CN106407998A - Probability time-varying seawater hydraulic pump fault prediction method - Google Patents

Abstract

The invention discloses a probability time-varying seawater hydraulic pump fault prediction method, which belongs to the field of deep sea seawater plunger pump fault prediction research. The probability time-varying seawater hydraulic pump fault prediction method comprises the steps of: selecting working volume efficiency of a deep sea seawater plunger pump as a fault prediction characteristic quantity, and acquiring seawater numerical values sent out by the plunger pump to form a plunger pump fault prediction original data sequence; carrying out Laplace wavelet extraction and then Hilbert envelope demodulation for processing the original data sequence, so as to obtain a new data sequence; conducting reconstruction and recombination based on multi-scale support vector machine data processing, so as to obtain accurate initial data; introducing the processed data into a service life prediction method based on time-varying state transition, and training the data to obtain whole life historical data and health state models of a system; subjecting current monitoring data to multi-scale support vector machine data processing, comparing the processed data with the health state models, recognizing the current health state of the system and calculating a state transition coefficient; and finally calculating remaining duration and remaining life of the system in the current state.

Description

A kind of seawater hydraulic pump failure prediction method of probability time-varying

Technical field

The invention belongs to deep sea water plunger pump trouble predictive study field and in particular to Laplace wavelet filtering extract, Multiple dimensioned support vector machine, the life-span prediction method based on time-varying state transfer.

Background technology

The raising deep sea equipment task rate of attendance and equipment availability being required with modern marine development of resources, surveys in addition Examination technology, signal analysis technology and computer technology develop rapidly, and abroad advanced deep sea equipment all employs complete event at present Barrier prognostics and health management system is to realize status monitoring, fault diagnosis and biometry, thus effectively reducing deep sea equipment thing Therefore rate, save maintenance cost.Wherein failure predication technology is the focus in prognostic and health management system and difficulties.

Deep sea water plunger displacement pump is the power source of deep-sea underwater performance equipment, and deep sea water plunger displacement pump is directly inhaled from sea water Water, high pressure sea water directly flows back to marine environment after doing work in system again, can not only realize the automatic compensation of seawater pressure also Will not pollution of marine environment, be particluarly suitable under big depth marine environment application, be the whole deep-sea underwater performance equipment " heart Dirty ", there is vital effect.For deep sea water hydraulic system, high reliability, long-life deep sea water plunger Pump manufacture and design one of most important precondition, how to realize predicting and realizing of accurate fault diagnosis and performance degradation The abundant use in life-span is the huge challenge facing, and the failure predication technical research of therefore deep sea water plunger displacement pump has weight Big theory and using value.

Zhao et al. is directed to the fault that the single plunger ball of hydraulic planger pump is increased with piston shoes gap, has invented one kind and has been based on The state identification method of discontinuous chaology.The method adopts the outlet pressure signal of chaotic oscillator theory analysis plunger displacement pump, And then realize the identification of plunger displacement pump health status.Using chaos method and wavelet method each leisure condition monitoring and fault diagnosis field In advantage, carry out system state machine monitoring and fault diagnosis.By WAVELET PACKET DECOMPOSITION is carried out to experiment gained vibration signal With reconstruct, obtain reflecting the characteristic signal of target component, by means of chaos method, after-treatment has been carried out to signal, so right The variation tendency of gained characteristic quantity is compared research it is determined that respectively monitoring the running status at position, and achieves fault location. But it is collected signal and is unable to real-time update, replace old data, update new data.And environment is complex in sea water, health State probability may change suddenly, and this method does not consider this emergency situations.

Gray system theory by Central China University of Science and Technology Deng Julong teach in nineteen eighty-two found, with " partial information oneself know, part Information is unknown " " little sample trees ", " lean information " uncertain system be object of study, " partial information oneself know, partial information is not Know " system be referred to as gray system.It is with gray model, gray system to be carried out with quantitation in advance based on the Predicting Technique of gray theory Survey.The differential equation pattern type of gray theory is referred to as GM model. and theory and practice proves, gray prediction can process lean information system System. it only requires that less initial data can model, and sample requirement is little；Grey forecasting model expression formula and calculating letter simultaneously Single, algorithm rapidity is preferable.

Have scholar at present gray prediction theory is applied in plunger displacement pump biometry, achieve preferable effect, but It is to still suffer from defect：Select outlet of plunger pump flow as biometry characteristic quantity, and the change of rate of discharge not only with plunger Pump inter deterioration is relevant, also closely related with the working condition of plunger displacement pump it is impossible to very accurately characterize plunger displacement pump life-span feelings Condition；Pretreatment is not carried out to improve data light slippery to original predictive data, and numerous studies prove, grey forecasting model Precision is heavily dependent on the smoothness of original data sequence；Employ residual GM GM (1,1) model, this modeling side The new state information constantly producing during plunger displacement pump performance degradation cannot be taken into account by method, therefore can only carry out short-term forecast, And for the deep sea water plunger displacement pump that the life-span is up to thousands of hours, medium-term and long-term biometry is more crucial.

