CN108757426A - Oilfield water filling plunger pump trouble diagnostic method - Google Patents
Oilfield water filling plunger pump trouble diagnostic method Download PDFInfo
- Publication number
- CN108757426A CN108757426A CN201810724488.0A CN201810724488A CN108757426A CN 108757426 A CN108757426 A CN 108757426A CN 201810724488 A CN201810724488 A CN 201810724488A CN 108757426 A CN108757426 A CN 108757426A
- Authority
- CN
- China
- Prior art keywords
- water filling
- plunger pump
- fault
- width
- oilfield water
- 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.)
- Granted
Links
Classifications
-
- F—MECHANICAL ENGINEERING; LIGHTING; HEATING; WEAPONS; BLASTING
- F04—POSITIVE - DISPLACEMENT MACHINES FOR LIQUIDS; PUMPS FOR LIQUIDS OR ELASTIC FLUIDS
- F04B—POSITIVE-DISPLACEMENT MACHINES FOR LIQUIDS; PUMPS
- F04B51/00—Testing machines, pumps, or pumping installations
-
- 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
- G06N3/04—Architecture, e.g. interconnection topology
- G06N3/045—Combinations of networks
Landscapes
- Engineering & Computer Science (AREA)
- Physics & Mathematics (AREA)
- Theoretical Computer Science (AREA)
- General Engineering & Computer Science (AREA)
- General Health & Medical Sciences (AREA)
- Molecular Biology (AREA)
- Biophysics (AREA)
- Computational Linguistics (AREA)
- Data Mining & Analysis (AREA)
- Evolutionary Computation (AREA)
- Artificial Intelligence (AREA)
- Biomedical Technology (AREA)
- Computing Systems (AREA)
- Life Sciences & Earth Sciences (AREA)
- General Physics & Mathematics (AREA)
- Mathematical Physics (AREA)
- Software Systems (AREA)
- Health & Medical Sciences (AREA)
- Mechanical Engineering (AREA)
- Testing Of Devices, Machine Parts, Or Other Structures Thereof (AREA)
Abstract
The present invention relates to oilfield water filling plunger pump trouble diagnostic methods.The fault mode of all kinds of known fault types of oilfield water filling plunger pump is predefined first, and the one-dimensional original vibration signal of each known fault type is acquired by acceleration transducer；Secondly minimum entropy deconvolution is done to one-dimensional original vibration signal to be filtered, and according to filtered one-dimensional vibration signal, pass through direct truncation, establish two-dimensional matrix, then convolution sum down-sampling operation is carried out to the two-dimensional matrix of foundation using convolutional neural networks, the convolutional neural networks diagnostic model that the minimum entropy deconvolution to obtain known fault pattern enhances；Finally utilize pattern-recognition of the model realization to the oilfield water filling plunger pump unknown failure type for needing to monitor.The automatic identification of multiple faults in oilfield water filling plunger pump may be implemented in the present invention, the problem of efficiently avoiding for dependences such as index and Feature Selections, can be widely applied to the fault diagnosis of a variety of oilfield water filling plunger pumps.
Description
Technical field
The invention belongs to mechanical fault diagnosis fields, are related to a kind of based on minimum entropy deconvolution enhancing convolutional Neural net
The oilfield water filling plunger pump trouble diagnostic method of network.
Background technology
With the fast development of Chinese national economy, the demand of oil is growing day by day, recovery of subterranean crude oil, improves oil
Yield is extremely urgent.Oil-field flooding is one of important link of oil recovery process, as the key equipment in the injecting process -- oil
Mining site water filling plunger pump, operating mode is severe, and work return water water quality inferiority, water temperature are high, easily to valve body, plunger, bearing and close
Packing material etc. causes to corrode, the service life of serious curtailment part, and great economic loss is brought even to lead to serious accident.Cause
This carries out it monitoring in real time and targetedly repairs, avoided there is an urgent need to study the fault diagnosis technology of water filling plunger pump
Burst accident improves production efficiency.
