CN102606557A - Health evaluation method for hydraulic system based on fault observer and SOM (self-organized mapping) - Google Patents

Abstract

The invention provides a health evaluation method for a hydraulic system based on a fault observer and SOM (self-organized mapping). The method includes: firstly, training the fault observer by the aid of input and output signals of the hydraulic system in a normal operation state to acquire residual signals in the normal state; secondly, acquiring residual signals of a certain component in a degenerate state through the fault observer by the aid of input and output signals of the degenerate component in the system; thirdly, extracting time-domain features of the residual signals in the two states and building a self-organized mapping model by the aid of residual features in the normal state; and finally, inputting residual features of the degenerate component to the self-organized mapping model to acquire the health degree of the hydraulic system in the corresponding states, and judging whether the current hydraulic system can continue running or not by the aid of the lowest health threshold value. The health evaluation method is used for evaluating health of the hydraulic control system in real time, is fine in evaluation effect, reduces dependency on historical data, and has engineering practicability.

Description

A kind of hydraulic system health evaluating method based on fault visualizer and SOM

Technical field

The invention belongs to the fault diagnosis technology field of hydraulic system, be specifically related to a kind of health evaluating method of the hydraulic system that combines based on fault visualizer and self-organization mapping (SOM).

Background technique

Since improving constantly of scientific and technological level, the automation of machinery and intelligent level phenomenal growth, mechanical structure becomes and becomes increasingly complex, and reliability plays an important role in industrial products.But no matter reliability of products has many height, because the influence of excessive load, abominable factors such as work condition environment, As time goes on and constantly the performance of machine can degenerate, and equipment finally disabler can occur.The machinery equipment maintenance can guarantee to a certain extent that as a kind of effective and efficient manner machinery equipment has lasting high reliability.But this simple dependence maintenance mode improves the reliability of equipment, in reality, usually can face the appearance that the equipment mistake is repaiied, situation such as former equipment life-span are repaiied, influenced in the leakage of bad equipment, also can increase maintenance cost.Therefore how the current health status of equipment is assessed, known the residual life of equipment indirectly, become one of present equipment health status Study of Monitoring focus.

Visualizer belongs to wherein a kind of method of carrying out faut detection based on analytical model; And based on the basic thought of analytical model method be: before and after fault takes place; System is in normal state and fault state respectively; The analytical model of system when setting up normal state, the measurement output through comparison model output and real system produces the fault indication device that is called " residual error ".If fault takes place, can influence the measurement output of system, cause residual error to change, this residual error is estimated, but both faut detections can be assessed the degenerate state of system again.

Self-organization mapping (SOM) is based on a kind of important kind of the neuron network of unsupervised learning method, and is one of research field the most glamorous of neuron network.It can be imported sample association through it and detect its regularity and input sample relation each other, and according to these information self-adapting adjustment networks of importing samples, later response of network and input sample is adapted.Not only can learn the distribution situation of input vector, can also learn the topological structure of input vector, its single neuron does not play a decisive role to pattern classification, and will lean on a plurality of neuronic synergies could accomplish pattern classification.

At present, be in the starting stage, and most existing appraisal procedure all is on the basis of the feature identification of fault, current state to be assessed for the health evaluating of hydraulic system.One of key of this method is to need the historical failure data of equipment to be assessed, needs could realize the health status assessment through the history feature pattern that compares current state and various fault degrees because have only.Yet, the historical failure data of equipment often difficulty obtain, especially to some equipment of newly installing and using, do not have the historical failure data at all, assessment weak effect, poor robustness, therefore the application of the method has received very big restriction in actual engineering.

Summary of the invention

The objective of the invention is in order to solve when carrying out the assessment of hydraulic system health status; Existing method is assessed weak effect, poor robustness and is depended on the problem of historical data, proposes a kind of hydraulic system health evaluating method that combines with the self-organization mapping based on the fault visualizer.

The present invention is a kind of based on fault visualizer and the hydraulic system health evaluating method that the self-organization mapping combines, and specifically may further comprise the steps:

Step 1, foundation and training fault visualizer; Described fault visualizer adopts neuron network to realize, is input as the input and output signal of hydraulic system, imports after the sluggish link of output signal through a true hydraulic system lag output function of simulation; Obtain the historical input and output signal under the hydraulic system normal working; Training fault visualizer, training input sample is the historical input and output signal that obtains, the training output sample is the history output signal that obtains; After the training of fault visualizer is accomplished, obtain hydraulic system residual signals in normal operation; Described residual signals is done difference by the estimated output signal of fault visualizer with the output signal of corresponding hydraulic system and is obtained.

