the nonlinear system to be diagnosed under each fault type;
and training a preset regression random forest fault identifier based on the fault type training set, and optimizing the structural parameter combination of the preset regression random forest fault identifier by using a sparrow optimization algorithm to obtain the trained regression random forest fault identifier corresponding to the optimal structural parameter combination.
3. The method of claim 1, wherein computing sequential probability ratio statistics for each fault type based on the probability outputs comprises:
calculating a sequential probability ratio statistic for each fault type by:
Figure 843361DEST_PATH_IMAGE001
wherein the content of the first and second substances,
Figure 623098DEST_PATH_IMAGE002
sequential probability ratio statistics of k fault types occurring at the time t of the nonlinear system to be diagnosed,
Figure 18307DEST_PATH_IMAGE003
for the nonlinear system to be diagnosed at t-tsSequential probability ratio statistic, t, for the k-th fault type occurring at a timesIs the residual signal to be measuredInterception time interval of numbers, Pk(Xt) Probability P of occurrence of k-th fault type at t moment of the nonlinear system to be diagnosed0(Xt) And the probability that the nonlinear system to be diagnosed is normal at the moment t is obtained.
4. A real-time fault diagnosis system based on sequential random forest is characterized by comprising: the system comprises an acquisition module, an extraction module, a fault separation module and a fault identification module; wherein the content of the first and second substances,
the acquisition module is used for acquiring a residual signal to be detected of the nonlinear system to be diagnosed;
the extraction module is used for extracting the time-frequency characteristic vector of the residual signal to be detected based on wavelet packet decomposition;
the fault separation module is used for substituting the time-frequency characteristic vector into a trained random forest fault separator and carrying out sequential probability ratio inspection on the nonlinear system to be diagnosed to obtain target fault information; the target fault information comprises a target fault type and target fault occurrence time;
the fault identification module is used for substituting the time-frequency characteristic vector and the target fault type into a trained regression random forest fault identifier to obtain fault size information corresponding to the target fault type;
the fault separation module is also used for substituting the time-frequency characteristic vector into a trained random forest fault separator, and obtaining the probability output of each fault type corresponding to the current state according to the forest voting result; calculating sequential probability ratio statistics for each fault type based on the probability outputs; performing sequential probability ratio test on the nonlinear system to be diagnosed based on the sequential probability ratio statistic of each fault type to obtain target fault information;
the obtaining module is further configured to: establishing a nominal model of the nonlinear system to be diagnosed, and acquiring a nominal system residual error signal of the nominal model;
acquiring an observation residual signal of the nonlinear system to be diagnosed;
obtaining the residual signal to be detected based on the nominal system residual signal and the observation residual signal;
the obtaining module is further configured to: obtaining the residual signal to be measured based on the nominal system residual signal and the observation residual signal, including:
determining the residual signal to be measured by the following formula: e.g. of the typel=es-g(en,tr,tc) (ii) a Wherein e islFor the residual signal to be measured, esFor the observed residual signal, enFor the nominal system residual signal, g (e)n,tr,tc) As a nominal residual injection function, tr,tcRespectively changing time of the command signal and nominal residual injection duration;
the system further includes a first training module to: acquiring a residual error sample training set of the nonlinear system to be diagnosed; the residual sample training set comprises residual sample data in a normal mode and residual sample data in a fault mode; and training a preset random forest fault separator based on the residual error sample training set, and optimizing the structural parameter combination of the preset random forest fault separator by using a sparrow optimization algorithm to obtain the trained random forest fault separator corresponding to the optimal structural parameter combination.
5. An electronic device comprising a memory, a processor and a computer program stored on the memory and executable on the processor, wherein the steps of the method of any of the preceding claims 1 to 3 are implemented when the computer program is executed by the processor.
6. A computer-readable medium having non-volatile program code executable by a processor, wherein the program code causes the processor to perform the method of any of claims 1-3.
CN202111058850.3A 2021-09-10 2021-09-10 Real-time fault diagnosis method and system based on sequential random forest Active CN113505445B (en)

Priority Applications (1)

Application Number Priority Date Filing Date Title
CN202111058850.3A CN113505445B (en) 2021-09-10 2021-09-10 Real-time fault diagnosis method and system based on sequential random forest

Applications Claiming Priority (1)

Application Number Priority Date Filing Date Title
CN202111058850.3A CN113505445B (en) 2021-09-10 2021-09-10 Real-time fault diagnosis method and system based on sequential random forest

