CN105467975B - A kind of equipment fault diagnosis method - Google Patents
A kind of equipment fault diagnosis method Download PDFInfo
- Publication number
- CN105467975B CN105467975B CN201511004767.2A CN201511004767A CN105467975B CN 105467975 B CN105467975 B CN 105467975B CN 201511004767 A CN201511004767 A CN 201511004767A CN 105467975 B CN105467975 B CN 105467975B
- Authority
- CN
- China
- Prior art keywords
- fault
- sample
- data
- failure
- equipment
- Prior art date
- Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
- Active
Links
- 238000003745 diagnosis Methods 0.000 title claims abstract description 66
- 238000000034 method Methods 0.000 title claims abstract description 62
- 238000012549 training Methods 0.000 claims abstract description 7
- 238000011017 operating method Methods 0.000 claims abstract description 6
- 230000011218 segmentation Effects 0.000 claims description 46
- 239000013598 vector Substances 0.000 claims description 44
- 238000000605 extraction Methods 0.000 claims description 27
- 239000011159 matrix material Substances 0.000 claims description 22
- 238000006243 chemical reaction Methods 0.000 claims description 17
- 238000012423 maintenance Methods 0.000 claims description 12
- 230000008569 process Effects 0.000 claims description 11
- 230000008030 elimination Effects 0.000 claims description 8
- 238000003379 elimination reaction Methods 0.000 claims description 8
- 238000012545 processing Methods 0.000 claims description 8
- 238000003860 storage Methods 0.000 claims description 8
- 241001269238 Data Species 0.000 claims description 7
- 238000004364 calculation method Methods 0.000 claims description 7
- 238000011160 research Methods 0.000 claims description 7
- 238000012216 screening Methods 0.000 claims description 6
- 230000015556 catabolic process Effects 0.000 claims description 3
- 230000002035 prolonged effect Effects 0.000 claims description 2
- 238000004422 calculation algorithm Methods 0.000 abstract description 10
- 239000000284 extract Substances 0.000 abstract description 4
- 238000012631 diagnostic technique Methods 0.000 abstract description 2
- 230000000877 morphologic effect Effects 0.000 abstract description 2
- 238000005457 optimization Methods 0.000 abstract description 2
- 238000005516 engineering process Methods 0.000 description 18
- 238000001914 filtration Methods 0.000 description 15
- 238000011161 development Methods 0.000 description 8
- 230000018109 developmental process Effects 0.000 description 8
- 230000000916 dilatatory effect Effects 0.000 description 6
- 230000000694 effects Effects 0.000 description 6
- 238000003909 pattern recognition Methods 0.000 description 6
- 238000004458 analytical method Methods 0.000 description 5
- 238000013528 artificial neural network Methods 0.000 description 5
- 230000004888 barrier function Effects 0.000 description 5
- 238000009826 distribution Methods 0.000 description 5
- 238000013461 design Methods 0.000 description 4
- 238000005259 measurement Methods 0.000 description 4
- 230000002159 abnormal effect Effects 0.000 description 3
- 230000005856 abnormality Effects 0.000 description 3
- 230000001419 dependent effect Effects 0.000 description 3
- 230000004927 fusion Effects 0.000 description 3
- 238000012544 monitoring process Methods 0.000 description 3
- 230000035772 mutation Effects 0.000 description 3
- 238000005086 pumping Methods 0.000 description 3
- XLYOFNOQVPJJNP-UHFFFAOYSA-N water Substances O XLYOFNOQVPJJNP-UHFFFAOYSA-N 0.000 description 3
- 230000008859 change Effects 0.000 description 2
- 238000010276 construction Methods 0.000 description 2
- 230000008878 coupling Effects 0.000 description 2
- 238000010168 coupling process Methods 0.000 description 2
- 238000005859 coupling reaction Methods 0.000 description 2
- 238000013075 data extraction Methods 0.000 description 2
- 238000000354 decomposition reaction Methods 0.000 description 2
- 238000001514 detection method Methods 0.000 description 2
- 238000012886 linear function Methods 0.000 description 2
- 238000004519 manufacturing process Methods 0.000 description 2
- 238000011089 mechanical engineering Methods 0.000 description 2
- 239000000203 mixture Substances 0.000 description 2
- 238000012986 modification Methods 0.000 description 2
- 230000004048 modification Effects 0.000 description 2
- 238000002360 preparation method Methods 0.000 description 2
- 230000003252 repetitive effect Effects 0.000 description 2
- 238000005070 sampling Methods 0.000 description 2
- 238000009825 accumulation Methods 0.000 description 1
- 125000002015 acyclic group Chemical group 0.000 description 1
- 230000006978 adaptation Effects 0.000 description 1
- 230000003044 adaptive effect Effects 0.000 description 1
- 238000013459 approach Methods 0.000 description 1
- 238000013473 artificial intelligence Methods 0.000 description 1
- 238000009412 basement excavation Methods 0.000 description 1
- 238000007906 compression Methods 0.000 description 1
- 230000006835 compression Effects 0.000 description 1
- 238000000205 computational method Methods 0.000 description 1
- 239000000470 constituent Substances 0.000 description 1
- 238000001816 cooling Methods 0.000 description 1
- 238000013144 data compression Methods 0.000 description 1
- 238000007418 data mining Methods 0.000 description 1
- 238000013499 data model Methods 0.000 description 1
- 230000007812 deficiency Effects 0.000 description 1
- 238000009432 framing Methods 0.000 description 1
- 238000007689 inspection Methods 0.000 description 1
- 238000003062 neural network model Methods 0.000 description 1
- 238000010606 normalization Methods 0.000 description 1
- 230000035939 shock Effects 0.000 description 1
- 230000009897 systematic effect Effects 0.000 description 1
- 238000012546 transfer Methods 0.000 description 1
- 238000013024 troubleshooting Methods 0.000 description 1
- 238000005303 weighing Methods 0.000 description 1
Classifications
-
- G—PHYSICS
- G05—CONTROLLING; REGULATING
- G05B—CONTROL OR REGULATING SYSTEMS IN GENERAL; FUNCTIONAL ELEMENTS OF SUCH SYSTEMS; MONITORING OR TESTING ARRANGEMENTS FOR SUCH SYSTEMS OR ELEMENTS
- G05B23/00—Testing or monitoring of control systems or parts thereof
- G05B23/02—Electric testing or monitoring
- G05B23/0205—Electric testing or monitoring by means of a monitoring system capable of detecting and responding to faults
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F16/00—Information retrieval; Database structures therefor; File system structures therefor
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06Q—INFORMATION AND COMMUNICATION TECHNOLOGY [ICT] SPECIALLY ADAPTED FOR ADMINISTRATIVE, COMMERCIAL, FINANCIAL, MANAGERIAL OR SUPERVISORY PURPOSES; SYSTEMS OR METHODS SPECIALLY ADAPTED FOR ADMINISTRATIVE, COMMERCIAL, FINANCIAL, MANAGERIAL OR SUPERVISORY PURPOSES, NOT OTHERWISE PROVIDED FOR
- G06Q50/00—Information and communication technology [ICT] specially adapted for implementation of business processes of specific business sectors, e.g. utilities or tourism
- G06Q50/06—Energy or water supply
Landscapes
- Engineering & Computer Science (AREA)
- General Physics & Mathematics (AREA)
- Business, Economics & Management (AREA)
- Physics & Mathematics (AREA)
- Theoretical Computer Science (AREA)
- Economics (AREA)
- Health & Medical Sciences (AREA)
- Human Resources & Organizations (AREA)
- Marketing (AREA)
- Automation & Control Theory (AREA)
- Data Mining & Analysis (AREA)
- Public Health (AREA)
- Water Supply & Treatment (AREA)
- Databases & Information Systems (AREA)
- General Engineering & Computer Science (AREA)
- General Health & Medical Sciences (AREA)
- Primary Health Care (AREA)
- Strategic Management (AREA)
- Tourism & Hospitality (AREA)
- General Business, Economics & Management (AREA)
- Management, Administration, Business Operations System, And Electronic Commerce (AREA)
- Testing Of Devices, Machine Parts, Or Other Structures Thereof (AREA)
Abstract
A kind of equipment fault diagnosis method, include fault diagnosis training step and fault diagnosis operating procedure successively, set about data shape feature essence, based on equipment fault data, the dynamic time warping algorithm for combining multidimensional piecewise fitting algorithm and optimization realizes pattern expression and the distance threshold abstraction function of modeling, and extracts morphological feature by the unit exception data to discovery and carry out pattern match and reach the function of device data fault type recognition and cause diagnosis solving the problems, such as to be difficult to efficiently and accurately portray in current failure diagnostic techniques similarity degree between fault data.
Description
Technical field
The present invention relates to fault diagnosis technology fields, and in particular to a kind of equipment fault diagnosis method.