At set intervals, obtain the input signal r (t) and the output signal y of the hydraulic system under the current time t working state _r(t), the input fault visualizer obtains the estimated output signal of fault visualizer

Further obtain the residual signals ε (t) of hydraulic system under the current t moment working state; Described residual signals ε (t) is by the estimated output signal of fault visualizer

Output signal y with corresponding hydraulic system _r(t) doing difference obtains.

Step 2, to each residual signals, be divided into the k group, in each group, extract n sample, to each group extraction characteristic quantity, obtain k characteristic quantity; Described characteristic quantity adopts average, effective value, kurtosis or shape factor to represent.

Step 3, structure self organizing maps model; The neural network structure that model is an input layer+output layer is insinuated in described self-organization; Before training, set frequency of training, initial health degree and the total neuronic number of whole neuron network; Use the input vector of k the characteristic quantity composition input layer that the residual signals under the hydraulic system normal working extracts then, carry out the training of self organizing maps model.

Step 4, hydraulic system is carried out real-time health evaluating: the self organizing maps model that will train in the characteristic quantity input of the residual signals of the hydraulic system under the moment t working state obtains the health degree A of the hydraulic system under the t working state constantly.

The minimum health degree B of the hydraulic system of step 5, the health degree A that incites somebody to action the hydraulic system under the moment t working state and setting compares, if A＞B, then hydraulic system operational excellence; If A≤B; Then there is fault in hydraulic system, and hydraulic system out of service is carried out fault and got rid of.

Advantage of the present invention and good effect are:

(1) made full use of the powerful cluster function of self-organizing map neural network, realized the health status of hydraulic system is assessed, this method strong robustness, the assessment effect is remarkable;

(2) residual error of introducing the fault visualizer of hydraulic system is carried out health evaluating, can obtain the health status of current hydraulic system in real time;

(3) defined the healthy threshold value of hydraulic system before total failure, overcome system and crossed maintenance and leak the problem of keeping in repair;

(4) adopt based on fault visualizer and the health evaluating method that self-organization mapping (SOM) combines, solved at present hydraulic system is assessed the problem that weak effect and existing method are not suitable for this hydraulic system;

(5) the inventive method only needs the time-domain signal under the hydraulic system normal state, can set up health evaluation model, realizes the health evaluating of hydraulic system, has reduced the dependence to historical data, has very high practical applications property;

(6) compare with existing equipment health evaluating method, the inventive method has significantly improved versatility and precision.

Description of drawings

Fig. 1 is the whole flow chart of steps of the health evaluating method of hydraulic system of the present invention;

Fig. 2 is that the structure of fault visualizer in the step 1 is set up schematic representation;

Fig. 3 is a SOM mesh topology schematic representation;

Fig. 4 is the concrete training process schematic representation of self organizing maps model in the step 3 of the present invention;

Fig. 5 is the input signal schematic representation of hydraulic system under normal state in the embodiment of the invention;

Fig. 6 is the output signal schematic representation of hydraulic system under normal state in the embodiment of the invention;

Fig. 7 is the system residual signals schematic representation of hydraulic system under normal state in the embodiment of the invention;

Fig. 8 is the system residual signals schematic representation of hydraulic system under the electron-amplifier degenerate state in the embodiment of the invention;

Fig. 9 is the schematic representation of hydraulic system effective value of residual signals under normal state in the embodiment of the invention;

Figure 10 is the schematic representation of hydraulic system effective value of residual signals under the electron-amplifier degenerate state in the embodiment of the invention;

Figure 11 is the training process of self-organization mapping assessment models in the embodiment of the invention;

Figure 12 is the plotted curve that the health degree of hydraulic system under the electron-amplifier degenerate state reduces in the embodiment of the invention.

Embodiment

To combine accompanying drawing and embodiment that the present invention is done further detailed description below.