Therefore the present invention has taken into full account that in sea water, environment is unstable, the feelings that deep sea water plunger displacement pump working environment changes at any time Condition, and take into account that workpiece is tired suddenly, does not meet the special circumstances of the health status suddenly change of Fatigue Regularity, make whole therefore Barrier prediction process from initial data begin to accurately, constantly update, then by the rationally rigorous longevity shifted based on time-varying state Life Forecasting Methodology obtains more accurately predicting the outcome.

Content of the invention

The purpose of the present invention is to overcome defect present in current deep sea water plunger pump trouble prediction process, and provides one Plant the deep sea water plunger pump trouble Forecasting Methodology combining time-varying state transfer based on multiple dimensioned support vector machine data reconstruction.

The accurate extraction to initial data, the process to time-varying state transition probability, substantially increase medium-term and long-term residual life The precision of prediction.

The present invention chooses deep sea water plunger pump displacement efficiency as failure predication characteristic quantity, gathers plunger displacement pump true The sea water numerical value sent forms plunger pump trouble prediction original data sequence；Extracted by Laplace small echo and carry out Hilbert again Envelope demodulation processes original data sequence, obtains new data sequence；Carry out further according to multiple dimensioned support vector machine data processing Reconstruct restructuring, obtains more accurate primary data.After processing, data introduces the biometry side based on time-varying state transfer In method, training obtains system life-cycle historical data and each health status model, and the currently monitored data is passed through multiple dimensioned Contrast with each health status model after holding vector machine data processing, identify current system health status and calculate state transfer Coefficient, final computing system is in remaining persistent period of current state and residual life.The Forecasting Methodology that the present invention provides is used for The failure predication of deep sea water plunger displacement pump, brings into operation until end-of-life from plunger displacement pump, circulation executes following failure predications Each step of algorithm, and export biometry value and status praesenss persistent period, this meets deep sea water plunger displacement pump health pipe The technical requirements of reason system.

The idiographic flow of the deep sea water plunger pump trouble Forecasting Methodology that the present invention provides is as follows：

The first step, is extracted the impact shock signal of plunger displacement pump, is converted into input signal by Laplace wavelet filtering.Right Input signal carries out the impact that Hilbert envelope demodulation eliminates other coupled vibrations signals, and finally gives Hilbert envelope spectrum In at each times of frequency side frequency interval relative energy and, as primary data.

Second step, primary data is introduced multiple dimensioned support vector machine data processing, is decomposed by multi-Scale Data and phase Space Reconstruction is theoretical, and time serieses y are resolved into s component：The x that decomposition is obtained₁,x₂,…,x_s, sought using FPE criterion Look for smallest embedding dimension number k₁,k₂,…,k_s, set up forecast model using support vector machine, obtain anticipation function f₁,f₂,…,f_s, then Final predictive value p_y,

3rd step, obtains data by multiple dimensioned support vector machine data processing, under each health status of plunger displacement pump Historical data under historical data and life-cycle state is trained, and obtains model and the life-cycle mould of each health status Type.

4th step, in the model under each health status of the data input after current time is monitored and processes, calculates Probability P in each health status model for the Current observation sequence (O | S), by compare obtain each P (O | S_m)(1≤m≤ N), P (O | S) is compared with three health status models, close put under P (O | S_m) in (1≤m≤N)；M is health status quantity； N span is the quantity of health status, completes the fitness evaluation to each health status model, chooses P (O | S) value maximum That model, determine system current state be the corresponding system health status of this model, the state recognition of completion system.Shape State identification is divided into steady catagen phase, uniform catagen phase, accelerates catagen phase.

(1) steadily the state probability of catagen phase describes

In steady catagen phase, state transition probability is fixing over time, back derives along the time, obtains The state transition probability of current time and the relational expression between state transition probability when having just enter into this health status.

(2) uniformly the state probability of catagen phase describes

In uniform catagen phase, state transition probability is linearly increasing over time, back derives along the time, Relational expression between state transition probability when obtaining the state transition probability of current time and having just enter into this health status.

(3) accelerate the state probability description of catagen phase

Accelerating catagen phase, state transition probability is by exponential form change over time, past along the time Push back and lead, the relation table between state transition probability when obtaining the state transition probability of current time and having just enter into this health status Reach formula.

Original state transition probability matrix A₀Obtained by training historical data.

Combined with matrix using the relational expression between state transition probability during health status, and then obtain accelerating Catagen phase experiences the state-transition matrix after moment t=k Δ t.After calculating the value of the transfer ratio that does well using EM algorithm, Just calculate the state transition probability of three kinds of catagen phases, rest on the probability of current state by comparison systemBe transferred to Other shape probability of statesSize, whenWhen it is believed that system performance degradation state occur turn Move, other states q are transferred to by current state p, thus calculating time-varying state transfer square from different state transfer ratio θ Battle array.

