CN104503420A - Non-linear process industry fault prediction method based on novel FDE-ELM and EFSM - Google Patents

Abstract

The invention provides a non-linear process industry fault prediction and inference method. The method comprises: a data pre-processing process, an extended finite state machine (EFSM) model building process, a variable prediction process based on a feedback differential evaluation optimized extreme learning machine (FDE-ELM), and a fault inference process based on the EFSM. The method is advantaged by high model building stability, high fault prediction precision, low algorithm complexity, automatic fault inferring, and visual inference process. The method provides help for ensuring security of process industry, ensuring property and personnel safety, and saving hardware cost.

Description

A kind of industry of the non-linear process based on novel FDE-ELM and time delay EFSM failure prediction method

Technical field

The present invention with non-linear process industrial system for object, to improve modeling stability and failure prediction precision for target, propose one and comprise complex process industrial data preconditioning technique, feedback differential optimization extreme learning machine (Feedback Differential Evolution Optimized Extreme Learning Machine, FDE-ELM) technology, and the industrial failure prediction method of delay spread finite state machine (Extended Finite State Machine, EFSM) technology.

Background technology

Process industrial comprises the basic industry of petrochemical industry, metallurgy, papermaking Deng Shi China, and it relates to the every aspect of national life.Process industrial possesses that production scale is large, complex process, solve nonlinear, danger coefficient high, thus becomes an important research field of failure prediction and diagnosis.China's relevant industries Frequent Accidents in the last few years, wherein particularly severe in chemical industry situation, 2013 Qingdao petrochemical industry blasts cause that 62 people are dead, 136 people are injured, and this makes the safety guarantee of intensifies process production run become extremely urgent.Cause serious consequence in order to what prevent this type of accident, high-precision failure prediction and diagnosis are that one effectively evades fault method.Therefore, research improves non-linear process industry failure prediction and fault diagnosis precision and model stability, has important theory significance and practical value.

Artificial neural network is the mathematical model that a kind of biomimetic structure is formed, can according to the data provided, and by training and learn the inner link that finds in data, thus builds the mathematical model of destination object.The method is a kind of method that technical data drives, practical application does not rely on priori Sum fanction, and there is very strong None-linear approximation ability, be thus widely used in the fields such as the parameter estimation of complex industrial object, operation optimization, failure prediction.In existing various algorithm, extreme learning machine (Extreme Learning Machine, ELM) be a kind of fast parameter training algorithm of single hidden layer feedforward neural network, it can solve most of traditional algorithm speed of convergence and be absorbed in the problems such as local minimum slowly, easily.But ELM network is in order to reach reasonable effect, usually needs a large amount of hidden layer nodes, which increasing the complexity of algorithm.Meanwhile, the input layer weights due to ELM learning algorithm are random generations, and random weights are not optimum usually, and therefore the precision of e-learning is not high.In order to solve this two problems, difference (differential evolution, DE) optimized algorithm can be introduced to find the input layer weights of network optimum, improve study precision, also can reduce node in hidden layer simultaneously, thus reach the object reducing algorithm complex.But in process industrial, DE-ELM does not consider the time delay of data and the inner link of system variable, and therefore often prediction effect is not good.

EFSM is a kind of microtomy very ripe in computer realm application, and it effectively can simplify Complex System Models and contribute to analysis and the understanding of system model.EFSM is made up of three parts, State-dependence figure, data dependence graph, migration table, just the internal relations between system and State Transferring can be characterized clearly by this three part.Therefore can be introduced into process industrial field, be applied in systematic analysis and fault reasoning.But traditional EFSM directly cannot introduce process industrial, wherein also there are some insurmountable problems: first, traditional E FSM data dependence graph is built and is relied on mathematical model completely, but in process industrial system, mathematical model is often very complicated, cannot Precise Representation; Secondly, often there is time delay in the variable between process industrial system, traditional E FSM does not consider this delay problem, and the analysis of relationship between variables thus can be caused inaccurate.