Equipment is carried out a kind of important branch and the research focus that real-time health evaluating has become fault diagnosis at present.More existing appraisal procedures all are that the trouble signal according to system carries out health evaluating, but these trouble signals are faint and be difficult to obtain.The present invention is directed to the structure and the data characteristic thereof of hydraulic system, propose a kind of hydraulic system health evaluating method that combines with self-organization mapping (SOM) based on the fault visualizer.The core concept of the inventive method be through calculating the current running state of hydraulic system and normal operation state respectively the contact ratio in characteristic of correspondence space come the assessed for performance degree of degeneration.The inventive method is transferred to detection and quantificational description to the performance degradation degree from traditional state recognition to fault; Through calculating equipment running status to be assessed and the normal state drift rate between the characteristic of correspondence space respectively, realize the purpose that the health status of equipment is assessed.

The present invention is a kind of structural feature to hydraulic system, the health evaluating method that adopts fault visualizer and self-organizing map neural network to combine, and as shown in Figure 1, concrete steps are following:

Step 1, set up the fault visualizer, obtain the residual signals of hydraulic system.Described fault visualizer adopts neuron network to realize, is also referred to as Neural Network Observer.

It is as shown in Figure 2 specifically to set up process; Make up the hydraulic system model according to actual use condition; Historical input and output signal under the hydraulic system model normal working as training data, is utilized neuron network strong non-linear match function, training fault visualizer.Hydraulic system in the reality is a feedback control system, so the output of hydraulic system also is incorporated into the input end of Neural Network Observer.Consider that simultaneously should there be certain hysteresis in output, therefore before the input end that is incorporated into Neural Network Observer, add a Z ^-1Link is approached the operative scenario of real hydraulic system, wherein Z ^-1It is the sluggish link that to simulate true hydraulic system lag output function.Set up after the fault visualizer, when breaking down as if a certain parts in the hydraulic system, then the output signal of hydraulic system model can change.According to the fault visualizer estimated output signal of this moment, obtain the estimated output signal of fault visualizer and the output signal residual error between the two of hydraulic system again, this will be as the important parameter of hydraulic system overall performance assessment.

The concrete embodiment of the invention adopts RBF (Radial Basis Function, RBF is called for short) neuron network to make up the fault visualizer, and obtain the residual signals of hydraulic system through following process:

At first; Obtain the historical input and output signal under the hydraulic system normal working; And input signal that obtains and output signal be put in the vector training input sample as the RBF Neural Network Observer, with the history output signal of the hydraulic system of obtaining training output sample as the RBF Neural Network Observer; Before training, need handle between [1,1], configure the basic parameter of RBF neuron network then, begin training the normalization of training input and output sample.After training Neural Network Observer; Can obtain the estimated output signal of RBF Neural Network Observer; With the estimated output signal of fault visualizer and the output signal subtraction of hydraulic system model, can obtain the residual signals under the hydraulic system normal working again.

Secondly, when certain parts in hydraulic system take place to degenerate, obtain the input signal r (t) and the output signal y of current t hydraulic system model constantly _r(t), and the input and output signal that obtains is placed in the vector, is transported in the fault visualizer that has trained after this vector normalization is handled, obtain the estimated output signal of the fault visualizer of hydraulic system state this moment

With the estimated output signal that obtains

Output signal y with the hydraulic system model _r(t) subtract each other, obtain the residual signals ε (t) under certain component degradation state in the hydraulic system.

The characteristic quantity of step 2, extraction residual signals.

The residual signals that the fault visualizer obtains belongs to time-domain signal, need carry out temporal signatures to this signal and extract.Temporal signatures commonly used has average, effective value, kurtosis, shape factor etc., selects to meet the temporal signatures of this residual signals characteristics.What the inventive method was chosen is that effective value is as the characteristic quantity of treating health evaluating.The formula of confirming effective value is:

$\mathrm{rms}=\sqrt{\frac{x_1^2+x_2^2+...+{x_p}^2}{p}}---\left(1\right)$

Rms representes the effective value of residual signals, and p is illustrated in the sample size of choosing in this residual signals, x ₁, x ₂X _pRepresent the 1st, the 2nd respectively ... The value of p sample.

In the inventive method,, it is divided into the k group to each residual signals; In each group, extract p sample; Each group is extracted characteristic, obtain k characteristic quantity, a resulting k characteristic quantity is input in the self organizing maps model constructed in the step 3 as sample.

