CN104091035A - Health monitoring method for effective loads of space station based on data-driven algorithm - Google Patents

Abstract

The invention provides a health monitoring method for effective loads of a space station based on a data-driven algorithm. In the design stage, after historical data of the effective loads are subjected to state vector construction, parameter standardization and weight processing, training samples are obtained; then, clustering learning is performed on the training samples, and different working condition data classifications can be obtained. In the running stage, after real-time downlink test data of the effective loads are processed, through the working conditions obtained through clustering learning, the downlink data are monitored in real time, if abnormal data occur, it shows that new working conditions happen to the loads, a fault may happen or is about to happen probably, finally, the abnormal data are detected in combination with a fault diagnosis tree method, and the position of the fault is determined. Through machine learning of the historical data, a system health knowledge base is formed, the abnormal state of the loads is found through calculation of the distance value of outliers, real-time monitoring on the health state of the loads is achieved, fault detection and positioning of the loads can be supported, and prediction to a certain extent is achieved.

Description

A kind of space station useful load health monitor method based on data-driven algorithm

Technical field

The invention belongs to space station application system useful load fault diagnosis and health control technical field, be specifically related to a kind of space station useful load health monitor method based on data-driven algorithm.

Background technology

Space mission has that political fallout is great, risk is high, investment is large and the feature such as the cycle is long, and therefore, the smooth enforcement that guarantees space mission is an important goal of country.

For guaranteeing the smooth enforcement of space mission, in technology, one of way of conventionally taking is now: adopt station, High Reliability Design mode design space useful load.Yet, due to the complicacy of space environment and the limitation of ground test condition, space station useful load time still there will be fault in operation, therefore, how efficiently and effectively space station useful load is carried out to failure prediction and fault diagnosis has important practical significance.

Summary of the invention

The defect existing for prior art, the invention provides a kind of space station useful load health monitor method based on data-driven algorithm, in order to address the above problem.

The technical solution used in the present invention is as follows:

The invention provides a kind of space station useful load health monitor method based on data-driven algorithm, comprise the following steps:

S1, for the space station useful load by health monitoring, described useful load has H test point, that is: every of described useful load descending test data is the test data that comprises H test parameter;

Set up the historical test database of described useful load; Wherein, described historical test database is for storing the descending test data of some history of described useful load; Wherein, the descending test data of described history is historical descending non-fault test data;

S2, when needs carry out health monitoring to described useful load, reads described historical test database, obtains many descending test datas of history;

Then the descending test data of resulting history is carried out to pre-service, be met the historical test data of n bar of requirement; Wherein, the historical test data of described n bar is the test data of described useful load normal operation process;

S3 based on default selection principle, chooses m test parameter and monitors the factor as key from a described H test parameter, obtains the state vector of m dimension; Wherein, m≤H;

S4, is configured to n*m matrix by the state vector of n m dimension, and n is matrix line number, and m is matrix columns; Wherein, m data in every a line are the data of m the crucial monitoring factor comprising in the historical test data that a S2 obtains; The data of each row are the same crucial monitoring factor in the test data that homogeneous test does not obtain;

S5, is normalized by row n*m matrix, and unified the drawing of the codomain of each crucial monitoring factor arrived to same interval;

S6, determines each crucial weight of monitoring the factor, then the matrix after normalization is weighted to processing, obtains weighting matrix;

S7, using each row vector of described weighting matrix as training sample, has n training sample; A described n training sample is carried out to cluster, obtain a plurality of clusters corresponding with different loads nominal situation difference;

S8, when described useful load is carried out to health monitoring, receives the real-time descending test data that described useful load sends; Then choose the m identical with a S3 test parameter and monitor the factor as key, form state vector;

S9, by the method identical with normalization processing method in S5, is normalized described state vector; Method by identical with adding authority processing method in S6, is weighted processing to the state vector after normalization, obtains weighting state vector;

S10, compare the training sample in each cluster that described weighting state vector and S7 obtain, judge whether to exist the specific training sample identical with described weighting state vector, if existed, draw the conclusion of the normal operation of described useful load, the corresponding load working condition of cluster at described specific training sample place is the current operating mode of described useful load, exports the current operating mode of described useful load, process ends; If there is no, carry out S11;

S11, the distance of each cluster that more described weighting state vector and S7 obtain, obtains apart from the shortest specific cluster of described weighting state vector, and establishing bee-line is D; Then compare D and predeterminable range critical value R, if D≤R, draws the conclusion of the normal operation of described useful load, the corresponding load working condition of described specific cluster is the current operating mode of described useful load, export the current operating mode of described useful load, process ends; If D > is R, show that described weighting state vector does not belong to any known Historic Clustering, described weighting state vector is abnormality vector; Further show that the current operating mode of described useful load does not belong to any normal operating mode, the conclusion that described useful load current time may occur or be about to break down; Finally preserve described abnormality vector, process ends.

Preferably, in S2, the descending test data of described history is carried out to pre-service, is specially:

By deficiency of data, abnormal data and format error data are referred to as bad data; Bad data in the descending test data of described history is rejected, obtain the pretreated historical test data meeting the demands.

