CN101299004B - Vibrating failure diagnosis method based on determined learning theory - Google Patents

Abstract

The present invention discloses a vibration fault diagnosis method based on confirming the learning theory. The method comprises the following steps: (1) executing learning exercise to the normal mode and fault mode of the diagnosed system; (2) establishing a mode database (comprising a normal mode and a fault mode); (3) establishing a dynamic estimator; (4) comparing the state of the dynamic estimator with the state of the detected system to generate a residual error; and (5) evaluating the residual error thereby discovering and eliminating the fault. The method is suitable for diagnosing the fault of complicated unknown non-linear vibration system. The unknown normal mode and fault mode can be studied to establish a mode database thereby executing fast failure discovery and elimination.

Description

A kind of based on the vibrating failure diagnosis method of determining the theories of learning

Technical field

The invention belongs to the system fault diagnosis field, be specifically related to a kind of based on the vibrating failure diagnosis method of determining the theories of learning.

Background technology

Fault diagnosis has crucial meaning to the modern project technological system.Glitch can make that production process interrupts, product is destroyed, and major break down can cause catastrophic effect, as injury to personnel, systemic breakdown etc.At present, about the method for fault diagnosis can be divided into method based on analytic model, based on method for processing signals with based on the method (seeing Zhou Donghua, Ye Yinzhong work, " modern fault diagnosis and fault-tolerant control ", publishing house of Tsing-Hua University, 2000) of knowledge.Need set up the mathematical model of monitored system based on the method for analytic model, but in fact, the system particularly mathematical model of complication system is to be difficult to obtain.Do not need the mathematical model of system based on method for processing signals, but just lacked deep understanding like this, particularly to the understanding of system's non-linear dynamic characteristic the system failure.Based on the method for knowledge, also there is bigger difficulty aspect automatic learning knowledge, the automatic working knowledge as expert system.

Radially base (radial basis function) neural network is called for short the RBF neural network, has the functional approximation capability and the optimal approximation characteristic of arbitrary accuracy, has obtained more application in fault diagnosis field.The main shortcoming of present this method is the physical significance of indigestion neural network, is difficult to guarantee that neural network weight converges to optimal value, and is difficult to guarantee that neural network really approaches dynamic system and then fault is carried out modeling.

Vibration extensively is present in the in service of engineering system, and vibration can be divided into random vibration and determination vibration.Random vibration is random, and determination vibration is clocklike, is produced by deterministic system.

The determinacy theories of learning are being used to nonlinear dynamic system in the recent period in the evaluation of cycle or shuttling movement.By adopting local RBF neural network, the proof part continues incentive condition and can be satisfied, as returning the lasting incentive condition of subvector along some of circulation system track among the RBF, this part continues incentive condition can be stable so that satisfy index along the error identification systems of circulation system track.Therefore, can obtain at regional area the accurate neural network of system dynamics is approached along the systemic circulation track.

The state trajectory of system is meant the track that the state vector of system forms over time in state space.The system that the present invention considered is the system that total state can be surveyed.

Lasting incentive condition is an important notion in System Discrimination field, has only the satisfied incentive condition that continues can estimate that just it is worth most to distinguishing of unknown parameter.

From in essence, the basic problem of fault diagnosis is to be difficult to system is carried out modeling.Just diagnose if avoid modeling problem, then cause erroneous judgement to fail to judge easily, can't guarantee the reliability of diagnostic result from symptom, characteristic parameter.Simultaneously, the behavior of most systems is a kind of dynamic nonlinear behaviors, and it is carried out modeling is the key of understanding fault behavior and character.

Summary of the invention

The objective of the invention is to overcome the shortcoming and defect of above-mentioned prior art, provide a kind of based on the vibrating failure diagnosis method of determining the theories of learning, this method can be at the fault modeling that carries out of the non-linear vibrating system of the unknown.Traditionally, modeling is a mathematical model of setting up system, but modeling is a difficult problem to the nonlinear system systems engineering.Here, we set up a library to the nonlinear system that produces oscillation behavior.Pattern when this library has comprised the operation of a large amount of systems, comprise under the normal condition and various failure conditions under pattern.The all corresponding a kind of pattern of each normal condition or failure condition.The pattern here is meant dynamic mode, constitutes by the factor of two aspects, and the one, the track of system, the 2nd, along the internal dynamic of system trajectory.Pattern can obtain from historical data or real time data.Can carry out the study and the upgrade mode storehouse of new model when having new fault to take place.Modeling process to various patterns is according to determining the theories of learning.