Step 3, structure self organizing maps model.

Self-organization mapping (SOM) method is a kind of clustering method of the tutor's of nothing study, and its network structure generally is made up of input layer and competition layer, does not have hidden layer, and the neuron between two-layer is realized two-way connection.With basic competition network difference be that its competition layer can be made up of one dimension or two-dimensional grid matrix-style, and the strategy of weights correction is also different.Fig. 3 is the SOM mesh topology of a two-dimensional grid.

Can see that from Fig. 3 the SOM neural network structure is the structure of level type, the typical structure form is: input layer+output layer.Input layer is to be used for accepting external information, and input pattern to the competition layer transmission, is play a part to observe, and through weight vector external information is pooled to each neuron of output layer, and the input nodal point number is identical with the dimension of input vector.Output layer also is competition layer, is to be responsible for input pattern is analyzed comparison, seeks rule, thus the effect of sorting out.Each neuron of output layer connects with other neuron side direction around it, is arranged in the checkerboard plane; Input layer is arranged for the individual layer neuron.Each node of input layer has been realized connecing with the totally interconnected of all output nodes; Just mean that also the dimension of the connection weight vector of each node of output layer equals the number of input layer node; And each input vector during by each text subject notion scanning input node coupling obtain; Therefore the dimension of input vector equals to import the node number and also equals to connect weight vector dimension (reference: Cai Lihong etc.; The improvement of SOM clustering algorithm and the application study in text mining [D] thereof. Nanjing Aero-Space University, 2011).

Why the SOM network is called the Feature Mapping neuron network; Be because network passes through the repetition learning to input pattern; Can be so that the space distribution density of connection power and the probability distribution of input pattern reach unanimity, the space distribution that promptly connects weight vector can reflect the statistical nature of input pattern.

In competition layer, neuronic competition is performed such: for that neuron G that wins, and N around it _GThe zone in, neuron is all obtaining excitement in varying degrees, and at N _GNeuron beyond the zone has all obtained inhibition, and promptly " with the triumph neuron is the center of circle, neighbour's neuron is shown excitability side feedback, and neighbour's far away neuron is shown inhibition side feedback, and the mutual excitation of neighbour person's phase, adjacent person far away suppresses each other ".It is big to show intermediate intensity on the whole, decays gradually in both sides, and the deep trend that is suppressed.

After learning the working principle of self-organization mapping, some basic parameters of setting network are such as neuronic number total in health degree corresponding under the frequency of training of network, the normal state, the network etc.Frequency of training all is artificial setting with the setting of initial health degree, and the setting meaning of initial health degree is in order to demarcate the health degree of hydraulic system under normal state.The frequency of training that is provided with here in the embodiment of the invention is 100, and initial health degree is 0.95, and total neuronic number d can set by empirical correlation according to sample size k in the network:

$d=\left[\sqrt{5*\sqrt{k}}\right]---\left(2\right)$

Sample size k value is exactly the quantitative value k of the group of in the step 2 each residual signals being divided, and the value of sample is exactly a characteristic quantity determined every group in the step 2.The input nodal point number of input layer is identical with the dimension of input vector; If under the hydraulic system normal working, only have a residual signals; K the characteristic quantity that then extracts from this residual signals formed an input vector of input layer; The dimension of this input vector is 1, and the input nodal point number of input layer is 1; If under the hydraulic system normal working, have plural residual signals; Then k characteristic quantity of each residual signals extraction formed an input vector of input layer; The dimension of input vector is identical with the number of residual signals in normal operation, and the input neuron number of the input layer also number with residual signals in normal operation is identical.

After basic parameter configured, the effective value of the residual signals under the promptly available hydraulic system normal working carried out the training of self organizing maps model.The concrete training process of SOM method is as shown in Figure 4.Each step in the flow process, details are as follows:

Step (1) initialization.