Publications (2)

Publication Number Publication Date
CN113505445A CN113505445A (en) 2021-10-15
CN113505445B true CN113505445B (en) 2022-03-01

Family

ID=78016548

Family Applications (1)

Application Number Title Priority Date Filing Date
CN202111058850.3A Active CN113505445B (en) 2021-09-10 2021-09-10 Real-time fault diagnosis method and system based on sequential random forest

Country Status (1)

Country Link
CN (1) CN113505445B (en)

Citations (2)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN108388860A (en) * 2018-02-12 2018-08-10 大连理工大学 A kind of Aeroengine Ball Bearings method for diagnosing faults based on power entropy-spectrum-random forest
CN108594788A (en) * 2018-03-27 2018-09-28 西北工业大学 A kind of aircraft actuator fault detection and diagnosis method based on depth random forests algorithm

Family Cites Families (3)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
US11907833B2 (en) * 2018-11-27 2024-02-20 The Boeing Company System and method for generating an aircraft fault prediction classifier
CN112036440A (en) * 2020-07-31 2020-12-04 山东科技大学 Satellite attitude control system fault diagnosis and early warning method based on random forest
CN112327219B (en) * 2020-10-29 2024-03-12 国网福建省电力有限公司南平供电公司 Distribution transformer fault diagnosis method with automatic feature mining and parameter automatic optimizing functions

Patent Citations (2)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN108388860A (en) * 2018-02-12 2018-08-10 大连理工大学 A kind of Aeroengine Ball Bearings method for diagnosing faults based on power entropy-spectrum-random forest
CN108594788A (en) * 2018-03-27 2018-09-28 西北工业大学 A kind of aircraft actuator fault detection and diagnosis method based on depth random forests algorithm

Also Published As

Publication number Publication date
CN113505445A (en) 2021-10-15

Similar Documents

Publication Publication Date Title
CN106197424B (en) The unmanned plane during flying state identification method of telemetry driving
CN107608335A (en) A kind of UAV Flight Control System fault detect and the data-driven method of fault reconstruction
CN108594788A (en) A kind of aircraft actuator fault detection and diagnosis method based on depth random forests algorithm
CN110070060B (en) Fault diagnosis method for bearing equipment
Liang et al. Data-driven fault diagnosis of FW-UAVs with consideration of multiple operation conditions
CN111176263B (en) Online identification method for thrust fault of aircraft based on BP neural network
CN111221345B (en) Decision tree-based aircraft power system fault online identification method
CN111460676B (en) Method for evaluating health performance of multi-rotor aircraft under atmospheric turbulence disturbance
EP2642362A2 (en) System monitoring
Sindhwani et al. Unsupervised anomaly detection for self-flying delivery drones
Ferdaus et al. Online identification of a rotary wing unmanned aerial vehicle from data streams
Wang et al. Data-driven anomaly detection of UAV based on multimodal regression model
Dutta et al. Statistical residual-based time series methods for multicopter fault detection and identification
CN113505445B (en) Real-time fault diagnosis method and system based on sequential random forest
CN112836581B (en) Sensitive fault feature extraction method and device based on correlation analysis
Cheng et al. Application of self-organizing map on flight data analysis for quadcopter health diagnosis system
Battipede et al. Neural networks for air data estimation: Test of neural network simulating real flight instruments
CN113850138B (en) Unmanned aerial vehicle sensor fault detection method, system and device based on data driving
CN115456171A (en) Aircraft attitude estimation method based on BP neural network model
CN115205582A (en) Intelligent drawing-simulating aviation sensor fault detection and classification method
Farhadi et al. Estimation of the lateral aerodynamic coefficients for Skywalker X8 flying wing from real flight-test data
Puchalski et al. UAV Fault Detection Methods, State-of-the-Art, Drones, 2022, 6, 330
CN114757332A (en) Intelligent fault detection method for aircraft pneumatic sensor
Dutta et al. Rotor Fault Detection and Identification for a Hexacopter Based on Control and State Signals via Statistical Learning Methods
CN114035536A (en) Flight control system robustness assessment method based on Monte Carlo method

Legal Events

Date Code Title Description
PB01 Publication
PB01 Publication
SE01 Entry into force of request for substantive examination
SE01 Entry into force of request for substantive examination
GR01 Patent grant
GR01 Patent grant