State-transition matrix is obtained by the state transfer ratio introducing for system different degraded stage, this state shifts Matrix changes over.

5th step, obtains average and the variance of state duration using improved Forward-backward algorithm, and calculation interval Persistent period and residual life；

System, during the ultimate failure that comes into operation, can experience multiple health status, and its remaining life is equal to System rests on time and the persistent period sum in each state follow-up of current state.Using the training of life-cycle historical data Obtain model, obtain the probability distribution in each state duration for the system.Generally, it is believed that system is in each health status Persistent period Gaussian distributed, so needing to obtain average and the variance of persistent period Gauss distribution, before improved Obtain average and the variance of state duration to-backward algorithm.

Average according to the state duration tried to achieve and variance can obtain the persistent period in each state i for the system. State duration is as the change of state-transition matrix and changes, i.e. time-varying state-transition matrix, and is supervised according to online Survey data and constantly update state transition probability, the change system with state transition probability also could in the persistent period of current state Change, provide more remaining failure predication value RUL^t.

Compared with prior art, the present invention has the advantages that.

1st, choose the volumetric efficiency of deep sea water plunger displacement pump as failure predication characteristic quantity, this parameter can fully react plunger The current operation conditions of pump, and conveniently carry out failure predication, more rationally have for the prediction of deep sea water plunger pump trouble Effect；

2nd, use Laplace small echo to extract impact shock signal, and then reconciled by Hilbert envelope, make primary data More accurate.

3rd, to Real-time Monitoring Data, many chis are all carried out from historical data training storehouse to the operational data of deep sea water plunger displacement pump Degree support vector machine data processing, and carry out failure predication by based on time-varying state transfer life-span prediction method, it is greatly improved Life prediction precision.

Brief description

Fig. 1 is deep sea water plunger pump trouble Forecasting Methodology schematic flow sheet；

Fig. 2 extracts impact shock signal flow graph for Laplace wavelet filtering；

Fig. 3 is that the vibration signal extracting is carried out with the impact stream that Hilbert envelope demodulation eliminates other coupled vibrations signals Cheng Tu.

Specific embodiment

Below the present invention is combined based on multiple dimensioned support vector machine data reconstruction with the deep sea water of time-varying state transfer Plunger pump trouble Forecasting Methodology is described in detail.

As shown in figure 1, the flow process of the present invention is as follows：

The first step, the volumetric efficiency choosing deep sea water plunger displacement pump, as failure predication characteristic quantity, gathers plunger displacement pump true The sea water numerical value sent, forms plunger pump trouble prediction original data sequence.

The life-span of plunger displacement pump refers to that plunger displacement pump inner body is damaged or worn out making plunger displacement pump lose what serviceability was experienced Time, develop with sea water plunger pump high pressure, high speed direction, the life-span of plunger displacement pump depends on the abrasion of internal key friction pair Life-span, and there is no effectively practical means accurately to measure wear extent at present.When the friction pair abrasion in deep sea water plunger displacement pump adds Weight, can cause plunger displacement pump internal leakage to increase, return flow dramatically increases, volumetric efficiency declines.When volumetric efficiency exceedes setting During value, that is, think plunger displacement pump end-of-life.So the volumetric efficiency change of plunger displacement pump is that the abrasion of these main friction pairs is common making Result, it can fully react the current life status of plunger displacement pump, is rationally effective as failure predication characteristic quantity.

Because marine environment is complicated, operating mode complicated, ocean flow pulsation is big, to be carried using Laplace wavelet filteration method Take the sea water numerical value that plunger displacement pump is truly sent.

Plunger pump impulse is extracted by Laplace wavelet filtering and hits vibration signal, Hilbert is carried out to the vibration signal extracting Envelope demodulation eliminates the impact of other coupling traffic signals, and finally gives side frequency area at each times of frequency in Hilbert envelope spectrum Between relative energy and as primary data, the sea water numerical value truly sent as deep sea water plunger displacement pump, acquisition interval is t_sLittle When, that is, every t_sThe sea water numerical value that hour plunger displacement pump of collection is truly sent, and it is designated as x⁽⁰⁾(i) L/min (i=1,2,3 ... N), represent the numerical value of i ＆ lt collection.Acquisition interval t_sSelection should not be too little or too big, interval is too little, gathers excessively intensive, Redundant data occurs, interval too conference causes important information during the sea water change in value truly sent to lose, be unfavorable for The accurate foundation of forecast model afterwards.Acquisition interval t_sSelection principle be the sea water truly sent in this time interval inner plunger pump Numerical value has obvious change.