Background technology
In recent years, with the industrialized fast development of continuous innovation growth and society of science and technology, it is more and more
Industrial industrialization system be loaded in large scale industry production occasion, created under this application environment can not be substituted it is huge
Productivity.Thus the maintenance work of these equipment, which seems, is even more important and arduous.But the influence factor often implied too much
It is likely to that the generation of these system equipment failures is caused even to fail, therefore country and enterprises and institutions increasingly pay close attention to pair
Work in terms of the state-detection and fault diagnosis of these important system equipment.
In general, these device structures are increasingly sophisticated careful, the effect between different parts mutually involves, and coupling is very
By force.By dismounting, the approach of disassembled equipment, i.e., time-consuming and laborious, diagnosis effect is often not satisfactory, and is easy to cause equipment performance not
Stablize.The state of equipment significant points is read in real time by installing monitoring sensor device, with observation device relevant parameter
This method simple, intuitive that operation trend compares metrics-thresholds is highly reliable, at present in productions such as all kinds of power plant, automobile, satellites
It is popularized than wide under occasion.But the efficiency of this method be still limited to plant maintenance personnel working experience ability and
Energy, the unknown situation processing of complexity that the personnel being lacking in experience highlight equipment are more had a heart but no strength.
The above method is hardly formed the diagnostics architecture of specification due to the predicament that Heuristics is expressed, and then one kind is based on
The method for diagnosing faults of mathematics digging technology is gradually introduced in the middle of the fault diagnosis work of these large-scale and complicated devices.Number
It learns digging technology and has merged such as Comtemporary Control Theory, computer science, artificial intelligence, signal processing, pattern-recognition, statistical mathematics
Section's knowledge, by researching and analysing the historical data and current real-time data of equipment, excavate lie in inside data be conducive to therefore
Hinder the information of status analysis.Expert system, correlation rule, neuroid, Bayesian network etc. are some applications than wide
Mathematics method for digging, these methods gradually formed it is a set of be divided into accident analysis, fault modeling, fault detect, failure infer,
The research system of the several respects such as failure decision, failure decision.
In paper《Merge the car fault diagnosis expert system of example and rule-based reasoning》(《Mechanical engineering journal》, 2002
Year, the 7th phase:In 91-95), article proposes a kind of hybrid inference method of completely new fusion example and rule, and it is whole to establish one
Cover the expert system for being difficult to exchange information for solving the problems, such as different diagnosis units.It can be seen that expert system can be in specific neck
In domain, goes to solve the troubleshooting issue in related field with the ability similar to human expert's level, be mainly characterized in that
Oneself distinctive rule is formed by existing human expert's Heuristics to carry out the analysis and solution of problem.This is expert system
The characteristics of system, but mankind's professional knowledge is limited in the shortcoming of tera incognita experience, often there is self-teachings for expert system
Inferior capabilities, diagnosis success rate be not high, systematic knowledge obtains a series of problems, such as difficult, and there is still a need for expert system for these problems
This method is further deeply probed into and is optimized.
In paper《Electric network failure diagnosis based on correlation rule data mining technology》(《Electric power system protection and control》,
2009, the 9th phase:In 8-14), association rule algorithm is applied in electric network failure diagnosis by article.Referred to according to fault characteristic
Fixed corresponding decision attribute and conditional attribute and the foundation for completing original decision table, while association rule-based algorithm is to decision table
Data carry out the excavation of frequent item set and the screening of strong rule, finally realize pushing away to the fault message of Various Complex situation
Reason diagnosis.But the technology there are problems that, if such as potential implicit fault mode discrimination, large-scale data
FAQs, these still follow-up related works such as processing capacity is poor, correlation rule storage is low with update efficiency are solved.
In paper《Research on Fault Diagnosis Technology based on BP neural network》(《Computer and modernization》, 2009, the 7th
Volume) in, article selects general BP neural network to be applied in fault diagnosis scene, in reasonable construction BP neural network model
In the case of, according to equipment level characteristic distributions successfully by Application of Neural Network in fault diagnosis, improve the effect of diagnosis
Rate and reliability.Neural network maximum feature be in the case of large sample size can unlimited None-linear approximation original data model,
But the features such as its intrinsic over-fitting, numerical value randomness, unstable training, limits neural network in fault diagnosis field
Application range, some other optimization algorithm need to be coordinated to carry out diagnostic application.
In paper《The Bayesian network Research on fault diagnosis method of multicharacteristic information fusion》(《China Mechanical Engineering》,
2010, the 8th phase:In 940-945), article is to pump class vibration signal as research object, using Bayes's Parameter Estimation Method to letter
The various faults feature such as frequency domain, time domain of number extraction is merged into row information, is reconstructed Bayesian network later and is established complete event
Hinder grader, the identification of fault type is carried out by the calculating of maximum a-posteriori estimation value.The structure of Bayesian network needs
The priori statistical knowledge of great amount of samples data, while Bayesian network directed acyclic manifestation mode is there are the risk of error accumulation,
These aspects are that there is an urgent need for concerns for Bayesian network method for diagnosing faults.
In view of above-mentioned several common faults diagnostic techniques there are the problem of and risk, how the present invention around fully excavating number
According to the potential intuitive expression-form and measurement fault mode difference validity problem of sample, attempt by setting up a kind of new mould
Formula matching technique preferably solves the problems, such as that the inefficiency of fault diagnosis, recognition accuracy be not high, good to be conducive to equipment
The duration of good state is safeguarded.
Invention content
It is an object of the invention to overcome the deficiencies of the prior art and provide a kind of equipment fault diagnosis methods, solve current
It is difficult to efficiently and accurately to portray the problem of similarity degree between fault data in fault diagnosis technology.
The present invention provides a kind of equipment fault diagnosis methods, in turn include the following steps:
Step 1:Fault diagnosis training step:Fault sample information is obtained from database, after calculation processing respectively
The eigenmatrix and pattern distance threshold value of the fault sample for the dimension that is eliminated, the two is associated with each other, and storage generates event
Hinder pattern repository;
Step 2:Fault diagnosis operating procedure:Exceptional sample information is obtained from real-time data base, is obtained after calculation processing
Off-note information, then calculates the sample characteristics in fault mode knowledge base that be converted to pattern after pattern distance similar successively
Degree, exports final fault diagnosis result.
Further, the step 1 the specific steps are:
(1.1) fault sample information is obtained from database;
(1.2) sectional linear fitting is carried out to each fault sample successively;
(1.3) feature extraction is carried out to every segment data of fault sample, obtains the eigenmatrix of fault sample;
(1.4) fault signature conversion is carried out, feature dimension, the eigenmatrix of the fault sample for the dimension that is eliminated are eliminated;
(1.5) pattern distance threshold value is calculated;
(1.6) fault signature matrix and pattern distance threshold value is associated with each other, storage generates fault mode knowledge base.
Further, the step (1.1) the specific steps are:It selects fault type number P >=2 and each failure is sent out
Raw number T >=2 meet the requirements can research equipment, the observation point N of enough numbers, wherein N >=10 are selected, to equipment long enough
The history run status data of time carries out failure logging lookup, is plucked from failure logging using the screening rule of setting and selects event
The useful information of the related measuring point information of barrier, the beginning and ending time of failure process and breakdown maintenance measure record, according to useful information
The read failure sample data from power plant real-time data base PI, wherein:
Measuring point number is n, the time counts out for m fault sample data, regard a n as in whole measuring point datas at j moment
The column vector of dimension, is expressed as:
u(tj)=[uj1, uj2, uj3..., ujn]
The sample data file saves as the matrix form of m × n, and concrete form is as follows:
Wherein row represents m fault time, and row represent n equipment observation point, and ranks m, n between each fault sample
Two values are not quite similar, while assigning its fault type for each fault sample and identifying ID, and fault type mark ID determines that method is
If in whole samples including X kind failures, the numberical range of fault type mark ID is:1-X.
Further, the step (1.2) the specific steps are:
(1.2.1) mean filter operates:Elimination is filtered to the noise pollution being entrained in sample data;
The segmentation initialization of (1.2.2) fault sample:Segmentation initialization is carried out to the fault sample after being filtered;
(1.2.3) is by initialization data section joint account error of fitting two-by-two;
(1.2.4) determines the segmentation cut point of fault sample, and the state that adaptivity is carried out to fault sample is divided.
