CN103353752B - Based on the aircraft environmental control system Control Component method for diagnosing faults of level Four RBF neural - Google Patents
Based on the aircraft environmental control system Control Component method for diagnosing faults of level Four RBF neural Download PDFInfo
- Publication number
- CN103353752B CN103353752B CN201310240505.0A CN201310240505A CN103353752B CN 103353752 B CN103353752 B CN 103353752B CN 201310240505 A CN201310240505 A CN 201310240505A CN 103353752 B CN103353752 B CN 103353752B
- Authority
- CN
- China
- Prior art keywords
- rbf
- fault
- follower
- control component
- output
- 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.)
- Expired - Fee Related
Links
- 230000007613 environmental effect Effects 0.000 title claims abstract description 62
- 238000000034 method Methods 0.000 title claims abstract description 55
- 230000001537 neural effect Effects 0.000 title claims abstract description 43
- 238000002955 isolation Methods 0.000 claims abstract description 63
- 230000003044 adaptive effect Effects 0.000 claims abstract description 45
- 239000000284 extract Substances 0.000 claims abstract description 17
- 238000001514 detection method Methods 0.000 claims abstract description 11
- 239000013598 vector Substances 0.000 claims description 87
- 238000012549 training Methods 0.000 claims description 72
- 238000013528 artificial neural network Methods 0.000 claims description 19
- 238000004088 simulation Methods 0.000 claims description 15
- 238000010606 normalization Methods 0.000 claims description 13
- 230000008569 process Effects 0.000 claims description 12
- 230000000052 comparative effect Effects 0.000 claims description 6
- 238000000605 extraction Methods 0.000 claims description 3
- 238000003064 k means clustering Methods 0.000 claims description 3
- 238000003745 diagnosis Methods 0.000 description 16
- 238000010586 diagram Methods 0.000 description 12
- 230000006870 function Effects 0.000 description 8
- 238000013178 mathematical model Methods 0.000 description 6
- 238000005070 sampling Methods 0.000 description 6
- 238000011160 research Methods 0.000 description 5
- 230000008901 benefit Effects 0.000 description 4
- 238000001816 cooling Methods 0.000 description 4
- 230000008859 change Effects 0.000 description 3
- 230000000694 effects Effects 0.000 description 3
- 238000005516 engineering process Methods 0.000 description 3
- 238000005192 partition Methods 0.000 description 3
- 230000003068 static effect Effects 0.000 description 3
- 238000012360 testing method Methods 0.000 description 3
- 238000004458 analytical method Methods 0.000 description 2
- 238000006243 chemical reaction Methods 0.000 description 2
- 238000012407 engineering method Methods 0.000 description 2
- 230000007257 malfunction Effects 0.000 description 2
- 230000009466 transformation Effects 0.000 description 2
- 241001269238 Data Species 0.000 description 1
- 239000012141 concentrate Substances 0.000 description 1
- 230000001276 controlling effect Effects 0.000 description 1
- 230000008878 coupling Effects 0.000 description 1
- 238000010168 coupling process Methods 0.000 description 1
- 238000005859 coupling reaction Methods 0.000 description 1
- 230000003111 delayed effect Effects 0.000 description 1
- 238000013461 design Methods 0.000 description 1
- 238000011161 development Methods 0.000 description 1
- 230000007274 generation of a signal involved in cell-cell signaling Effects 0.000 description 1
- 238000012905 input function Methods 0.000 description 1
- 238000004519 manufacturing process Methods 0.000 description 1
- 238000005457 optimization Methods 0.000 description 1
- 230000009467 reduction Effects 0.000 description 1
- 230000001105 regulatory effect Effects 0.000 description 1
- 238000000926 separation method Methods 0.000 description 1
Landscapes
- Feedback Control In General (AREA)
Abstract
The invention discloses a kind of aircraft environmental control system Control Component method for diagnosing faults based on level Four RBF neural, concrete steps are as follows: step one, set up and train one-level RBF neural observer; Step 2, foundation train secondary RBF adaptive threshold generator; Step 3, foundation train three grades of RBF fault followers, extract its network parameter; Step 4, foundation train level Four RBF fault isolation device; Step 5, real-time fault detection is carried out to aircraft environmental control system Control Component; Step 6, real time fail isolation is carried out to aircraft environmental control system Control Component; The inventive method adopts the method for diagnosing faults based on level Four RBF neural, for aircraft environmental control system Control Component provides fault detect and the isolation scheme of complete set, has very high practical engineering application and is worth.
Description
Technical field
The invention belongs to the fault diagnosis technology field of aircraft environmental control system, be specifically related to a kind of aircraft environmental control system Control Component method for diagnosing faults based on level Four RBF neural.
Background technology
The safety problem of plane environmental control system has been subject to people in recent years and has more and more paid close attention to.Good cabin atmosphere not only to the comfort of pilot and personal safety very important, and be the important guarantee that under different flight state, multiple air environment normally runs.Control Component is the important component part of aircraft environmental control system, and plays vital effect to the reliability service of aircraft environmental control system.Any fault of Control Component all directly may have influence on the stable regulation of electronics bay temperature, and then causes as problems such as the wasting of resources, reductions equipment life, and the more serious life security of airborne personnel of will giving brings threat.Therefore, the fault diagnosis research tool carrying out aircraft environmental control system Control Component is of great significance.
The existing research about aircraft environmental control system mainly concentrates on the design of system, control and system optimization, and the research for aircraft environmental control system fault diagnosis is less, and the fault diagnosis research of relevant controlling assembly is then less.Day by day complicated aircraft environmental control system and the fault diagnosis technology of fast development make to develop a set of effectively, the aircraft environmental control system Control Component method for diagnosing faults of system necessary all the more with may.Have researchist to propose a kind of air temperature control system fault diagnosis model based on observer, but their unpromising fault detect sets a suitable threshold value, and it is not high to carry out fault diagnosis reliability based on single residual signals.