Summary of the invention

The object of the invention is to: overcome the shortcoming that conventional failure precision of prediction is low, model stability is not high, and for prior art exist the problems referred to above, propose a kind of based on FDE-ELM and time delay EFSM non-linear process industry failure prediction method.Artificial neural network and EFSM are introduced process industrial failure prediction field by the method, build the State-dependence figure based on time delay EFSM and data dependence graph respectively, and based on the variable prediction model of FDE-ELM technology, propose the failure prediction method that a kind of precision of prediction is high, model stability good, algorithm complex is low, for reduction accident rate, saving hardware cost, raising production safety grade provide technical support.

The invention provides a kind of failure prediction for non-linear process industry and inference method, it is characterized in that described method comprises:

Process of data preprocessing: noise reduction process is carried out to industrial data;

Time delay delay spread finite state machine EFSM model construction process: use time delay mutual information TDMI to carry out calculating time delay and correlation analysis to pretreated data, build data dependence graph, and build State-dependence figure and migration table by priori with to the Analysis on Mechanism of model;

The variable prediction process of extreme learning machine FDE-ELM is optimized: build FDE-ELM network based on feedback differential, the key variables of selecting system as the output node of network, and obtain the input node corresponding with output node by the association between variables that described data dependence graph is set up;

Fault reasoning process based on time delay EFSM: when described FDE-ELM variable prediction the output of process predict the outcome the control threshold range exceeding setting time, by described predict the outcome import time delay EFSM carry out fault reasoning; Be specially, carry out reasoning according to pre-set migration table, when meeting transition condition when predicting the outcome, state changes, and when state no longer occurs to change, the state of output is the fault type of generation.

In described process of data preprocessing, Wavelet Denoising Method is adopted to carry out pre-service to data.

Building of described data dependence graph specifically comprises: use TDMI to calculate related coefficient between itself and its dependent variable and time delay for each variable in system, and the most Two Variables corresponding to short delaing time of gained is the Two Variables corresponding to the shortest travel path; Connect described Two Variables and form travel path; Calculate the interferencing propagation direction between described Two Variables, determine final between Two Variables time delay t and correlation coefficient r to go forward side by side rower note, form complete data dependence graph.

Described State-dependence figure and building of migration table specifically comprise: by the analysis of mechanism model and the state of priori certainty annuity, and creation state dependency graph is for characterizing the contact between each state; According to t and r in described data dependence graph between variable, must do well and relation between variable, thus obtain the migration rules that moves between state corresponding in State-dependence figure, and then build migration table.

The described variable prediction process based on FDE-ELM specifically comprises: add the time serial message that feedback layer carrys out Storage Estimation variable, and the relationship between variables after simultaneously being extracted by EFSM introduces FDE-ELM; Using the input of variable adjacent with target prediction variable in described data dependence graph as network.

The described fault reasoning process based on time delay EFSM specifically comprises: when target variable predict the outcome beyond setting control limit scope, then by key variables all in FDE-ELM network predict the outcome import EFSM carry out fault reasoning; Described fault reasoning progressively carries out according to the transition condition in migration table, be specially, under current state, whether search current state meets as the transition condition that original state is corresponding, if meet, migration is activated, state is moved, and exports corresponding Fault Identification result when state no longer occurs to move.

Described reasoning process is realized by manual operation, or is realized by process programming robotization.

The present invention's innovative point is compared with prior art:

(1) invention increases a kind of new neural network Forecasting Methodology---FDE-ELM, the method can under the prerequisite of Stochastic choice input layer weights, utilize difference optimization to find optimum input layer weights to send into network, thus improve training and the Generalization accuracy of network.The introducing of feedback, can the time sequence information of data in the continuous operational process of storage system effectively, improves the problem that the Generalization accuracy that causes due to time delay is not high.The method has many good characteristics such as learning rate is fast, results of learning stable, Generalization accuracy is high, model stability is strong, for process industrial failure prediction provides new approaches.