Further, the step (1.3) the specific steps are:
Fault sample is divided into through step (1.2)This x segmentation data segment, fromData segment starts
Feature extraction is carried out, concrete operations are as follows:
Feature extraction is carried out according to the dimension (row) of measuring point, feature has:Slope k, duration l, mean value m, tetra- kinds of variance v;
Linear fit is carried out to this vector according to principle of least square method, fitting result is linear function p (x)=a0i+
a1iX, so slope characteristics k=a1i, wherein a0i、a1iFor fitting constant;
Duration characteristics l isVector length h, i.e. l=h;
Characteristics of mean m isVectorial all numerical value are averaged, i.e.,
Variance feature v isAll fluctuating ranges the sum of of the numerical value from mean value of vector, i.e.,N is vector
Sum;
It is remainingFeature extraction is carried out according to above-mentioned operating method, converts original time domain data matrix to
Mathematical feature matrix, it is final to divide data segment setIt is converted into:
Further, the step (1.4) the specific steps are:Slope k is converted to inclination angle alpha, when duration l is converted into
Between ratio, characteristics of mean m is normalized, by the eigenmatrix of fault sample
Further, the step 2 the specific steps are:
(2.1) exceptional sample information is obtained from real-time data base;
(2.2) sectional linear fitting is carried out to current exceptional sample;
(2.3) off-note extraction and conversion are carried out to the data after piecewise fitting;
(2.4) pattern distance is calculated successively to the sample characteristics in fault mode knowledge base with off-note information;
(2.5) pattern distance is converted to Pattern similarity using the drop ridge method of fractional steps;
(2.6) final fault diagnosis result is exported.
Further, the step (2.1) the specific steps are:
There is certain unknown abnormality in equipment state early warning system discovering device according to power plant, carries out following phase
Close operation:
(2.1.1) determines device alarm generation time t from early warning system1With deadline t2;
(2.1.2) determines the dependent observation point wp=[x of device alarm from early warning system1, x2..., xn′];
(2.1.3) is according to generation time t1With deadline t2And database sampling frequency fs, obtain time points m=
fs×(t2-t1), equipment measure-point amount n=length ([x1, x2..., xn′]), wherein length () is computational length letter
Number);
(2.1.4) obtain exceptional sample data be measuring point number be n ', the time counts out as the sample data of m, at the j moment
Whole measuring point datas regard the column vector of a n ' dimension as, are expressed as:
v(tj)=[vj1, vj2, vj3..., vjn′]
Sample data file is stored as to the matrix form of m × n ', concrete form is as follows:
Wherein row represents m fault time, and row represent a equipment observation points of n '.
The step (2.2) the specific steps are:
(2.2.1) mean filter operates:According to mean filter principle to being entrained in each measuring point in exceptional sample data
Noise pollution is filtered elimination;
The segmentation initialization of (2.2.2) exceptional sample:Segmentation initialization is carried out to the exceptional sample after being filtered;
(2.2.3) is by initialization data section joint account error of fitting two-by-two;
(2.2.4) determines that exceptional sample is segmented cut point based on the mode that recurrence merges.
Further, the step (2.5) the specific steps are:
Pattern distance is converted to Pattern similarity according to following formula according to the principle of the drop ridge method of fractional steps
Wherein,It represents exceptional sample and carries out pattern-recognition calculating with j-th of fault sample under fault type i
Pattern distance, and ThiRepresent the pattern distance threshold value under fault type i;
According to the dimension of fault type, by each distance vector
It is converted into similarity vector
The similarity set of exceptional sample FT and fault knowledge library whole sample are finally obtained, as follows:
Further, the step (2.6) the specific steps are:
According to similarity aggregated result ρ Sets, according to the rule output fault diagnosis result of agreement, wherein meeting rule:
Rule 1:If have with exceptional sample similarity be more than 90% fault sample and occur it is multiple, by whole samples
Serial number exports, while exporting the fault type belonging to fault sample, the confidence level of affiliated fault type, fault progression stage as most
The best match position of big similarity sample exports the maintenance measures that suggestion is taken if confidence level is more than 50%;
Rule 2:If not being more than that 90% fault sample and maximum similarity numerical value are more than etc. with exceptional sample similarity
In 60%, then the serial number of maximum similarity sample is exported, while exporting the fault type belonging to the fault sample, affiliated failure
The confidence level of type, the best match position that the fault progression stage is maximum similarity sample, if confidence level is more than 50%,
The maintenance measures taken are suggested in output;
Rule 3:If not being more than 90% fault sample with exceptional sample similarity and maximum similarity numerical value is less than
60%, then it is uncommon operating mode or unknown failure type by output abnormality sample, continues to pay close attention to this abnormal development.
The equipment fault diagnosis method of the present invention, may be implemented:
1. the present invention is a kind of completely new independent of fault diagnosis technology.Traditional at present or existing diagnosis skill
Art, they are intended to that dependent on the participation of expertise and prior information fault diagnosis work could be unfolded, and the present invention is from analysis
The math-morphological features of historical failure sample data are set about, to current exceptional sample determine in time to it is faulty similar
Degree and its most probable Developing Trend.
2. fault diagnosis work can be accurately and accurately unfolded in the present invention, by the way that sample data is converted from time domain space
To math-morphological features space, that is, data information compression goal is realized, and eliminate the influence of redundant data noise, to greatly
Improve the efficiency of fault diagnosis.
3. for the validity of Assured Mode expression, the present invention establishes a kind of multidimensional segmentation based on data state of development
Approximating method.In view of the unsharp feature of coupling between each variable in sample data, the present invention uses adaptive algorithm
The optimal segmenting threshold of current sample data is excavated, and the split position of all variables is consistent, to make each stage of segmentation
Data state of development has more hierarchy.
4. the present invention is low to the sample data requirement property for carrying out fault diagnosis, without carrying out complex data pretreatment work.
Existing the relevant technologies generally cannot achieve the sample data pair inconsistent from the sample duration in knowledge base, variables number is different
Fault diagnosis, the present invention trimmed by variable, the methods of dynamic time warping can solve the problems, such as this.
5. accident analysis of the present invention within second grade, can not only realize full failure with the fault identification part speed of service
The sample of mode carries out fault diagnosis, moreover it is possible to carry out fault location to partial fault mode.Therefore the present invention can transplant completely
To being applied in equipment on-line condition detecting system.
Description of the drawings
The general flowchart of Fig. 1 fault diagnosis model training stage and operation phase
Fig. 2 dynamic time warping distance design sketch
The crucial measuring point tendency chart of three kinds of failures of Fig. 3 heat pumps fore pump
Fig. 4 exceptional sample on-line fault diagnosis design sketch
Specific implementation mode
The following detailed description of the specific implementation of the present invention, it is necessary to it is indicated herein to be, implement to be only intended to this hair below
Bright further explanation, should not be understood as limiting the scope of the invention, and field technology skilled person is according to above-mentioned
Some nonessential modifications and adaptations that invention content makes the present invention, still fall within protection scope of the present invention.
It is individually to isolate each variable dimension of fault data to drive capable calculating into have most of the relevant technologies, and this hair
It is bright consider between sample data difference variable relevant information there are the problem of, realize a kind of quasi- based on bottom-up segmentation
Conjunction technology takes into account the numerical characteristics of fault data while extracting a variety of configuration features of fault data.The present invention is logical
Being effectively compressed for large-scale data can be realized by crossing this method, reduce failure diagnosis time, and is fundamentally improved failure and examined
Disconnected precision, therefore the present invention has rapid failure diagnosis ability.The present invention carries out fault sample in dynamic time warping method
Feature gap calculates, and can solve the problems, such as that the fault sample similarity degree of different length is difficult to weigh.The fault diagnosis of the present invention
Distance is not used as final measurement standard to method in mode, but is compared with similar percentage, has unified the weighing apparatus of fault identification
Amount standard, and then fault diagnosis precision is improved again.
It is that one kind setting about data shape the present invention provides a kind of equipment fault diagnosis method based on multidimensional piecewise fitting
The completely new fault diagnosis technology of feature essence, it is based on equipment fault data, combines multidimensional piecewise fitting algorithm and excellent
The dynamic time warping algorithm of change realizes pattern expression and the distance threshold abstraction function of modeling, and different by the equipment to discovery
Regular data extraction morphological feature carries out pattern match and reaches device data fault type recognition and the function of cause diagnosis.This method
Main includes establishing two processes of model and moving model.
Fig. 1 (left side) is the flow chart that the present invention establishes model, and entire modeling process mainly includes the following steps that:
Step 1:Fault sample information is obtained from system database.
Condition Detection system of the fault sample information now obtained from power plant, rough operating process are as follows:It is first
First, what one fault type number P (>=2 P) of selection and each failure frequency T (>=2 T) were met the requirements can research
Equipment selectes the observation point N (>=10 N) of enough numbers, and event is carried out to equipment history run status data prolonged enough
Hinder record search;Then it is plucked from failure logging using the screening rule of setting and selects failure correlation measuring point information, failure process
Beginning and ending time and breakdown maintenance measure record etc. useful informations;Last foundation above- mentioned information is from power plant real-time data base PI
Read failure sample data.
