CN110163075A - A kind of multi-information fusion method for diagnosing faults based on Weight Training - Google Patents
A kind of multi-information fusion method for diagnosing faults based on Weight Training Download PDFInfo
- Publication number
- CN110163075A CN110163075A CN201910156004.1A CN201910156004A CN110163075A CN 110163075 A CN110163075 A CN 110163075A CN 201910156004 A CN201910156004 A CN 201910156004A CN 110163075 A CN110163075 A CN 110163075A
- Authority
- CN
- China
- Prior art keywords
- weight
- sensor
- signal
- fault
- training
- 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.)
- Pending
Links
Classifications
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F18/00—Pattern recognition
- G06F18/20—Analysing
- G06F18/21—Design or setup of recognition systems or techniques; Extraction of features in feature space; Blind source separation
- G06F18/214—Generating training patterns; Bootstrap methods, e.g. bagging or boosting
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F18/00—Pattern recognition
- G06F18/20—Analysing
- G06F18/25—Fusion techniques
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N3/00—Computing arrangements based on biological models
- G06N3/02—Neural networks
- G06N3/08—Learning methods
- G06N3/084—Backpropagation, e.g. using gradient descent
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F2218/00—Aspects of pattern recognition specially adapted for signal processing
- G06F2218/08—Feature extraction
Landscapes
- Engineering & Computer Science (AREA)
- Theoretical Computer Science (AREA)
- Data Mining & Analysis (AREA)
- Physics & Mathematics (AREA)
- Artificial Intelligence (AREA)
- General Engineering & Computer Science (AREA)
- Life Sciences & Earth Sciences (AREA)
- Evolutionary Computation (AREA)
- General Physics & Mathematics (AREA)
- Computer Vision & Pattern Recognition (AREA)
- Bioinformatics & Computational Biology (AREA)
- Bioinformatics & Cheminformatics (AREA)
- Evolutionary Biology (AREA)
- Biophysics (AREA)
- Computing Systems (AREA)
- Mathematical Physics (AREA)
- Software Systems (AREA)
- Molecular Biology (AREA)
- General Health & Medical Sciences (AREA)
- Computational Linguistics (AREA)
- Biomedical Technology (AREA)
- Health & Medical Sciences (AREA)
- Testing And Monitoring For Control Systems (AREA)
- Testing Or Calibration Of Command Recording Devices (AREA)
Abstract
The invention discloses a kind of multi-information fusion method for diagnosing faults based on Weight Training, including collection step off-line sensors signal and corresponding states information;Carry out WAVELET PACKET DECOMPOSITION processing;Calculate each sensor signal fault feature vector;Weight Training is carried out using BP neural network Weight Training module;Obtain each sensor weight;The various types of signal of measuring device in operation;Carry out WAVELET PACKET DECOMPOSITION processing;Calculate each sensor signal fault feature vector;Calculate each sensor fault degree of membership;Failure degree of membership obtained by each sensor is obtained into weighted evidence multiplied by weight;Data fusion is carried out to weighted evidence, obtains fault diagnosis result.The present invention realizes the acquisition of the off-line training and weighted evidence of multi-information fusion weight using BP neural network, improves the precision and validity of sensor signal mapped device state, effectively increases the accuracy of system diagnostics.
Description
Technical field
The invention belongs to on-line monitoring and fault diagonosing technical fields, and in particular to it is a kind of based on neural network weight training
Multi-information fusion method for diagnosing faults.
Background technique
In order to guarantee that important equipment is safely and reliably run, on the one hand can be realized by improving the reliability of equipment,
On the other hand scientific and reasonable maintenance can be arranged equipment.The maintenance mode of equipment is broadly divided into three kinds at present, i.e., subsequent inspection
It repairs, scheduled overhaul and repair based on condition of component.According to traditional scheduled overhaul method, since service personnel can not understand the reality of equipment in time
Border operating status, operation blindness is strong, be easy to cause maintenance superfluous or maintenance deficiency.In addition, equipment rebuilding needs to disintegrate, the time
Cost and economic cost are high, while disintegrating and ressembling and being likely to result in new defect, reduce the reliability of equipment instead.And
The information characteristics of equipment under operation are obtained by the method for repair based on condition of component, determine whether equipment is sent out by analyzing relatively
Raw failure or defect, break down or the component locations of defect, have very strong real-time and specific aim.Therefore, for equipment
On-line monitoring and fault diagnosis have become essential link.
During fault diagnosis, information fusion can be applied to the place of the processing of original data layer, feature abstraction layer
Reason, decision-making level each stratum such as processing.Correspondingly, applying different mathematical algorithms during different levels fusion treatment
To solve the problems, such as to encounter in fusion process.The influence of the problems such as due to sensor self performance, external environment interference, so that passing
The information of sensor acquisition has uncertainty.Using multisensor carry out information fusion can by the unascertained information of acquisition into
Row is complementary, reasonably makes inferences decision.
There is no consider that sensor is credible when carrying out fault diagnosis in pervious multiple sensor information amalgamation method
Degree, i.e., the weight that each sensor occupies in failure diagnostic process.And D-S evidence theory requires sternly the independence of evidence
Lattice, and can generate antinomy when evidence generates when conflicting information uses composition rule and lead to Fusion failure.
Summary of the invention
Goal of the invention: for overcome the deficiencies in the prior art, the present invention provides a kind of precise and high efficiency, can effectively improve equipment
The multi-information fusion method for diagnosing faults based on weight off-line training of reliability.