Usually, method for diagnosing faults may be summarized to be the method based on model, the method based on data-driven and Knowledge based engineering method.Method for diagnosing faults based on model depends on mathematical model accurately, but due to Control Component unintentional nonlinearity feature, accurate mathematical model is often difficult to obtain.The method for diagnosing faults of Corpus--based Method depends on a large amount of experimental datas, and for aircraft environmental control system Control Component, experimental data is difficult to obtain in practice.As a kind of Knowledge based engineering method for diagnosing faults, neural network is easy to realize fault diagnosis that is non-linear and robustness.Compare with other feedforward networks, RBF neural has better approximation capability, faster learning ability, better robustness and does not have local minimum, may be used for the change of accurate tracking control system model, and adaptively modifying self neural network parameter, thus realize the fault diagnosis of aircraft environmental control system Control Component.
Meanwhile, in fault diagnosis, threshold value directly can affect the effect of fault detection and diagnosis.Threshold value is excessive may can't detect fault, too small, may cause false-alarm.Owing to being subject to the impact of the factors such as random disturbance, operating mode disturbance, system input and current system conditions, traditional fault detect based on fixed threshold cannot meet practical application request.Consider the learning ability of neural network (NN) and good robustness, some use the adaptive threshold based on neural network to be achieved the simulation study that Control Component carries out fault detect in recent years.
Although have some fault detects based on neural network and partition method to be successfully applied to commercial production, but seldom have the fault diagnosis research about aircraft environmental control system Control Component, more lack the Control Component fault detect of aircraft environmental control system and the partition method of complete set, comprise residual error generation, adaptive threshold detection, online fault tracking and fault isolation.
Summary of the invention
The object of the invention is to solve in aircraft environmental control system fault diagnosis field, still lack the Control Component fault detect of aircraft environmental control system and the partition method of complete set, existing fault detection method poor real, false alarm rate are high, failure separation method is practical engineering application problem not reliably, the advantage that fault diagnosis possesses is carried out according to RBF neural, propose the aircraft environmental control system Control Component method for diagnosing faults based on level Four RBF neural, realize aircraft environmental control system Control Component real-time fault detection and isolation.
The present invention is based on the aircraft environmental control system Control Component method for diagnosing faults of level Four RBF neural, specifically comprise the following steps:
Step one, set up and train one-level RBF neural observer.
Gather the inputoutput data training RBF observer under the various normal condition of Control Component system.Its training input vector is Control Component input temp signal and system delay output temperature signal, and wherein system delay output temperature signal is obtained by the sluggish link of a Reality simulation Control Component lag output function by the actual output of system.RBF observer training output vector is the actual output temperature signal of system.
Step 2, foundation train secondary RBF adaptive threshold generator.
First the system input under various for system normal condition and system delay are exported the one-level RBF neural observer sent into and train, obtain observer and estimate to export.Estimate to export by comparative observation device and export with system is actual the residual error obtained under system normal condition, this residual error is defined as benchmark residual error sta_threshold.Again system input signal and observer estimated output signal are trained secondary RBF adaptive threshold generator as input vector.Its output vector is that the benchmark residual error sta_threshold obtained under system normal condition adds modifying factor β.Wherein, β considers unknown system interference and modeling error, is obtained by emulation.
Step 3, set up three grades of RBF fault followers, extract its network architecture parameters.
Inject fault at moment t to system, schedule time length is h.Then (t-h+2, t+1), (t-h+3, t+2), (t-h+4, t+3) ..., the system inputoutput data in (t-h+n+1, t+n) time interval length is respectively used to on-line training RBF follower.Wherein input vector is that the input of Control Component system and system delay export, and output vector is the actual output of system.Extract RBF fault follower configuration parameter by pseudoinverse technique, obtain n group RBF neural structural parameters altogether, for the fault isolation of Control Component.
Step 4, foundation train level Four RBF fault isolation device.
Residual error effective value in the n group RBF fault follower configuration parameter obtained in step 3 and corresponding time interval is merged into n group input vector, as the training input vector of level Four RBF fault isolation device, training RBF fault isolation device.Its training output vector is that the theory of Control Component system under different faults pattern exports.
Step 5, real-time fault detection is carried out to aircraft environmental control system Control Component.
The input temp signal of Real-time Collection Control Component system and output temperature signal.System input signal and delay output signal are sent into as input vector the RBF neural observer trained in step one, obtains the system output that observer is estimated in real time.By estimation output temperature signal and the corresponding real system output temperature signal of comparative observation device, obtain the residual signals of Control Component.Again system input signal and observer estimated output signal are sent into the RBF adaptive threshold generator trained in step 2, obtain the adaptive threshold of current time.By comparing residual sum threshold value, judge whether system there occurs fault.Consider the impact of unknown system interference, assuming that when residual error continues to exceed threshold value p point, think and produce system jam and report to the police.
Step 6, real time fail isolation is carried out to aircraft environmental control system Control Component.
Once system jam be detected, then on-line training RBF fault follower.System inputoutput data before the extraction system warning moment in schedule time length, training RBF fault follower.Training input vector is that the input of Control Component system and system delay export, and output vector is the actual output of system.Extract the residual error effective value in the structural parameters of neural network and corresponding time interval, merged into input vector and send into the RBF fault isolation device trained in step 4, obtain fault isolation result.
Advantage of the present invention and good effect are:
(1) the inventive method adopts the method for diagnosing faults based on level Four RBF neural, for aircraft environmental control system Control Component provides fault detect and the isolation scheme of complete set, has very high practical engineering application and is worth;
(2) the inventive method takes full advantage of the powerful learning ability of RBF neural, has good robustness to aircraft environmental control system Control Component unintentional nonlinearity feature;
(3) the inventive method have employed the adaptive threshold generator based on RBF neural, overcomes wrong report and false-alarm that traditional fixed threshold may cause in process fault detection;
(4) the inventive method changes the feature of own net parameter with taking full advantage of RBF neural energy system for tracking state self-adaption, extract the network parameter of RBF neural fault follower, send into RBF fault isolation device and carry out fault isolation together with residual error feature, improve the fiduciary level that traditional simple dependence system residual error carries out fault isolation;
(5) the inventive method can carry out system state tracking, fault detect and isolation to aircraft environmental control system Control Component in real time, has very high engineering practicability;
Accompanying drawing explanation
Fig. 1 is aircraft environmental control system Control Component method for diagnosing faults process flow diagram of the present invention;
Fig. 2 is secondary RBF adaptive threshold generator training process flow diagram;
Fig. 3 (a) is aircraft environmental control system structure of the present invention;
Fig. 3 (b) is aircraft environmental control system Control Component schematic diagram of the present invention;
Fig. 3 (c) is aircraft environmental control system Control Component mathematical model of the present invention;
Fig. 4 is the aircraft environmental control system Control Component model in the embodiment of the present invention under Matlab Simulink environment;
Fig. 5 is the stuck fault simulation result of valve in the embodiment of the present invention;
Fig. 6 is sensor constant gain failures simulation result in the embodiment of the present invention;
Fig. 7 is the stuck test process simulation result of valve in the embodiment of the present invention;
Fig. 8 is sensor constant gain failures test process simulation result in the embodiment of the present invention;
Embodiment
Below in conjunction with drawings and Examples, the present invention is described in further detail.