Citing, a measuring point number is n, the time counts out for m fault sample data, whole measuring point datas at j moment can
The column vector for regarding n dimensions as, is expressed as:
u(tj)=[uj1, uj2, uj3..., ujn]
The sample data file saves as the matrix form of m × n, and concrete form is as follows:
According to the method described above, the storage form of all fault samples:This dimension of row represents m fault time, arranges this
Dimension represents n equipment observation point, and two value of ranks dimension (m, n) between each fault sample may be not quite similar.Simultaneously
Its fault type, which is assigned, for each fault sample identifies ID.Fault type mark ID determines method:If including X kinds in whole samples
Failure, the then numberical range that fault type identifies ID are:1~X.
Step 2:Multidimensional sectional linear fitting is carried out to each fault sample successively
This step be this fault diagnosis invention key technology part, mainly according to fault sample overall development trend into
The state of row adaptivity is divided, and data prediction preparation is carried out for the extraction of subsequent fault signature.
Step 2.1 mean filter operates:
In order to effectively carry out the segmentation of fault sample data, first have to carry out the noise pollution being entrained in sample data
Filtering is eliminated.Fast and effeciently to realize that noise data filters, the linear filtering algorithm of we selected typical:Mean filter.
According to mean filter principle, specific filtering operations are:First observation of fault sample data F is extracted first
Point column vectors data u-1, form is as follows:
By column vector u-1See Serial No., Filtering Template size as(round is round
Function), if L is even number, need to add 1 again.After having determined Filtering Template size, substantive filtering operation is proceeded by:
1. the center of Filtering Template is placed in u11Place, by within template all data summation divided by L, institute's value
For u11The output numerical value of filtering operation be:
2. the center of Filtering Template continues to move forward, output is filtered until all numbers according to above-mentioned identical operation
According to being fully completed filtering operation, such as in u1iThe output of position, mean filter is:
3. the column vector data u finally obtained-1Filtering data be also a column vector, concrete form is as follows:
All measuring points of fault sample data complete filtering operation, such fault sample data F according to above-mentioned 3 step successively
It is converted into F*, form is:
The segmentation initialization of step 2.2 fault sample
Segmentation initialization is carried out to the fault sample F* after being filtered, every 2 data is divided into 1 data segment, due to event
Barrier sample data length is m, then is classified asA most careful data segment;If m is odd number, final stage is by 3 datas
Composition, is finally divided intoA most careful data segment.The initialization effect of segmentation is:
Step 2.3 initialization data section joint account error of fitting two-by-two
The main purpose of this step is to being generated after the initialization segmentation of upper step(or) a data segment, each
Data segment and its right adjacent data segment merge, while calculating the error of fitting of generation.
This is clearly one and seeks I=I (a0, a1) extreme value quadratic programming problem.According to Lagrangian principle, obtain
Wherein, m=4 n=1
Above formula is a0, a1About system of linear equations, be expressed in matrix as
Above formula becomes normal equation group, and coefficient matrix is a symmetric positive definite matrix, therefore existence and unique solution.From above formula
Solve ak(k=0,1), so as to obtain multinomial:
P (x)=a0+a1x
Since the mean error that fitting generates is known as error of fitting, error of fitting is denoted as:
After all measuring point fittings are completed, the sum of error of fitting of generation(N is measure-point amount).
Step 2.4 determines the segmentation cut point of fault sample
This step main purpose is that the determination of fault sample split position is realized by way of recurrence merging.
1. a numerical value MAX_ERROR is rule of thumb arranged as the index judged error of fitting in the present invention.If quasi-
Close error vectorIn all numerical value when being all higher than maximum error of fitting threshold value MAX_ERROR, then execute operation 1:ConfirmIt is exactly
Final sample decomposition location sets;Otherwise operation 2 is executed:Merge error of fitting vectorMiddle minimum value position J and its right neighbour
The segmentation data segment of position J+1, and update segmentation data segment set F*, error of fitting vectorWith sample entirety split position to
Amount
Its specific update, which operates, is:F*Length subtracts 1, and preceding J-1 is remained unchanged, and the numerical value after J moves forward 1, F ' on the positions JJ
=[FJ;FJ+1];According to updated F*Determine new split position vector Length subtracts 1, and preceding J-1 is remained unchanged, after J
Numerical value moves forward 1, on the positions JMerge is to generate asking for design according to least square method to be fitted mistake
Difference function;
2. carrying out size with maximum error of fitting threshold value MAX_ERROR to numerical quantity to updated error of fitting again to sentence
It is disconnected, if all greater than threshold value MAX_ERROR operation 1 will be executed, it otherwise will jump out circulate operation stopping;
3. the segmentation data segment that loop fusion minimum error of fitting position and its right ortho position are set, until last error of fitting
VectorAll greater than MAX_ERROR numerical value, recurrence merges cycle and stops.
After above-mentioned steps, after whole segmentation data segments is completed by way of recurrence merging, three can be generated
It is a as a result, one be segmentation data segment setSecond is that error of fitting is vectorialThird,
Sample entirety split position vector(wherein x is the number that sample data is finally divided).So far it just realizes
The target that the state segmentation of adaptivity is carried out to fault sample, is ready for the extraction of lower step fault signature.
Step 3:Feature extraction is carried out to fault sample segmentation data segment;
This step is realized each divides the numbers such as data segment extraction slope, time scale, mean value, variance to current failure sample
Learn feature.
For example, by fault sample F*It is divided into through step 2This x segmentation data segment.FromData segment is opened
Begin to carry out feature extraction.
The concrete operations of feature extraction are as follows:
Feature extraction is carried out according to the dimension (row) of measuring point, main feature has:Slope k, duration l, mean value m, variance v tetra-
Kind.
Linear fit is carried out to this vector according to principle of least square method, fitting result is linear function p (x)=a0i+
a1ix.So slope characteristics k=a1i;
Duration characteristics l isVector length, i.e. l=h;
Characteristics of mean m isVectorial all numerical value are averaged, i.e.,
Variance feature v isAll fluctuating ranges the sum of of the numerical value from mean value of vector, i.e.,
It is remainingFeature extraction is carried out according to above-mentioned operating method, converts original time domain data matrix to
Mathematical feature matrix.
Final segmentation data segment setIt is converted into such form:
Step 4:Fault signature conversion is carried out, feature dimension is eliminated;
1 slope k is converted into inclination angle alpha:Since the value range of k is [- ∞ ,+∞], similar two events in order to prevent
For barrier sample since the excessive difference that the presence of error of fitting generates slope numerical value eventually leads to erroneous judgement, we pass through arctan function
Slope numerical value is converted to angular values, to realize the elimination of feature dimension.Slope characteristics conversion formula is:αij=arctan
(kij) wherein i ∈ [1, x];J ∈ [1, n;
2 duration l are converted into time scale p:In the case of fault sample length difference, segmentation duration can not represent very
Real effective segmentation stage length, also easy tos produce erroneous judgement, so should divided by original fault sample entire length, when eliminating
The dimension of long feature.Duration characteristics conversion formula is:Wherein i ∈ [1, x];
3 characteristics of mean m normalization:In the case of fault sample divides number difference, mean value size can not represent very
Real effective segmentation stage numerical value feature, also easy tos produce erroneous judgement, so should divided by original fault sample all numerical value total amounts,
To eliminate the dimension of characteristics of mean.Duration characteristics conversion formula is:Wherein i ∈ [1, x];J ∈ [1, n];
Such matrix form;
Step 5:Calculate pattern distance threshold value;
The purpose of this step is to seek whole samples acquired in step 1 pattern distance between any two, and calculate
The distance threshold of each fault category.
By the four above steps, all fault sample data { F1, F2..., FsIt is completely converted into similar formula (1)
Eigenmatrix form set { fm1, fm2..., fms}.Next it needs to calculate similar event using dynamic time warping method (DTW)
The pattern distance of different samples, foundation result of calculation determine the pattern distance threshold value per class failure two-by-two in barrier sample.
The concrete operations introduced below that pattern distance is calculated by dynamic time warping (DTW) method:
1 assumes that there are two sample characteristics matrix I under fault type IAAnd IB, concrete form is as follows:
Wherein, ai(i=1,2 ..., x), bj(j=1,2 ..., x ') is respectively two sample characteristics matrix IAAnd IBRow
Vector represents mathematical feature performance of each of respective fault sample segmentation data segment in all measuring point dimensions.