Technical solution: a kind of multi-information fusion method for diagnosing faults based on Weight Training, including weight off-line training and
Monitor two parts on-line;The specific steps of the diagnostic method include:
Weight off-line training part includes:
(1) off-line sensors signal and corresponding states information are collected;
(2) WAVELET PACKET DECOMPOSITION processing is carried out;
(3) each sensor signal fault feature vector is calculated;
(4) Weight Training is carried out using BP neural network Weight Training module;
(5) each sensor weight is obtained;
On-line monitoring part includes:
(6) sensor, the various types of signal of measuring device in operation are installed on monitored equipment;
(7) WAVELET PACKET DECOMPOSITION processing is carried out;The letter that each sensor is collected respectively using WAVELET PACKET DECOMPOSITION theory
Number carry out WAVELET PACKET DECOMPOSITION;The feature for the signal that specific WAVELET PACKET DECOMPOSITION function is collected by sensor in actual use Lai
It determines;
(8) each sensor signal fault feature vector is calculated;The signal of each sensor is calculated after WAVELET PACKET DECOMPOSITION
The energy value of each reproducing sequence is normalized the energy of extracted each reproducing sequence for the data for reacting fault message
Processing, obtains fault feature vector;
(9) each sensor fault degree of membership is calculated;Calculate the fault feature vector of the signal of collected each sensor
The Euclidean distance of the feature vector of existing typical fault status signal corresponding with Mishap Database, distance metric is bigger, table
Bright degree of membership between the two is weaker;Calculate the fault feature vector presence corresponding with Mishap Database of each sensor signal
Each typical fault status signal feature vector Euclidean distance normalization reciprocal as a result, obtaining each sensor pair
The failure degree of membership answered, i.e. evidence;
(10) failure degree of membership obtained by each sensor is obtained into weighted evidence multiplied by weight;
(11) data fusion is carried out to weighted evidence using D-S evidence theory rule of combination, obtains fault diagnosis result.
Preferably, WAVELET PACKET DECOMPOSITION WPD frequency division when full frequency band is all with higher described in step (2), step (7)
Resolution has finer localization performance.Frequency band is carried out to multi-level division, the height not segmented to multiresolution analysis
Frequency part is further decomposed, and according to the feature of analyzed signal, adaptive selection frequency band is allowed to and signal spectrum phase
Matching, to improve time frequency resolution.(A indicates that low frequency component, D indicate high frequency to three layers of WAVELET PACKET DECOMPOSITION exploded relationship such as following formula
Component, end serial number indicate the number of plies of WAVELET PACKET DECOMPOSITION):
S=AAA3+DAA3+ADA3+DDA3+AAD3+DAD3+ADD3+DDD3。
Preferably, the specific calculation of the step (3), step (8) fault feature vector are as follows:
If Ej *For the energy of wavelet reconstruction sequence, it is expressed as
In formula:For k-th of component of j-th of wavelet package reconstruction sequence;N is sequence djThe number of middle component;N is small echo
Packet Decomposition order.If the energy value to each frequency range that acquisition signal carries out n-layer wavelet decomposition is
Return place one to change reason feature vector can be obtained the vector being
In formula:
The corresponding one group of fault feature vector of the signal that each sensor collects.
Preferably, the training process of step (4) the BP neural network Weight Training module are as follows:
1. collecting off-line sensors signal and corresponding states information;
2. carrying out WAVELET PACKET DECOMPOSITION processing;
3. calculating each sensor signal fault feature vector;
4. carrying out data prediction;The equipment operation sensor signal and corresponding equipment state that each sensor acquires offline
For priori data, output needed for data prediction obtains trained BP neural network is carried out for the priori data of each sensor,
Classify with data weight corresponding to feature vector and exports;
5. carrying out neural metwork training and parameter optimization;Using back-propagation algorithm training neural network, selected
The random initializtion neural network number of plies between the network number of plies, every layer of neuron number, each neuron weight and every layer of biasing,
By the neural network after feature vector input initialization, the weight classification O of network output is obtainedm, calculate output and data weight
Cross entropy C between classification O, calculation are as follows:
Given threshold, when intersecting entropy lower than threshold value, neural network is met the requirements, and saves network, is unsatisfactory for, and is used
Stochastic gradient algorithm update, neuron weight, every layer biasing, if the number of iterations cannot still be converged in 100 intersection entropy it is low
When threshold value, updates the neural network number of plies and every layer of neuron number and initialize neuron weight and every layer of biasing, sentence
It is disconnected whether to select in section, if then resetting the number of iterations, if not then termination algorithm, completes neural metwork training.Finally
It is chosen from the neural network of preservation and intersects the smallest neural network of entropy as selection neural network.
6. obtaining BP neural network Weight Training module.