The initial data collecting is charged to ordered series of numbers x⁽⁰⁾={ x⁽⁰⁾(1),x⁽⁰⁾(2),x⁽⁰⁾(3),x⁽⁰⁾(4),…,x⁽⁰⁾ (m) }, this ordered series of numbers is referred to as plunger pump trouble prediction original data sequence abbreviation original data sequence.

Second step, the original data sequence counting in the first step is carried out multiple dimensioned support vector machine data processing.

Original data sequence is as follows,

x⁽⁰⁾={ x⁽⁰⁾(1),x⁽⁰⁾(2),x⁽⁰⁾(3),x⁽⁰⁾(4),…,x⁽⁰⁾(m)}

Decomposed by multi-Scale Data and Phase-space Reconstruction, time serieses y are resolved into s component：For decomposition The x obtaining₁,x₂,…,x_s, smallest embedding dimension number k is found using FPE criterion₁,k₂,…,k_s, set up prediction using support vector machine Model, obtains anticipation function f₁,f₂,…,f_s, then final predictive value p_yFor：Carry out non-thread from RBF kernel function Property mapping：K(x,x_i)=exp-γ | x-x_i|²}.

γ be the weight coefficient of function in order to find the γ of optimum, using grid optimizing, that is,：γ is made to take within the specific limits Centrifugal pump, takes the parameter making the most last model of final training set precision of prediction highest γ.Obtaining support vector machine After excellent parameter, that is, it is predicted.

3rd step, obtains data by multiple dimensioned support vector machine data processing, under each health status of plunger displacement pump Historical data under historical data and life-cycle state is trained, and obtains model and the life-cycle mould of each health status Type.

4th step, equally carries out multiple dimensioned support vector machine data processing to the currently monitored data, inputs each afterwards and is good for In model under health state, calculate probability P in each health status model for the Current observation sequence (O | S), by comparing Arrive each P (O | S_i) (1≤i≤N) fitness evaluation of completing to each health status model, choose P (O | S) value maximum Model, and determine that system current state is the corresponding system health status of this model, the state recognition of completion system.

Have in the present invention three health status P (O | S₁)、P(O|S₂)、P(O|S₃), three health status are corresponding steady respectively Degenerate state, uniform degenerate state and acceleration degenerate state.

(1) steadily the state probability of catagen phase describes

In steady catagen phase, state transition probability is fixing, that is, over time

a_ii(t)-a_ii(t+ Δ t)=θ₁

Wherein θ₁For constant and θ₁>=0, Δ t are the fixed interval between the observation moment twice.Because,Institute With variable θ₁Need to distribute to a_ij(t+ Δ t), according to it is assumed that then next observation when etching system state transition probability be：

Back derive along the time, state transition probability and shape when having just enter into this health status of current time can be obtained Relational expression between state transition probability.

(2) uniformly the state probability of catagen phase describes

In uniform catagen phase, state transition probability is linearly increasing, that is, over time

Wherein θ₂For constant and θ₂≥0.According to it is assumed that next observation when etching system state transition probability be：

Back derive along the time

Relation between state transition probability when obtaining the state transition probability of current time and having just enter into this health status Expression formula.

(3) accelerate the state probability description of catagen phase

Accelerating catagen phase, state transition probability is to change by exponential form, that is, over time

Wherein θ₃For constant and θ₃≥0.According to it is assumed that then next observation when etching system state transition probability be：

Back derive along the time, obtain the state transition probability of current time and state when having just enter into this health status turns Move the relational expression between probability.

Original state transition probability matrix A₀Obtained by training historical data.Under practical situation, in system operation process If do not keeped in repair to it, its performance is gradually degenerated in time, only can proceed to worse health status, therefore, when 1≤ During i ＜ j≤N, a_ij=0, original state shift-matrix A₀It is described as：

By each health status expression formula respectively with matrix A₀In conjunction with obtain accelerate catagen phase experience moment t=k Δ t State-transition matrix afterwards.It is possible to calculate three kinds of catagen phases calculate the value of the transfer ratio that does well using EM algorithm after State transition probability, rest on the probability of current state by comparison systemBe transferred to other shape probability of statesSize, whenWhen it is believed that system performance degradation state shifts, shifted by current state i To other states j, thus calculating time-varying state transfer matrix from different state transfer ratio θ.

5th step, can obtain average and the variance of state duration, and calculate using improved Forward-backward algorithm Window duration and residual life；

System, during the ultimate failure that comes into operation, can experience multiple health status, its remaining life etc. Rest on time and the persistent period sum in each state follow-up of current state in system.Using life-cycle historical data instruction The model getting, obtains the probability distribution in each state duration for the system.Generally, it is believed that system is in each health The persistent period Gaussian distributed of state, so need to obtain average and the variance of persistent period Gauss distribution, using improvement Forward-backward algorithm obtain average and the variance of state duration.