Water filling plunger pump is served in the coupled system of electromechanical hydraulic pressure, and vibration signal is a variety of driving source comprehensive functions
As a result, it is extremely difficult therefrom to detach the driving source that is out of order.Further, since oilfield water filling plunger pump working environment itself is answered
Polygamy causes its fault type various, and failure cause is complicated, and the failure mechanism of many fault types is indefinite.In addition, unit goes out
When existing failure, same fault characteristic frequency corresponds to several different fault types, and fault diagnosis is difficult.
Convolutional neural networks (Convoutional neural network, CNN) are a kind of deep learnings of data-driven
Network, feature that can be in automatic learning data pass through convolution (convolution) and down-sampling (sub-sampling) replace
It operates, the characteristic component in retention data, reduces higher-dimension interference, data characteristics is finally regenerated by full articulamentum, to real
The distributed nature expression of existing data.However, due to the special Service Environment of oilfield water filling plunger pump, vibration signal without
It can avoid, by the interference of each road transmission path of mechanical electronic hydraulic and polluting for ambient noise, having seriously affected data characteristics and having learned automatically
The effect of habit.
Invention content
The purpose of the invention is to overcome shortcoming and defect of the existing technology, and provide a kind of based on minimum entropy solution
Convolution enhances the oilfield water filling plunger pump trouble diagnostic method of convolutional neural networks.
To achieve the above object, the technical scheme is that including the following steps：
S1：The fault mode of all kinds of known fault types in predefined oilfield water filling plunger pump, and acquire each event
Hinder the one-dimensional original vibration signal of type；
S2:Minimum entropy deconvolution is done to original vibration signal to be filtered, and two-dimensional matrix is established to filtered signal,
Training sample as model；
S3:Convolution sum down-sampling operation is carried out to two-dimensional matrix using convolutional neural networks, completely reservation table reference number
The part of feature reduces the higher-dimension interference component in signal, solves the automatic problem concerning study of feature of fault-signal, obtains known event
The convolutional neural networks diagnostic model based on the enhancing of minimum entropy deconvolution of barrier pattern；
S4:Minimum entropy deconvolution filtering is carried out to the fault-signal that needs monitor, two-dimensional matrix is established, constructs test specimens
This, inputs known diagnosis model, finally determines fault type.
Further setting is the one-dimensional original that the step S1 acquires each known fault type by acceleration transducer
Beginning vibration signal.
Further setting is that the step S2 includes：
(1) select the kurtosis of output signal y as the object function of filtering：
(2) relationship between y and f is constructed, can be summarized as：
In formula, x (n) (n=1,2 ..., N) is input signal x by the list entries after data sampling, f (l) (l=
1,2 ..., L) be filter first of weight coefficient, the functional relation of output sequence and filter weight can be expressed as：
(3) object function for maximizing f (l), can obtain：
(4) it is constantly iteratively solved by f (l), update filter weight f：
(5) the optimum filter weight f obtained by formula (9), and formula (6) is substituted into, impact enhancing is calculated
Output signal y, and build two-dimensional matrix using it.
Further setting is using two-dimensional matrix constructed by step S3 as the input picture of convolutional neural networks, the step
Suddenly the convolution operation in S3 includes：
The characteristic dimension of picture and the consistent convolution operation of original image after holding convolution are：It is grasped by a convolution
Picture size after work is：
In formula (10), inputmaplengthIndicate the length of the input picture of convolutional layer, inputmapwidthIndicate convolutional layer
Input picture width；stridelengthIndicate the length of convolution kernel moving step length, stridewidthIndicate convolution kernel movement step
Long width；outputmaplengthIndicate that the output picture length after a convolution operation, size are
inputmaplength/stridelength, and round up；outputmapwidthIndicate the output figure after a convolution operation
Piece width, size inputmapwidth/stridewidth, and round up；
This setting is to can also be that the convolution operation in the step S3 includes：
The consistent convolution operation of the characteristic dimension and original image that need not keep the picture after convolution is：If passed through
Picture size after convolution operation is：
outputmapwidth=[(inputmapwidth-filterwidth)+1/stridewidth]
In formula (11), filter indicates convolution kernel, filterlengthIndicate the length of convolution kernel, filterwidthIndicate volume
The width of product core.