Preferably, the 4. carry out data prediction, method particularly includes: the failure degree of membership of sensor is set as m
(a)=(a1,a2,…,ak) and sensor be w (x)=(x to the weight of state in fault database1,x2,…,xk);Degree of membership highest
Corresponding states be exactly the sensor diagnostic result, degree of membership it is higher explanation it is higher to the certainty of the state.According to each
The degree of membership result of sensor handles each sensing data, and for any sensor signal and label is any state
The feature vector of i (i-th kind of state of k state in fault database), lays down a regulation as follows:
(i) number of states is k, the failure degree of membership of sensor are as follows: m in fault database1(a)=(a1,a2,…,ak), sensing
Device is to state weight in fault database are as follows: w1(x)=(x1,x2,…,xk).Weight output area [0,10] is divided into p class: [0,
10/p), [2,2+10/p) ..., [10-10/p, 10] }, corresponding classification sequence number 1,2 ..., p is divided when one group of data corresponds to weight
When for class l, then the output of corresponding training BP neural network is O=(01,02,…,1l,…,0p-1,0p), other weight classification are defeated
Out and so on;
(ii) data label is state i, for sensor:
Diagnostic result is consistent with data label:
If (a) ai>=0.7, then xiFor class p, if a1||a2||…||ai-1||ai+1||…||ak< 0.1, then by weight x1||
x2||…||xi-1||xi+1||…||xkIt is set as class p, otherwise by weight x1||x2||…||xi-1||xi+1||…||xkIt is set as classification
(1/a1)||(1/a2)||…||(1/ai-1)||(1/ai+1)||…||(1/ak);
If (b) ai< 0.7, then weight xiFor classification 10aiIf a1||a2||…||ai-1||ai+1||…||ak< 0.1 is right by its
Weight is answered to be divided into class p, if a1||a2||…||ai-1||ai+1||…||ak> 0.4, then its weight is divided into class 1, otherwise by weight
It is set as classification (1/a1)||(1/a2)||…||(1/ai-1)||(1/ai+1)||…||(1/ak);
(iii) this group of data diagnosis result and label are inconsistent, then by weight xiIt is divided into class 1, if max (a1,…,ak)=
am, then by the corresponding weight x of maximum valuemIt is divided into class 1, if while a1||a2||…||ai-1||ai+1||…||ak< 0.1, its is right
Answer weight x1||x2||…||xi-1||xi+1||…||xkIt is divided into class p, if a1||a2||…||ai-1||ai+1||…||ak> 0.4, it will
Corresponding weight x1||x2||…||xi-1||xi+1||…||xkIt is divided into class 1, if being all unsatisfactory for weight x otherwise1||x2||…||xi-1
||xi+1||…||xkIt is set as classification (1/a1)||(1/a2)||…||(1/ai-1)||(1/ai+1)||…||(1/ak)。
Above-mentioned rule is arranged in data when to acquisition from all the sensors, is pre-processed, and obtains training BP neural network
Required data weight classification O (t)=(t1,t2,…,tp)。
Preferably, the specific calculation of step (9) the failure degree of membership is as follows:
In formula: E is each frequency energy normalization of WAVELET PACKET DECOMPOSITION of collected signal as a result, i.e. collected signal
Feature vector;EfFor each frequency range of the WAVELET PACKET DECOMPOSITION of typical fault status signal or normal state signal in fault signature database
Energy normalized as a result, i.e. typical fault or normal condition feature vector;N is the WAVELET PACKET DECOMPOSITION number of plies.
Preferably, Euclidean distance is in inverse relation between failure degree of membership and signal fault feature in step (9), failure is subordinate to
Category degree is assigned following formula:
In formula:
∑1/dk=1/d1+1/d2+...+1/dk
Wherein dkFor the Euclidean distance of signal characteristic in collected signal fault feature and fault database;K is in fault database
The number of fault type;M (a) is failure degree of membership.
Preferably, the specific calculation of step (10) weighted evidence are as follows:
In conjunction with the weight that offline Weight Training obtains, the calculation of weighted evidence are as follows:
M (a)=w1×m1(a)+w2×m2(a)+...+wn×mn(a)
In formula: m (a) is the weighted evidence obtained;wiFor the weight of each sensor required by front;miIt (a) is each biography
The failure degree of membership that sensor obtains, as original evidence;N is number of sensors.
The utility model has the advantages that a kind of multi-information fusion method for diagnosing faults based on Weight Training of the invention, has following excellent
Point:
1. the uncertain information of acquisition can be carried out to complementary, standard present invention employs multi-sensor information fusion technology
Decision reasonably really is made inferences to information.
2. the present invention establishes the relationship for acquiring each sensor signal and each sensor weight using BP neural network, offline to instruct
Practice each sensor weight, using weights are weighted failure degree of membership, effectively improve the accuracy of system diagnostics.
3. the method that the present invention uses on-line condition monitoring obtains the information characteristics of equipment under operation, has very
Strong real-time and specific aim can accurately carry out the fault diagnosis of monitored equipment, shorten maintenance and search the time, improve dimension
Repair efficiency.
4. diagnostic result precise and high efficiency of the present invention is a kind of fault diagnosis side that can effectively improve equipment safety and reliability
Method.
Detailed description of the invention
Fig. 1 is the multi-information fusion method for diagnosing faults structural block diagram based on Weight Training in the present invention;
Fig. 2 is the training flow chart of BP neural network Weight Training module in the present invention;
Fig. 3 is data prediction flow chart in the present invention;
Fig. 4 is neural metwork training and parameter optimization flow chart in the present invention;
Fig. 5 is fault diagnosis system schematic diagram in the present invention.
Specific embodiment
Technical solution of the present invention is described further with reference to the accompanying drawing.