The present invention is directed in aircraft environmental control system fault diagnosis field, lack a set of effectively and the present situation of the aircraft environmental control system Control Component method for diagnosing faults of system, according to structure and the data characteristics of aircraft environmental control system Control Component, propose a kind of aircraft environmental control system Control Component method for diagnosing faults based on level Four RBF neural.Figure 1 shows that the aircraft environmental control system Control Component method for diagnosing faults process flow diagram based on level Four RBF neural, concrete steps are as follows:
Step one, set up and train one-level RBF neural observer.
Considering to export exist certain delayed, therefore before being incorporated into the input end of Neural Network Observer, adds a Z
-1link, the operative scenario of the aircraft environmental control system Control Component of approaching to reality, wherein Z
-1be one can the sluggish link of Reality simulation Control Component lag output function.
Gather system input temp signal r (t) under Control Component various normal operating conditions (t=2,3,4 ..., n) with output temperature signal y
r(t) (t=2,3,4 ..., n), output temperature signal is through sluggish link Z
-1obtain system delay afterwards and export y'
r(t) (t=1,2,3 ..., n-1), by obtain system input signal r (t) (t=2,3,4 ..., n) with delay output signal y'
r(t) (t=1,2,3 ..., n-1) and be put into the training input amendment as RBF neural observer in a vector, by the Control Component system output signal y obtained
r(t) (t=2,3,4 ..., n) as the training output sample of RBF neural observer.Needed, between the process to [-1,1] of training input and output samples normalization, then to set the basic parameter of RBF neural, start training before training.The RBF neural observer trained is preserved when having trained.
Step 2, foundation train secondary RBF adaptive threshold generator.
The training process flow diagram of secondary RBF adaptive threshold generator as shown in Figure 2.First, input temp signal r (t) under the various normal condition of acquisition system (t=2,3,4 ..., n) and system output temperature signal y
r(t) (t=2,3,4 ..., n), output signal and obtain system delay output y' after sluggish link
r(t) (t=1,2,3 ..., n-1).By input temp signal r (t) (t=2,3,4 ..., n) and delay output signal y'
r(t) (t=1,2,3 ..., n-1) and as input vector, after normalization, send into the RBF observer trained in step one, the observer obtained under system fault condition is estimated to export
(t=2,3,4 ..., n).Observer is estimated to export
(t=2,3,4 ..., n) export y with system is actual
r(t) (t=2,3,4 ..., n) poor, can obtain residual signals ε (t) under system normal condition (t=2,3,4 ..., n), this residual signals is defined as benchmark residual error sta_threshold.
Again said system is inputted r (t) (t=2,3,4 ..., n) and observer estimate system export
(t=2,3,4 ..., n) after normalization as network input vector, training secondary RBF adaptive threshold generator.Its training output vector is adaptive threshold: adap_thrshold=sta_threshold+ β.Wherein, β considers unknown system interference and modeling error, by emulating the modifying factor obtained.The RBF adaptive threshold generator trained is preserved when having trained.
Step 3, foundation train three grades of RBF fault followers, extract its network parameter.
When system jam, then on-line training RBF fault follower.Due to when the fault mode of system changes, RBF neural can accurately follow the tracks of this change, and adaptively modifying inherent parameters.Therefore, the structural parameters of RBF fault follower may be used for the fault isolation of Control Component.Here, adopt K means clustering algorithm training RBF fault follower, obtain the structural parameters of neural network by pseudoinverse technique.The RBF fault follower network parameter extracted is: the band fat vector σ of RBF follower i-th node
i, the center vector c of RBF follower i-th node
i, the connection weight w of RBF follower
i.
The sample of RBF fault follower on-line training is the system input and output data before the fault generation moment in schedule time length.Inject fault at moment t to system, schedule time length is h.Then (t-h+2, t+1), (t-h+3, t+2), (t-h+4, t+3) ..., the system inputoutput data in (t-h+n+1, t+n) time interval length is respectively used to on-line training RBF follower.Wherein input vector is that the input of Control Component system and system delay export, and output vector is the actual output of system.After extracting network architecture parameters by pseudoinverse technique, obtain n group RBF neural structural parameters altogether, for the fault isolation of Control Component.Meanwhile, because residual error can show different features under different faults pattern, therefore, the residual error effective value in corresponding time interval is also used to the fault isolation realizing Control Component, total n group residual error effective value.They and RBF fault follower configuration parameter form n group input vector, jointly for training level Four RBF fault isolation device.
Step 4, foundation train level Four RBF fault isolation device.
N group RBF fault follower configuration parameter is obtained by step 3, and the residual error effective value in corresponding time interval.They are merged into n group input vector, as the training input vector of level Four RBF fault isolation device, training RBF fault isolation device.Training input vector is defined as follows:
Z＝[z
1z
2z
3z
4]
T＝[σ
ic
iw
iε]
T
Training input vector Z comprises four characteristic quantity z
1, z
2, z
3, z
4, corresponding σ respectively
i, c
i, w
i, ε.Wherein σ
ifor the band fat vector of RBF follower i-th node, c
ifor the center vector of RBF follower i-th node, w
ifor the connection weights of RBF follower, ε is the effective value (RMS) of residual error in schedule time length before the fault generation moment.
RBF fault isolation device training output vector is that the theory of Control Component system under different faults pattern exports.For the m kind fault mode of Control Component system, RBF fault isolation device training output vector is:
The target of table 1.RBF fault isolation device exports
Step 5, real-time fault detection is carried out to aircraft environmental control system Control Component.