Preferably, in S7, a described n training sample is carried out to cluster, obtains a plurality of clusters corresponding with different loads nominal situation difference, be specially:

S7.1, judges that whether classifying rules is known, if classifying rules is completely known, carries out S7.2; If classifying rules is completely unknown, carry out S7.3; If classifying rules part is known, carry out S7.4;

S7.2, described classifying rules is known referring to completely: for the m dimension state vector being comprised of m the crucial monitoring factor, classifying rules is: each crucial monitoring factor is all bound several state grades, and the data interval that each state grade is corresponding is all known; The various combination of the state grade of each crucial monitoring factor forms different load nominal situations;

Clustering method is: whether training of judgement sample belongs to the load working condition that classifying rules is known, if belonged to, this training sample is included into corresponding load working condition, and each training sample that belongs to same load working condition forms a cluster; If do not belonged to, using this training sample as a new cluster, deposit in database; Wherein, the corresponding a kind of new load working condition of new cluster;

S7.3, the complete the unknown of described classifying rules refers to: for the m dimension state vector being comprised of m the crucial monitoring factor, each crucial monitoring factor is several state grades of end binding all;

Clustering method is:

The first step, judges whether existence foundation class, if existed, carries out second step; If there is no, given foundation class is determined cluster centre and the cluster radius of each cluster in described foundation class simultaneously by following formula;

Cluster centre: x _center=(x _max+ x _min)/2; (3)

Cluster radius: d _min=D (x _max, x _min)/2; (4)

Wherein, x _max, x _maxbe respectively the upper limit and the lower limit of training sample in same cluster; D is vector x _maxand vector x _minbetween Euclidean distance;

Second step, constantly expands foundation class, until each cluster obtaining comprises all training samples;

The 3rd step, is optimized all clusters that obtain, and the operating mode using each cluster finally obtaining as useful load deposits database in;

S7.4, known the referring to of described classifying rules part: for the m dimension state vector being comprised of m the crucial monitoring factor, exist M1 the crucial monitoring factor to bind several state grades, the data interval that each state grade is corresponding is all known; Exist M2 the crucial monitoring factor not bind several state grades; Wherein, M1+M2=m; M1 >=1; M2 >=1; M1 and M2 are natural number; : M1 the crucial monitoring factor is the factor that classifying rules is known; M2 the crucial monitoring factor is the factor of classifying rules the unknown;

Clustering method is: for n training sample, the factor that each training sample is comprised is according to its represented physical significance classification; The factor that the classifying rules of take is known is classification foundation, by the clustering method in S7.2, n training sample is carried out to cluster for the first time, obtains several original clusters;

For each original cluster, by the clustering method in S7.3, carry out cluster for the second time, obtain some sub-clusters;

Judge whether resulting each sub-cluster meets operating mode and divide requirement, if met, using each sub-cluster as final cluster result, deposits in database; If do not met, selected M1 the crucial monitoring factor while changing for the first time cluster, re-starts cluster and for the second time cluster for the first time, and circulation said process, until resulting each the sub-cluster of cluster meets operating mode division requirement for the second time.

Preferably, in the 3rd step in S7.3, all clusters that obtain are optimized, are specially:

For foundation class C={C ₁, C ₂..., C _q, suppose that wherein min cluster is C _i(1≤i≤q), in this min cluster, element bound is respectively: x _max=(a ₁, a ₂..., a _p) and x _min=(b ₁, b ₂..., b _m), the Euclidean distance between note bound element is d _min, that is: d _min=D (x _max, x _min); To arbitrary extension class C' _j, remember that the Euclidean distance between its bound element is d' _j;

If cancel this extension class, each state vector being comprised is included into closes on class;

If retain this extension class.

Preferably, in S7.2, using this training sample as a new cluster, deposit in database, be specially: establishing this training sample is x ₀={ a ₁, a ₂... .., a _m;

Calculate x ₀and the distance d between nearest operating mode border, by x ₀as the center of new cluster, the radius apart from d as new cluster, new cluster table is shown:

C _new＝{x,D(x,x ₀)≤d} (2)

Wherein, C _newrepresent new cluster, x represents to belong to the free position vector of new cluster; D represents compute vector x and x ₀between Euclidean distance.

Preferably, after S11, also comprise:

S12, constantly deposits the real-time descending test data of useful load in described historical test database, and described historical test database is constantly updated; Then, the historical test database based on after upgrading, every the fixing cycle, up-to-date training sample set carries out self-teaching, forms up-to-date cluster, is specially:

1) existing class is carried out to new cluster, according to formula (5), redefine its cluster centre;

$c_i=\frac{1}{\left|A_i\right|}{\mathrm{\Σ}}_{x\∈A_i}x,i=\mathrm{1,2},...\left(5\right)$

Wherein, c _irepresent the cluster centre after upgrading for the i time, A _irepresent the cluster after upgrading for the i time, x represents A _ielement, | A _i| represent cluster A _inumber of elements;

2) for the class of new formation, it is redefined to its cluster radius and cluster centre as known operating mode, complete the renewal to new class.

Preferably, after S11, also comprise:

S12, after showing that described weighting state vector is the conclusion of abnormality vector, first determines whether false-alarm, if be defined as false-alarm, and direct process ends; If not, continue to judge whether this abnormality vector exists exception monitoring factor values, if there is no, illustrate that useful load current working is abnormal, the probability that is in fault critical point is higher; If existed, find the corresponding useful load test point of this exception monitoring factor, in conjunction with fault diagnosis tree method, useful load fault is detected and isolated, parts and failure mode information thereof that final output is broken down.

Preferably, in S12, by repeatedly testing and determine whether false-alarm.

Beneficial effect of the present invention is as follows:

A kind of space station useful load health monitor method based on data-driven algorithm provided by the invention, machine learning by historical data forms system health knowledge base, distance value based on outlier calculates the abnormality of finding load, the Real-Time Monitoring of realization to load health status, fault detect and location that can sustained load, and prediction to a certain degree.

Accompanying drawing explanation

Fig. 1 is the overall flow schematic diagram of the space station useful load health monitor method based on data-driven algorithm provided by the invention;

Fig. 2 is the schematic flow sheet of payload data pretreatment stage;

Fig. 3 is under classifying rules known case, training sample clustering method schematic flow sheet;

Fig. 4 is under the complete unknown situation of classifying rules, training sample clustering method process flow diagram;

Fig. 5 is for expanding the schematic flow sheet of foundation class;

Fig. 6 is under classifying rules part known case, training sample clustering method process flow diagram;

Fig. 7 is for carrying out the process flow diagram of Real-Time Monitoring process to real-time descending test data;

Fig. 8 is the schematic flow sheet of fault detect and isolation processes;

Fig. 9 is the measuring point distribution schematic diagram of space station cooling system;

Figure 10 is that the embodiment of the present invention is to historical test data normalized result figure;

Figure 11 is embodiment of the present invention producing condition classification process flow diagram;

Figure 12 is clustering method while treating direction finding amount data monitoring, and the test datas before 1300 show figure;

Figure 13 is clustering method while treating direction finding amount data monitoring, and 1000 later test datas show figure.