(2) EFSM applying to computer realm is successfully introduced process industrial failure prediction field by the present invention.Meanwhile, in traditional E FSM, introduce TDMI, obtain Delay and related coefficient between variable and define a kind of new time delay EFSM.By the method, we can be well understood to the internal relations between system variable, choose the variable input prediction network that correlativity is high, thus avoid the impact of a large amount of redundant informations on forecasting process.Experiment shows, the introducing of time delay EFSM can increase substantially the precision of variable prediction.

(3) the present invention uses time delay EFSM, establishes system state dependency graph and migration table, defines a set of new transition condition expression method, efficiently solves the problem of process industrial mathematical model complexity.

(4) time delay EFSM and FDE-ELM organically combines by the present invention, FDE-ELM is predicted the variate-value of gained imports FDE-ELM and applies to fault reasoning process, has given full play to the technical advantage of both, obtained good technique effect.

Accompanying drawing explanation

Fig. 1 is the workflow diagram of the method for the invention;

Fig. 2 is TE process process flow diagram;

Fig. 3 is the time delay EFSM data dependence graph of TE process;

Fig. 4 is the time delay EFSM State-dependence figure of TE process;

Fig. 5 is FDE-ELM network structure;

Fig. 6 is that TE process variable predicts the outcome.

Embodiment

As shown in Figure 1, be the workflow diagram of the method for the invention.(1) process of data preprocessing: this process mainly carries out noise reduction process to industrial data, avoids because noise affects the accuracy of subsequent operation result.(2) time delay EFSM model construction process: this process mainly uses TDMI to carry out calculating time delay and correlation analysis to pretreated data, build data dependence graph, and build State-dependence figure and migration table by priori with to the Analysis on Mechanism of model, thus the process industrial object of complexity is about kept to simple model, clearly represent the internal connection between system variable, the internal connection between state and connecting each other between state and variable.(3) based on the variable prediction process of FDE-ELM: this process is contingent exception in using FDE-ELM neural network forecast process industrial to run.When building FDE-ELM network, the key variables of selecting system (generally using during exception can the variable-definition that occurs of direct causing trouble be key variables) as the output node of network, and obtain the input node corresponding with output node by the association between variables that data dependence graph is set up.Difference optimization and feedback play a role simultaneously in the training process, and stability and the reliability predicted the outcome of Logistics networks, make extensive result have less error.(4) based on the fault reasoning process of time delay EFSM: when FDE-ELM export predict the outcome the control threshold range exceeding setting time, by FDE-ELM export predict the outcome import time delay EFSM carry out fault reasoning.Reasoning is carried out according to pre-set migration table, and when meeting transition condition when predicting the outcome, state changes.Finally, when state no longer occurs to change, the state of output is contingent fault type.

In order to the detailed process of the method is clearly described, we select TE process (Tennessee Eastman Process) as simulation object, and its concrete technology process flow diagram as shown in Figure 2.TE standard procedure is an actual nonlinear industrial processes, is often used to the effect of validation fault prediction and fault diagnosis.TE process is a kind of analogue simulation of actual chemical process, is proposed, be widely used in the research of process control technology by J.J.Downs and E.F.Vogel of process control group of Tennessee Eastman chemical company of the U.S..TE process comprises 12 manipulated variables and 41 measurands, as shown in table 1.Meanwhile, in standard TE process, contain 20 kinds of faults, be divided into random fault and the large class of both phase step fault two.We choose both phase step fault (fault 1, fault 2, fault 3, fault 4, fault 5 and fault 7) wherein as research object, as shown in table 2.

Table 1TE process variable

Table 2TE procedure fault

Data prediction described in step (1), is embodied in:

The data of emulation are containing much noise, and therefore we adopt Wavelet Denoising Method to carry out pre-service to data.