Input temp signal r (t) of Real-time Collection Control Component system (t=2,3,4 ..., n) with output temperature signal y
r(t) (t=2,3,4 ..., n), output temperature signal is through sluggish link Z
-1obtain system delay afterwards and export y '
r(t) (t=1,2,3 ..., n-1), by obtain system input signal r (t) (t=2,3,4 ..., n) with delay output signal y '
r(t) (t=1,2,3 ..., n-1) be put in a vector, send into the RBF neural observer trained in step one after normalization, obtain the system output that observer is estimated in real time
(t=2,3,4 ..., n).By the estimation output temperature signal of comparative observation device
(t=2,3,4 ..., n) with corresponding real system output temperature signal y
r(t) (t=2,3,4 ..., n), the residual signals of current t Control Component can be obtained.
Again by system input signal r (t) (t=2,3,4 ..., n) and observer estimated output signal
(t=2,3,4 ..., n) send into the RBF adaptive threshold generator trained in step 2, obtain the adaptive threshold of current time.
Under normal circumstances, residual error is less than threshold value to system, close to zero.When a certain component failure in Control Component system, residual error will increase, and exceed adaptive threshold.Consider the impact of unknown system interference, set when residual error continues to exceed threshold value p point, think and produce system jam and report to the police.
Step 6, real time fail isolation is carried out to aircraft environmental control system Control Component.
Once system jam be detected, then on-line training RBF fault follower.Extraction system warning moment t
1system inputoutput data before in schedule time length h, for training RBF fault follower.Training input vector is that the input of Control Component system and system delay export, and output vector is the actual output of system.The structural parameters of neural network are obtained by pseudoinverse technique.The RBF fault follower network parameter extracted is: the band fat vector σ of RBF follower i-th node
i, the center vector c of RBF follower i-th node
i, the connection weight w of RBF follower
i.Meanwhile, extract the residual error effective value in corresponding time interval, itself and RBF fault follower network architecture parameters merged into fault isolation input vector:
Z＝[z
1z
2z
3z
4]
T＝[σ
ic
iw
iε]
T
Sent into the RBF fault isolation device trained in step 4, obtained fault isolation result.Here Z is after system jam being detected, during for carrying out fault isolation, sends into the input vector of the RBF fault isolation device trained; And the Z in step 4 is the RBF fault isolation device training input vector in order to train RBF fault isolation device to extract.
Embodiment:
This example adopt the inventive method to environmental control system Control Component valve the stuck and permanent gain of sensor two kinds of faults carry out fault detect and isolation, to set forth invented content, and further illustrate the use procedure of content of the present invention.。
Fig. 3 (a) is depicted as aircraft environmental control system structural drawing of the present invention.The air of High Temperature High Pressure is flowed out from the compressor of aircraft engine.These discharge air after primary cooler cooling, are divided into two parts: a part enters the hot road conduit with by-pass valve control, and another part enters cooling duct, is cooled by heat exchanger and turbine.The air entering cooling duct is initially cooled by ram-air through the cold junction of over-heat-exchanger, again cools after then expanding in turbine, drives the coupling shaft between fan and turbine to make fan produce ram-air simultaneously.Cryogenic air is discharged in turbine gas outlet.By the temperature that by-pass valve control adjustment can make electronics bay reach suitable from the air ratio of cooling duct and Re Lu conduit.
Control Component is the chief component of aircraft environmental control system, is also very important in temperature adjustment process.Fig. 3 (b) is depicted as aircraft environmental control system Control Component schematic diagram of the present invention.In this control loop, a given preset temperature, amplifier will be converted to corresponding voltage signal temperature signal, as the input of actuator.Then, actuator is converted to corresponding valve corner to control the temperature for air-flow and electronics bay voltage signal.Meanwhile, the temperature signal that temperature sensor gathers in cabin feeds back as control, to guarantee that temperature reaches preset value quickly and accurately.
Fig. 3 (c) is depicted as aircraft environmental control system Control Component mathematical model of the present invention.Wherein, the relation of k representation temperature and engine input voltage, corresponding diagram 3(b) in amplifier;
represent the relation of valve corner and engine input voltage, f represents valve corner and for the relation of air-flow, their corresponding diagram 3(b) in actuator;
for the transport function of temperature sensor, corresponding diagram 3(b) in temperature sensor;
for simplify electronics bay mathematical model, corresponding diagram 3(b) in electronics bay.
Wherein: set the input function of a system as x (t), output function is y (t), then the business of Laplace transformation Y (s) of y (t) and Laplace transformation X (s) of x (t): W (s)=Y (s)/X (s) is called this system transter.K
m, k
z, k
Φ, T
w, T is parameter in transport function.
Fig. 4 is the aircraft environmental control system Control Component model in the embodiment of the present invention under Matlab Simulink environment, and simulation parameter is as shown in table 2.System is input as sine wave, and the MATLAB power function module registration string wave amplitude embedded by is changed, thus realizes the change of operating mode.Add random noise at the input end of system, average is 0.00001, and variance is 0.05.Amplifier magnification ratio K=60, the Gain in corresponding diagram.Be filled with two kinds of faults in the embodiment of the present invention aircraft environmental control system Control Component: the stuck gain permanent in sensor of valve.When valve is stuck, valve is in certain aperture, can not play regulatory role, and comparatively the long-time interior existence of big error can not be eliminated.By aforesaid analysis, in Simulink realistic model, inject the stuck fault of valve, by switch structure, realize the conversion of normal module and malfunctioning module.A branch road is malfunction above, the angle that the MATLAB power function module governor valve embedded by is stuck; A branch road is normal condition below, k
m=2.5, f=0.0133, Gain1 in corresponding diagram.During sensor constant gain failures, the feedback signal of sensor has certain deviation, causes control signal generation deviation, finally causes system to occur static difference (sensor front signal).By aforesaid analysis, in Simulink realistic model, inject sensor constant gain failures, by switch structure, realize the conversion of normal module and malfunctioning module.A branch road is malfunction above, and the MATLAB power function module embedded by regulates the degrees of offset of sensor gain; A branch road is normal condition below, k
z=1, T
w=0.6.
for electronics bay mathematical model.Can observing system input and system output by scope1 and scope2.
Table 2. environmental control system Control Component simulation parameter
Concrete steps are as follows:
Step one, set up and train one-level RBF neural observer.