Embodiment

Below in conjunction with accompanying drawing, the present invention is described in detail:

The research of domestic space product health monitoring technology is also in the starting stage, and in engineering, the downlink data of useful load was only done simple processing on ground in the past, and the health and fitness information wherein comprising is excavated fully.The present invention is using for reference on the basis of data-driven method, from aspects such as payload data pre-service, Data classification and clusters to original method, improve and integrate, and in conjunction with fault mode and the failure mechanism of useful load, formed the space station useful load health monitor method based on data-driven algorithm, the present invention roughly thinking is: importation comprises effective load histories test data and the real-time descending test data of useful load, respectively as the input of design phase and the input of operation phase.(1), in the design phase, mainly include effect load histories test data pre-service and the load data producing condition classification based on clustering algorithm.The historical test data of useful load is after rough handling, and the parametric configuration state vector that first selection will detect, after state vector generates, carries out standardization and weight processing to the element in vector, removes bad data, obtains training sample.Then training sample is carried out to clustering learning, obtain different floor data classification.(2), in the operation phase, mainly include the real-time descending test data pre-service of effect load, data monitoring and fault detect and isolation.The real-time descending test data of useful load, through after removing bad data, standardization and weight processing, enters the monitoring stage, and normal data, deposits database in if.If there is extremely, output abnormality data, on the other hand, report to the police to possible fault, and it are carried out to fault detect and isolation on the one hand, deposit diagnostic result in database.

As shown in Figure 1, the present invention mainly divides quinquepartite: monitoring, the fault detect based on fault tree algorithm and isolation and the load health knowledge learning process of payload data pre-service, the classification of the floor data based on clustering algorithm, the real-time downlink data of useful load.Below this five part is introduced respectively in detail:

(1) payload data pretreatment stage

Payload data pretreatment stage is divided into the historical descending test data of useful load is carried out to pre-service, and the real-time descending test data of useful load is carried out to pre-service, by pre-service, obtains being applicable to cluster desired data.

As shown in Figure 2, be the schematic flow sheet of payload data pretreatment stage;

(1) for the historical descending test data of useful load, mainly divide into data pre-service, data normalization and parameter weighting and process three parts:

(1.1) data pre-service:

For the space station useful load by health monitoring, useful load has H test point, that is: every of useful load descending test data is the test data that comprises H test parameter; Server is set up the historical test database of useful load; Wherein, historical test database is for storing the descending test data of some history of useful load; For example, H test point is 6 test points, wherein has two test points to be respectively used to test traffic, and other 4 test points are respectively used to measuring current, voltage, temperature, pressure.Every descending test data is the test data that comprises 6 test parameters.

When needs carry out health monitoring to described useful load, read described historical test database, obtain many descending test datas of history; Then the descending test data of resulting history is carried out to pre-service, be met the historical test data of n bar of requirement; Wherein, the historical test data of described n bar is the test data of described useful load normal operation process.

Pre-service in this step is specially: judge that whether historical test data is intact, not intact data definition is bad data, and it is removed; Wherein, bad data mainly comprises deficiency of data, abnormal data and format error data.

Because every descending test data of useful load relates to various test parameters, as electric current, voltage, temperature, flow, pressure etc., still, for avoiding the element in test vector too much to cause calculation of complex; And, in follow-up cluster process, for the known situation of classifying rules, need to be according to the interval judgement operating mode of each test parameter, directly select the test parameter can be more simply credible as analytic target.Therefore, selected part test parameter carries out cluster conventionally.

In the present invention, from a described H test parameter, choose m test parameter and monitor the factor as key, obtain the state vector of m dimension; Wherein, m≤n.

(1.2) data normalization

Because state vector to be tested comprises many kinds of parameters, each test parameter unit is different, so the numerical value codomain of each test parameter is also not quite similar.For example, certain state vector is A={a ₁, a ₂... .., a _m, wherein, a ₁∈ [0,1], a ₂∈ [0,100].Choose two groups of data A of state vector A ₁and A ₂, A ₁=[0.01,1 ...., 20], A ₂=[0.9,10 ...., 20], except the first two element, all the other elements are all consistent.Although element a ₁between Euclidean distance gap (0.89) much smaller than element a ₂(9), but from codomain span, element a ₁variable quantity is: $\frac{0.9-0.01}{1}\×100\%=89\%,$ And element a ₂variable quantity is: $\frac{10-1}{100}\×100\%=9\%.$ So the normalized of vector element is very necessary, the unification of parameters codomain is divided into same interval, for example, in [0,1].

When having the state vector of n m dimension, can construct n*m matrix, n is matrix line number, m is matrix columns; Wherein, m data in every a line are the data of m the crucial monitoring factor comprising in the historical test data that a S2 obtains; The data of each row are the same crucial monitoring factor in the test data that homogeneous test does not obtain; N*m matrix is normalized by row, unified the drawing of the codomain of each crucial monitoring factor arrived to same interval.