Connection weights W with network _IjTax is provided with the radius R (0) of initial neighborhood with [0,1] interval random value, and obtaining initial field is N _c(0), confirms learning rate η (0), 0＜η (0)＜1; It is T that the total study number of times of network is set, and a counter l is set writes down current study number of times, initial l=0.And the normalized input vector

and connection weight vector

$\stackrel{\‾}{X_k}=\frac{X_k}{\left|\right|X_k\left|\right|}---\left(3\right)$

$\stackrel{\‾}{W_j}=\frac{W_j}{\left|\right|W_j\left|\right|}=\frac{(W_{1j},W_{2j},...,W_{\mathrm{ij}},...,W_{\mathrm{nj}})}{\sqrt{{\left(W_{1j}\right)}^2+...+{\left(W_{\mathrm{ij}}\right)}^2+...+{\left(W_{\mathrm{nj}}\right)}^2}}---\left(4\right)$

Input vector X _kBe the vector of the capable k row of n, n is the number of input layer, and k is a sample size, X _k=(X _1k, X _2K..., X _Ik..., X _Nk) ^Tr, X _IkFor having the row vector of k data, under the hydraulic system normal working, only has in the residual signals X _k=X _1k|| || expression is asked mould, matrix norm to be defined as in the matrix quadratic sum of all elements and is opened radical sign and obtain.W _jBe connected weight vector between expression competition layer neuron j and the input layer, be expressed as (W _1j, W _2j..., W _Ij..., W _Nj), W _IjBe connected weight vector between expression input layer i and the competition layer neuron j, W is the abbreviation of weight weight, j=1, and 2 ..., m, m are the neuron number on the competition layer.Under the hydraulic system normal working, only has in the residual signals W _j=W _1j

Step (2) is accepted input.

after the normalization as input vector, offered the input layer of network.

Step (3) is sought the triumph neuron.

Calculate input vector

Be connected weight vector

Between Euclidean distance d _Jk, and the neuron c of the minimum competition layer of chosen distance is the triumph neuron, that is:

$\left|\right|\stackrel{\‾}{X_k}-{\stackrel{\‾}{W}}_c\left|\right|=\min \left\{d_{\mathrm{jk}}\right\}---\left(5\right)$

Wherein, W _cThe connection weight vector of the c of output layer epineural unit after the expression normalization.

Step (4) parameter adjustment.

All neuronic connections in competition layer triumph neuron c and the triumph field thereof are acted temporarily as following adjustment, and the neuron weights outside the field remain unchanged.

W _ij(l+1)＝W _ij(l)+η(l)h _c，j(l)(X _i-W _ij(l)) (6)

Wherein η (l) is the l time a learning rate, 0＜η (l)＜1, h _{C, j}(l) be the field function, learning rate and field function are all along with the time reduces gradually; W _Ij(l+1) be connected weights between the input layer i that learns for the l+1 time of expression and the competition layer neuron j.

Step (5) is upgraded learning rate and field function.

η (l) and h _{C, j}(l) all be to begin to get greatly, diminish h gradually along with the increase of study number of times _{C, j}(l) adopt Gaussian function usually, as follows:

$h_{c,j}\left(l\right)=\exp (-\frac{d_{\mathrm{cj}}^2}{{2R}^2\left(l\right)})---\left(7\right)$

Wherein, d _CjEuclidean distance in expression triumph neuron c and the field between arbitrary output layer neuron j that is activated, R (l) is the field radius of the l time study, and the field radius is the same with learning rate also all to diminish along with the increase of study number of times l gradually, and adjustment law is following:

$R(l+1)=\mathrm{INT}((R\left(l\right)-1)\×(1-\frac{l}{T}))+1$ (INT is a bracket function) (8)

Learning rate η (l) adjusts through following formula:

$\mathrm{\η}(l+1)=\mathrm{\η}\left(l\right)-\frac{\mathrm{\η}\left(0\right)}{T}---\left(9\right)$

Step (6) circulation study.If it is a plurality of that input vector has, then need judge current whether all input vector global learnings being finished, if not, then next input vector is input to input layer, return step (3) then and carry out, finish up to all input vector global learnings; If all input vector study finishes, then order study number of times l=l+1 returns step (2) and carries out, till l=T or network convergence.

Step 4, carry out real-time health evaluating.

At first, set the minimum health degree B under the hydraulic system normal state, the setting of this minimum health degree B should be to obtain according to a large amount of real hydraulic system data statistics.If lack the hydraulic system True Data, so initial minimum health degree B sets by expertise, can be set in 0.2～0.3 scope such as B, along with the operation of actual hydraulic pressure system, this minimum health degree B can adjust according to the real data of hydraulic system.