The input signal of known environmental control system Control Component is sinusoidal.Under Matlab Simulink environment, acquisition system is input as input temp signal r (t) (t=2,3,4 under 15+15sin (π/50) t and 3+3sin (π/50) t two kinds of nominal situations respectively,, 1001) and system output temperature signal y
r(t) (t=2,3,4 ..., 1001), output signal and obtain system delay output y' after sluggish link
r(t) (t=1,2,3 ..., 1000).Inputoutput data under two kinds of operating modes respectively gathers 1000 groups, and sampling rate is 1s.With the inputoutput data Training RBF Neural Network observer collected.Training input vector be system input signal r (t) (t=2,3,4 ..., 1001) and export y' through the system delay of sluggish link
r(t) (t=1,2,3 ..., 1000), training output vector is system output signal y'
r(t) (t=2,3,4 ..., 1001).Needed, between the process to [-1,1] of training input and output samples normalization, then to set the basic parameter of RBF neural, start training before training.The RBF neural observer trained is preserved when having trained.
Step 2, foundation train secondary RBF adaptive threshold generator.
Resurvey system be input as input temp signal r (t) under 15+15sin (π/50t) and 3+3sin (π/50) t two kinds of nominal situations (t=2,3,4 ..., 1001) and system output temperature signal y
r(t) (t=2,3,4 ..., 1001), output signal and obtain system delay output y' after sluggish link
r(t) (t=1,2,3 ..., 1000).Under two kinds of operating modes, inputoutput data respectively gathers 1000 groups, and sampling rate is 1s.By input temp signal r (t) (t=2,3,4 ..., 1001) and delay output signal y'
r(t) (t=1,2,3 ..., 1000) and as input vector, after normalization, send into the RBF observer trained in step one, the observer obtained under system fault condition is estimated to export
(t=2,3,4 ..., 1001).Observer is estimated to export y
(t=2,3,4 ..., 1001) and export y with system is actual
r(t) (t=2,3,4 ..., 1001) and poor, the residual signals under system normal condition can be obtained, this residual signals is defined as benchmark residual error sta_threshold.
Again said system is inputted r (t) (t=2,3,4 ..., 1001) and observer estimate system export
(t=2,3,4 ..., 1001) after normalization as network input vector, training secondary RBF adaptive threshold generator.Its training output vector is that benchmark residual error sta_threshold adds modifying factor β, i.e. adaptive threshold: adap_thrshold=sta_threshold+ β.Wherein, β considers unknown system interference and modeling error, is obtained, get β=0.3 here by emulation.The RBF adaptive threshold generator trained is preserved when having trained.
Step 3, foundation train three grades of RBF fault followers, extract its network parameter.
Environmental control system Control Component model is run, acquisition system inputoutput data under Matlab Simulink environment.The simulation run time is 2000s, and sampling rate is 1s.The stuck fault of valve is injected to Control Component model at 1000s.Fig. 5 is the stuck fault simulation result of valve in the embodiment of the present invention.Fig. 5 (a) is system input waveform, and the mathematical formulae of its correspondence as shown in Equation (1).
Fig. 5 (b) is system output waveform, and as can be seen from the figure, after injecting the stuck fault of valve, system exports and increases gradually, finally trends towards a steady state value.Fig. 5 (c) is system residual error and adaptive threshold waveform, and wherein solid line represents residual error, represented by dotted arrows adaptive threshold.As can be seen from the figure, under system normal condition, residual error is close to 0 and be less than threshold value; After injecting the stuck fault of valve, system residual error increases gradually, and exceedes adaptive threshold, shows that system there occurs fault.
Schedule time length is 300s.Extract 702s ~ 1001s respectively, 703s ~ 1002s ..., the system inputoutput data of 901s ~ 1200s, adopts K means clustering algorithm on-line training RBF follower.Training input vector is that the input of Control Component system and system delay export, and output vector is the actual output of system.The band fat vector σ of RBF follower i-th node is extracted by pseudoinverse technique
i, the center vector c of RBF follower i-th node
i, the connection weight w of RBF follower
i.Obtain 200 groups of RBF fault follower configuration parameters altogether.Extract the residual error effective value in the corresponding time period simultaneously, itself and RBF fault follower configuration parameter are formed 200 groups of input vectors, jointly for training level Four RBF fault isolation device.
Again under Matlab Simulink environment, environmental control system Control Component model is run, acquisition system inputoutput data.The simulation run time is 2000s, and sampling rate is 1s.Sensor constant gain failures is injected to Control Component model at 1000s.Fig. 6 is sensor constant gain failures simulation result in the embodiment of the present invention.Fig. 6 (a) is system input waveform, and the mathematical formulae of its correspondence as shown in Equation (1).Fig. 6 (b) is system output waveform, and as can be seen from the figure, after injecting sensor constant gain failures, static difference appears in system.Fig. 6 (c) is system residual error and adaptive threshold waveform, and wherein solid line represents residual error, represented by dotted arrows adaptive threshold.As can be seen from the figure, under system normal condition, residual error is close to 0 and be less than threshold value; After injecting sensor constant gain failures, system residual error increases gradually, and eventually exceeds adaptive threshold, shows that system there occurs fault.
On-line training RBF fault follower, extracts its network architecture parameters in the same way, and the residual error effective value in the corresponding time period.Obtain 200 groups of input vectors altogether, for training level Four RBF fault isolation device.
Step 4, foundation train level Four RBF fault isolation device.
System is obtained at valve under the stuck and permanent gain of sensor two kinds of faults, RBF fault follower configuration parameter and each 200 groups of corresponding residual error effective value by step 3.Input vector is it can be used as to train level Four RBF fault isolation device.Training output vector is that the theory of Control Component system under different faults pattern exports, as shown in table 3.After having trained, preserve RBF fault isolation device neural network.
The target of table 3.RBF fault isolation device exports
Step 5, real-time fault detection and isolation are carried out to the stuck fault of aircraft environmental control system Control Component valve.