(1.3) parameter weighting is processed

Parameter weighting processing refers to according to parameter significance level gives corresponding weight, just makes parameter unit same all parameter normalizeds of vector to be measured, but does not consider that between different parameters, significance level may be different.Equally, suppose state vector A={a ₁, a ₂... .., a _m, suppose a ₁, a ₂with representing flow, that different is a ₁represent main line discharge, and a ₂what represent is a certain branch road discharge.Through after normalized, two groups of test vector A of A _iand A _jbe respectively A _i=[0.3,0.2 ...., 0.2], A _j=[0.7,0.9 ...., 0.2], suppose that all the other elements are all consistent except the first two element.Element a ₁between Euclidean distance poor be 0.4, and a ₂between Euclidean distance poor be 0.7, although a ₂be greater than a ₁but, a ₁main line flow, a ₂variable quantity be a ₁a part for variable quantity, fluctuations in discharge amount, a ₁0.4 of variable quantity will comprise a ₂0.7 of variable quantity.So, to parameter, give corresponding weight, to the accuracy of cluster result, be very necessary.For convenience of calculation, on the other hand, between vectorial element, there is stronger independence on the one hand, can suppose that parameters is identical through weight after normalized.

Concrete, can determine each crucial weight of monitoring the factor, then the matrix after normalization is weighted to processing, obtain weighting matrix.

Vector to be measured is after parameter normalization and weight processing, and data preprocessing phase finishes.Through pretreated data, next step producing condition classification will be entered as training sample.Meanwhile, deposit the state vector after processing in database, in order to reuse later.

(2) for the real-time descending test data of useful load, need to carry out pre-service equally, as shown in Figure 2, real-time descending test data preprocessing process and historical descending test data preprocessing process are basically identical; That is: choose an identical m test parameter and monitor the factor as key, form state vector; Adopt identical normalization processing method, described state vector is normalized; Adopt the identical authority processing method that adds, the state vector after normalization is weighted to processing, obtain weighting state vector.

Real-time descending test data preprocessing process is only from the different of historical descending test data preprocessing process, and after through bad data screening, first real-time descending test data will deposit database in, and parametric configuration desired data reads from database again.

(2) the floor data sorting phase based on clustering algorithm

If after data pre-service, obtain altogether n state vector, this n state vector, as n training sample, is carried out cluster to a described n training sample, obtains a plurality of clusters corresponding with different loads nominal situation difference; As shown in Figure 1, whether known by classifying rules, divide into altogether three kinds of situations:

(1) classifying rules is completely known

Classifying rules is completely known, be that the different floor data threshold values of load are determined, be specially: for the m dimension state vector being formed by m the crucial monitoring factor, classifying rules is: each crucial monitoring factor is all bound several state grades, and the data interval that each state grade is corresponding is all known; The various combination of the state grade of each crucial monitoring factor forms different load nominal situations;

For example,, for state vector A={a ₁, a ₂, a ₃, have 3 crucial monitoring factors, represent respectively water temperature, wind-warm syndrome and discharge, suppose corresponding 3 state grades of water temperature, be respectively: high, medium and low; Corresponding 3 state grades of wind-warm syndrome, are respectively: high, medium and low; Corresponding 3 state grades of discharge, are respectively: high, medium and low; Have 3 ³=27 load nominal situations.For example, if the corresponding state grade tool of water temperature, wind-warm syndrome and discharge given data is interval: for water temperature, known high-grade data interval is 40～50; The data interval of middle grade is 30～40; Low-grade data interval is 20～30; This kind of situation is classifying rules known situation completely.

Under classifying rules known case, as shown in Figure 3, whether training of judgement sample belongs to the load working condition that classifying rules is known to training sample clustering method, if belonged to, this training sample is included into corresponding load working condition, each training sample that belongs to same load working condition forms a cluster; If do not belonged to, using this training sample as a new cluster, deposit in database; Wherein, the corresponding a kind of new load working condition of new cluster;

Wherein, do not belong to the situation of existing any operating mode for training sample, using it as new operating mode, given rational threshold value, deposits database in.Threshold is: suppose certain training sample x ₀={ a ₁, a ₂... .., a _mdo not belong to any known operating mode, calculate x ₀the operating mode frontier distance d nearest with it, by x ₀as the center of new cluster, the radius apart from d as new cluster, new cluster table is shown:

C _new＝{x,D(x,x ₀)≤d} (2)

Wherein, C _newrepresent new cluster, x represents to belong to the free position vector of new cluster.D represents compute vector x and x ₀between Euclidean distance.

Wherein, establish vectorial X={x ₁, x ₂..., x _m, vectorial Y={y ₁, y ₂... .., y _m, between X and Y, Euclidean distance calculates by following formula:

$D(X,Y)=\sqrt{\underset{i=1}{\overset{m}{\mathrm{\Σ}}}{(y_i-x_i)}^2}---\left(1\right)$

(2) classifying rules is completely unknown

The complete the unknown of classifying rules refers to: for the m dimension state vector being comprised of m the crucial monitoring factor, each crucial monitoring factor is several state grades of end binding all;

As shown in Figure 4, under the complete unknown situation of classifying rules, training sample clustering method process flow diagram, comprises the following steps:

The first step, judges whether existence foundation class, if existed, carries out second step; If there is no, provide a foundation class: foundation class is given by means of the linkage () function in Matlab, to part training sample, according to distance size between training sample, carries out preliminary grouping, and hard clustering center and cluster radius.

Cluster centre: x _center=(x _max+ x _min)/2; (3)

Cluster radius: d _min=D (x _max, x _min)/2; (4)

Wherein, x _max, x _maxbe respectively the vectorial upper limit and lower limit in cluster; Wherein, the upper limit of vector refers to: for a certain cluster, and vector corresponding to vectorial mould maximal value; The lower limit of vector refers to: for a certain cluster, and vector corresponding to vectorial mould minimum value.The computing method of vector mould are: for example, for vector (3,3,3), its mould is: $\left|\left(\mathrm{3,3,3}\right)\right|=\sqrt{3^2+3^2+3^2}=3\sqrt{3}$

Second step, constantly expands foundation class, until the cluster obtaining comprises all training samples, the method that expands foundation class is similar to the method for the new class of the known middle interpolation of classifying rules, as shown in Figure 5, and for expanding the schematic flow sheet of foundation class;

For training sample A _i, calculate A _iand the distance between current each cluster, obtains bee-line; Judge whether value of breaking bounds of bee-line, if exceeded, by A _ias new cluster; Then process next training sample; If do not exceeded, by training sample A _ibe included in the cluster of bee-line; Then process next training sample.