At set intervals; Obtain the input and output signal of the hydraulic system under the current time t working state in real time; Through the fault visualizer that has trained, obtain when the estimation of prior fault visualizer output, the output signal of current hydraulic system is exported to subtract each other with this estimation obtain residual epsilon (t).Extract the effective value rms of this residual error, be input to the self-organization mapping assessment models that has trained, the output of self-organization mapping assessment models is exactly the health degree A of the hydraulic system under the current time t working state.

Step 5, thrashing are judged.

Hydraulic system health degree A under the current time t working state that obtains according to step 4 combines the minimum health degree B of the hydraulic system that experience in the past sets again, knows the remaining time of current health degree A before hydraulic system total failure.As the hydraulic system health degree A that obtains during greater than the minimum health degree B that sets, then the hydraulic system under the current time t working state can also continue operation; Otherwise there is fault in current hydraulic system, need be out of service, fix a breakdown.Like this, just can obtain the health status of hydraulic system in real time, fundamentally eliminate the situation of crossing maintenance or leaking maintenance, improve the reliability of hydraulic system.

Embodiment:

This instance takes the emulated data of hydraulic system (Hydrauservo System) to verify.Use Hydrauservo System sample signal normal and that its electron-amplifier performance takes place under the decline two states to carry out detection validation to the present invention is based on the fault visualizer with the hydraulic system health evaluating method that the self-organization mapping combines respectively, concrete steps are following:

Step 1, set up the fault visualizer, obtain the residual signals of hydraulic system.

The input signal of known hydraulic is sinusoidal, and amplitude is 28, and frequency is 4 π, is [0.2,0.2] between the hydraulic system noise range.According to the hydraulic system Simulation model, can obtain the output signal of hydraulic system.Data volume is bigger, is the input signal of hydraulic system like Fig. 5, is the output signal of hydraulic system like Fig. 6, because hydraulic system belongs to followup system, so its output and input signal are more or less the same.Fig. 5 representes the amplitude size that sampled point and sampled point are corresponding respectively with horizontal stroke, y coordinate among Fig. 6.

Train normal fault visualizer (Neural Network Observer) with the input and output signal of hydraulic system, what the neuron network at this place was used is the RBF network, through circuit training, obtains the estimation output of fault visualizer.It is poor that the normal output of resulting estimation output and hydraulic system is done, and obtains hydraulic system system's residual signals in normal operation, as shown in Figure 7.Horizontal stroke among Fig. 7, y coordinate are represented the amplitude size that sampled point is corresponding with sampled point respectively.

For the electron-amplifier in the hydraulic system model injects degradation filture; And the input and output signal under this state is sent in the fault visualizer that trains; Obtain the estimation output under the degenerate state; Further obtain the residual signals under the electron-amplifier degenerate state in the hydraulic system, as shown in Figure 8.The horizontal stroke of Fig. 8, y coordinate are represented the amplitude size that sampled point is corresponding with sampled point respectively.

Step 2, residual signals is extracted characteristic quantity.

Utilize formula (1) give hydraulic system normal with electron-amplifier degeneration two states under residual signals carry out the temporal signatures extraction.Each has 4000 numbers residual signals, these data is divided into 80 groups at present, every group of 50 data, and 50 data of every group are got an effective value, obtain the effective value under the normal and electron-amplifier degeneration two states of hydraulic system, respectively like Fig. 9 and shown in Figure 10.

The basic parameter of step 3, setting network makes up self organizing maps model.

Set frequency of training and initial health degree earlier, the frequency of training T here is 100, and initial health degree is 0.95, and neuronic number can be set by empirical correlation (2) according to sample size k=80, tries to achieve neuron number d=6.With the training sample of the effective value rms under the normal state as self-organization mapping assessment models, the topological diagram in the training process is shown in figure 11, shows the essential information in the training process, like training step number, used time etc.

Step 4, carry out real-time health evaluating.

Effective value rms under the electron-amplifier performance degradation state in the hydraulic system is input to last one goes on foot in the self-organization mapping assessment models that has trained, shown in figure 12.This curve can representative system along with the degeneration gradually of electron-amplifier, the health degree of hydraulic system is also decreasing, and so just can reach the purpose of real-time assessment.Horizontal y coordinate among Figure 12 is represented the healthy amplitude that the sampled point of health degree is corresponding with sampled point respectively.