As shown in Figure 1, a kind of multi-information fusion method for diagnosing faults based on Weight Training, including weight off-line training and
Monitor two parts on-line.The weight off-line training includes the signal and corresponding failure letter of offline n sensor of acquisition
Sensor signal is passed through WAVELET PACKET DECOMPOSITION and then calculates feature vector, by feature vector and its corresponding fault message by breath
BP neural network Weight Training module, the exportable each sensor power of BP neural network Weight Training module are input to as input quantity
Value, i.e. realization weight off-line training.The on-line monitoring is multiple sensors by being mounted on monitored equipment, to setting
It is standby to be monitored on-line, wavelet packet decomposition algorithm processing is carried out to each sensor signal obtained in real time and calculate feature to
It measures, typicalness feature vector in combination failure database, the failure for calculating each sensor for each typicalness of equipment is subordinate to
Failure degree of membership is arrived the weighted evidence of each sensor by degree, i.e. evidence multiplied by the output of weight off-line training.Based on adding
Warrant completes the fusion of multi-sensor information according to application D-S evidence theory, finally obtains the result of decision, as monitored equipment event
Hinder the status information of diagnostic result and output.
As shown in Fig. 2, the training process of the BP neural network Weight Training module, comprising:
(1) off-line sensors signal and corresponding states information are collected;
(2) WAVELET PACKET DECOMPOSITION processing is carried out.WAVELET PACKET DECOMPOSITION (WPD) in full frequency band time frequency resolution all with higher,
With finer localization performance.Frequency band is carried out to multi-level division, the radio-frequency head not segmented to multiresolution analysis
Divide and further decompose, and according to the feature of analyzed signal, adaptive selection frequency band is allowed to and signal spectrum phase
Match, improves time frequency resolution.(A indicates that low frequency component, D indicate high fdrequency component, end to three layers of WAVELET PACKET DECOMPOSITION exploded relationship such as following formula
The number of plies of tail serial number expression WAVELET PACKET DECOMPOSITION):
S=AAA3+DAA3+ADA3+DDA3+AAD3+DAD3+ADD3+DDD3
(3) each sensor signal fault feature vector is calculated.The signal of each sensor is calculated after WAVELET PACKET DECOMPOSITION
The energy value of each reproducing sequence is normalized the energy of extracted each reproducing sequence, obtains data characteristics vector.
If Ej *For the energy of wavelet reconstruction sequence, it is expressed as
In formula:For k-th of component of j-th of wavelet package reconstruction sequence;N is sequence djThe number of middle component;N is small echo
Packet Decomposition order.
If the energy value to each frequency range that acquisition signal carries out n-layer wavelet decomposition is
Return place one to change reason feature vector can be obtained the vector being
In formula:
(4) off-line data pretreatment is carried out;
(5) neural metwork training and parameter optimization are carried out;
(6) BP neural network Weight Training module is obtained.
It is illustrated in figure 3 the data prediction process, comprising: set the failure degree of membership of sensor as m (a)=(a1,
a2,…,ak) and sensor be w (x)=(x to the weight of state in fault database1,x2,…,xk).The highest corresponding states of degree of membership
It is exactly the diagnostic result of the sensor, degree of membership is higher, and explanation is higher to the certainty of the state.According to the person in servitude of each sensor
Category degree result handles each sensing data, and for any sensor signal and label is any state i (in fault database
I-th kind of state of k state) feature vector, lay down a regulation as follows:
(i) number of states is k, the failure degree of membership of sensor are as follows: m in fault database1(a)=(a1,a2,…,ak), sensing
Device is to state weight in fault database are as follows: w1(x)=(x1,x2,…,xk).Weight output area [0,10] is divided into p class: [0,
10/p), [2,2+10/p) ..., [10-10/p, 10] }, corresponding classification sequence number 1,2 ..., p is divided when one group of data corresponds to weight
When for class l, then the output of corresponding training BP neural network is O=(01,02,…,1l,…,0p-1,0p), other weight classification are defeated
Out and so on;
(ii) data label is state i, for sensor:
Diagnostic result is consistent with data label:
If (a) ai>=0.7, then xiFor class p, if a1||a2||…||ai-1||ai+1||…||ak< 0.1, then by weight x1||
x2||…||xi-1||xi+1||…||xkIt is set as class p, otherwise by weight x1||x2||…||xi-1||xi+1||…||xkIt is set as classification
(1/a1)||(1/a2)||…||(1/ai-1)||(1/ai+1)||…||(1/ak);
If (b) ai< 0.7, then weight xiFor classification 10aiIf a1||a2||…||ai-1||ai+1||…||ak< 0.1 is right by its
Weight is answered to be divided into class p, if a1||a2||…||ai-1||ai+1||…||ak> 0.4, then its weight is divided into class 1, otherwise by weight
It is set as classification (1/a1)||(1/a2)||…||(1/ai-1)||(1/ai+1)||…||(1/ak);
(iii) this group of data diagnosis result and label are inconsistent, then by weight xiIt is divided into class 1, if max (a1,…,ak)=
am, then by the corresponding weight x of maximum valuemIt is divided into class 1, if while a1||a2||…||ai-1||ai+1||…||ak< 0.1, its is right
Answer weight x1||x2||…||xi-1||xi+1||…||xkIt is divided into class p, if a1||a2||…||ai-1||ai+1||…||ak> 0.4, it will
Corresponding weight x1||x2||…||xi-1||xi+1||…||xkIt is divided into class 1, if being all unsatisfactory for weight x otherwise1||x2||…||xi-1
||xi+1||…||xkIt is set as classification (1/a1)||(1/a2)||…||(1/ai-1)||(1/ai+1)||…||(1/ak)。
Above-mentioned rule is arranged in data when to acquisition from all the sensors, is pre-processed, and obtains training BP neural network
Required data weight classification O (t)=(t1,t2,…,tp);
It is illustrated in figure 4 the neural metwork training and parameter optimization process, comprising: between the selected network number of plies
The random initializtion neural network number of plies, every layer of neuron number, each neuron weight and every layer of biasing, by feature vector
Neural network after input initialization obtains the weight classification O of network outputm, calculate between output and data weight classification O
Cross entropy C, calculation are as follows:
Given threshold, when intersecting entropy lower than threshold value, neural network is met the requirements, and saves network, is unsatisfactory for, and is used
Stochastic gradient algorithm update, neuron weight, every layer biasing, if the number of iterations cannot still be converged in 100 intersection entropy it is low
When threshold value, updates the neural network number of plies and every layer of neuron number and initialize neuron weight and every layer of biasing, sentence
It is disconnected whether to select in section, if then resetting the number of iterations, if not then termination algorithm, completes neural metwork training.Finally
It is chosen from the neural network of preservation and intersects the smallest neural network of entropy as selection neural network
It is illustrated in figure 5 fault diagnosis system, comprising: exist by multi-sensor collection monitored equipment such as high-voltage circuitbreaker
The signals such as vibration, travel displacement, divide-shut brake coil current during divide-shut brake are input to after the conditioning of conditioning circuit
In data collecting card, host computer is combined logical by reading signal data in data collecting card by built-in diagnosis algorithm processing
The weight that off-line training obtains is crossed, fault diagnosis result is obtained.