The 3rd step, is optimized all clusters that obtain, and the operating mode using each cluster finally obtaining as useful load deposits database in;

Optimization method includes but not limited to following mode:

Class and neighborhood that threshold value is very little merge, and merge rule as follows:

For foundation class C={C ₁, C ₂..., C _q, suppose that wherein min cluster is C _i(1≤i≤q), in this min cluster, element bound is respectively: x _max=(a ₁, a ₂..., a _p) and x _min=(b ₁, b ₂..., b _m), the Euclidean distance between note bound element is d _min, that is: d _min=D (x _max, x _min); To arbitrary extension class C' _j, remember that the Euclidean distance between its bound element is d' _j;

If cancel this extension class, each state vector being comprised is included into closes on class (being generally included into the class that threshold value is less); If retain this extension class.

(3) classifying rules part is known

In most cases, load data classifying rules is simple unknown or unknown, and such as certain cooling system has two temperature sensors, Data classification rule is known.Add on this basis a flow sensor, if there is no historical reference, now vector fractional integration series rule-like belongs to the known situation of part again.

Concrete, classifying rules part is known to be referred to: for the m dimension state vector being comprised of m the crucial monitoring factor, exist M1 the crucial monitoring factor to bind several state grades, the data interval that each state grade is corresponding is all known; Exist M2 the crucial monitoring factor not bind several state grades; Wherein, M1+M2=m; M1 >=1; M2 >=1; M1 and M2 are natural number; : M1 the crucial monitoring factor is the factor that classifying rules is known; M2 the crucial monitoring factor is the factor of classifying rules the unknown;

The present invention adopts the method for double classification.Classification is to classify according to the known parameter of classifying rules in training sample for the first time, and classification is for the second time by the method for cluster, training sample further to be divided on the basis of classification for the first time.

Training sample generally comprises a plurality of parameters, and each parameter unit and represented physical significance are not quite similar.As shown in Figure 6, under classifying rules part known case, training sample clustering method process flow diagram, comprises the following steps:

For n training sample, the factor that each training sample is comprised is according to its represented physical significance classification; The factor that the classifying rules of take is known is classification foundation, by the clustering method in S7.2, n training sample is carried out to cluster for the first time, obtains several original clusters;

For each original cluster, by the clustering method in S7.3, carry out cluster for the second time, obtain several more detailed sub-clusters;

Judge whether resulting each sub-cluster meets operating mode and divide requirement, if met, using each sub-cluster as final cluster result, deposits in database; If do not met, selected M1 the crucial monitoring factor while changing for the first time cluster, re-starts cluster and for the second time cluster for the first time, and circulation said process, until resulting each the sub-cluster of cluster meets operating mode division requirement for the second time.

The sorting technique of secondary classification method and classifying rules the unknown is similar, difference be secondary classification for be through the subclass data after a subseries, be not for total data.

For example, suppose that vectorial A comprises parameter { a ₁, a ₂... .., a _n, all parameters all belong to following four classes { water temperature, wind-warm syndrome, discharge, wind flow }.According to the discharge classification in { water temperature, wind-warm syndrome, discharge, wind flow }, if current divide two-way,

Discharge={ tributary a _p, tributary a _q, 1≤p, q≤n

Suppose that each branch road flow nominal situation is divided three classes: high, medium and low, the represented state of two path water flow has 9 classes, as shown in the table:

Therefore, if according to discharge division rule the division foundation as vectorial A, tentatively vectorial A is divided into 9 class operating modes.On this basis, respectively for each operating mode, in conjunction with clustering method, obtain for the more detailed division of each class.The same above example that continues for sub-operating mode { height, height }, continues to be divided into K sub-operating mode on this operating mode basis, is respectively: and height, height, sub-operating mode 1}, height, and height, sub-operating mode 2} ...., { height, height, sub-operating mode K}.Finally realize the division completely of operating mode.

After producing condition classification Rulemaking, according to regular Preliminary division operating mode classification, then in conjunction with clustering method, obtain finer Data classification.If producing condition classification is undesirable after cluster, as intersection appears in cluster, by reselecting new vector parameter, reformulate classifying rules, until meet the requirements.

(3) monitoring of the real-time downlink data of useful load

Through two steps above, obtained cluster corresponding under useful load nominal situation.The cluster that these have been obtained monitors real-time descending test data as foundation class, if there is the data that do not belong to any cluster, occurs, illustrates that new working condition (if not false-alarm, being exactly new fault mode so) has appearred in load.

Real-time descending test data after processing through pre-service, normalization and weighting is called weighting state vector, and as the input of Real-Time Monitoring process, testing process as shown in Figure 7, comprises the following steps:

(1) compare the training sample in each cluster that described weighting state vector and abovementioned steps obtain, judge whether to exist the specific training sample identical with described weighting state vector, if existed, draw the conclusion of the normal operation of described useful load, the corresponding load working condition of cluster at described specific training sample place is the current operating mode of described useful load, export the current operating mode of described useful load, process ends, completes the observation process to this real-time descending test data; If there is no, carry out (2);

(2) distance of more described weighting state vector and each cluster, obtains apart from the shortest specific cluster of described weighting state vector, and establishing bee-line is D; Then compare D and predeterminable range critical value R, if D≤R, draws the conclusion of the normal operation of described useful load, the corresponding load working condition of described specific cluster is the current operating mode of described useful load, export the current operating mode of described useful load, process ends; If D > is R, show that described weighting state vector does not belong to any known Historic Clustering, described weighting state vector is abnormality vector; Further show that the current operating mode of described useful load does not belong to any normal operating mode, the conclusion that described useful load current time may occur or be about to break down; Finally preserve described abnormality vector, useful load observation process finishes.