Average according to the state duration tried to achieve and variance can obtain the persistent period in each state i for the system. State duration is as the change of state-transition matrix and changes, i.e. time-varying state-transition matrix, and is supervised according to online Survey data and constantly update state transition probability, the change system with state transition probability also could in the persistent period of current state Change, provide more remaining failure predication value：

Wherein,Represent the remaining life after system operation t, D (j) expression system holding in j state The continuous time,Residence time under state i after expression system operation t, it is affected by time-varying state transition probability, It is the numerical value of a dynamic change, its computing formula is：

During whole Technical Design, the first step have chosen the true sea water numerical value discharged of deep sea water plunger displacement pump and makees For failure predication characteristic quantity, and gathering the true sea water numerical value discharged during plunger displacement pump performance degradation, to form plunger pump trouble pre- Survey original data sequence；Extracted by Laplace small echo and carry out Hilbert envelope demodulation process original data sequence again, obtain New data sequence is as primary data；Multiple dimensioned support vector machine data processing is used by primary data in second stepMulti-resolution decomposition to Space Reconstruction；Base is employed in 3rd step Life-span prediction method in time-varying state transfer is trained to primary data；

4th step is to Real-time Monitoring Data sequence x⁽⁰⁾={ x⁽⁰⁾(1),x⁽⁰⁾(2),x⁽⁰⁾(3),x⁽⁰⁾(4),…,x⁽⁰⁾(m)} Equally carry out multiple dimensioned support vector machine data processing, then input in the model under each health status, obtained by comparing Each P (O | S_i) (1≤i≤N) fitness evaluation of completing to each health status model, choose P (O | S) the maximum mould of value Type, determines that the current state of system is the corresponding system health status of this model, the state recognition of completion system.

5th step can obtain average and the variance of state duration by improved Forward-backward algorithm, and calculates Window duration and residual life.After each step above, design terminates.

Claims (2)

1. a kind of seawater hydraulic pump failure prediction method of probability time-varying it is characterised in that：The idiographic flow of this method is as follows：

The first step, is extracted the impact shock signal of plunger displacement pump, is converted into input signal by Laplace wavelet filtering；To input Signal carries out the impact that Hilbert envelope demodulation eliminates other coupled vibrations signals, and finally gives each in Hilbert envelope spectrum At times frequency side frequency interval relative energy and, as primary data；

Second step, primary data is introduced multiple dimensioned support vector machine data processing, is decomposed by multi-Scale Data and phase space Re-construction theory, time serieses y are resolved into s component：The x that decomposition is obtained₁,x₂,…,x_s, found using FPE criterion Good Embedded dimensions k₁,k₂,…,k_s, set up forecast model using support vector machine, obtain anticipation function f₁,f₂,…,f_s, then finally Predictive value p_y,

3rd step, obtains data by multiple dimensioned support vector machine data processing, to the history under each health status of plunger displacement pump Historical data under data and life-cycle state is trained, and obtains model and the life-cycle model of each health status；

4th step, in the model under each health status of the data input after current time is monitored and processes, calculates current Probability P in each health status model for the observation sequence (O | S), by compare obtain each P (O | S_m) (1≤m≤N), P (O | S) is compared with three health status models, close put under P (O | S_m) in (1≤m≤N)；M is health status quantity；N takes Value scope is the quantity of health status, completes the fitness evaluation to each health status model, chooses P (O | S) value maximum That model, determines that the current state of system is the corresponding system health status of this model, the state recognition of completion system；State Identification is divided into steady catagen phase, uniform catagen phase, accelerates catagen phase；

(1) steadily the state probability of catagen phase describes

In steady catagen phase, state transition probability is fixing over time, back derives along the time, obtains current The state transition probability in moment and the relational expression between state transition probability when having just enter into this health status；

(2) uniformly the state probability of catagen phase describes

In uniform catagen phase, state transition probability is linearly increasing over time, back derives along the time, obtains The state transition probability of current time and the relational expression between state transition probability when having just enter into this health status；

(3) accelerate the state probability description of catagen phase

Accelerating catagen phase, state transition probability is by exponential form change over time, along the time toward pushing back Lead, the relationship expression between state transition probability when obtaining the state transition probability of current time and having just enter into this health status Formula；

Original state transition probability matrix A₀Obtained by training historical data；

Combined with matrix using the relational expression between state transition probability during health status, and then obtain accelerating to degenerate Stage experiences the state-transition matrix after moment t=k Δ t；After calculating the value of the transfer ratio that does well using EM algorithm, just count Calculate the state transition probability of three kinds of catagen phases, rest on the probability of current state by comparison systemBe transferred to other Shape probability of stateThe size of (1≤p ≠ q≤N), whenWhen it is believed that system performance degradation state shifts, by work as Front state p is transferred to other states q, thus calculating time-varying state transfer matrix from different state transfer ratio θ；