Technical scheme of the present invention is realized by following steps:

A kind of described vibration is meant determination vibration based on the vibrating failure diagnosis method of determining the theories of learning, and the state trajectory of system was cycle, the track class cycle or chaos when this vibration took place, and this method comprises the steps:

(1) learning training of vibration mode: adopt the RBF neural network, to vibrational system under normal circumstances with various failure conditions under each vibration mode carry out learning training, described learning training adopts based on the learning method of Liapunov and according to determining the theories of learning, realizes the weight convergence of RBF neural network and RBF neural network approaching the internal dynamic of system vibration pattern;

(2) set up library: the average of getting each weights in a period of time behind the weight convergence described in the step (1) is stored in it in library as the learning training result;

(3) set up dynamic estimator: utilize the weights in the described library of step (2) to set up constant RBF neural network, utilize this constant RBF neural network to set up dynamic estimator then, the corresponding a kind of vibration mode of each dynamic estimator, when the pattern of dynamic estimator correspondence takes place, constant RBF neural network can be recalled the knowledge of having acquired fast, and the internal dynamic information of this vibration mode is provided;

(4) structure residual error: the state of each dynamic estimator is compared with the state of detected system respectively, with the error between them as residual error;

(5) residual error assessment: set a threshold value, if have only a residual error in greater than the time of one-period less than this threshold value, judge that then the pairing dynamic estimator of this residual error during this period of time with by diagnostic system is complementary, if have a plurality of residual errors with the section time in less than this threshold value, then each residual error is averaged l ₁Norm is selected wherein minimum average l ₁Pairing dynamic estimator conduct of norm and the dynamic estimator that is complementary by diagnostic system, if detected system with represent the dynamic estimator of normal mode to be complementary, just illustrate that detected system works is normal, if detected system with represent the dynamic estimator of certain fault to be complementary, the generation of this fault just is described.

In the said method, weight convergence has two kinds of situations in the step (1):

A kind of is that its weight convergence is to optimal value along the satisfied incentive condition that continues of neuron of the RBF neural network of system trajectory; Another kind is that its weights are zero substantially away from the neuron of the RBF neural network of system trajectory excited target not.

In the said method, approaching described in the step (1) is along the approaching of the internal dynamic of system trajectory, and is not approached away from the internal dynamic of system trajectory.

In the said method, the described learning training of step (1) is the process that knowledge obtains, and described knowledge is expressed with constant RBF neural network weight, and every group of corresponding a kind of vibration mode of weights expressed as the static state of this vibration mode.

In the said method, the described dynamic estimator of step (3) is the utilization again to described knowledge, as the dynamic expression of vibration mode, the dynamic behaviour of reproducing pairing vibration mode.

In the said method, if the described residual error evaluation process of step (5) does not have dynamic estimator and is complementary by diagnostic system, then thinking has new fault to take place, and at this moment the described learning training process of setting up procedure (1) is learnt new fault once more.

In the said method, step (3), (4) and (5) are the observation processes to the system vibration fault, to the monitoring of all faults be walk abreast, dynamic and real-time process.

In the said method, the described coupling of step (5) is meant to be had similarity between diagnostic system and the dynamic estimator, and the measurement factor of its similarity is: i) by the difference of the state of the state of diagnostic system and dynamic estimator; Ii) along by the track of diagnostic system, the dynamic difference between dynamic by the internal system of diagnostic system with the internal system of dynamic estimator.

The present invention compared with prior art has following advantage:

1, it is more feasible than setting up mathematical model to set up library, because each pattern is corresponding a kind of system action, library is described by the behavior of a large amount of patterns to system, and library can also be dealt with the generation of new fault by the upgrading that does not stop simultaneously.And mathematical model is just by the behavior that model comes expression system, data when this requires training need travel through the running space of total system, such data are to be difficult to obtain in real process, and may be unnecessary, because system may just move on one or several track.On the other hand, if data are enough abundant, data volume will be very big, and training simultaneously will be very difficult thing.

2, with based on method of characteristic parameters compare, more fully stored data based on the vibrating failure diagnosis method of determining the theories of learning.Extract characteristic parameter, certainly will lose bulk information, comprising Useful Information.In based on the method for determining the theories of learning, neural network is by determining study, and it is dynamic in the internal system of normal and various failure conditions lower edge system trajectory accurately to approach unknown system.The action process of system can be got off by complete preservation like this.

3, be a kind of method that system is carried out dynamic monitoring based on the vibrating failure diagnosis method of determining the theories of learning, can reflect system transients behavior and nonlinear characteristic.