Under Matlab Simulink environment, run environmental control system Control Component model, real-time acquisition system inputoutput data, sampling rate is 1s.The stuck fault of valve is injected to Control Component model at 1200s.By obtain system input signal r (t) (t=2,3,4 ..., n) with delay output signal y'
r(t) (t=1,2,3 ..., n-1) be put in a vector, send into the RBF neural observer trained in step one after normalization, obtain the system output that observer is estimated in real time
(t=2,3,4 ..., n).By the estimation output temperature signal of comparative observation device
(t=2,3,4 ..., n) with corresponding real system output temperature signal y
r(t) (t=2,3,4 ..., n), the residual signals of current t Control Component can be obtained.
Again by system input signal r (t) (t=2,3,4 ..., n) and observer estimated output signal
(t=2,3,4 ..., n) send into the RBF adaptive threshold generator trained in step 2, obtain the adaptive threshold of current time.Fig. 7 is the stuck test process simulation result of valve in the embodiment of the present invention.Fig. 7 (a) is system output waveform figure, injects fault at 1200s to system, and system exports and starts to be tending towards constant; Fig. 7 (b) is system residual sum adaptive threshold oscillogram.Solid line represents residual error, represented by dotted arrows adaptive threshold.Consider the impact of unknown system interference, in the embodiment of the present invention, setting is when residual error continues to exceed threshold value 25 points, thinks and produces system jam and report to the police.As can be seen from Fig. 7 (b), after 1200s injects the stuck fault of valve to system, residual error increases, and continues to exceed threshold value 25 points in 1261s residual error, and system produces reports to the police.Fig. 7 (c) reports to the police before the moment for system generation, within schedule time length 300s, i.e. and the system residual sum adaptive threshold waveform of 962s ~ 1261s.
Because schedule time length is 300s, after system is reported to the police, extract the system inputoutput data on-line training RBF fault follower of 962s ~ 1261s.Training input vector is that the input of Control Component system and system delay export, and output vector is the actual output of system.The structural parameters of neural network are extracted: the band fat vector σ of RBF follower i-th node by pseudoinverse technique
i, the center vector c of RBF follower i-th node
i, the connection weight w of RBF follower
i.Meanwhile, extract the residual error effective value in 962s ~ 1261s, itself and RBF fault follower network architecture parameters are merged into input vector, send into the RBF fault isolation device trained in step 4, obtain fault isolation result as shown in table 4.
The stuck fault isolation result of table 4. valve
As can be seen from Table 4, the actual output of neural network is similar to theoretical output, illustrates that the aircraft environmental control system Control Component method for diagnosing faults based on level Four RBF neural has good validity.
Step 6, real-time fault detection and isolation are carried out to aircraft environmental control system Control Component sensor constant gain failures.
Under Matlab Simulink environment, run environmental control system Control Component model, real-time acquisition system inputoutput data, sampling rate is 1s.Sensor constant gain failures is injected to Control Component model at 1200s.By obtain system input signal r (t) (t=2,3,4 ..., n) with delay output signal y'
r(t) (t=1,2,3 ..., n-1) be put in a vector, send into the RBF neural observer trained in step one after normalization, obtain the system output that observer is estimated in real time
(t=2,3,4 ..., n).By the estimation output temperature signal of comparative observation device
(t=2,3,4 ..., n) with corresponding real system output temperature signal y
r(t) (t=2,3,4 ..., n), the residual signals of current t Control Component can be obtained.
Again by system input signal r (t) (t=2,3,4 ..., n) and observer estimated output signal
(t=2,3,4 ..., n) send into the RBF adaptive threshold generator trained in step 2, obtain the adaptive threshold of current time.As can be seen from Figure 8, under normal circumstances, residual error is less than threshold value to system, close to zero.When injecting sensor constant gain failures, residual error increases, and exceedes adaptive threshold.Continue to exceed threshold value 25 points in 1336s residual error, system is reported to the police.Fig. 8 (a) is system output waveform figure, injects fault at 1200s to system, and system exports and starts to occur static difference; Fig. 8 (b) is system residual sum adaptive threshold oscillogram.Solid line represents residual error, represented by dotted arrows adaptive threshold.Consider the impact of unknown system interference, think when residual error continues to exceed threshold value 25 points equally, system jam, produce and report to the police.As can be seen from Fig. 8 (b), after 1200s injects the stuck fault of valve to system, residual error increases gradually, continues to exceed threshold value 25 points in 1336s residual error, and system produces reports to the police.Fig. 8 (c) reports to the police before the moment for system generation, within schedule time length 300s, i.e. and the system residual sum adaptive threshold waveform of 1037s ~ 1336s.
Because schedule time length is 300s, after system is reported to the police, extract the system inputoutput data on-line training RBF fault follower of 1037s ~ 1336s.Training input vector is that the input of Control Component system and system delay export, and output vector is the actual output of system.The structural parameters of neural network are extracted: the band fat vector σ of RBF follower i-th node by pseudoinverse technique
i, the center vector c of RBF follower i-th node
i, the connection weight w of RBF follower
i.Meanwhile, extract the residual error effective value in 1037s ~ 1336s, itself and RBF fault follower network architecture parameters are merged into input vector, send into the RBF fault isolation device trained in step 4, obtain fault isolation result as shown in table 5.
Table 5. sensor constant gain failures isolation result
As can be seen from Table 5, the actual output of neural network is similar to theoretical output, further illustrates the validity of the inventive method for aircraft Control Component fault diagnosis.