(4) fault detect based on fault tree algorithm and isolation

After showing that described weighting state vector is the conclusion of abnormality vector, abnormality vector, as the input of fault detect and isolation processes, as shown in Figure 8, is the schematic flow sheet of fault detect and isolation processes; First determine whether false-alarm, if be defined as false-alarm, direct process ends; If not, continue to judge whether this abnormality vector exists exception monitoring factor values, if there is no, illustrate that useful load current working is abnormal, may be in fault critical point, load probably breaks down within the time soon; If existed, find the corresponding useful load test point of this exception monitoring factor, in conjunction with fault diagnosis tree (Fault Diagnosis Tree) method, useful load fault is detected and isolated, parts and failure mode information thereof that final output is broken down.

(5) load health knowledge learning process

Through step 5, the detection and diagnosis of abnormal data is finished, this step object is to improve monitoring algorithm.

The increase of data volume has produced the impact of two aspects on cluster result:

1) generated new cluster;

2) data volume that original cluster comprises increases.

Therefore,, in order to improve the accuracy of cluster result, need to periodically upgrade cluster result.Be specially:

1) to original poly-, due to the increase of data, its cluster centre unavoidably changes, and the cluster centre new to it upgrades according to formula (5) method.

$c_i=\frac{1}{\left|A_i\right|}{\mathrm{\Σ}}_{x\∈A_i}x,i=\mathrm{1,2},...\left(5\right)$

Wherein, c _irepresent the cluster centre after upgrading for the i time, A _irepresent the cluster after upgrading for the i time, x represents A _ielement, | A _i| represent cluster A _inumber of elements.

2) for the class of new formation, according to the given method of formula (2), complete determining of Dui Xinlei center and cluster radius.

Embodiment

The present embodiment is with reference to the amended case of space station cooling system, and this cooling system comprises three water branch roads, all has corresponding load on every branch road, all has temperature point before and after load, after every tributary flow valve also to there being discharge measuring point.According to previous experiences data, the producing condition classification rule of data on flows determined, the data on flows of each branch road is divided into { height, in, low } three classes according to size; And the classifying rules of temperature data is unknown.

As shown in Figure 9, be the measuring point distribution schematic diagram of space station cooling system, suppose that temperature sensor is T ₁～T ₆, flow sensor is T ₇～T ₉.The space station useful load health monitor method based on data-driven that utilizes the present invention to provide below, by sensor T ₁～T ₉descending test data is carried out Treatment Analysis, completes the monitoring to the health status of this cooling system.Step is as follows:

Step 1: the historical test data pre-service of useful load

According to data processing requirement, historical test data is carried out to bad data rejecting, normalized, suppose that element importance degree is identical.Because load data is for just, normalized is:

$x^{\′}=\frac{x-x_{\min }}{x_{\max }-x_{\min }}$

Wherein, x is former data, and x' is data after normalization, and with branch road 1 outlet water temperature (T4) data instance, normalized result as shown in figure 10 now.

Obtaining vector T is:

T＝{T ₁,T ₂,T ₃,T ₄,T ₅,T ₆,T ₇,T ₈,T ₉}

Wherein, T ₁, T ₂, T ₃, T ₄, T ₅, T ₆for water temperature measuring point, T ₇, T ₈, T ₉for discharge measuring point, measuring point distribution table is as shown in the table:

Test point numbering	Variable	Sensor	Way	Remarks
					T1	Load 1 inlet water temperature	1
T2	Load 2 inlet water temperatures	1
					T3	Load 3 inlet water temperatures	1
T4	Load 1 outlet water temperature	1
					T5	Load 2 outlet water temperatures	1
T6	Load 3 outlet water temperatures	1
					T7	Load 1 water branch flow	1
T8	Load 2 water branch flows	1
					T9	Load 3 water branch flows	1

Afterwards, testing data is divided into two classes: the first kind, for method study, completes the classification of operating mode, the second step as learning data for method; Equations of The Second Kind, for checking, is used for the 4th step of method as real-time downlink data.

Step 2: load working condition classification

Data on flows T in case ₇～T ₉classifying rules is known, and temperature data T ₁～T ₆classifying rules is unknown, belongs to the known situation of classifying rules part.According to given above method, we choose flow parameter and carry out producing condition classification one time, and flow control valve adjust flux divides three states: high, medium and low, a producing condition classification of three water tributary combinations is total: 3 ³=27 classes.

On a basis of classification, each class data is divided more accurately according to the given method of the present invention, i.e. secondary cluster.Here, with operating mode 1 data instance of a subseries, describe, producing condition classification process flow diagram as shown in figure 11.

By Matlab middle distance function pdist () and classification function linkage () function, first kind partial data is processed, obtain 4 groups of foundation classs, with reference to formula (3), (4), determined cluster centre and the cluster radius of each class.

The remaining data of the first kind is carried out to secondary cluster according to Fig. 6 method, for example certain test data t={t ₁, t ₂..., t ₉, make Euclidean distance with all cluster centres, that is:

$d_k=d(x_k,t)=\sqrt{\underset{i=1}{\overset{9}{\mathrm{\Σ}}}{(x_{k,i}-t_i)}^2},k=\mathrm{1,2,3,4}$

Wherein, x _k, k=1,2,3,4 represent the centre coordinate of initial classes;

Get minimum distance d _m=d _min={ d ₁, d ₂, d ₃, d ₄, compare d _mthe cluster radius R corresponding with it _mif, d _m≤ R _m, these data belong to this foundation class; If d _m> R _m, data t does not belong to existing any class, sets up a new class deposit database according to formula (2), and producing condition classification flow process is as shown in figure 11.

Step 3: the real-time downlink data pre-service of useful load

The real-time downlink data pre-service of useful load refer step one and Fig. 2.