Characteristics of variables in step (2) extracts, and is embodied in:

Be be mutually related between variablees a large amount of in system, we need these variablees to carry out TDMI to calculate the feature extracted between them.To any Two Variables X _i, X _j(i=1 ~ 53, j=1 ~ 53, i ≠ j), its time delay information entropy is defined as:

${\mathrm{TDMI}}_{X_i{X}_j}\left(\mathrm{\τ}\right)=-\underset{i=1}{\overset{n_{x_i}}{\mathrm{\Σ}}}\underset{j=1}{\overset{n_{y_j}}{\mathrm{\Σ}}}{p}_{\mathrm{ij}}(X_i\left(t\right),X_j(t+\mathrm{\τ}))\ln \frac{p_{\mathrm{ij}}(X_i\left(t\right),X_j(t+\mathrm{\τ}))}{p_{\mathrm{ij}}\left(X_i\left(t\right)\right)p_{\mathrm{ij}}\left(X_j(t+\mathrm{\τ})\right)}---\left(1\right)$

${\mathrm{TDMI}}_{X_j{X}_i}\left(\mathrm{\τ}\right)=-\underset{j=1}{\overset{n_{y_j}}{\mathrm{\Σ}}}\underset{i=1}{\overset{n_{x_i}}{\mathrm{\Σ}}}{p}_{\mathrm{ji}}(X_j\left(t\right),X_i(t+\mathrm{\τ}))\ln \frac{p_{\mathrm{ji}}(X_j\left(t\right),X_i(t+\mathrm{\τ}))}{p_{\mathrm{ji}}\left(X_j\left(t\right)\right)p_{\mathrm{ij}}\left(X_i(t+\mathrm{\τ})\right)}---\left(2\right)$

Pij represents joint probability density, X _it () represents the value of i-th variable in t, X _i(t+ τ) represents the value of i-th variable in the t+ τ moment, and wherein τ is time delay, and t is current sample time.(X, Y, τ, t all do not define, defined) represent X _i(τ) and X _jdelay entropy between (t+ τ), represent X _j(τ) and X _idelay entropy between (t+ τ) is for identical one group of variable X _i, X _j, set different τ and calculate different with and first of acquired results maximum value is designated as respectively with then the dependence of X and Y is calculated:

$D_{X_i\→X_j}={\mathrm{TDMI}}_{X_i{X}_j}^{*}\left(\mathrm{\τ}\right)-{\mathrm{TDMI}}_{X_j{X}_i}^{*}\left(\mathrm{\τ}\right)---\left(3\right)$

If be greater than zero, then the direction of information flow is from X _ipoint to X _jif, be less than zero, then the direction of information flow is from X _jpoint to X _i.When when equalling zero, then represent X _iwith X _jseparate.After determining dependence, we just can determine related coefficient between Two Variables and time delay.Choose with larger value is denoted as correlation coefficient r between Two Variables ', the τ of its correspondence be designated as delay time T between Two Variables '.

Data dependence graph in step (2) builds by the following method:

A) variable X is chosen _i(1≤i≤n, n is the sum of variable), calculates this variable X _iwith another variable X _jthe delay time T of (1≤j≤n, j ≠ i) ' _ij, repeat this operation, obtain variable X _iand the time delay in system between remaining variables.

B) variable X is found _iminimum delay time τ ' _imincorresponding variable X _m.X _mcan be described as the correlated variables of variable X i, and connect this Two Variables with straight line.

C) this operation is repeated to variablees all in system, obtain each variable correlated variables, and connect, finally arrange the relational network obtained between a variable.

D) according to characteristics of variables extract in describe dependence computing method, the dependence be connected in calculated relationship network between variable, and according to dependence add on line before upward arrow characterization information stream direction (if then arrow is by X _ipoint to X _mif, then arrow is by X _mpoint to X _i).

E) mark out on the line of band arrow and to be extracted correlation coefficient r between calculated Two Variables ' and delay time T ' by characteristics of variables, just complete the structure (as shown in Figure 3) of data dependence graph.