Step 5, hydraulic system fails are judged.

The health degree of hydraulic system is reducing gradually, needs minimum health degree B of setting to characterize hydraulic system and when reaches total failure.Here,, set 0.25 and be the value of minimum health degree B according to experience in the past, shown in figure 12.When the health degree value of hydraulic system arrives minimum health degree threshold value, show that hydraulic system can not work on, need keep in repair immediately.

Through above appraisal procedure and result's detailed description, visible health evaluating method of the present invention can realize the real-time assessment of the health status of hydraulic system, assesses effectively, has tangible actual application value.

Claims (6)

1. a hydraulic system health evaluating method that combines based on fault visualizer and self-organization mapping is characterized in that, specifically comprises the steps:

Step 1, foundation and training fault visualizer; Described fault visualizer adopts neuron network to realize, is input as the input and output signal of hydraulic system, imports after the sluggish link of output signal through a true hydraulic system lag output function of simulation; Obtain the historical input and output signal under the hydraulic system normal working; Training fault visualizer, training input sample is the historical input and output signal that obtains, the training output sample is the history output signal that obtains; After the training of fault visualizer is accomplished, obtain hydraulic system residual signals in normal operation; Described residual signals is done difference by the estimated output signal of fault visualizer with the output signal of corresponding hydraulic system and is obtained;

Then, at set intervals, obtain the input signal r (t) and the output signal y of the hydraulic system under the current time t working state _r(t), the input fault visualizer obtains the estimated output signal of fault visualizer

Further obtain the residual signals ε (t) of hydraulic system under the current t moment working state; Described residual signals ε (t) is by the estimated output signal of fault visualizer Output signal y with corresponding hydraulic system _r(t) doing difference obtains;

Step 2, to each residual signals, be divided into the k group, in each group, extract n sample, to each group extraction characteristic quantity, obtain k characteristic quantity; Described characteristic quantity adopts average, effective value, kurtosis or shape factor to represent;

Step 3, structure self organizing maps model; The neural network structure that model is an input layer+output layer is insinuated in described self-organization; Before training, set frequency of training, initial health degree and the total neuronic number of whole neuron network; Use the input vector of k the characteristic quantity composition input layer that the residual signals under the hydraulic system normal working extracts then, carry out the training of self organizing maps model;

Step 4, hydraulic system is carried out real-time health evaluating: the self organizing maps model that will train in the characteristic quantity input of the residual signals of the hydraulic system under the moment t working state obtains the health degree A of the hydraulic system under the t working state constantly;

The minimum health degree B of the hydraulic system of step 5, the health degree A that incites somebody to action the hydraulic system under the moment t working state and setting compares, if A＞B, then hydraulic system operational excellence; If A≤B; Then there is fault in hydraulic system, and hydraulic system out of service is carried out fault and got rid of.

2. a kind of hydraulic system health evaluating method that combines with the self-organization mapping based on the fault visualizer according to claim 1; It is characterized in that the fault visualizer described in the step 1 adopts the RBF neuron network to set up, before the fault visualizer is trained; Normalization is handled training input sample and is trained output sample to [1; 1] between, after training is accomplished, obtains the estimated output signal of fault visualizer output; It is poor that the estimated output signal and the output signal of corresponding hydraulic system are done, and obtains the residual signals of the hydraulic system under the normal working; When certain parts in current hydraulic system take place to degenerate; Obtain the input and output signal of current hydraulic system, this input and output signal is placed in the vector, and this vector is done normalization handle; Be transported in the fault visualizer that has trained; The fault visualizer is exported the estimated output signal of current hydraulic system, with current estimated output signal that obtains and the output signal subtraction that obtains, obtains the residual signals of hydraulic system under certain component degradation state.

3. according to claim 1 a kind of based on fault visualizer and the hydraulic system health evaluating method that the self-organization mapping combines, it is characterized in that the eigenvalue described in the step 2 preferably adopts the effective value rms of residual signals to represent:

$\mathrm{rms}=\sqrt{\frac{x_1^2+x_2^2+...+x_p^2}{p}}$

Wherein, x ₁, x ₂X _pRepresent respectively to be extracted in this residual signals group data the 1st, the 2nd ... The value of p sample.