Step 4: the monitoring of the real-time downlink data of useful load

Through step 2, obtained the producing condition classification under useful load normal mode.Utilize these operating modes to implement monitoring to downlink data, if there is the data that do not belong to any operating mode, illustrate that new working condition has appearred in load, and these data are deposited in to database as abnormal data.Utilize the cluster that step 2 generates to monitor secondary sources, as shown in Figure 12 and Figure 13, wherein, Figure 12 is clustering method while treating direction finding amount data monitoring, and the test datas before 1300 show figure; Figure 13 is clustering method while treating direction finding amount data monitoring, and 1000 later test datas show figure;

Upper figure draws, since the 1300th vector data, occurred departing from normal class (surpass 1 class be exception class), and this is indicating load or be about to and breaks down.

Step 5: useful load fault detect and isolation

Abnormal data to step 4 output, first judges whether it is false-alarm (by repeatedly testing and determine whether false-alarm).If determine it is abnormal data, according to Fig. 8 fault detect flow process, first judge that whether each element value of this abnormal data is in normal range, if there is element value to break bounds, illustrate that the sensor that this element is corresponding spreads out of abnormal data, according to fault diagnosis tree method, determine the position (load) that fault occurs.

Step 6: load health knowledge learning process

Every the fixing cycle, load health knowledge carries out self-teaching, to improve the accuracy of algorithm.Be mainly manifested in 2 points:

1) existing class is carried out to new cluster, according to formula (5), redefine its cluster centre.

2) for the class of new formation, according to step 2, it is redefined to its cluster radius and cluster centre as known operating mode, complete the renewal to new class.

In sum, space station useful load health monitor method based on data-driven algorithm provided by the invention, by useful load real-time descending test data in space station is analyzed, the Real-Time Monitoring of realization to load health status, fault detect and location that can sustained load, and prediction to a certain degree.In addition, in load design and development stage, its mode of operation and failure mode have been carried out to sufficient analysis, carried out on this basis testability design, object is to make the data that gather with transmitting be conducive to carry out fault diagnosis and health control.The method has changed the descending test data of load in the past and has only done the way that shows and transfinite judgement, machine learning by historical data forms system health knowledge base, distance value based on outlier calculates the abnormality of finding load, can before true fault occurs, provide prompting.

The above is only the preferred embodiment of the present invention; it should be pointed out that for those skilled in the art, under the premise without departing from the principles of the invention; can also make some improvements and modifications, these improvements and modifications also should be looked protection scope of the present invention.

Claims (8)

1. the space station useful load health monitor method based on data-driven algorithm, is characterized in that, comprises the following steps:

S1, for the space station useful load by health monitoring, described useful load has H test point, that is: every of described useful load descending test data is the test data that comprises H test parameter;

Set up the historical test database of described useful load; Wherein, described historical test database is for storing the descending test data of some history of described useful load; Wherein, the descending test data of described history is historical descending non-fault test data;

S2, when needs carry out health monitoring to described useful load, reads described historical test database, obtains many descending test datas of history;

Then the descending test data of resulting history is carried out to pre-service, be met the historical test data of n bar of requirement; Wherein, the historical test data of described n bar is the test data of described useful load normal operation process;

S3 based on default selection principle, chooses m test parameter and monitors the factor as key from a described H test parameter, obtains the state vector of m dimension; Wherein, m≤H;

S4, is configured to n*m matrix by the state vector of n m dimension, and n is matrix line number, and m is matrix columns; Wherein, m data in every a line are the data of m the crucial monitoring factor comprising in the historical test data that a S2 obtains; The data of each row are the same crucial monitoring factor in the test data that homogeneous test does not obtain;

S5, is normalized by row n*m matrix, and unified the drawing of the codomain of each crucial monitoring factor arrived to same interval;

S6, determines each crucial weight of monitoring the factor, then the matrix after normalization is weighted to processing, obtains weighting matrix;

S7, using each row vector of described weighting matrix as training sample, has n training sample; A described n training sample is carried out to cluster, obtain a plurality of clusters corresponding with different loads nominal situation difference;

S8, when described useful load is carried out to health monitoring, receives the real-time descending test data that described useful load sends; Then choose the m identical with a S3 test parameter and monitor the factor as key, form state vector;

S9, by the method identical with normalization processing method in S5, is normalized described state vector; Method by identical with adding authority processing method in S6, is weighted processing to the state vector after normalization, obtains weighting state vector;

S10, compare the training sample in each cluster that described weighting state vector and S7 obtain, judge whether to exist the specific training sample identical with described weighting state vector, if existed, draw the conclusion of the normal operation of described useful load, the corresponding load working condition of cluster at described specific training sample place is the current operating mode of described useful load, exports the current operating mode of described useful load, process ends; If there is no, carry out S11;

S11, the distance of each cluster that more described weighting state vector and S7 obtain, obtains apart from the shortest specific cluster of described weighting state vector, and establishing bee-line is D; Then compare D and predeterminable range critical value R, if D≤R, draws the conclusion of the normal operation of described useful load, the corresponding load working condition of described specific cluster is the current operating mode of described useful load, export the current operating mode of described useful load, process ends; If D > is R, show that described weighting state vector does not belong to any known Historic Clustering, described weighting state vector is abnormality vector; Further show that the current operating mode of described useful load does not belong to any normal operating mode, the conclusion that described useful load current time may occur or be about to break down; Finally preserve described abnormality vector, process ends.

2. the space station useful load health monitor method based on data-driven algorithm according to claim 1, is characterized in that, in S2, the descending test data of described history is carried out to pre-service, is specially:

By deficiency of data, abnormal data and format error data are referred to as bad data; Bad data in the descending test data of described history is rejected, obtain the pretreated historical test data meeting the demands.