State-dependence figure in step (2) builds by the following method:

By building State-dependence figure (as shown in Figure 4) to the Analysis on Mechanism of TE system and priori.In figure, Sk represents the state of system, sets, Tk (k=1,2,3 by the analysis of TE process flow process and the needs of failure prediction ...) be migration between two states.When migration that and if only if is activated, state just can shift.

Migration rules table in step (2) builds by the following method:

Migration rules table is set up for the migration between each state of TE, as shown in table 3.In State-dependence figure, there are two migration paths in certain class state (as S3, S5, S7), and these two migration paths have identical original state, but dbjective state is different.There is two state model in this kind of state correspondence in migration table, mates different transition conditions and different dbjective states respectively.In table, Sk (cv1) represents first state model (Sk is state corresponding in Fig. 4) of Sk, and Sk (cv2) represents second state model of Sk.Here state X _iset(i=1 ~ 53) representative system variable X _iinitial set value, X _iLrepresent variable X _icontrol bottom threshold, X _iHrepresent variable X _icontrol upper threshold, t _setthe representative system sampling interval cycle.Symbol ' ∨ ' presentation logic relation "or", concrete meaning is that this migration is activated when any one condition of ' ∨ ' both sides meets.Symbol ' ∧ ' presentation logic relation "AND", concrete meaning is that this migration is just activated when ' ∧ ' both sides condition all meets.Symbol (+) represents that variate-value has exceeded the control threshold value preset and reached the standard grade, and symbol (-) represents that variate-value is lower than the control bottom threshold preset simultaneously.

The migration table of table 3TE process

Step (3), based on the failure prediction of FDE-ELM, is embodied in:

The network structure being FDE-ELM shown in Fig. 5, from figure, we can see by adding feedback layer in output layer and input layer, the network output valve in the upper moment during input of feedback layer, the output of feedback layer is connected to the input layer of master network, so just can carry out iterative computation, thus the time sequence information of storage of variables, raising study precision.

For convenience of description, we represent predictive variable (X with y (t) _i, i=1 ~ 53) and in t-1 moment neural network forecast output valve, the variable of y (t+1) representative prediction is in t neural network forecast output valve.Meanwhile, for the input of FDE-ELM network, x is used _it () represents i-th variable X of t network _ithe sampling actual value of current time.As shown in Figure 5, y (t) represents the network output valve in a moment on predictive variable, y (t) is fed back to the input network of current time simultaneously, and y (t+1) represents the network output valve (i.e. predicted value) of predictive variable current time.As follows corresponding to TE model FDE-ELM specific algorithm:

(1) controling parameters (scale factor F, crossover probability CR, maximum iteration time and population quantity Np) in training set and difference optimized algorithm (DE) is inputted.

(2) initialization of population, setting current iteration number of times j=1.

(3) object vector a is tieed up from solution space stochastic generation d _i(j) and b _ij () forms size is the initial population of N.Setting function evaluates number of times FES=Np.

(4) from i=1toNp, following steps are performed: (for convenience of describing, in this step, parameter a, b are unified substitutes with alphabetical v)

(4.1) make a variation: for target individual, by following formula by random for original individuality according to a certain percentage by variation to a new individuality, thus generate the vectorial u of variation _i(j+1)=[u _i1(j+1), u _i2(j+1) ..., u _id(j+1)].

u _i(j+1)＝v _best(j)+F(v _r1(j)-v _r2(j))(4)

Wherein r1, r2 ∈ 1,2 ... Np} is random, and r1 ≠ r2 ≠ i.Scale factor F ∈ (0,2] be a constant, control Different Variation v by it _r1(j)-v _r2the multiple of (j).V _bestj () is base vector, be classic individuality in current population, can carry out shared best information between population by it.Because mutation operation uses the difference between two individualities of Stochastic choice to calculate, its result of calculation may occur that mutated individual exceeds the situation in setting search territory, and therefore we need to be limited it.As Optimal Parameters u _iq(j+1), when exceeding setting search territory, be defined as follows:

$u_{\mathrm{iq}}(j+1)=\left\{\begin{array}{cc}{u}_{\max },& \mathrm{if}{u}_{\mathrm{iq}}(j+1)>u_{\max }\\ -u_{\max }& \mathrm{if}{u}_{\mathrm{iq}}(j+1)<-u_{\max },q=\mathrm{1,2},...d\\ u_{\mathrm{iq}}(j+1),& \mathrm{otherwise}\end{array}\right.---\left(5\right)$

(4.2) intersect: after mutation operation, application interlace operation increases the diversity of population.Each object vector v _ij trial vector k that () is corresponding _i(j+1)=[k _i1(j+1), k _i2(j+1) ..., k _id(j+1)] can be intersected by binomial (formula 6) generate.

$k_{\mathrm{iq}}(j+1)=\left\{\begin{array}{cc}{u}_{\mathrm{iq}}(j+1),& \mathrm{ifrand}\left(q\right)<\mathrm{CR\; or\; q}=\mathrm{randn}\left(i\right)\\ v_{\mathrm{iq}}\left(j\right),& \mathrm{otherwise}\end{array},q=\mathrm{1,2},...,d\right.---\left(6\right)$

Wherein, rand (q) represents that independent random between [0,1] is uniformly distributed, randn (i) represent variable i be from set 1,2 ... d} Stochastic choice, CR ∈ [0,1] represents crossover probability.

(4.3) select: after crossover process completes, carry out selection operation and decide trial vector k _i(j+1) j+1 can whether be become for the individuality in population.This is an optimization problem, needs to use greedy selection algorithm by trial vector k _i(j+1) the individual v with original object _ij () compares, obtain of future generation individual, detailed process is as follows:

$v_i(j+1)=\left\{\begin{array}{cc}{k}_i(j+1),& \mathrm{if\; f}(k_i(j+1)<v_i\left(j\right))\\ v_i\left(j\right),& \mathrm{otherwise}\end{array}\right.---\left(7\right)$

$f(\&CenterDot;)=\sqrt{\frac{{\mathrm{\&Sigma;}}_{i=1}^N{|Y_i-{T^{\&prime;}}_i|}^2}{\mathrm{Np}}}$ For fitness function, wherein Y=[y ₁, y ₂..., y _n] ^t, $T^{\&prime;}={[t_{11},t_{21},...,t_{N_{1}1}]}^T=\mathrm{H\&beta;}.$

(5) FES=FES+Np is made.

(6) if FES=FES _max, jump to step 8, otherwise make j=j+1, jump to step 4.

(7) according to fitness function f () minimum principle, optimum a and b is obtained.

(8) according to obtaining optimum a and b, hidden layer output valve h is calculated _ns=g{a _s[x _n(t), o (t)]+b _swherein, the activation function that g () is hidden layer neuron.The hidden layer output valve of all training samples about individual neural network m is formed a hidden layer output matrix H:

$H=\left[\begin{array}{ccc}{h}_{11}& ...& h_{1s}\\ .& & .\\ .& ...& .\\ .& & .\\ h_{n1}& ...& h_{\mathrm{ns}}\\ h_{y1}& ...& h_{\mathrm{ys}}\end{array}\right]=\left[\begin{array}{ccc}g(a_1{x}_1+b_1)& ...& g(a_s{x}_1+b_s)\\ .& & .\\ .& ...& .\\ .& & .\\ g(a_1{x}_n+b_1)& ...& g(a_s{x}_n+b_s)\\ g(a_{1}o+b_1)& ...& g(a_{s}o+b_s)\end{array}\right]$

(9) utilize Moore-Penrose generalized inverse to calculate neural network output layer weight vector: β=(H) ⁺y, wherein Y=[y ₁, y ₂..., y _n] ^t, (H) ⁺for the Moore-Penrose generalized inverse of H.