4, the vibrating failure diagnosis method based on definite theories of learning can utilize neural network to learn automatically, and diagnoses automatically with minimum residual method, thereby improves the automaticity of failure diagnostic process greatly.

5, compare with traditional neural network learning, neural network in this method has concrete physical significance, in fact it is the mathematical model of having set up along track, so it is a dynamic process of accurately and comprehensively having described system, can help people that the mechanism of system and fault is carried out deep understanding, judge the reliability of diagnostic procedure simultaneously.And traditional neural network learning does not have physical significance, can't reach such effect.

6, because neural network has powerful approximation capability, and, can carry out accurate modeling, therefore can find small fault the dynamic mode of system by determining study.The diagnosis of glitch is a crucial thing in the fault diagnosis, because many catastrophic faults are all grown up by glitch.Only the commitment that takes place in fault is found just can the avert a calamity generation of incidents of fault.

7, compare with the learning training of traditional neural network, the acquisition that the present invention is real knowledge, and effectively utilized knowledge.The knowledge of describing the unknown failure pattern obtains by determining study, and this knowledge is expressed with the constant neural network weight.The process of utilizing knowledge is a kind of parallel, process dynamically and fast.

Description of drawings

Fig. 1 is the system architecture sketch of RBF neural network to unknown dynamic system study.

Fig. 2 is a RBF neural network synoptic diagram of the present invention.

Fig. 3 utilizes dynamic estimator to produce the structure diagram of residual error in the embodiment of the invention.

Fig. 4 is the state trajectory of normal mode among the embodiment.

Fig. 5 is the state trajectory of fault 1 among the embodiment.

Fig. 6 is the state trajectory of fault 2 among the embodiment.

Fig. 7 is the convergence situation of indivedual weights during the normal mode learning training among the embodiment.

Fig. 8 be among the embodiment behind the normal mode learning training RBF neural network approach by the internal dynamic of diagnostic system along system trajectory.

Fig. 9 is the internal dynamic during by the diagnostic system operate as normal and the situation of output of constant RBF neural network of representing the dynamic estimator of normal mode among the embodiment.

Figure 10 a represents the dynamic estimator of normal mode and by the time-domain diagram of the residual error between the diagnostic system among the embodiment.

Figure 10 b represents the dynamic estimator of fault 1 and by the time-domain diagram of the residual error between the diagnostic system among the embodiment.

Figure 10 c represents the dynamic estimator of fault 2 and by the time-domain diagram of the residual error between the diagnostic system among the embodiment.

Embodiment

Below in conjunction with embodiment and accompanying drawing, the specific embodiment of the present invention is further described.

Embodiment

Consider following by diagnostic system:

x ₁＝x ₂，s＝0，1，2 (1)

$x_2={\mathrm{\φ}}_2^s(x_1,x_2)$

X wherein ₁, x ₂Be system state, ${\mathrm{\φ}}_2^0(x_1,x_2)=1.08(0.95x_1^2-1)x_2-0.81x_1$ Unknown internal dynamic under the expression normal condition, ${\mathrm{\φ}}_2^1(x_1,x_2)=1.08(0.95x_1^2-1)x_2-0.81x_1+0.75\sin \left(x_1\right),$

${\mathrm{\φ}}_2^1(x_1,x_2)=1.08(0.95x_1^2-1)x_2-0.81x_1+0.75\cos \left(x_1\right)$ Represent respectively system break down 1 and fault 2 situations under unknown internal dynamic.The state trajectory of normal mode, fault 1 and fault 2 is respectively as Fig. 4 figure, 5 and shown in Figure 6 in the present embodiment.X among the figure ₁, x ₂It is system state.

Employing is as follows based on the step of the vibrating failure diagnosis method of determining the theories of learning:

(1) learning training of vibration mode:

Adopt neural network as shown in Figure 1 to the system architecture sketch of unknown system learning training.Wherein x is a system state, Be the output of RBF neural network learning device.

With the input of measured system state as neural network, neural network adopts radially base net network of Gauss, as shown in Figure 2, and x wherein ₁, x ₂..., x _nBe the input of neural network, y is the output of neural network, has only x in the present embodiment ₁, x ₂Two inputs.

RBF neural network learning device is by following The Representation Equation:

${\stackrel{\·}{\hat{x}}}_2=-a({\hat{x}}_2-x_2)+{\hat{W}}_{2}S(x_1,x_2)---\left(2\right)$

Wherein

Be the state of RBF neural network learning device, x ₁, x ₂Be the state of system (1), a=5 is the constant of design,

Be the RBF neural network, be used for unknown function φ in the system of approaching (1) ₂ ^s(x ₁, x ₂).