State-transition matrix is obtained by the state transfer ratio introducing for system different degraded stage, this state-transition matrix Change over；

5th step, obtains average and the variance of state duration using improved Forward-backward algorithm, and calculation interval continues Time and residual life；

System, during the ultimate failure that comes into operation, can experience multiple health status, and its remaining life is equal to system Rest on time and the persistent period sum in each state follow-up of current state；Obtained using the training of life-cycle historical data Model, obtains the probability distribution in each state duration for the system；Generally, it is believed that system holding in each health status Continuous time Gaussian distributed, so needing to obtain average and the variance of persistent period Gauss distribution, using improved forward direction-after Obtain average and the variance of state duration to algorithm；

Average according to the state duration tried to achieve and variance can obtain the persistent period in each state i for the system；State Persistent period is as the change of state-transition matrix and changes, i.e. time-varying state-transition matrix, and according to on-line monitoring number According to constantly updating state transition probability, the change system with state transition probability also can occur in the persistent period of current state Change, provides more remaining failure predication value RUL^t.

2. a kind of probability time-varying according to claim 1 seawater hydraulic pump failure prediction method it is characterised in that：

The first step, the volumetric efficiency choosing deep sea water plunger displacement pump, as failure predication characteristic quantity, gathers plunger displacement pump and truly sends Sea water numerical value, formed plunger pump trouble prediction original data sequence；

The life-span of plunger displacement pump refers to that plunger displacement pump inner body is damaged or worn out making plunger displacement pump lose the time that serviceability is experienced, Develop with sea water plunger pump high pressure, high speed direction, the life-span of plunger displacement pump depends on the wear-out life of internal key friction pair, And there is no effectively practical means accurately to measure wear extent at present；When the friction pair increased wear in deep sea water plunger displacement pump, meeting Plunger displacement pump internal leakage is caused to increase, return flow dramatically increases, volumetric efficiency declines；When volumetric efficiency exceedes setting value, Think plunger displacement pump end-of-life；So the volumetric efficiency change of plunger displacement pump is these main friction pairs coefficient knot of abrasion Really, it can fully react the current life status of plunger displacement pump, is rationally effective as failure predication characteristic quantity；

Because marine environment is complicated, operating mode complicated, ocean flow pulsation is big, extraction column is come using Laplace wavelet filteration method The sea water numerical value that plug pump is truly sent；

Plunger pump impulse is extracted by Laplace wavelet filtering and hits vibration signal, Hilbert envelope is carried out to the vibration signal extracting Demodulation eliminates the impact of other coupling traffic signals, and finally gives side frequency interval phase at each times of frequency in Hilbert envelope spectrum To energy with as primary data, the sea water numerical value truly sent as deep sea water plunger displacement pump, acquisition interval is t_sHour, that is, Every t_sThe sea water numerical value that hour plunger displacement pump of collection is truly sent, and it is designated as x⁽⁰⁾I (i=1,2,3 ... n), table for () L/min Show the numerical value of i ＆ lt collection；Acquisition interval t_sSelection should not be too little or too big, interval is too little, gathers excessively intensive, occurs Redundant data, during interval too conference causes the sea water change in value truly sent, important information is lost, pre- after being unfavorable for Survey the accurate foundation of model；Acquisition interval t_sSelection principle be the sea water numerical value truly sent in this time interval inner plunger pump There is obvious change；

The initial data collecting is charged to ordered series of numbers x⁽⁰⁾={ x⁽⁰⁾(1),x⁽⁰⁾(2),x⁽⁰⁾(3),x⁽⁰⁾(4),…,x⁽⁰⁾(m) }, should Ordered series of numbers is referred to as plunger pump trouble prediction original data sequence abbreviation original data sequence；

Second step, the original data sequence counting in the first step is carried out multiple dimensioned support vector machine data processing；

Original data sequence is as follows,

x⁽⁰⁾={ x⁽⁰⁾(1),x⁽⁰⁾(2),x⁽⁰⁾(3),x⁽⁰⁾(4),…,x⁽⁰⁾(m)}

Decomposed by multi-Scale Data and Phase-space Reconstruction, time serieses y are resolved into s component：Decomposition is obtained X₁,x₂,…,x_s, smallest embedding dimension number k is found using FPE criterion₁,k₂,…,k_s, set up prediction mould using support vector machine Type, obtains anticipation function f₁,f₂,…,f_s, then final predictive value p_yFor：Carried out non-linear from RBF kernel function Mapping：K(x,x_i)=exp-γ | x-x_i|²}；

γ be the weight coefficient of function in order to find the γ of optimum, using grid optimizing, that is,：γ is made to take within the specific limits discrete Value, takes the parameter making the most last model of final training set precision of prediction highest γ；Obtaining the optimum ginseng of support vector machine After number, that is, it is predicted；