By taking high-voltage circuitbreaker matches and takes k=3 there are three sensor and three failure typicalness as an example, illustration is used for
The multi-sensor Fusion Algorithm of high-voltage circuitbreaker.
Table 1 has weight and the multi-sensor information fusion comparison without weight
Table 1 is to have weight and the multi-sensor information fusion comparison without weight, including three sensor degrees of membership, three biographies
The weight that sensor off-line training obtains, three sensor degrees of membership pass more multiplied by the weighted evidence obtained after weight, no weight
Sensor information fusion results and there is weight multi-sensor information fusion result.The high-voltage circuitbreaker of the state that is practically in 1 is adopted
Collect related data, is calculated, obtain each sensor degree of membership and weighted evidence.In table 1, the degree of membership of the acquisition of sensor 2
With the degree of membership that sensor 1,3 obtains have it is obvious conflict, and sensor 1 is close to the degree of membership of state 1 and state 2, by
The result obtained after no weight multi-sensor information fusion is (0.398166,0.417337,0.184497), if setting judgement failure
The threshold value of generation is 0.5, then can not be differentiated without weight multi-sensor information fusion, due to the presence of contradiction evidence, even if
Do not take threshold value also can false judgment high-voltage circuitbreaker be in state 2.And after being merged multiplied by the weighted evidence of the weight of off-line training
The result of acquisition is (0.528130,0.368132,0.103738).For state 1, change to the higher weight of sensor 2
Kind evidence contradiction, can accurately be differentiated, obtain correct diagnostic result: i.e. high-voltage circuitbreaker is in state 1.
It should be pointed out that for those skilled in the art, without departing from the principle of the present invention,
Several improvements and modifications can also be made, these modifications and embellishments should also be considered as the scope of protection of the present invention.In the present embodiment not
The available prior art of specific each component part is realized.
Claims (8)
1. a kind of multi-information fusion method for diagnosing faults based on Weight Training, it is characterised in that: including weight off-line training and
Monitor two parts on-line;The specific steps of the diagnostic method include:
Weight off-line training part includes:
(1) off-line sensors signal and corresponding states information are collected;
(2) WAVELET PACKET DECOMPOSITION processing is carried out;
(3) each sensor signal fault feature vector is calculated;
(4) Weight Training is carried out using BP neural network Weight Training module;
(5) each sensor weight is obtained;
On-line monitoring part includes:
(6) sensor, the various types of signal of measuring device in operation are installed on monitored equipment;
(7) WAVELET PACKET DECOMPOSITION processing is carried out;The signal each sensor collected respectively using WAVELET PACKET DECOMPOSITION theory into
Row WAVELET PACKET DECOMPOSITION;The feature for the signal that specific WAVELET PACKET DECOMPOSITION function is collected by sensor in actual use is Lai really
It is fixed;
(8) each sensor signal fault feature vector is calculated;It is each heavy after WAVELET PACKET DECOMPOSITION to calculate the signal of each sensor
The energy value of structure sequence is normalized the energy of extracted each reproducing sequence for the data for reacting fault message,
Obtain fault feature vector;
(9) each sensor fault degree of membership is calculated;Calculate fault feature vector and the event of the signal of collected each sensor
Hinder the Euclidean distance of the feature vector of corresponding existing typical fault status signal in database, distance metric is bigger, shows two
Degree of membership between person is weaker;The fault feature vector for calculating each sensor signal is corresponding with Mishap Database existing each
The normalization reciprocal of the Euclidean distance of the feature vector of a typical fault status signal is as a result, to obtain each sensor corresponding
Failure degree of membership, i.e. evidence;
(10) failure degree of membership obtained by each sensor is obtained into weighted evidence multiplied by weight;
(11) data fusion is carried out to weighted evidence using D-S evidence theory rule of combination, obtains fault diagnosis result.