(10) checking sample set is got

x _n=[x _n1, x _n2..., x _nP] ^t∈ R ^p; Y _n=[y _n1] ^t∈ R ¹, N ₁for the sum of sample set, P is the input number of nodes of network.First according to the input layer weight vector a produced _swith hidden layer threshold value b _s, calculate the hidden layer output matrix of neural network

$\tilde{H}=\left[\begin{array}{ccc}g(a_1{x}_{n_1}+b_1)& ...& g(a_s{x}_{n_1}+b_s)\\ .& & .\\ .& ...& .\\ .& & .\\ g(a_1{x}_{n_{N_1}}+b_1)& ...& g(a_s{x}_{n_{N_1}}+b_s)\\ g(a_1{o}_n+b_1)& ...& g(a_s{o}_n+b_s)\end{array}\right]$

Then according to the output valve T of all checking samples of following formulae discovery in neural network.

$T={[t_{11},t_{21},...,t_{N_{1}1}]}^T=\tilde{H}\mathrm{\&beta;}$

(11) the root-mean-square error RMSE of neural network is calculated.Wherein, root-mean-square error computing formula is:

$\mathrm{RMSE}=\frac{1}{N_1}\sqrt{\underset{n=1}{\overset{N_1}{\mathrm{\&Sigma;}}}{(t_{n1}-y_{n1})}^2}$

In order to verify the validity of this method, for TE process method FDE-ELM in conjunction with TD-EFSM, DE-ELM in conjunction with TD-EFSM, single FDE-ELM and single DE-ELM tests respectively, and its extensive error is as shown in table 4.What calculate extensive error here is mean square deviation, and it can not only represent that the height of precision also can be used for the stability of judgment models.

The extensive application condition of the different model of table 4

From table 4, we can find out that FDE-ELM is minimum in conjunction with the method error of TD-EFSM, and precision of prediction has had and increases substantially compared with its excess-three kind method.We it can also be seen that the raising contribution of TD-EFSM to precision of prediction is maximum in addition, because TD-EFSM only chooses the high variable of degree of correlation as input, and the accuracy of the interference prediction of the bulk redundancy information avoided.That feeds back adds the time sequence information considering variable, to certain raising that precision also has, serves the effect of further optimization method.Both actings in conjunction, obtain a kind of precision of prediction the highest, the Forecasting Methodology that model is the most stable.

Step (4), based on the fault reasoning process of time delay EFSM, is embodied in:

Observe predicting the outcome of FDE-ELM, if all variablees are all in normal range, then temporarily non-fault occurs.If there is variable to exceed the control threshold range of setting, then system may have occurred fault.Then, in order to infer fault type further, we introduce the migration rules of TD-EFSM by predicting the outcome.Predict the outcome and to be closely connected by migration rules and process status, so carry out depth-first search by migration rules in TD-EFSM State-dependence figure to identify the source of trouble.Here we are for fault 3, show the whole flow process of Fault Identification and reasoning.

First supposing the system is in normal operating condition (S4) at present, in order to the performance of verification method, adds fault IDV (3) artificially in systems in which.Control threshold value to be determined by the average (μ) of variable each under nominal situation and variance (σ), controlling upper threshold is μ+3 σ, and control bottom threshold is μ-3 σ.In order to the predicted data unusual condition of clearer displaying FDE-ELM, predicted data is normalized.

(1) t=kt is worked as _settime, be activated according to the known migration 6 (T6) of the rule judgment in migration table.Then, system state moves to S3 by S4.Read in the variable prediction result of variable current time, and be normalized as shown in Figure 6.

(2) state 3 times, system, according to the discovery that predicts the outcome, exist abnormal, and the currency of variable X 45 is in normal range between variable X 51 and variable X 52.From transition condition, migration 8 (T8) are activated, and system state moves to S6 by S3.