2 the same as two sample S under fault typeAAnd SBEigenmatrix IA、IBPattern distance, in conjunction with attached drawing 2 according to dynamic
Time Warp principle determines its optimum way, by IAAnd IBRow vector is replicated and is stretched, and is reached and is met two sample characteristics squares
I under conditions of array length alignment is equalAAnd IBPattern distance result of calculation it is minimum.SAAnd SBPattern distance calculation formula
For:
Wherein, T [i:] indicate the Multiphase sequences that i-th of row vector of eigenmatrix T is formed to a last row vector;
MDbase(ai, bj) indicate and segmentation data segment Euclidean distance, circular is as follows:
Wherein, eud (X, Y) indicates to seek the Euclidean distance of two vectors X, Y, i.e.,ε, λ, ρ, ω tetra-
Variable represents tetra- inclination angle alpha, time scale p, average ratio T, variance v features power shared in calculating pattern distance
Weight values.The sum of tetra- variables of ε, λ, ρ, ω are 1, and 0.25 is all arranged in the present invention
According to DTW computational methods described above, two sample S can be acquiredAAnd SBPattern distance be
According to aforesaid way can calculate all fault types pattern distance threshold value, compositional model distance threshold vector:
Thsets=[Th1, Th2..., Thx]
Step 6:It integrates useful information and generates fault mode knowledge base;
Useful information should include following sections in fault knowledge library:
The eigenmatrix set FM={ fm that 1 all fault samples extract1, fm2..., fms(s is the number of fault sample
Mesh);
2 each relevant measuring point name set PT={ pt of fault type1, pt2..., ptx(x is the type of fault type
Number);
3 each corresponding pattern distance threshold vector Thsets=[Th of fault type1, Th2..., Thx] (x is failure
The type number of type).
In order to which two parts information (fault signature matrix and pattern distance threshold value) is associated with each other, the present invention according to
Fault message is stored and generates fault knowledge library by under type:
Fig. 1 (right side) is the flow chart of model running of the present invention, and whole service process includes mainly following 6 steps:
Step 1:Exceptional sample information is obtained from real-time data base;
Equipment state early warning system according to power plant finds that certain unknown abnormality occurs in the equipment, carries out following
Relevant operation:1 determines device alarm generation time t from early warning system1With deadline t2;
2 determine the dependent observation point wp=[x of the device alarm from early warning system1, x2..., xn′];
According to generation time t1With deadline t2And database sampling frequency fs, can obtain the time points m=fs ×
(t2-t1), equipment measure-point amount n=length ([x1, x2..., xn′]) (length computational lengths function).
In this way, the exceptional sample data obtained are a measuring point numbers is n ', the time counts out as the sample data of m, at the j moment
Whole measuring point datas can regard the column vector of a n ' dimension as, be expressed as:
v(tj)=[vj1, vj2, vj3..., vjn′]
The sample data file saves as the matrix form of m × n ', and concrete form is as follows:
According to the method described above, the exceptional sample storage form of acquisition:This dimension of row represents m fault time, arranges this
Dimension represents a equipment observation points of n '.
Step 2:Sectional linear fitting is carried out to current exceptional sample;
The state that adaptivity is mainly carried out according to exceptional sample overall development trend is divided, and is subsequent fault signature
Extraction carries out data prediction preparation.This step is operated according to sectional linear fitting the method for model training stage:
2.1 mean filters operate
Exceptional sample is set0.1 times of sample length be used as Filtering Template length, according to mean filter principle to doping
The noise pollution of each measuring point is filtered elimination in exceptional sample.The visible model training stage step 2.1 of concrete operations.
All measuring points of exceptional sample data complete filtering operation, such exceptional sample data according to corresponding steps successivelyTurn
It turns toForm is:
The segmentation initialization of 2.2 exceptional samples
To the exceptional sample after being filteredSegmentation initialization is carried out, every 2 data is divided into 1 data segment, due to different
Normal sample data length is m, then is classified asA most careful data segment;If m is odd number, final stage is by 3 datas
Composition, is finally divided intoA most careful data segment.The initialization effect of segmentation is:
2.3 initialization data sections joint account error of fitting two-by-two
To what is generated after the segmentation initialization of step 2.2 exceptional sample(or) a data segment, enable each data segment
It is fitted according to principle of least square method with its right adjacent data segment and is merged with realizing, and calculate the quasi- of all measuring points generations
Close error and.
2.4 recurrence, which merge, determines that exceptional sample is segmented cut point
This step concrete operations reference model builds the stage, the above segmentationA data segment is by way of recurrence merging
Constantly adjacent two data segment for generating minimum error of fitting is merged, will produce updated 3 as a result, 1 is segmentation number
Gather according to sectionSecond is that error of fitting is vectorialThird, sample entirety framing bits
Set vector(wherein x is the number that sample data is finally divided).So far be achieved that exceptional sample into
The target of the state segmentation of row adaptivity,For the operation object of lower step off-note extraction.
Step 3:Off-note extraction and conversion are carried out to the data after piecewise fitting;
This step realizes each data segment extraction slope after dividing to current exceptional sample, time span, mean value, variance
4 mathematical features.
For example, by exceptional sampleIt is divided into through step 2This x segmentation data segment.Feature extraction
Concrete operations can build the step 3 in stage with reference model.
When all segmentation data segments of exceptional sample carry out feature extraction in the manner described above, length is the segmentation number of x
According to vector paragraphFollowing eigenmatrix form can be converted into:
In order to prevent when calculating pattern distance, exceptional sample is with similar fault sample due to following
Presence can lead to final erroneous judgement:
The presence of 1 error of fitting leads to the excessive difference of slope numerical value, and settling mode is by arctan function by slope
Numerical value k is converted to angular values α, to realize the elimination of feature dimension;
2 in the case of fault sample length difference, and segmentation duration l can not be represented shared by the authentic and valid segmentation stage
Ratio easy tos produce erroneous judgement, should divided by former fault sample entire length m, to eliminate the dimension of duration characteristics;
3 in the case of fault sample divides number difference, and mean value size can not represent the authentic and valid segmentation stage
Numerical value feature also easy tos produce erroneous judgement, so should divided by former fault sample all numerical value total amounts, to eliminate characteristics of mean
Dimension.
Eliminate the feature conversion formula of the step 4 in the concrete operation method visible model construction stage of dimension.Pass through in this way
Conversion, Feature Conversion are following Dimensionless Form:
Exceptional sampleIt is converted into features described above matrixAfterwards, so that it may carry out fault knowledge library sample mode distance and calculate
Step.
Step 4:Pattern distance is calculated successively to the sample characteristics in fault mode knowledge base with off-note information;
First, the storage information faultMessageSets in extraction fault knowledge library, concrete mode are as follows:
Then, by the eigenmatrix FT of exceptional sample and measuring point information vector wp and first kind fault message V1={ fm11,
fm12..., fm1, c1, pt1, Th1Carry out fault identification work.Concrete operations are as follows:
Two exceptional sample eigenmatrix FT carry out pattern as follows with each eigenmatrix under fault type 1 successively
Distance values dist1iIt calculates:
Since exceptional sample may be certain class full failure sample, it is also possible to for the part event in certain class evolution
Hinder sample.Accurately to calculate dist1iNumerical value, this step carry out the accurate of trouble location using dilating window slip scan technology
Positioning, that is, realize the pattern recognition function to exceptional sample, and capture out the developing stage information of exceptional sample.
2.1 first, determines the numerical value that dilating window WT is possible to take.The following contents is set in the present invention:
Fault sample in 1 fault knowledge library is entirely the sample of complete procedure;
The scaling of 2 dilating windows is ξ=20%;
Then the value range of WT can be set to round (length (FT) * 80%)~round (length (FT) *
120%), setting WT=[n1, n2..., nx];
Wherein, round is round function;Length is the function for calculating sample decomposition hop count, i.e. feature square
The line number of battle array.
Then 2.2 be n from dilating window1Start, in each sample characteristics matrix fm of fault type 11iOn slided
Calculate pattern distance.
If fm1iLength be m, then the sliding number of dilating window be, fault sample carry out pattern-recognition ranging from:
[1, n1]~[m-n1+ 1, m].
Exceptional sample from fault sample [1, n1] position starts to calculate pattern distance, each sliding step L is set to 1, under
Face is to calculate FT and fm1i[1:n1] pattern distance formula:
dist11=κ * MDDTW(FT, fm1i[1:n1]
Wherein, if repetitive rate η is equal to 0, κ=2;If repetitive rate η is not equal to 0,MDDTW(FT, fm1i[1,
n1]) it is to ask FT and fm1i[1, n1] DTW distance functions.
Residue [2, n2]~[m-n1+ 1, m] range calculated according to above-mentioned 2.1-2.2 steps, can finally obtain one
A pattern distance vectorIt is n in window size1In the case of, local optimum
Pattern distance is dist11=min (distVector1), local optimum matching position is:
L11=[nmin-n1+ 1, nmin]。
2.3 x-1 value condition n of other last dilating windows2~nx, calculate the part under each window value condition
Optimal mode distance dist1iWith local best match position L1i=[nx-ni+ 1, nx]。
After 2.4 whole window value matchings finish, following two important results can be generated:
1 local mode distance vector
2 local optimum positions vector
Pattern distance minimum value is picked out from local mode distance vectorAnd it is corresponding
Best match positionAs exceptional sample, the eigenmatrix FT and fm of fault sample1iPattern-recognition it is final
As a result.