3rd step, obtains data by multiple dimensioned support vector machine data processing, to the history under each health status of plunger displacement pump Historical data under data and life-cycle state is trained, and obtains model and the life-cycle model of each health status；

4th step, equally carries out multiple dimensioned support vector machine data processing to the currently monitored data, inputs each healthy shape afterwards In model under state, calculate probability P in each health status model for the Current observation sequence (O | S), obtained by comparing Each P (O | S_i) (1≤i≤N) fitness evaluation of completing to each health status model, choose P (O | S) the maximum mould of value Type, and determine that system current state is the corresponding system health status of this model, the state recognition of completion system；

Have in this method three health status P (O | S₁)、P(O|S₂)、P(O|S₃), three health status corresponding steadily degenerations respectively State, uniform degenerate state and acceleration degenerate state；

(1) steadily the state probability of catagen phase describes

In steady catagen phase, state transition probability is fixing, that is, over time

a_ii(t)-a_ii(t+ Δ t)=θ₁

Wherein θ₁For constant and θ₁>=0, Δ t are the fixed interval between the observation moment twice；Because,So becoming Amount θ₁Need to distribute to a_ij(t+ Δ t), according to it is assumed that then next observation when etching system state transition probability be：

$\left\{\begin{array}{c}{a}_{ii}(t+\mathrm{\Δ}t)=a_{ii}\left(t\right)-{\mathrm{\θ}}_1\\ a_{ij}(t+\mathrm{\Δ}t)=a_{ij}\left(t\right)+\frac{{\mathrm{\θ}}_1\·a_{ij}\left(t\right)}{\underset{j=i+1}{\overset{N}{\Σ}}{a}_{ij}\left(t\right)}\end{array}\right.$

Back derive along the time, the state transition probability of current time can be obtained and state when having just enter into this health status turns Move the relational expression between probability；

$\left\{\begin{array}{c}{a}_{ii}\left(t=k\mathrm{\Δ}t\right)=a_{ii}\left(t=0\right)-{\mathrm{k\θ}}_1\\ a_{ij}\left(t=k\mathrm{\Δ}t\right)=a_{ij}\left(t=0\right)+\frac{k\·{\mathrm{\θ}}_1\·a_{ij}\left(t=0\right)}{\underset{j=i+1}{\overset{N}{\mathrm{\Σ}}}{a}_{ij}\left(t=0\right)}\end{array}\right.$

(2) uniformly the state probability of catagen phase describes

In uniform catagen phase, state transition probability is linearly increasing, that is, over time

$\frac{a_{ii}\left(t\right)-a_{ii}(t+\mathrm{\Δ}t)}{a_{ii}\left(t\right)}={\mathrm{\θ}}_2$

Wherein θ₂For constant and θ₂≥0；According to it is assumed that next observation when etching system state transition probability be：

Back derive along the time；

Relationship expression between state transition probability when obtaining the state transition probability of current time and having just enter into this health status Formula；

$\left\{\begin{array}{c}{a}_{ii}\left(t=k\mathrm{\Δ}t\right)={\left(1-{\mathrm{\θ}}_2\right)}^k{a}_{ii}\left(t=0\right)\\ a_{ij}\left(t=k\mathrm{\Δ}t\right)=a_{ij}\left(t=0\right)+\frac{{\mathrm{\θ}}_2\·a_{ii}\left(t=0\right)\·a_{ij}\left(t=0\right)}{\underset{j=i+1}{\overset{N}{\mathrm{\Σ}}}{a}_{ij}\left(t\right)}\underset{j=1}{\overset{N}{\mathrm{\Σ}}}{\left(1-{\mathrm{\θ}}_2\right)}^{k-1}\end{array}\right.$

(3) accelerate the state probability description of catagen phase

Accelerating catagen phase, state transition probability is to change by exponential form, that is, over time

$\frac{a_{ii}(t+\mathrm{\Δ}t)}{a_{ii}\left(t\right)}=a_{ii}^{{\mathrm{\θ}}_3}\left(t\right)$

Wherein θ₃For constant and θ₃≥0；According to it is assumed that then next observation when etching system state transition probability be：

$\left\{\begin{array}{c}{a}_{ii}\left(t+\mathrm{\Δ}t\right)={\[a_{ii}\left(t\right)\]}^{\left(1+{\mathrm{\θ}}_3\right)}\\ a_{ij}\left(t+\mathrm{\Δ}t\right)=a_{ij}\left(t\right)+\frac{\[a_{ii}\left(t\right)-a_{ii}\left(t+\mathrm{\Δ}t\right)\]\·a_{ij}\left(t\right)}{\underset{j=i+1}{\overset{N}{\mathrm{\Σ}}}{a}_{ij}\left(t\right)}\end{array}\right.$

Back derive along the time, the state transition probability obtaining current time is general with state transfer when having just enter into this health status Relational expression between rate；