4. a kind of hydraulic system health evaluating method that combines based on fault visualizer and self-organization mapping according to claim 1 is characterized in that the frequency of training described in the step 3 is set to 100, and initial health degree is set to 0.95, neuronic number $d=\left[\sqrt{5*\sqrt{k}}\right].$

5. according to claim 1 a kind of based on fault visualizer and the hydraulic system health evaluating method that the self-organization mapping combines, it is characterized in that the concrete grammar of the described structure self organizing maps model of step 3 is:

Step 1, initialization: the connection power W that self-organization is insinuated the network of model _IjTax is with [0,1] interval random value, W _IjBe connected power between expression input layer i and the output layer neuron j, i=1,2 ...; N, j=1,2; ..., m, n are the neuronic number on the input layer; M is the neuronic number on the output layer, and the total study number of times of radius R (0), network that initial neighborhood is set is T and learning rate η (0), 0＜η (0)＜1; A counter l is set writes down current study number of times, initial L=0; To import to X _kNormalization obtains

$\stackrel{\‾}{X_k}=\frac{X_k}{\left|\right|X_k\left|\right|}$

Wherein, input vector X _k=(X _1k, X _2k..., X _Ik..., X _Nk) ^Tr, be the vector of the capable k row of n, k is the number of the characteristic quantity that residual signals extracted, X _IkBeing the input vector of input layer i, is the row vector with k data, || || mould is asked in expression;

With being connected weight vector W between output layer neuron j and the input layer _jNormalization obtains

$\stackrel{\‾}{W_j}=\frac{W_j}{\left|\right|W_j\left|\right|}=\frac{(W_{1j},W_{2j},...,W_{\mathrm{ij}},...,W_{\mathrm{nj}})}{\sqrt{{\left(W_{1j}\right)}^2+...+{\left(W_{\mathrm{ij}}\right)}^2+...+{\left(W_{\mathrm{nj}}\right)}^2}}$

Step 2;

as input vector, offers the input layer of network with normalized input vector;

Step 3 is sought the triumph neuron: calculate input vector

Be connected weight vector

Between Euclidean distance d _Jk, and the neuron c of the output layer of chosen distance minimum is the triumph neuron:

Wherein, W _cThe connection weight vector of the c of output layer epineural unit after the expression normalization;

Step 4, parameter adjustment: keep the value of the neuronic connection power outside the triumph neuron c triumph field constant, the value of all neuronic connection power in triumph neuron c and the triumph field thereof is done following adjustment:

W _ij(l+1)＝W _ij(l)+η(l)h _c，j(l)(X _i-W _ij(l))

Wherein, the learning rate of the l time study of η (l) expression, 0＜η (l)＜1, h _{C, j}(l) be the field function of the l time study;

Step 5 is upgraded learning rate and field function: field function h _{C, j}(l) upgrade through following formula:

$h_{c,j}\left(l\right)=\exp (-\frac{d_{\mathrm{cj}}^2}{2R^2\left(l\right)})$

Wherein, d _CjEuclidean distance between the neuron j in expression triumph neuron c and the triumph field on arbitrary output layer that is activated, R (l) is the field radius of the l time study, the field radius upgrades through following formula:

$R(l+1)=\mathrm{INT}((R\left(l\right)-1)\×(1-\frac{l}{T}))+1$ (INT is a bracket function)

Learning rate upgrades according to following formula:

$\mathrm{\η}(l+1)=\mathrm{\η}\left(l\right)-\frac{\mathrm{\η}\left(0\right)}{T}$

Step 6, circulation study: if input vector more than two, then need be judged current whether all input vector global learnings being finished; If not; Then next input vector is input to input layer, returns step 3 then and carry out, finish up to all input vector global learnings; If all input vector study finishes, then order study number of times l=l+1 returns step 2 and carries out, until till the l=T or till the network convergence.

6. according to claim 1 a kind of based on fault visualizer and the hydraulic system health evaluating method that the self-organization mapping combines, it is characterized in that the span of the minimum health degree B of the hydraulic system in the step 5 is 0.2～0.3.

Images