After three exceptional sample eigenmatrix FT and all eigenmatrixes of remaining fault sample match one by one, generate pattern away from
It measures descriscentAnd position vector
Step 5:Pattern distance is converted to Pattern similarity using drop ridge distribution;
It is essentially Euclidean distance that the pattern distance that upper step acquires, which is using the distance that DTW algorithms acquire, and Europe
It is relevant with sample time length that distance is obtained in several, and time span is bigger, and Euclidean distance also can correspondingly increase, and cause
We can not intuitively judge the similitude of two samples, therefore this step is acquired upper step using drop ridge distribution
Pattern distance be converted into Pattern similarity after could obtain final fault diagnosis result.
The obtained pattern distance set constituent of step 4 is abc pattern distancesPattern distance is according to drop ridge point
The principle of footwork is converted to similarity according to following formula
Wherein,It represents exceptional sample and carries out pattern-recognition calculating with j-th of fault sample under fault type i
Pattern distance, and ThiRepresent the pattern distance threshold value under fault type i.
In this way according to the dimension of fault type, by each distance vector
It is converted into similarity vector
The similarity set of exceptional sample FT and fault knowledge library whole sample can be finally obtained, as follows:
Step 6:Export final fault diagnosis result.
The similarity aggregated result ρ Sets that the present invention is obtained according to step 5, according to the rule of agreement export it is most effective, can
The fault diagnosis result leaned on:
Rule 1:If have with exceptional sample similarity be more than 90% fault sample and occur it is multiple, by whole samples
Serial number exports, while exporting the confidence level of the fault type typeVector belonging to fault sample, affiliated fault type
If confidenceLevel, fault progression stage are the best match position Location of maximum similarity sample, confidence level
ConfidenceLevel is more than 50%, then exports the maintenance measures repairAction that suggestion is taken;
Rule 2:If not being more than that 90% fault sample and maximum similarity numerical value are more than etc. with exceptional sample similarity
In 60%, then the serial number of maximum similarity sample is exported, while exporting the fault type belonging to the fault sample
TypeVector, the confidence level confidenceLevel of affiliated fault type, fault progression stage are maximum similarity sample
Best match position Location;If confidence level confidenceLevel is more than 50%, it is proposed that the maintenance measures taken
repairAction;
Rule 3:If not being more than 90% fault sample with exceptional sample similarity and maximum similarity numerical value is less than
60%, then will export the exceptional sample is uncommon operating mode or unknown failure type, it is proposed that continues to pay close attention to this abnormal development;
Wherein, typeVector numerical value is to meet above-mentioned ruleIn i values;Certain fault type
The fault sample that confidenceLevel numerical value is equal to the fault type number divided by the fault type that meet above-mentioned rule is total
Number;Maintenance measures repairAction can be obtained according to fault type by inquiring the operating standard of the equipment.
Using the heat pump fore pump of northern certain thermal power plant #2 units as status monitoring object, heat pump fore pump be steam-operating to
The important component of water pump, its function is important, maintains the performance safety of boiler plant.Since this equipment belongs to exposed facility,
Observation position is more and easily sends out failure various, this feature is suitble to the equipment fault diagnosis method that the present invention designs.By this implementation
Example elaborates, the implementation process further illustrated the present invention.
The embodiment of the present invention is as follows to the implementation steps of certain power plant's heat pump fore pump equipment fault diagnosis:
One, the fault diagnosis knowledge base building process of the preposition pumping unit of heat pump
Step 1:Heat pump fore pump fault sample information is obtained from power plant's PI databases;
It chooses and the relevant key parameter of the preposition pump operation of heat pump 21, including real hair power (MW), motor drive end diameter
Every observation data to a watt temperature (DEG C), thrust bearing shoe valve temperature (DEG C), current of electric (A) etc., therefore the equipment are 21 dimension row vectors:
u(tj)=[uj1, uj2, uj3..., uj21]
The failure logging of the equipment is carried out from October, 2013 in October, 2015 this period of history running state data
It searches, screening rule is set to failure mode P and is more than or equal to 2 more than or equal to 2 and failure frequency T, then according to screening rule
Following information is extracted from equipment fault record:This period occur altogether 3 kinds of failures (radial watt of motor drive end temperature jumps,
Thrust bearing shoe valve temperature bust, current of electric mutation), the number that breaks down is that 86 times (radial watt of motor drive end temperature jumps 25 times, thrust
Watt warm bust 28 times, current of electric are mutated 33 times), failure correlation measuring point is different, and (radial watt of motor drive end temperature jumps correlative measurement
Point is 6~measuring point of measuring point 14, thrust bearing shoe valve temperature bust correlation measuring point is 15~measuring point of measuring point 17, the related measuring point of current of electric mutation is
1~measuring point of measuring point 6 and 18~measuring point of measuring point 21), trouble duration ranging from 2 hours~14 hours.;Finally according to above-mentioned letter
Breath reads whole fault sample data from power plant real-time data base PI.
The fault sample finally obtained include fault type number information (radial watt of motor drive end temperature jump number be 1,
Thrust bearing shoe valve temperature bust number is 2, current of electric mutation number be 3), fault sample data, fault type correlation measuring point information three
Partial content, wherein the sequence per class fault sample is ranked up according to the priority of time of failure:
Wherein,iFjConcrete form be matrix form, row m representative sample length, row represent measuring point.As follows:
Step 2:Sectional linear fitting is carried out to each fault sample in Fault Sets;
From sample1F1Start until sample3F33, each fault sample is using sectional linear fitting technology, and realization is according to event
Hinder the state segmentation that sample overall development trend carries out adaptivity, it is accurate to carry out data prediction for the extraction of subsequent fault signature
It is standby.
First, mean filter operation is carried out.It extracts successivelyiFjEach measuring point column vector u-i, according in specific embodiment
Mean filter principle operating procedure is filtered elimination to the noise pollution being entrained in sample data, obtains filtering dataiFj *。
Finally by each sample of heat pump fore pumpiFjAll it is processed into
Step 3:Feature extraction is carried out per segment data to each fault sample after segmentation;
Since sample data duration is in hour rank, quickly and effectively run for ease of fault diagnosis of the present invention, it will pair event
Hinder every segment data extraction slope k of sample, duration l, mean value m, the progress data compression of the mathematical features such as variance v and feature space
Conversion.
After feature extractioniFjEvery segment dataIt is converted into four-dimensional element (k, l, m, v).Specific feature extraction is public
The modelling phase step 3 of the visible specific embodiment of formula.Fault sample each so just realizes sample and goes to spy by time domain space
Levy the process in space:
Step 4:Fault signature conversion is carried out, feature dimension is eliminated;
In order to prevent when calculating pattern distance, different faults sample occur causes to judge by accident due to dimension, needs
Convert slope, duration, mean value these three features to inclination angle, time scale, average ratio.Eliminate the concrete operations of dimension
The visible modelling phase step 4 feature conversion formula of method.By conversion, Feature Conversion is following Dimensionless Form:
So far, whole samples just realizes conversion of the time domain space to feature space, fault sample set Fault
Sets's is converted into fault sample characteristic set:
Step 5:Calculate the pattern distance threshold value under each fault type;
By above 4 step, all fault sample data are completely converted into eigenmatrix form aggregate form.This step
It needs using the pattern distance of different samples, foundation are counted two-by-two in the similar fault sample of dynamic time warping method (DTW) calculating
Calculate the pattern distance threshold value that result determines every class failure.
For example, having 25 sample characteristics matrixes in the 1st class failure in fault sample, eigenmatrix calculates DTW two-by-two
Distance obtains a vector DtwVector1=[Md1, Md2..., Md25].The pattern distance threshold value Th of fault type 11=min
(DtwVector1).The pattern distance threshold value Th of fault type 2,3 can be obtained in the same fashion2、Th3。
Step 6:Generate heat pump makings failure of pump pattern repository;
Summarize the information obtained by above-mentioned steps, should include following 3 partial information in heat pump fore pump fault knowledge library:
1. the eigenmatrix set FaultFeatureSets that all fault samples extract.
2. the correlative measurement point name set PT={ pt of three classes failure1, pt2, pt3};
3. the corresponding pattern distance threshold value Thsets=[Th of three classes failure1, Th2..., Thx]。
It integrates above 3 partial informations and generates complete heat pump fore pump fault knowledge library, form is as follows:
Two, the on-line fault diagnosis process of the preposition pumping unit of heat pump
December in 2014, the preposition pumping unit measuring point of level of factory monitoring system discovery heat pump of 15 03Shi Gai power plant occurred
Unknown exception.In order to more preferably help professional carry out inspection and maintenance, plant personnel according to known exception beginning and ending time with
And relative alarm measuring point information pt transfers exceptional sample data from real-time data baseAnd with the fault diagnosis side of the present invention
Method is prejudged in advance.