Further setting is use maximum value in the step S3 or carries out down-sampling operation using average value, to contract
The size of small picture.
Further setting is after convolution operation and down-sampling operation are completed, and the Dropout by full articulamentum is random
Some hidden neurons in convolutional neural networks are deleted on ground, and keep input and output neuron number constant, and full articulamentum will
The output picture of its last layer connects into an one-dimensional vector, and penalty values are calculated by loss function, to continuous to network
Iteration updates, and finally obtains the ideal weighted value W of convolutional neural networks and biasing b, that is, obtains the minimum entropy of known fault pattern
The convolutional neural networks diagnostic model of deconvolution enhancing,
The loss function is the function evaluated with actual result for model output result, is selected as mean square error
Difference function：
In formula (12), z is desired output；F (wq+b) is the reality output of model, and q is the input of neuron, and w is power
Weight values, b are biasing, and f () is activation primitive.
Further setting is that the step S4 is：To the one-dimensional original of oilfield water filling plunger pump unknown failure type
Vibration signal carries out minimum entropy deconvolution and is filtered, and according to filtered one-dimensional vibration signal, passes through direct truncated position
Reason, establishes two-dimensional matrix, as test sample, inputs the convolutional Neural of the minimum entropy deconvolution enhancing of the known fault pattern
Network diagnosis model, finally determines fault type.
The present invention is directed to provide a kind of diagnostic method of multiple faults type for oilfield water filling plunger pump, to solve its event
The above-mentioned problem occurred in barrier diagnosis.
For this purpose, the present invention proposes a kind of minimum entropy deconvolution enhancing convolution of oilfield water filling plunger pump trouble diagnosis
Neural network method adaptively finds filter using minimum entropy deconvolution (Minimum Entropy Deconvolution, MED)
Wave device weight carries out de-noising to original plunger pump trouble signal, improves signal-to-noise ratio, while realizing data using convolutional neural networks
The problems such as feature learns automatically, and solution failure mechanism is indefinite, achieve very satisfied failure modes as a result, related this respect
Research, there is no report at present.
It is an advantage of the invention that：
On the one hand the method for the present invention carries out adaptive-filtering using minimum entropy deconvolution to primary fault signal, eliminate each
The transmission path interference of class fault mode original signal, noise background pollution etc., improve the signal-to-noise ratio of failure original signal, and can be compared with
The completely key message in stick signal；On the other hand, using the feature learning ability of convolutional neural networks brilliance, ideally
The feature by the filtered fault mode of minimum entropy deconvolution is characterized, and successfully identifies each fault type.
Description of the drawings
In order to more clearly explain the embodiment of the invention or the technical proposal in the existing technology, to embodiment or will show below
There is attached drawing needed in technology description to be briefly described, it should be apparent that, the accompanying drawings in the following description is only this
Some embodiments of invention, for those of ordinary skill in the art, without having to pay creative labor, according to
These attached drawings obtain other attached drawings and still fall within scope of the invention.
Fig. 1 flow charts of the method for the present invention；
The method schematic diagram of Fig. 2 present invention；
The minimum entropy deconvolution filtering figure of Fig. 3 present invention；
One oilfield water filling plunger pump trouble signal waveforms of Fig. 4 embodiment of the present invention case；
One oilfield water filling plunger pump trouble feature learning comparative result figure of Fig. 5 embodiment of the present invention case；
One oilfield water filling plunger pump multistream heat exchanger comparative result figure of Fig. 6 embodiment of the present invention case
Two oilfield water filling plunger pump trouble signal waveforms of Fig. 7 embodiment of the present invention case；
Two oilfield water filling plunger pump trouble feature learning Comparative result of Fig. 8 embodiment of the present invention case；
Two oilfield water filling plunger pump multistream heat exchanger Comparative result of Fig. 9 embodiment of the present invention case.
Specific implementation mode
To make the object, technical solutions and advantages of the present invention clearer, the present invention is made into one below in conjunction with attached drawing
Step ground detailed description.