Claims (1)
1., based on the aircraft environmental control system Control Component method for diagnosing faults of level Four RBF neural, concrete steps are as follows:
Step one, set up and train one-level RBF neural observer;
Z is increased before the input end of Neural Network Observer
-1link, Z
-1for the sluggish link of Reality simulation Control Component lag output function;
Gather system input temp signal r (t) under the various normal operating conditions of Control Component and output temperature signal y
r(t), t=2,3,4 ..., n, output temperature signal is through sluggish link Z
-1obtain system delay afterwards and export y'
r(t), t=1,2,3 ..., n-1, by system input signal r (t) and the delay output signal y' that obtain
rt () is put into the training input amendment as RBF neural observer in a vector, by the Control Component system output signal y obtained
rt () is as the training output sample of RBF neural observer; Between the process to [-1,1] of training input and output samples normalization, the basic parameter of setting RBF neural, starts training, preserves the RBF neural observer trained when having trained;
Step 2, foundation train secondary RBF adaptive threshold generator;
First, input temp signal r (t) and system output temperature signal y under the various normal condition of acquisition system
rt (), outputs signal and obtain system delay output y' after sluggish link
r(t); By input temp signal r (t) and delay output signal y'
rt (), as input vector, sends into the RBF observer trained in step one after normalization, the observer obtained under system fault condition is estimated to export
t=2,3,4 ..., n; Observer is estimated to export
y is exported with system is actual
rt () is poor, obtain residual signals ε (t) under system normal condition, t=2,3,4 ..., n, is defined as benchmark residual error sta_threshold by this residual signals;
The system again said system being inputted r (t) and observer estimation exports
as network input vector after normalization, training secondary RBF adaptive threshold generator; Its training output vector is adaptive threshold: adap_thrshold=sta_threshold+ β; Wherein, β considers unknown system interference and modeling error, by emulating the modifying factor obtained; The RBF adaptive threshold generator trained is preserved when having trained;
Step 3, foundation train three grades of RBF fault followers, extract its network parameter;
When system jam, on-line training RBF fault follower, adopt K means clustering algorithm training RBF fault follower, obtain the structural parameters of neural network by pseudoinverse technique, the RBF fault follower network parameter extracted is: the band fat vector σ of RBF follower i-th node
i, the center vector c of RBF follower i-th node
i, the connection weight w of RBF follower
i;
The sample of RBF fault follower on-line training is the system input and output data before the fault generation moment in schedule time length; Inject fault at moment t to system, schedule time length is h; Then (t-h+2, t+1), (t-h+3, t+2), (t-h+4, t+3) ..., the system inputoutput data in (t-h+n+1, t+n) time interval length is respectively used to on-line training RBF follower; Wherein input vector is that the input of Control Component system and system delay export, and output vector is the actual output of system; After extracting network architecture parameters by pseudoinverse technique, obtain n group RBF neural structural parameters altogether, meanwhile, the residual error effective value in corresponding time interval has n group residual error effective value; Residual error effective value and RBF fault follower configuration parameter form n group input vector jointly, for training level Four RBF fault isolation device;
Step 4, foundation train level Four RBF fault isolation device;
N group RBF fault follower configuration parameter is obtained by step 3, and the residual error effective value in corresponding time interval, merge into n group input vector, as the training input vector of level Four RBF fault isolation device, training RBF fault isolation device; Training input vector is defined as follows:
Z＝[z
1z
2z
3z
4]
T＝[σ
ic
iw
iε]
T
Training input vector Z comprises four characteristic quantity z
1, z
2, z
3, z
4, corresponding σ respectively
i, c
i, w
i, ε; Wherein σ
ifor the band fat vector of RBF follower i-th node, c
ifor the center vector of RBF follower i-th node, w
ifor the connection weights of RBF follower, ε is the effective value of residual error in schedule time length before the fault generation moment;
RBF fault isolation device training output vector is that the theory of Control Component system under different faults pattern exports, and for the m kind fault mode of Control Component system, RBF fault isolation device training output vector is:
The target of RBF fault isolation device exports
Step 5, real-time fault detection is carried out to aircraft environmental control system Control Component;
Input temp signal r (t) of Real-time Collection Control Component system and output temperature signal y
rt (), output temperature signal is through sluggish link Z
-1obtain system delay afterwards and export y'
rt (), by system input signal r (t) and the delay output signal y' that obtain
rt () is put in a vector, send into the RBF neural observer trained in step one after normalization, obtains the system output that observer is estimated in real time
by the estimation output temperature signal of comparative observation device
with corresponding real system output temperature signal y
rt (), obtains the residual signals of current t Control Component;
Again by system input signal r (t) and observer estimated output signal
send into the RBF adaptive threshold generator trained in step 2, obtain the adaptive threshold of current time;
Under normal circumstances, residual error is less than threshold value to system, and when a certain component failure in Control Component system, residual error will increase, and exceed adaptive threshold, and setting, when residual error continues to exceed threshold value p point, thinks system jam, generation warning;
Step 6, real time fail isolation is carried out to aircraft environmental control system Control Component;
Once system jam be detected, then on-line training RBF fault follower; Extraction system warning moment t
1system inputoutput data before in schedule time length h, for training RBF fault follower; Training input vector is that the input of Control Component system and system delay export, and output vector is the actual output of system; The structural parameters of neural network are obtained by pseudoinverse technique; The RBF fault follower network parameter extracted is: the band fat vector σ of RBF follower i-th node
i, the center vector c of RBF follower i-th node
i, the connection weight w of RBF follower
i; Meanwhile, extract the residual error effective value in corresponding time interval, itself and RBF fault follower network architecture parameters merged into fault isolation input vector:
Z＝[z
1z
2z
3z
4]
T＝[σ
ic
iw
iε]
T
Sent into the RBF fault isolation device trained in step 4, obtained fault isolation result; Here Z is after system jam being detected, during for carrying out fault isolation, sends into the input vector of the RBF fault isolation device trained.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201310240505.0A CN103353752B (en) | 2013-04-12 | 2013-06-18 | Based on the aircraft environmental control system Control Component method for diagnosing faults of level Four RBF neural |