$\left\{\begin{array}{c}{a}_{ii}(t=k\mathrm{\Δ}t)={\[a_{ii}(t=0)\]}^{{(1+{\mathrm{\θ}}_3)}^k}\\ a_{ij}(t=k\mathrm{\Δ}t)=a_{ij}(t=0)+\frac{\[a_{ij}(t=0)-a_{ii}{(t=0)}^{{(1+{\mathrm{\θ}}_3)}^k}\]\·a_{ij}(t=0)}{\underset{j=i+1}{\overset{N}{\Σ}}{a}_{ij}(t=0)}\end{array}\right.$

Original state transition probability matrix A₀Obtained by training historical data；Under practical situation, if not in system operation process It is keeped in repair, its performance is gradually degenerated in time, only can proceed to worse health status, therefore, when 1≤i ＜ j≤ During N, a_ij=0, original state shift-matrix A₀It is described as：

$A_0=\left[\begin{array}{cccc}{a}_{11}& a_{12}& ...& a_{1N}\\ 0& a_{22}& ...& a_{2N}\\ \begin{array}{c}.\\ .\\ .\end{array}& \begin{array}{c}.\\ .\\ .\end{array}& \begin{array}{c}\\ ...\\ \end{array}& \begin{array}{c}.\\ .\\ .\end{array}\\ 0& 0& ...& a_{NN}\end{array}\right]$

By each health status expression formula respectively with matrix A₀In conjunction with obtain accelerate catagen phase experience moment t=k Δ t after State-transition matrix；It becomes possible to calculate the shape of three kinds of catagen phases calculate the value of the transfer ratio that does well using EM algorithm after State transition probability, rests on the probability of current state by comparison systemBe transferred to other shape probability of states(1≤i≠j ≤ N) size, whenWhen it is believed that system performance degradation state shifts, other states are transferred to by current state i J, thus calculate time-varying state transfer matrix from different state transfer ratio θ；

5th step, can obtain average and the variance of state duration using improved Forward-backward algorithm, and calculation interval Persistent period and residual life；

System, during the ultimate failure that comes into operation, can experience multiple health status, and its remaining life is equal to and is System rests on time and the persistent period sum in each state follow-up of current state；Trained using life-cycle historical data The model arriving, obtains the probability distribution in each state duration for the system；Generally, it is believed that system is in each health status Persistent period Gaussian distributed, so needing to obtain average and the variance of persistent period Gauss distribution, before improved Obtain average and the variance of state duration to-backward algorithm；

Average according to the state duration tried to achieve and variance can obtain the persistent period in each state i for the system；State Persistent period is as the change of state-transition matrix and changes, i.e. time-varying state-transition matrix, and according to on-line monitoring number According to constantly updating state transition probability, the change system with state transition probability also can occur in the persistent period of current state Change, provides more remaining failure predication value：

${\mathrm{RUL}}_i^t=D_{\left(i\right)}^t+\underset{j=i+1}{\overset{N}{\Σ}}D\left(j\right)$

Wherein,Represent the remaining life after system operation t, D (j) represent system j state lasting when Between,Residence time under state i after expression system operation t, it is affected by time-varying state transition probability, is one The numerical value of individual dynamic change, its computing formula is：

$D_{\left(i\right)}^t=D\left(i\right)\[1-(1-a_{ii}^t)/\underset{o}{\overset{t}{\Π}}{a}_{ij}^t\]$

During whole Technical Design, the first step have chosen the true sea water numerical value discharged of deep sea water plunger displacement pump as event Barrier predicted characteristics amount, and gathering the true sea water numerical value discharged during plunger displacement pump performance degradation, to form plunger pump trouble prediction former Beginning data sequence；Extracted by Laplace small echo and carry out Hilbert envelope demodulation process original data sequence again, obtain new Data sequence is as primary data；Multiple dimensioned support vector machine data processing is used by primary data x in second step⁽⁰⁾={ x⁽⁰⁾ (1),x⁽⁰⁾(2),x⁽⁰⁾(3),x⁽⁰⁾(4),..,x⁽⁰⁾(m) } multi-resolution decomposition to Space Reconstruction；Employ in 3rd step and be based on The life-span prediction method of time-varying state transfer is trained to primary data；

4th step is to Real-time Monitoring Data sequence x⁽⁰⁾={ x⁽⁰⁾(1),x⁽⁰⁾(2),x⁽⁰⁾(3),x⁽⁰⁾(4),…,x⁽⁰⁾(m) } same Carry out multiple dimensioned support vector machine data processing, then input in the model under each health status, by comparing obtain each Individual P (O | S_i) (1≤i≤N) fitness evaluation of completing to each health status model, choose P (O | S) the maximum model of value, Determine system current state be the corresponding system health status of this model, the state recognition of completion system；

5th step can obtain average and the variance of state duration by improved Forward-backward algorithm, and calculation interval Persistent period and residual life；After each step above, design terminates.