Exceptional sample dataBy data processings behaviour such as sectional linear fitting, off-note extraction and the eliminations of feature dimension
After work, calculate with exceptional sample dataCorresponding eigenmatrixMatrix.The above concrete operations can be found in specific implementation
The correspondence step in example model running stage.
Exceptional sample informationAnd pt will match the sample information in fault mode knowledge base one by one, and will
Pattern distance is converted into corresponding failure similarity.
Exceptional sample is matched with whole fault samples in knowledge base, and the present invention is calculated using sliding extension time window method
DTW distances are as final pattern distance.It is as follows that distance set form can be obtained:
Between distance is mapped to [0,1] numerical value by operation drop ridge distribution, pattern distance converted by unified standard similar
Degree.The drop ridge distribution in specific embodiment model running stage is with herein.It is as follows that similarity aggregate form can be obtained:
The judgement result of fault diagnosis is exported according to the Failure Diagnostic Code of agreement:It is this time abnormal to be determined as motor drive in advance
Radial watt of temperature is held to jump failure, judgement reason is:With exceptional sample similarity it is more than 90% failure sample in heat pump fore pump
Originally have 15 (Deng), be diagnosed as the confidence level of fault type 1 is more than for 60%
Confidence threshold value, similarity maximum fault sample are that segmentation hop count is 7Best match position is 1~4 section, physical fault
Diagnosis effect figure is attached drawing 4.Two sample overall trends are consistent from starting to centre position to match, therefore by the fault progression stage
It is determined as fault progression mid-term, it is cold water shock motor drive end forced cooling side that maintenance measures are taken in inquiry operating standard suggestion
Method
It turns out that diagnostic result after waiting professionals to rush in time is radial watt of motor drive end temperature really jump therefore
Barrier, after forcing heat pump fore pump to cool to reasonable temperature section using cold water falling temperature method, heat pump fore pump equipment operation restores just
Often.
Although for illustrative purposes, it has been described that exemplary embodiments of the present invention, those skilled in the art
Member it will be understood that, can be in form and details in the case of not departing from the scope and spirit invented disclosed in appended claims
On the change that carry out various modifications, add and replace etc., and all these changes should all belong to appended claims of the present invention
Protection domain, and each step in the claimed each department of product and method, can be in any combination
Form is combined.Therefore, to disclosed in this invention the description of embodiment be not intended to limit the scope of the invention,
But for describing the present invention.Correspondingly, the scope of the present invention is not limited by embodiment of above, but by claim or
Its equivalent is defined.
Claims (1)
1. a kind of equipment fault diagnosis method, which is characterized in that in turn include the following steps:
Step 1:Fault diagnosis training step:Fault sample information is obtained from database, is respectively obtained after calculation processing
The eigenmatrix and pattern distance threshold value of the fault sample of dimension are eliminated, the two is associated with each other, and storage generates failure mould
Formula knowledge base;
Step 2:Fault diagnosis operating procedure:Exceptional sample information is obtained from real-time data base, and exception is obtained after calculation processing
Characteristic information is converted to Pattern similarity after then calculating pattern distance successively to the sample characteristics in fault mode knowledge base,
Export final fault diagnosis result;
Wherein, the step 1 the specific steps are:
(1.1) fault sample information is obtained from database;
(1.2) sectional linear fitting is carried out to each fault sample successively;
(1.3) feature extraction is carried out to every segment data of fault sample, obtains the eigenmatrix of fault sample;
(1.4) fault signature conversion is carried out, feature dimension, the eigenmatrix of the fault sample for the dimension that is eliminated are eliminated;
(1.5) pattern distance threshold value is calculated;
(1.6) fault signature matrix and pattern distance threshold value is associated with each other, storage generates fault mode knowledge base;
Wherein, the step (1.1) the specific steps are:Select fault type number P >=2 and each failure frequency T
>=2 meet the requirements can research equipment, select the observation point N of enough numbers, wherein N >=10 are prolonged enough to equipment
History run status data carries out failure logging lookup, is plucked from failure logging using the screening rule of setting and selects failure correlation
The useful information of measuring point information, the beginning and ending time of failure process and breakdown maintenance measure record, according to useful information from power plant
Read failure sample data in real-time data base PI,
Wherein:Measuring point number is n, the time counts out for m fault sample data, regard a n as in whole measuring point datas at j moment
The column vector of dimension, is expressed as:
u(tj)=[uj1,uj2,uj3,...,ujn]
The sample data saves as the matrix form of m × n, and concrete form is as follows:
Wherein row represents m fault time, and row represent n equipment observation point, and two value of ranks m, n between each fault sample
It is not quite similar, while assigning its fault type for each fault sample and identifying ID, if fault type mark ID determines that method is complete
Include X kind failures in portion's sample, then the numberical range of fault type mark ID is:1-X;
Wherein, the step (1.2) the specific steps are:
(1.2.1) mean filter operates:Elimination is filtered to the noise pollution being entrained in sample data;
The segmentation initialization of (1.2.2) fault sample:Segmentation initialization is carried out to the fault sample after being filtered;
(1.2.3) is by initialization data section joint account error of fitting two-by-two;
(1.2.4) determines the segmentation cut point of fault sample, and the state that adaptivity is carried out to fault sample is divided.
Priority Applications (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN201511004767.2A CN105467975B (en) | 2015-12-29 | 2015-12-29 | A kind of equipment fault diagnosis method |
Applications Claiming Priority (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN201511004767.2A CN105467975B (en) | 2015-12-29 | 2015-12-29 | A kind of equipment fault diagnosis method |
Publications (2)
Publication Number | Publication Date |
---|---|
CN105467975A CN105467975A (en) | 2016-04-06 |
CN105467975B true CN105467975B (en) | 2018-08-03 |
Family
ID=55605784
Family Applications (1)
Application Number | Title | Priority Date | Filing Date |
---|---|---|---|
CN201511004767.2A Active CN105467975B (en) | 2015-12-29 | 2015-12-29 | A kind of equipment fault diagnosis method |
Country Status (1)
Country | Link |
---|---|
CN (1) | CN105467975B (en) |
Cited By (1)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
RU2803903C1 (en) * | 2022-12-19 | 2023-09-21 | федеральное государственное бюджетное образовательное учреждение высшего образования "Уфимский университет науки и технологий" | Method for operational control of parameters and diagnosis of the technical condition of small unmanned aerial vehicles (uav) |
Families Citing this family (34)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN108151795B (en) * | 2016-12-02 | 2021-02-26 | Abb瑞士股份有限公司 | Method and system for configuring a condition monitoring device |
CN106697291A (en) * | 2016-12-22 | 2017-05-24 | 深圳万智联合科技有限公司 | Intelligent unmanned plane pesticide spraying device |
CN106527407B (en) * | 2016-12-22 | 2018-11-27 | 泰兴市东城水处理工程有限公司 | A kind of comprehensive waste treatment system |
CN106874896B (en) * | 2017-03-31 | 2021-03-02 | 阳江核电有限公司 | Auxiliary learning method and system for signal feature identification of nuclear power station primary loop component loosening diagnosis system |
EP3407273A1 (en) * | 2017-05-22 | 2018-11-28 | Siemens Aktiengesellschaft | Method and system for determining an abnormal state of a system |
JP6772963B2 (en) * | 2017-06-05 | 2020-10-21 | トヨタ自動車株式会社 | Abnormality diagnosis device and abnormality diagnosis method |
CN107065839B (en) * | 2017-06-06 | 2019-09-27 | 苏州大学 | A kind of method for diagnosing faults and device based on diversity recursion elimination feature |
CN107402064B (en) * | 2017-07-25 | 2020-04-03 | 上海控创信息技术股份有限公司 | Noise detection method and system |
CN113408899A (en) * | 2017-12-05 | 2021-09-17 | 北京绪水互联科技有限公司 | Detection method, early warning method and early warning system for early warning points of residual emergency maintenance time of equipment |