Applications Claiming Priority (3)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201310127894 | 2013-04-12 | ||
| CN201310127894.6 | 2013-04-12 | ||
| CN201310240505.0A CN103353752B (en) | 2013-04-12 | 2013-06-18 | Based on the aircraft environmental control system Control Component method for diagnosing faults of level Four RBF neural |
Publications (2)
| Publication Number | Publication Date |
|---|---|
| CN103353752A CN103353752A (en) | 2013-10-16 |
| CN103353752B true CN103353752B (en) | 2015-10-28 |
Family
ID=49310133
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN201310240505.0A Expired - Fee Related CN103353752B (en) | 2013-04-12 | 2013-06-18 | Based on the aircraft environmental control system Control Component method for diagnosing faults of level Four RBF neural |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN103353752B (en) |
Families Citing this family (10)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN104850889B (en) * | 2014-11-19 | 2017-02-22 | 北京航空航天大学 | Airplane rotation actuator drive unit adaptive fault detection, isolation and confidences assessment method |
| CN104536292A (en) * | 2014-12-05 | 2015-04-22 | 北京航空航天大学 | Method for conducting fault diagnosis on heat exchanger of aircraft environmental control system based on STF (Strong Tracking Filter) and MB |
| CN104390776B (en) * | 2014-12-10 | 2016-11-16 | 北京航空航天大学 | A kind of redundance aileron actuator fault detect, diagnosis and performance estimating method |
| CN106301610A (en) * | 2016-08-29 | 2017-01-04 | 北京航空航天大学 | The adaptive failure detection of a kind of superhet and diagnostic method and device |
| CN106571882A (en) * | 2016-09-06 | 2017-04-19 | 北京航空航天大学 | Superheterodyne receiver performance degradation assessment and prediction method and system thereof |
| CN108227676A (en) * | 2017-12-28 | 2018-06-29 | 浙江工业大学 | The online fault detect of valve-controlled cylinder electrohydraulic servo system, estimation and localization method |
| CN109447401A (en) * | 2018-09-18 | 2019-03-08 | 中国科学院自动化研究所 | Ground health control method, system and device based on adaptive resource allotment |
| CN109376413B (en) * | 2018-10-08 | 2022-08-02 | 四川大学 | Online neural network fault diagnosis method based on data driving |
| CN111308985B (en) * | 2020-02-18 | 2021-03-26 | 北京航空航天大学 | Performance degradation evaluation method for control assembly of airplane environmental control system based on NSCT and DM |
| CN115061447B (en) * | 2022-05-07 | 2024-10-11 | 河南科技大学 | Fault detection method for high-speed aircraft temperature control system |
Citations (3)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN101697079A (en) * | 2009-09-27 | 2010-04-21 | 华中科技大学 | Blind system fault detection and isolation method for real-time signal processing of spacecraft |
| CN102122133A (en) * | 2011-01-21 | 2011-07-13 | 北京工业大学 | Self-adaption wavelet neural network abnormity detection and fault diagnosis classification system and method |
| CN102681442A (en) * | 2012-05-28 | 2012-09-19 | 上海海事大学 | Intelligent fault-tolerant control system and control method for zonal power distribution of full-electric boat |
Family Cites Families (2)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| WO2006086894A1 (en) * | 2005-02-15 | 2006-08-24 | Abb Research Ltd | Diagnostic device for a process control system |
| EP1913506A4 (en) * | 2005-07-11 | 2008-08-13 | Brooks Automation Inc | Intelligent condition monitoring and fault diagnostic system for predictive maintenance |
-
2013
- 2013-06-18 CN CN201310240505.0A patent/CN103353752B/en not_active Expired - Fee Related
Patent Citations (3)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN101697079A (en) * | 2009-09-27 | 2010-04-21 | 华中科技大学 | Blind system fault detection and isolation method for real-time signal processing of spacecraft |
| CN102122133A (en) * | 2011-01-21 | 2011-07-13 | 北京工业大学 | Self-adaption wavelet neural network abnormity detection and fault diagnosis classification system and method |
| CN102681442A (en) * | 2012-05-28 | 2012-09-19 | 上海海事大学 | Intelligent fault-tolerant control system and control method for zonal power distribution of full-electric boat |
Also Published As
| Publication number | Publication date |
|---|---|
| CN103353752A (en) | 2013-10-16 |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN103353752B (en) | Based on the aircraft environmental control system Control Component method for diagnosing faults of level Four RBF neural | |
| CN104390776B (en) | A kind of redundance aileron actuator fault detect, diagnosis and performance estimating method | |
| CN108711005A (en) | Flight risk analysis method based on QAR data and Bayesian network | |
| Habibi et al. | Power maximization of variable-speed variable-pitch wind turbines using passive adaptive neural fault tolerant control | |
| CN111024143A (en) | Hypersonic aircraft sensor cascading failure diagnosis and fault-tolerant control method | |
| CN102929142A (en) | Method for designing controller of time-varying aircraft model with uncertainty | |
| CN104749958A (en) | Model-depended average dwell time asynchronous fuzzy control method of nonlinear switching system | |
| Li et al. | Performance simulation and diagnosis of faulty states in air-cycle refrigeration systems in civil aircrafts | |
| CN102929128A (en) | Method for designing controller of aircraft model with uncertainty | |
| Summers et al. | Towards cyber-physical special protection schemes: Design and development of a co-simulation testbed leveraging sceptre™ | |
| CN105205492B (en) | A kind of fuzzy arest neighbors fusion diagnosis method of therrmodynamic system failure | |
| CN102929143B (en) | Control design method for aircraft time lag model | |
| CN104023350A (en) | Self-healing method for wind turbine generator condition monitoring system | |
| Wang et al. | Research on health assessment and fault prediction method of aircraft electromechanical system based on fault tree | |
| CN108512312B (en) | A kind of substation interval primary equipment running state analysis method | |
| Li et al. | Research on system reliability of eight-rotor fire-fighting uavs | |
| Wang et al. | Self-adaptive RBF neural network PID control in exhaust temperature of micro gas turbine | |
| Wang et al. | Modeling and simulation in an aircraft safety design based on a hybrid AHP and FCA algorithm | |
| Xu et al. | Dynamic neuro-fuzzy control design for civil aviation aircraft in intelligent landing system | |
| Lin et al. | Simulation research in fault diagnosis of environmental control system for fighter aircraft | |
| CN114694441B (en) | Airborne self-defense interference pod maintenance training simulation cabin | |
| Boniface | Design and Development of High Performance Unmanned Autonomous Aerial Vehicle (UAAV) Using MATLAB/Simulink in Civil Applications | |
| Zhao et al. | Detection of spike-type stall of axial compressors based on dilated causal convolutional neural networks | |
| Zhao et al. | Fault line selection method of distribution network based on improved neural network | |
| Cheng et al. | Neural Network Incremental Dynamic Inversion-Based Trajectory Tracking Control of an Aircraft with Pilot Mishandling |
Legal Events
| Date | Code | Title | Description |
|---|---|---|---|
| C06 | Publication | ||
| PB01 | Publication | ||
| C10 | Entry into substantive examination | ||
| SE01 | Entry into force of request for substantive examination | ||
| C14 | Grant of patent or utility model | ||
| GR01 | Patent grant | ||
| CF01 | Termination of patent right due to non-payment of annual fee |
Granted publication date: 20151028 |
|
| CF01 | Termination of patent right due to non-payment of annual fee |