2. a kind of multi-information fusion method for diagnosing faults based on Weight Training according to claim 1, it is characterised in that:
WAVELET PACKET DECOMPOSITION WPD described in step (2), step (7) has finer in full frequency band time frequency resolution with higher
Localize performance.Frequency band is carried out to multi-level division, the high frequency section that multiresolution analysis does not segment further is decomposed,
And according to the feature of analyzed signal, adaptive selection frequency band is allowed to match with signal spectrum, to improve time-frequency
Resolution ratio.(A indicates that low frequency component, D indicate that high fdrequency component, end serial number indicate to three layers of WAVELET PACKET DECOMPOSITION exploded relationship such as following formula
The number of plies of WAVELET PACKET DECOMPOSITION):
S=AAA3+DAA3+ADA3+DDA3+AAD3+DAD3+ADD3+DDD3。
3. a kind of multi-information fusion method for diagnosing faults based on Weight Training according to claim 1, it is characterised in that:
The step (3), step (8) fault feature vector specific calculation are as follows:
If Ej *For the energy of wavelet reconstruction sequence, it is expressed as
In formula:For k-th of component of j-th of wavelet package reconstruction sequence;N is sequence djThe number of middle component;N is wavelet packet point
Solve the number of plies.If the energy value to each frequency range that acquisition signal carries out n-layer wavelet decomposition is
Return place one to change reason feature vector can be obtained the vector being
In formula:
The corresponding one group of fault feature vector of the signal that each sensor collects.
4. a kind of multi-information fusion method for diagnosing faults based on Weight Training according to claim 1, it is characterised in that:
The training process of step (4) the BP neural network Weight Training module are as follows:
1. collecting off-line sensors signal and corresponding states information;
2. carrying out WAVELET PACKET DECOMPOSITION processing;
3. calculating each sensor signal fault feature vector;
4. carrying out data prediction;The equipment operation sensor signal and corresponding equipment state that each sensor acquires offline are first
Test data, for each sensor priori data carry out data prediction obtain train BP neural network needed for output, i.e., with
The classification output of data weight corresponding to feature vector;
5. carrying out neural metwork training and parameter optimization;Using back-propagation algorithm training neural network, in selected network
The random initializtion neural network number of plies between the number of plies, every layer of neuron number, each neuron weight and every layer of biasing, will be special
Neural network after levying vector input initialization obtains the weight classification O of network outputm, calculate output and data weight classification O
Between cross entropy C, calculation are as follows:
Given threshold, when intersecting entropy lower than threshold value, neural network is met the requirements, and saves network, is unsatisfactory for then using random
Gradient algorithm updates, neuron weight, every layer of biasing, if the number of iterations cannot still be converged to lower than threshold in 100 intersection entropy
When value, updates the neural network number of plies and every layer of neuron number and initialize neuron weight and every layer of biasing, judgement are
It is no to select in section, if then resetting the number of iterations, if not then termination algorithm, completes neural metwork training.Finally from guarantor
It is chosen in the neural network deposited and intersects the smallest neural network of entropy as selection neural network.
6. obtaining BP neural network Weight Training module.
5. a kind of multi-information fusion method for diagnosing faults based on Weight Training according to claim 4, it is characterised in that:
4. the carry out data prediction, method particularly includes: the failure degree of membership of sensor is set as m (a)=(a1,a2,…,ak)
It is w (x)=(x with weight of the sensor to state in fault database1,x2,…,xk);The highest corresponding states of degree of membership is exactly the biography
The diagnostic result of sensor, degree of membership is higher, and explanation is higher to the certainty of the state.According to the degree of membership result of each sensor
Each sensing data is handled, for any sensor signal and label is any state i (k state in fault database
I-th kind of state) feature vector, lay down a regulation as follows:
(i) number of states is k, the failure degree of membership of sensor are as follows: m in fault database1(a)=(a1,a2,…,ak), sensor is to event
Hinder state weight in library are as follows: w1(x)=(x1,x2,…,xk).Weight output area [0,10] is divided into p class: [0,10/p),
[2,2+10/p) ..., [10-10/p, 10] }, corresponding classification sequence number 1,2 ..., p is divided into class l when one group of data corresponds to weight
When, then the output of corresponding training BP neural network is O=(01,02,…,1l,…,0p-1,0p), other weight classification outputs are with this
Analogize;
(ii) data label is state i, for sensor:
Diagnostic result is consistent with data label:
If (a) ai>=0.7, then xiFor class p, if a1||a2||…||ai-1||ai+1||…||ak< 0.1, then by weight x1||x2||…
||xi-1||xi+1||…||xkIt is set as class p, otherwise by weight x1||x2||…||xi-1||xi+1||…||xkIt is set as classification (1/a1)|
|(1/a2)||…||(1/ai-1)||(1/ai+1)||…||(1/ak);
If (b) ai< 0.7, then weight xiFor classification 10aiIf a1||a2||…||ai-1||ai+1||…||ak< 0.1 is corresponded to power
Value is divided into class p, if a1||a2||…||ai-1||ai+1||…||ak> 0.4, then it divides its weight into class 1, is otherwise set as weight
Classification (1/a1)||(1/a2)||…||(1/ai-1)||(1/ai+1)||…||(1/ak);
(iii) this group of data diagnosis result and label are inconsistent, then by weight xiIt is divided into class 1, if max (a1,…,ak)=am, then
By the corresponding weight x of maximum valuemIt is divided into class 1, if while a1||a2||…||ai-1||ai+1||…||ak< 0.1, corresponded to power
Value x1||x2||…||xi-1||xi+1||…||xkIt is divided into class p, if a1||a2||…||ai-1||ai+1||…||ak> 0.4, it will correspond to
Weight x1||x2||…||xi-1||xi+1||…||xkIt is divided into class 1, if being all unsatisfactory for weight x otherwise1||x2||…||xi-1||
xi+1||…||xkIt is set as classification (1/a1)||(1/a2)||…||(1/ai-1)||(1/ai+1)||…||(1/ak)。
Above-mentioned rule is arranged in data when to acquisition from all the sensors, is pre-processed, and obtains needed for training BP neural network
Want data weight classification O (t)=(t1,t2,…,tp)。
6. a kind of multi-information fusion method for diagnosing faults based on Weight Training according to claim 1, it is characterised in that:
Euclidean distance is in inverse relation between failure degree of membership and signal fault feature in step (9), and failure degree of membership is assigned following formula:
In formula:
∑1/dk=1/d1+1/d2+...+1/dk
Wherein dkFor the Euclidean distance of signal characteristic in collected signal fault feature and fault database;K is failure classes in fault database
The number of type;M (a) is failure degree of membership.