Estimation weight in the formula (2)

Use learning method based on Liapunov (Lyapunov):

${\stackrel{\·}{\hat{W}}}_2=-\mathrm{\ΓS}(x_1,x_2){\tilde{x}}_2-\mathrm{\σ\Γ}{\hat{W}}_2---\left(3\right)$

Wherein ${\tilde{x}}_2={\hat{x}}_2-{\mathrm{<figure-callout\; id="2"\; label="x"\; filenames="S2008100289815E00022.png,S2008100289815E00031.png"\; state="\{\{state\}\}">x</figure-callout>}}_2,$ Γ＝2，σ＝0.001。

According to definite theories of learning, as shown in Figure 7, continue incentive condition near the satisfied part of the neuronic weights of system trajectory, thus the optimal value of converging to, and very little away from the degree of the neuron excited target of system trajectory, remain essentially in the scope of very little value.On the other hand, along system trajectory, the RBF neural network is approached by the internal dynamic of diagnostic system, as shown in Figure 8.

(2) set up library:

When the RBF neural network weight was restrained, the weights average that gets a period of time after holding back was a learning outcome:

${\stackrel{\&OverBar;}{W}}_2={\mathrm{mean}}_{t\&Element;[t_a,t_b]}{\hat{W}}_2\left(t\right)---\left(4\right)$

T wherein _b＞t _aA period of time after＞0 this convergence process of expression.We just obtain the constant vector W of neural network weight like this ₂, it is stored in the library.

Originally implement three vibration modes, learning training is carried out to these three vibration modes in (1) (2) set by step, just obtains three groups of weights: W at last ₂ ^s, s=0,1,2.Every group of corresponding vibration mode of weights expressed as the static state of corresponding vibration mode.These three groups of weights constitute a library simultaneously.If the vibration mode of learning training is more, library just has more groups of weights.

(3) set up dynamic estimator:

From library, access weights, give the RBF neural network, so just set up three constant RBF neural networks these weights assignment.The expression-form of these three constant RBF neural networks is:

${\stackrel{\&OverBar;}{W}}_2^{\mathrm{sT}}S(x_1,x_2),s=\mathrm{0,1,2}$

Utilize these three constant RBF neural networks to set up dynamic estimator.Dynamic estimator can be described by following equation:

${\stackrel{\&CenterDot;}{\stackrel{\&OverBar;}{x}}}_2^s=-b({\stackrel{\&OverBar;}{x}}_2^s-x_2)+{\stackrel{\&OverBar;}{W}}_2^{\mathrm{sT}}S(x_1,x_2),s=\mathrm{0,1,2}---\left(5\right)$

X wherein ₂It is the state of dynamic estimator.x ₁, x ₂Be the state of system (1), b＞0 is usually less than a (a provides in (2)).W ₂ ^STS (x ₁, x ₂) be the constant RBF neural network that obtains by determinacy study, when the vibration mode of dynamic estimator correspondence took place, constant RBF neural network can be recalled the knowledge of having acquired fast, and the internal dynamic information of this vibration mode is provided.As shown in Figure 9, when by diagnostic system when working properly, by the internal dynamic of diagnostic system over time shown in solid line among the figure, the output of representing constant RBF neural network in the dynamic estimator of normal mode as shown in phantom in FIG., article two, line is approximate overlaps, and shows that the output of constant RBF neural network in the dynamic estimator of representing normal mode approaches by the internal dynamic of diagnostic system.

(4) structure residual error:

Synchronous error between dynamic estimator and the detected system is as residual error, and its system architecture sketch as shown in Figure 3.Wherein x is the state of system (1), x ^s(s=0,1,2) is the state of dynamic estimator, ${\tilde{x}}^s={\stackrel{\&OverBar;}{x}}^s-x,(s=\mathrm{0,1,2})$ It is the residual error between dynamic estimator and the detected system.

(5) residual error assessment:

Set a threshold value, this threshold value is followed according to dynamic estimator maximal value of residual error when being mated by diagnostic system and is set, and this enforcement is set at 0.1.

Figure 10 shows when fault 1 constantly took place at 20 seconds each dynamic estimator and by the situation of the residual error between the diagnostic system.Before 22 seconds, have only the residual error of dynamic estimator 0 can therefore during this period of time can not judge system's operate as normal greater than threshold value; The residual error of having only dynamic estimator 1 after 22 seconds can be greater than threshold value, therefore can judge 22 seconds after fault 1 take place.Therefore the situation that does not occur a plurality of residual errors all little threshold value in greater than the time of one-period in the present embodiment needn't further distinguish dynamic estimator.