CN108241894A (en) * | 2018-01-10 | 2018-07-03 | 湖南大唐先科技有限公司 | Fault Locating Method, equipment and storage medium |
CN108764341B (en) * | 2018-05-29 | 2019-07-19 | 中国矿业大学 | A kind of Fault Diagnosis of Roller Bearings under the conditions of variable working condition |
CN109239265B (en) | 2018-09-11 | 2021-01-22 | 清华大学合肥公共安全研究院 | Fault detection method and device for monitoring equipment |
CN109445406B (en) * | 2018-10-18 | 2021-05-18 | 西南交通大学 | Industrial control system safety detection method based on scene test and transaction search |
CN109855873B (en) * | 2018-12-12 | 2021-02-05 | 华润电力技术研究院有限公司 | Fault diagnosis method and device for turbine main shaft |
CN109631848B (en) * | 2018-12-14 | 2021-04-16 | 山东鲁能软件技术有限公司 | Transmission line foreign matter intrusion detection system and detection method |
CN111600735B (en) * | 2019-02-21 | 2021-08-03 | 烽火通信科技股份有限公司 | Sample data processing method, system and device |
CN109992600B (en) * | 2019-03-28 | 2021-09-07 | 佛山市百斯特电器科技有限公司 | Response method and device for equipment fault |
CN110275879A (en) * | 2019-05-16 | 2019-09-24 | 浙江浙能技术研究院有限公司 | A method of Trouble Match and early warning are carried out based on fault data state matrix |
TWI687783B (en) * | 2019-06-17 | 2020-03-11 | 臺灣塑膠工業股份有限公司 | Device abnormality detection method and system thereof |
CN110286330A (en) * | 2019-08-15 | 2019-09-27 | 莆田市烛火信息技术有限公司 | One kind being used for lithium battery fault detection system |
CN110501127B9 (en) * | 2019-08-28 | 2021-04-09 | 湘潭大学 | Equal-section beam damage identification method based on damage state inclination slope |
CN110608885B (en) * | 2019-09-09 | 2021-10-29 | 天津工业大学 | Method for diagnosing wear fault and predicting trend of inner ring of rolling bearing |
CN110470465B (en) * | 2019-09-19 | 2021-09-10 | 武汉市华英电力科技有限公司 | Circuit breaker testing method and system based on vibration signal analysis |
CN110672950A (en) * | 2019-10-08 | 2020-01-10 | 深圳海岸语音技术有限公司 | Power equipment fault sound image detection system and method |
CN110955226B (en) * | 2019-11-22 | 2021-04-13 | 深圳市通用互联科技有限责任公司 | Equipment failure prediction method and device, computer equipment and storage medium |
CN111222774B (en) * | 2019-12-30 | 2020-08-18 | 广州博士信息技术研究院有限公司 | Enterprise data analysis method and device and server |
CN111324637B (en) * | 2020-02-05 | 2021-01-15 | 北京工业大数据创新中心有限公司 | Fault symptom searching method and system for industrial time sequence data |
CN111896246A (en) * | 2020-07-29 | 2020-11-06 | 北京天地龙跃科技有限公司 | Health management verifies evaluation system |
CN112415331B (en) * | 2020-10-27 | 2024-04-09 | 中国南方电网有限责任公司 | Power grid secondary system fault diagnosis method based on multi-source fault information |
CN114325188A (en) * | 2021-12-28 | 2022-04-12 | 国网河北省电力有限公司经济技术研究院 | Fault detection method and device and server |
CN114444623B (en) * | 2022-04-11 | 2022-08-12 | 智昌科技集团股份有限公司 | Industrial robot-oriented anomaly detection and analysis method and system |
CN115956543A (en) * | 2022-07-15 | 2023-04-14 | 生态环境部南京环境科学研究所 | Fishing device and fishing method for wild giant salamander resource investigation |
CN116608549B (en) * | 2023-07-20 | 2023-10-03 | 上海东方低碳科技产业股份有限公司 | Intelligent detection method for energy-saving efficiency of heating ventilation air conditioner |
CN117389967B (en) * | 2023-12-13 | 2024-05-14 | 国网江西省电力有限公司电力科学研究院 | Multi-dimensional fault information extraction processing and diagnosis method and system for power transformation equipment |
Citations (5)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN101634605A (en) * | 2009-04-10 | 2010-01-27 | 北京工业大学 | Intelligent gearbox fault diagnosis method based on mixed inference and neural network |
CN102566555A (en) * | 2012-02-10 | 2012-07-11 | 安徽建筑工业学院 | Major appliance work state monitoring method based on pattern recognition |
CN103745229A (en) * | 2013-12-31 | 2014-04-23 | 北京泰乐德信息技术有限公司 | Method and system of fault diagnosis of rail transit based on SVM (Support Vector Machine) |
CN103822786A (en) * | 2012-11-16 | 2014-05-28 | 中国水利电力物资有限公司 | Wind turbine generator set mechanical equipment state diagnosis method based on multivariate statistical analysis |
CN104122086A (en) * | 2014-08-11 | 2014-10-29 | 长沙理工大学 | Kriging model based wind turbine gearbox fault diagnosis method |
-
2015
- 2015-12-29 CN CN201511004767.2A patent/CN105467975B/en active Active
Patent Citations (5)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN101634605A (en) * | 2009-04-10 | 2010-01-27 | 北京工业大学 | Intelligent gearbox fault diagnosis method based on mixed inference and neural network |
CN102566555A (en) * | 2012-02-10 | 2012-07-11 | 安徽建筑工业学院 | Major appliance work state monitoring method based on pattern recognition |
CN103822786A (en) * | 2012-11-16 | 2014-05-28 | 中国水利电力物资有限公司 | Wind turbine generator set mechanical equipment state diagnosis method based on multivariate statistical analysis |
CN103745229A (en) * | 2013-12-31 | 2014-04-23 | 北京泰乐德信息技术有限公司 | Method and system of fault diagnosis of rail transit based on SVM (Support Vector Machine) |
CN104122086A (en) * | 2014-08-11 | 2014-10-29 | 长沙理工大学 | Kriging model based wind turbine gearbox fault diagnosis method |
Non-Patent Citations (1)
Title |
---|
三相逆变器功率管故障智能诊断研究;刘艳等;《电测与仪表》;20130131;第50卷(第565期);第15-20页 * |
Cited By (1)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
RU2803903C1 (en) * | 2022-12-19 | 2023-09-21 | федеральное государственное бюджетное образовательное учреждение высшего образования "Уфимский университет науки и технологий" | Method for operational control of parameters and diagnosis of the technical condition of small unmanned aerial vehicles (uav) |
Also Published As
Publication number | Publication date |
---|---|
CN105467975A (en) | 2016-04-06 |
Similar Documents
Publication | Publication Date | Title |
---|---|---|
CN105467975B (en) | A kind of equipment fault diagnosis method | |
CN105548764B (en) | A kind of Fault Diagnosis for Electrical Equipment method | |
CN105574284B (en) | A kind of Fault Diagnosis for Electrical Equipment method based on trend feature point | |
Capozzoli et al. | Fault detection analysis using data mining techniques for a cluster of smart office buildings | |
CN115578015B (en) | Sewage treatment whole process supervision method, system and storage medium based on Internet of things | |
CN105631596B (en) | Equipment fault diagnosis method based on multi-dimensional piecewise fitting | |
JP5808605B2 (en) | Abnormality detection / diagnosis method and abnormality detection / diagnosis system | |
CN108985380B (en) | Point switch fault identification method based on cluster integration | |
CN111061620A (en) | Intelligent detection method and detection system for server abnormity of mixed strategy | |
CN110990461A (en) | Big data analysis model algorithm model selection method and device, electronic equipment and medium | |
WO2011043108A1 (en) | Equipment status monitoring method, monitoring system, and monitoring program | |
CN109829236A (en) | A kind of Compressor Fault Diagnosis method based on XGBoost feature extraction | |
CN111412579B (en) | Air conditioning unit fault type diagnosis method and system based on big data | |
CN106779200A (en) | Based on the Wind turbines trend prediction method for carrying out similarity in the historical data | |
CN113762329A (en) | Method and system for constructing state prediction model of large rolling mill | |
CN112464439A (en) | Three-layer data mining-based power system fault type analysis method and device | |
CN114881101B (en) | Bionic search-based power system typical scene association feature selection method | |
CN112906288A (en) | Switch cabinet state evaluation method based on normal sample data | |
CN113962308A (en) | Aviation equipment fault prediction method | |
CN110956281A (en) | Power equipment abnormity detection alarm system based on Log analysis | |
CN113505639B (en) | Rotary machine multi-parameter health state assessment method based on TPE-XGBoost | |
CN114239902A (en) | User track prediction and abnormal track detection method and system | |
CN117277566B (en) | Power grid data analysis power dispatching system and method based on big data | |
Ye et al. | A novel self-supervised learning-based anomalous node detection method based on an autoencoder for wireless sensor networks | |
CN116308883A (en) | Regional power grid data overall management system based on big data |
Legal Events
Date | Code | Title | Description |
---|---|---|---|
C06 | Publication | ||
PB01 | Publication | ||
C10 | Entry into substantive examination | ||
SE01 | Entry into force of request for substantive examination | ||
GR01 | Patent grant | ||
GR01 | Patent grant | ||
CP01 | Change in the name or title of a patent holder |
Address after: 250101 5th floor, block B, Yinhe building, 2008 Xinluo street, high tech Zone, Jinan City, Shandong Province Patentee after: Shandong luruan Digital Technology Co.,Ltd. Address before: 250101 5th floor, block B, Yinhe building, 2008 Xinluo street, high tech Zone, Jinan City, Shandong Province Patentee before: SHANDONG LUNENG SOFTWARE TECHNOLOGY Co.,Ltd. |
|
CP01 | Change in the name or title of a patent holder |