As shown in Figure 1 to Figure 2, it is to include the following steps in the embodiment of the present invention：
S1:The fault mode of all kinds of known fault types of predefined oilfield water filling plunger pump, and passed by acceleration
Sensor acquires the one-dimensional original vibration signal of each known fault type
It is predefined that known fault types all kinds of to oilfield water filling plunger pump carry out fault mode：Foundation is covered various
A self-contained mode (the unknown failure type monitored is needed to belong to the self-contained mode) and failure definition mould for fault type
Formula.Then the one-dimensional original vibration signal of each known fault type is acquired by acceleration transducer.
S2:Minimum entropy deconvolution is done to one-dimensional original vibration signal to be filtered.
Mixed signal c is made of impulse source o, random noise e and other interference d：
C=o+d+e (1)
In general, signaling path can simply be modeled as linearly invariant (Linear Time Invariant,
LTI) system transter.The signal x finally collected is the convolution of transmission function h and mixed signal c, this, which is equivalent to one, has
Limit the filtering of impulse response (Finite Impulse Response, FIR) filter：
X=c*h (2)
Entropy is generally understood as the measurement of description disturbance state.In most cases, we are to that may imply machinery
The impact ingredient (i.e. the signal with sharp pulse) of the initial failure of system is interested, i.e. the ingredient with smaller entropy.One
As for, random noise signal be intended to present Gaussian Profile.Meanwhile studies have pointed out that in all real number variances and mean value
In probability distribution, the entropy of Gaussian Profile is maximum.Therefore, entropy is minimized, can effectively enhances the impact in mixed signal
Property feature, highlights impact ingredient.
Minimum entropy deconvolution can be regarded as the inverse filtering process opposite with signal transduction process, and main target is to pass through formula
(3) a best inverse filter coefficient f is found, to restore the impact ingredient in mixed signal：
Y=f*x (3)
As shown in Fig. 2, in the case of no any priori, minimum entropy deconvolution filter can be defeated by optimizing
The object function for going out y adaptively to adjust filter, obtains optimal weight coefficient.
It is well known that statistical indicator (such as high-order statistic) is usually used in describing probability density function (Probability
Density Function, PDF) shape.For example, larger kurtosis (Fourth amount) value generally means that point is presented in PDF
Peak.Therefore, the feature that quantized signal is carried out using high-order statistic as object function can effectively eliminate the interference of mixed signal
Ingredient, prominent spike.For arbitrary output signal y, object functionIt can be summarized as：
In formula, y (n) (n=1,2 ..., N) is output signal y by the output sequence after data sampling, and r is expressed as r
Rank statistic, the usual values of s are 2, are Fourth amount (kurtosis) as r=4.
As previously mentioned, impact signal can be enhanced by finding with best deconvolution coefficient vector f, main mistake
Journey is summarized as follows：
(1) object function of the kurtosis of output y as filtering is selected：
(2) relationship between y and f is constructed, can be summarized as：
In formula, x (n) (n=1,2 ..., N) is input signal x by the list entries after data sampling, f (l) (l=
1,2 ..., L) be filter first of weight coefficient.The functional relation of output sequence and filter weight can be expressed as：
(3) object function for maximizing f (l), can obtain：
(4) it is constantly iteratively solved by f (l), update filter weight f：
(5) the optimum filter weight f obtained by formula (9), and formula (6) is substituted into, enhanced output letter is calculated
Number y can obtain impact after being filtered by minimum entropy deconvolution to the original vibration signal of oilfield water filling plunger pump
The signal y of enhancing builds two-dimensional matrix, the input picture as convolutional neural networks using it.
S3:Establish the convolutional neural networks diagnostic model of the minimum entropy deconvolution enhancing of known fault pattern.
Convolutional neural networks have many different frameworks (Architecture), including, LeNet-5, AlexNet, VGG,
ResNet etc., all convolutional neural networks frameworks, which all have, communicates operation, is respectively, convolutional layer, down-sampling layer and full connection
Layer, each layer have input picture inputmap and output outputmap.The different network architectures uses the different numbers of plies, swashs
Function and training method etc. living can select the above suitable convolutional neural networks framework according to actual needs.