7. a kind of multi-information fusion method for diagnosing faults based on Weight Training according to claim 1, it is characterised in that:
The specific calculation of step (10) weighted evidence are as follows:
In conjunction with the weight that offline Weight Training obtains, the calculation of weighted evidence are as follows:
M (a)=w1×m1(a)+w2×m2(a)+...+wn×mn(a)
In formula: m (a) is the weighted evidence obtained;wiFor the weight of each sensor required by front;miIt (a) is each sensor
The failure degree of membership of acquisition, as original evidence;N is number of sensors.
8. a kind of multi-information fusion method for diagnosing faults based on Weight Training according to claim 1, it is characterised in that:
The specific calculation of step (10) weighted evidence are as follows:
In conjunction with the weight that offline Weight Training obtains, the calculation of weighted evidence are as follows:
M (a)=w1×m1(a)+w2×m2(a)+...+wn×mn(a)
In formula: m (a) is the weighted evidence obtained;wiFor the weight of each sensor required by front;miIt (a) is each sensor
The failure degree of membership of acquisition, as original evidence;N is number of sensors.
Priority Applications (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN201910156004.1A CN110163075A (en) | 2019-03-01 | 2019-03-01 | A kind of multi-information fusion method for diagnosing faults based on Weight Training |
Applications Claiming Priority (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN201910156004.1A CN110163075A (en) | 2019-03-01 | 2019-03-01 | A kind of multi-information fusion method for diagnosing faults based on Weight Training |
Publications (1)
Publication Number | Publication Date |
---|---|
CN110163075A true CN110163075A (en) | 2019-08-23 |
Family
ID=67645411
Family Applications (1)
Application Number | Title | Priority Date | Filing Date |
---|---|---|---|
CN201910156004.1A Pending CN110163075A (en) | 2019-03-01 | 2019-03-01 | A kind of multi-information fusion method for diagnosing faults based on Weight Training |
Country Status (1)
Country | Link |
---|---|
CN (1) | CN110163075A (en) |
Cited By (8)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN110781612A (en) * | 2019-11-14 | 2020-02-11 | 哈工大机器人(山东)智能装备研究院 | Fault diagnosis method and device for ball screw, computer device and storage medium |
CN111678679A (en) * | 2020-05-06 | 2020-09-18 | 内蒙古电力(集团)有限责任公司电力调度控制分公司 | Circuit breaker fault diagnosis method based on PCA-BPNN |
CN111965246A (en) * | 2020-08-11 | 2020-11-20 | 太原理工大学 | Scraper machine fault detection method and detection system based on multi-information fusion |
CN112116003A (en) * | 2020-09-18 | 2020-12-22 | 中国人民解放军火箭军工程大学 | Complex sensor network self-adaptive dynamic fault diagnosis method and system |
CN112964246A (en) * | 2021-05-17 | 2021-06-15 | 北京三快在线科技有限公司 | Unmanned aerial vehicle multi-sensor data fusion method and device, storage medium and equipment |
CN113625103A (en) * | 2021-07-12 | 2021-11-09 | 广西电网有限责任公司 | Line selection method for single-phase earth fault of small current grounding system |
CN113688910A (en) * | 2021-08-26 | 2021-11-23 | 国网江苏省电力有限公司常州供电分公司 | Power grid operation multi-source information fusion method based on k-BP algorithm |
CN114722865A (en) * | 2022-03-27 | 2022-07-08 | 江苏东佳电气有限公司 | Explosion-proof motor fault diagnosis method based on data identification |
Citations (2)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN105956290A (en) * | 2016-05-06 | 2016-09-21 | 国网江苏省电力公司检修分公司 | High-voltage circuit breaker mechanical fault diagnosis method based on multi-data fusion technology |
CN107796602A (en) * | 2016-08-31 | 2018-03-13 | 华北电力大学(保定) | A kind of circuit breaker failure diagnostic method of sound and vibration signal fused processing |
-
2019
- 2019-03-01 CN CN201910156004.1A patent/CN110163075A/en active Pending
Patent Citations (2)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN105956290A (en) * | 2016-05-06 | 2016-09-21 | 国网江苏省电力公司检修分公司 | High-voltage circuit breaker mechanical fault diagnosis method based on multi-data fusion technology |
CN107796602A (en) * | 2016-08-31 | 2018-03-13 | 华北电力大学(保定) | A kind of circuit breaker failure diagnostic method of sound and vibration signal fused processing |
Non-Patent Citations (1)
Title |
---|
嵇斗: "基于D-S证据理论和BP算法的直流电机故障诊断研究", 《船电技术》 * |
Cited By (11)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN110781612A (en) * | 2019-11-14 | 2020-02-11 | 哈工大机器人(山东)智能装备研究院 | Fault diagnosis method and device for ball screw, computer device and storage medium |
CN111678679A (en) * | 2020-05-06 | 2020-09-18 | 内蒙古电力(集团)有限责任公司电力调度控制分公司 | Circuit breaker fault diagnosis method based on PCA-BPNN |