3. the space station useful load health monitor method based on data-driven algorithm according to claim 1, is characterized in that, in S7, a described n training sample is carried out to cluster, obtains a plurality of clusters corresponding with different loads nominal situation difference, is specially:

S7.1, judges that whether classifying rules is known, if classifying rules is completely known, carries out S7.2; If classifying rules is completely unknown, carry out S7.3; If classifying rules part is known, carry out S7.4;

S7.2, described classifying rules is known referring to completely: for the m dimension state vector being comprised of m the crucial monitoring factor, classifying rules is: each crucial monitoring factor is all bound several state grades, and the data interval that each state grade is corresponding is all known; The various combination of the state grade of each crucial monitoring factor forms different load nominal situations;

Clustering method is: whether training of judgement sample belongs to the load working condition that classifying rules is known, if belonged to, this training sample is included into corresponding load working condition, and each training sample that belongs to same load working condition forms a cluster; If do not belonged to, using this training sample as a new cluster, deposit in database; Wherein, the corresponding a kind of new load working condition of new cluster;

S7.3, the complete the unknown of described classifying rules refers to: for the m dimension state vector being comprised of m the crucial monitoring factor, each crucial monitoring factor is several state grades of end binding all;

Clustering method is:

The first step, judges whether existence foundation class, if existed, carries out second step; If there is no, given foundation class is determined cluster centre and the cluster radius of each cluster in described foundation class simultaneously by following formula;

Cluster centre: x _center=(x _max+ x _min)/2; (3)

Cluster radius: d _min=D (x _max, x _min)/2; (4)

Wherein, x _max, x _maxbe respectively the upper limit and the lower limit of training sample in same cluster; D is vector x _maxand vector x _minbetween Euclidean distance;

Second step, constantly expands foundation class, until each cluster obtaining comprises all training samples;

The 3rd step, is optimized all clusters that obtain, and the operating mode using each cluster finally obtaining as useful load deposits database in;

S7.4, known the referring to of described classifying rules part: for the m dimension state vector being comprised of m the crucial monitoring factor, exist M1 the crucial monitoring factor to bind several state grades, the data interval that each state grade is corresponding is all known; Exist M2 the crucial monitoring factor not bind several state grades; Wherein, M1+M2=m; M1 >=1; M2 >=1; M1 and M2 are natural number; : M1 the crucial monitoring factor is the factor that classifying rules is known; M2 the crucial monitoring factor is the factor of classifying rules the unknown;

Clustering method is: for n training sample, the factor that each training sample is comprised is according to its represented physical significance classification; The factor that the classifying rules of take is known is classification foundation, by the clustering method in S7.2, n training sample is carried out to cluster for the first time, obtains several original clusters;

For each original cluster, by the clustering method in S7.3, carry out cluster for the second time, obtain some sub-clusters;

Judge whether resulting each sub-cluster meets operating mode and divide requirement, if met, using each sub-cluster as final cluster result, deposits in database; If do not met, selected M1 the crucial monitoring factor while changing for the first time cluster, re-starts cluster and for the second time cluster for the first time, and circulation said process, until resulting each the sub-cluster of cluster meets operating mode division requirement for the second time.

4. the space station useful load health monitor method based on data-driven algorithm according to claim 3, is characterized in that, in the 3rd step in S7.3, all clusters that obtain is optimized, and is specially:

For foundation class C={C ₁, C ₂..., C _q, suppose that wherein min cluster is C _i(1≤i≤q), in this min cluster, element bound is respectively: x _max=(a ₁, a ₂..., a _p) and x _min=(b ₁, b ₂..., b _m), the Euclidean distance between note bound element is d _min, that is: d _min=D (x _max, x _min); To arbitrary extension class C' _j, remember that the Euclidean distance between its bound element is d' _j;

If cancel this extension class, each state vector being comprised is included into closes on class;

If retain this extension class.

5. the space station useful load health monitor method based on data-driven algorithm according to claim 4, is characterized in that, in S7.2, using this training sample as a new cluster, deposits in database, is specially: establishing this training sample is x ₀={ a ₁, a ₂... .., a _m;

Calculate x ₀and the distance d between nearest operating mode border, by x ₀as the center of new cluster, the radius apart from d as new cluster, new cluster table is shown:

C _new＝{x,D(x,x ₀)≤d} (2)

Wherein, C _newrepresent new cluster, x represents to belong to the free position vector of new cluster; D represents compute vector x and x ₀between Euclidean distance.

6. the space station useful load health monitor method based on data-driven algorithm according to claim 1, is characterized in that, after S11, also comprises:

S12, constantly deposits the real-time descending test data of useful load in described historical test database, and described historical test database is constantly updated; Then, the historical test database based on after upgrading, every the fixing cycle, up-to-date training sample set carries out self-teaching, forms up-to-date cluster, is specially:

1) existing class is carried out to new cluster, according to formula (5), redefine its cluster centre;

$c_i=\frac{1}{\left|A_i\right|}{\mathrm{\Σ}}_{x\∈A_i}x,i=\mathrm{1,2},...\left(5\right)$

Wherein, c _irepresent the cluster centre after upgrading for the i time, A _irepresent the cluster after upgrading for the i time, x represents A _ielement, | A _i| represent cluster A _inumber of elements;

2) for the class of new formation, it is redefined to its cluster radius and cluster centre as known operating mode, complete the renewal to new class.

7. the space station useful load health monitor method based on data-driven algorithm according to claim 1, is characterized in that, after S11, also comprises:

S12, after showing that described weighting state vector is the conclusion of abnormality vector, first determines whether false-alarm, if be defined as false-alarm, and direct process ends; If not, continue to judge whether this abnormality vector exists exception monitoring factor values, if there is no, illustrate that useful load current working is abnormal, the probability that is in fault critical point is higher; If existed, find the corresponding useful load test point of this exception monitoring factor, in conjunction with fault diagnosis tree method, useful load fault is detected and isolated, parts and failure mode information thereof that final output is broken down.

8. the space station useful load health monitor method based on data-driven algorithm according to claim 7, is characterized in that, in S12, by repeatedly testing and determine whether false-alarm.