(3) state 6 times, the unusual condition of further judgment variable X51 and variable X 52, finds that X51 has exceeded control upper threshold, is therefore labeled as X51 (+).Meanwhile, there is not any exception in variable X 52.Therefore, system can judge to show that migration 15 (T15) is activated according to transition condition, and system state moves to S7 by S6.

(4) state 7 times, further reasoning finds that system also exists other exceptional variable.Variable X 21, lower than control bottom threshold, is labeled as S21 (-), and other variable is all in normal range simultaneously.Now, migration 17 (T17) are activated, and system state moves to S12 by S7.

(5) because S12 is end state, therefore reasoning process terminates.

The concrete meaning of state S12 is that fault 3 occurs, and therefore fault 3 has been identified accurately.

This method promotes and achieves good effect in the precision and model stability of failure prediction in summary, and effectively can identify fault type, and the migration path of the state in visualization system fault reasoning identifying, has very strong novelty.

Claims (7)

1., for failure prediction and the inference method of non-linear process industry, it is characterized in that described method comprises:

Process of data preprocessing: noise reduction process is carried out to industrial data;

Time delay delay spread finite state machine EFSM model construction process: use time delay mutual information TDMI to carry out calculating time delay and correlation analysis to pretreated data, build data dependence graph, and build State-dependence figure and migration table by priori with to the Analysis on Mechanism of model;

The variable prediction process of extreme learning machine FDE-ELM is optimized: build FDE-ELM network based on feedback differential, the key variables of selecting system as the output node of network, and obtain the input node corresponding with output node by the association between variables that described data dependence graph is set up;

Fault reasoning process based on time delay EFSM: when described FDE-ELM variable prediction the output of process predict the outcome the control threshold range exceeding setting time, by described predict the outcome import time delay EFSM carry out fault reasoning; Be specially, carry out reasoning according to pre-set migration table, when meeting transition condition when predicting the outcome, state changes, and when state no longer occurs to change, the state of output is the fault type of generation.

2. method according to claim 1, is characterized in that in described process of data preprocessing, adopts Wavelet Denoising Method to carry out pre-service to data.

3. method according to claim 1, it is characterized in that building of described data dependence graph specifically comprises: use TDMI to calculate related coefficient between itself and its dependent variable and time delay for each variable in system, the most Two Variables corresponding to short delaing time of gained is the Two Variables corresponding to the shortest travel path; Connect described Two Variables and form travel path; Calculate the interferencing propagation direction between described Two Variables, determine final between Two Variables time delay t and correlation coefficient r to go forward side by side rower note, form complete data dependence graph.

4. method according to claim 3, it is characterized in that described State-dependence figure and building of migration table specifically comprise: by the analysis of mechanism model and the state of priori certainty annuity, creation state dependency graph is for characterizing the contact between each state; According to t and r in described data dependence graph between variable, must do well and relation between variable, thus obtain the migration rules that moves between state corresponding in State-dependence figure, and then build migration table.

5. method according to claim 1, it is characterized in that the described variable prediction process based on FDE-ELM specifically comprises: add the time serial message that feedback layer carrys out Storage Estimation variable, the relationship between variables after simultaneously being extracted by EFSM introduces FDE-ELM; Using the input of variable adjacent with target prediction variable in described data dependence graph as network.

6. method according to claim 1, it is characterized in that the described fault reasoning process based on time delay EFSM specifically comprises: predict the outcome beyond the control limit scope of setting when target variable, then predicting the outcome of key variables all in FDE-ELM network is imported EFSM and carry out fault reasoning; Described fault reasoning progressively carries out according to the transition condition in migration table, be specially, under current state, whether search current state meets as the transition condition that original state is corresponding, if meet, migration is activated, state is moved, and exports corresponding Fault Identification result when state no longer occurs to move.

7. the fault reasoning process approach based on time delay EFSM according to claim 6, be is characterized in that described reasoning process is realized by manual operation, or is realized by process programming robotization.