CN111965246A (en) * | 2020-08-11 | 2020-11-20 | 太原理工大学 | Scraper machine fault detection method and detection system based on multi-information fusion |
CN111965246B (en) * | 2020-08-11 | 2023-10-31 | 太原理工大学 | Scraper machine fault detection method and system based on multi-information fusion |
CN112116003A (en) * | 2020-09-18 | 2020-12-22 | 中国人民解放军火箭军工程大学 | Complex sensor network self-adaptive dynamic fault diagnosis method and system |
CN112116003B (en) * | 2020-09-18 | 2024-01-19 | 中国人民解放军火箭军工程大学 | Self-adaptive dynamic fault diagnosis method and system for complex sensor network |
CN112964246A (en) * | 2021-05-17 | 2021-06-15 | 北京三快在线科技有限公司 | Unmanned aerial vehicle multi-sensor data fusion method and device, storage medium and equipment |
CN113625103A (en) * | 2021-07-12 | 2021-11-09 | 广西电网有限责任公司 | Line selection method for single-phase earth fault of small current grounding system |
CN113625103B (en) * | 2021-07-12 | 2023-09-19 | 广西电网有限责任公司 | Line selection method for single-phase grounding fault of low-current grounding system |
CN113688910A (en) * | 2021-08-26 | 2021-11-23 | 国网江苏省电力有限公司常州供电分公司 | Power grid operation multi-source information fusion method based on k-BP algorithm |
CN114722865A (en) * | 2022-03-27 | 2022-07-08 | 江苏东佳电气有限公司 | Explosion-proof motor fault diagnosis method based on data identification |
Similar Documents
Publication | Publication Date | Title |
---|---|---|
CN110163075A (en) | A kind of multi-information fusion method for diagnosing faults based on Weight Training | |
CN106338406B (en) | The on-line monitoring of train traction electric drive system and fault early warning system and method | |
CN110334740A (en) | The electrical equipment fault of artificial intelligence reasoning fusion detects localization method | |
CN105956290B (en) | A kind of Mechanical Failure of HV Circuit Breaker diagnostic method based on multi-data fusion technology | |
CN101718634B (en) | Equipment state comprehensive dynamic alarming method based on multivariate probability model | |
CN109213121B (en) | Method for diagnosing clamping cylinder fault of fan braking system | |
CN102721941B (en) | Method for fusing and diagnosing fault information of circuit of electric meter on basis of SOM (self-organized mapping) and D-S (Dempster-Shafer) theories | |
CN113255848B (en) | Water turbine cavitation sound signal identification method based on big data learning | |
CN110161343A (en) | A kind of non-intrusion type real-time dynamic monitoring method of intelligence train exterior power receiving device | |
WO2022052510A1 (en) | Anomaly detection system and method for sterile filling production line | |
CN110376522A (en) | A kind of Method of Motor Fault Diagnosis of the deep learning network of data fusion | |
CN111679166A (en) | Switch cabinet partial discharge fault multi-source information fusion detection early warning system and method based on wireless transmission technology | |
CN106021771A (en) | Method and device for diagnosing faults | |
CN106951197A (en) | A kind of printing equipment remote fault diagnosis method and system based on signal testing | |
CN102122133A (en) | Self-adaption wavelet neural network abnormity detection and fault diagnosis classification system and method | |
CN108710756A (en) | The method for diagnosing faults of lower multicharacteristic information Weighted Fusion is analyzed based on spectral clustering | |
CN109389325B (en) | Method for evaluating state of electronic transformer of transformer substation based on wavelet neural network | |
CN112906764B (en) | Communication safety equipment intelligent diagnosis method and system based on improved BP neural network | |
CN109298633A (en) | Chemical production process fault monitoring method based on adaptive piecemeal Non-negative Matrix Factorization | |
CN108920426A (en) | A kind of method for diagnosing faults based on power equal operator and DS evidence theory | |
CN110375983A (en) | Failsafe valve real-time diagnosis system and diagnostic method based on time series analysis | |
CN106646068A (en) | Method for diagnosing defects of intelligent substation secondary system based on multi-parameter information fusion | |
CN111665066A (en) | Equipment fault self-adaptive upper and lower early warning boundary generation method based on convolutional neural network | |
CN115791174B (en) | Rolling bearing abnormality diagnosis method, system, electronic equipment and storage medium | |
CN110188837A (en) | A kind of MVB network fault diagnosis method based on fuzzy neural |
Legal Events
Date | Code | Title | Description |
---|---|---|---|
PB01 | Publication | ||
PB01 | Publication | ||
SE01 | Entry into force of request for substantive examination | ||
SE01 | Entry into force of request for substantive examination |