If the situation of a plurality of residual errors all little threshold value in greater than the time of one-period just need be averaged l to these residual errors ₁Norm.

Residual error is averaged l ₁Norm is calculated as follows:

${\left|\right|{\tilde{x}}_2^s\left(t\right)\left|\right|}_{1T}=\frac{1}{T}{\&Integral;}_{t-T}^t\left|{\tilde{x}}_2^s\left(\mathrm{\&tau;}\right)\right|\mathrm{d\&tau;},t>T,s=\mathrm{0,1,2}---\left(6\right)$

T is by the cycle of diagnostic system in the formula (6).

Claims (8)

1. one kind based on the vibrating failure diagnosis method of determining the theories of learning, and described vibration is meant determination vibration, is cycle, the track class cycle or chaos by the state trajectory of diagnostic system when this vibration takes place, and it is characterized in that this method comprises the steps:

(1) learning training of vibration mode: adopt the RBF neural network, to by diagnostic system under normal circumstances with various failure conditions under each vibration mode carry out learning training, described learning training adopts based on the learning method of Liapunov and according to determining the theories of learning, realizes the weight convergence of RBF neural network and RBF neural network approaching the internal dynamic of system vibration pattern;

(2) set up library: the average of getting each weights in a period of time behind the weight convergence described in the step (1) is stored in it in library as the learning training result;

(3) set up dynamic estimator: utilize the weights in the described library of step (2) to set up constant RBF neural network, utilize this constant RBF neural network to set up dynamic estimator then, the corresponding a kind of vibration mode of each dynamic estimator, when the pattern of dynamic estimator correspondence takes place, constant RBF neural network can be recalled the knowledge of having acquired fast, and the internal dynamic information of this vibration mode is provided;

(4) structure residual error: the state of each dynamic estimator is compared respectively with by the state of diagnostic system, with the error between them as residual error;

(5) residual error assessment: set a threshold value, if have only a residual error in greater than the time of one-period less than this threshold value, judge that then the pairing dynamic estimator of this residual error during this period of time with by diagnostic system is complementary, if have a plurality of residual errors with the section time in less than this threshold value, then each residual error is averaged l ₁Norm is selected wherein minimum average l ₁Pairing dynamic estimator conduct of norm and the dynamic estimator that is complementary by diagnostic system, if by diagnostic system with represent the dynamic estimator of normal mode to be complementary, just explanation is working properly by diagnostic system, if by diagnostic system with represent the dynamic estimator of certain fault to be complementary, the generation of this fault just is described.

2. method according to claim 1 is characterized in that weight convergence has two kinds of situations in the step (1):

A kind of is that its weight convergence is to optimal value along the satisfied incentive condition that continues of neuron of the RBF neural network of system trajectory; Another kind is that its weights are zero substantially away from the neuron of the RBF neural network of system trajectory excited target not.

3. method according to claim 1, it is characterized in that approaching described in the step (1) is along the approaching of the internal dynamic of system trajectory, and is not approached away from the internal dynamic of system trajectory.

4. according to each described method of claim 1～3, it is characterized in that the described learning training of step (1) is the process that knowledge obtains, described knowledge is expressed with constant RBF neural network weight, and every group of corresponding a kind of vibration mode of weights expressed as the static state of this vibration mode.

5. method according to claim 4 is characterized in that the described dynamic estimator of step (3) is the utilization again to described knowledge, as the dynamic expression of vibration mode, the dynamic behaviour of reproducing pairing vibration mode.

6. method according to claim 5, it is characterized in that if the described residual error evaluation process of step (5) does not have dynamic estimator and is complementary by diagnostic system, then thinking has new fault to take place, and at this moment the described learning training process of setting up procedure (1) is learnt new fault once more.

7. method according to claim 5 is characterized in that step (3), (4) and (5) are the observation processes to the system vibration fault, to the monitoring of all faults be walk abreast, dynamic and real-time process.

8. method according to claim 5, it is characterized in that the described coupling of step (5) is meant to be had similarity between diagnostic system and the dynamic estimator, the measurement factor of its similarity is: i) by the difference of the state of the state of diagnostic system and dynamic estimator; Ii) along by the track of diagnostic system, the dynamic difference between dynamic by the internal system of diagnostic system with the internal system of dynamic estimator.

Images