Convolution operation is part most crucial in convolutional neural networks, it is equivalent to a filter (filter), can be with
The critical data information in picture is extracted, and is stored in the form of weighted value, therefore, is generally called convolution kernel.Convolution
The size of core and artificial specified.Originally weighted value generates at random, and continued to optimize during successively training, may finally
Ideally characterize critical data information in picture.
For convolution operation, if it is desired to keep the characteristic dimension of the picture after convolution and consistent, the process of original image
Picture size after convolution operation is：
In formula (10), inputmaplengthIndicate the length of the input picture of convolutional layer, inputmapwidthIndicate convolutional layer
Input picture width；stridelengthIndicate the length of convolution kernel moving step length, stridewidthIndicate convolution kernel movement step
Long width；outputmaplengthIndicate that the output picture length after a convolution operation, size are
inputmaplength/stridelength, and round up；outputmapwidthIndicate the output figure after a convolution operation
Piece width, size inputmapwidth/stridewidth, and round up；
For convolution operation, if the characteristic dimension of the picture after convolution and consistent, the process of original image need not be kept
Picture size after convolution operation is：
In formula (11), filter indicates convolution kernel, filterlengthIndicate the length of convolution kernel, filterwidthIndicate volume
The width of product core.
Down-sampling operation is added after convolution, to reduce the size of picture.Selection that there are two types of down-sampling operations usually, one
Kind is to use maximum value, and another kind is to use average value, can select to use according to actual needs.
After convolution operation and down-sampling operation are completed, randomly deleted in network by the Dropout of full articulamentum
Some hidden neurons, and keep input and output neuron number constant, be traditionally arranged to be 0.5.Full articulamentum will thereon one
The output picture of layer connects into an one-dimensional vector, and penalty values are calculated by loss function, thus more to the continuous iteration of network
Newly, the ideal weighted value W of convolutional neural networks and biasing b are finally obtained, that is, obtains the minimum entropy deconvolution of known fault pattern
The convolutional neural networks diagnostic model of enhancing.
Loss function (loss fuction) is the function evaluated for model output result and actual result, can be with
It is selected as mean square error function：
In formula (12), z is desired output (i.e. fault mode good defined in step 1)；F (wq+b) is the reality of model
Output (i.e. by the model learning of known fault pattern after, obtained output fault type, q is the input of neuron, and w is
Weighted value, b are biasing, and f () is activation primitive.The convolutional Neural net of the frameworks such as AlexNet, VGG, ResNet is used in selection
Network when, activation primitive f () usually selects ReLU, keeps data smoothened, accelerate training process.
S4:Pattern-recognition to the oilfield water filling plunger pump unknown failure type that needs monitor, last diagnostic go out event
Hinder type.
Finally to the one-dimensional original vibration signal of oilfield water filling plunger pump unknown failure type, minimum entropy uncoiling is carried out
Product is filtered, and establishes two-dimensional matrix by direct truncation according to filtered one-dimensional vibration signal, as test
Sample inputs known diagnosis model, finally determines fault type.
Embodiment case one：
Take the fault data of certain Oil Field operation pump group, water filling plunger pump model：5ZB_14/38 is horizontal straight-through
Formula combines plate valve integral pump head, and pumped (conveying) medium is clear water, diameter of plunger 38mm, pump speed 260r/min, stroke 180mm.
The fault data of 7 kinds of oil-field flooding plunger pumps is taken in present case, fault type is respectively：(1) normal operating conditions makes a reservation for
Justice is fault mode 1；(2) crankcase bearing outer ring is worn, and is predefined as fault mode 2；(3) crankcase bearing inner ring is worn,
It is predefined as fault mode 3；(4) Phillips head bolts loosened screw is predefined as fault mode 4；(5) plunger connecting clamp pine
It is dynamic, it is predefined as fault mode 5；(6) tapping valve loosens, and is predefined as fault mode 6；(7) plunger wear is predefined as failure
Mode 7.Sample frequency 25.6k Hz.
Original signal waveform is as shown in Figure 4：(a) fault mode 1；(b) fault mode 2；(c) fault mode 3；(d) failure
Pattern 4；(e) fault mode 5；(f) fault mode 6；(g) fault mode 7.As seen from the figure, per a kind of fault-signal in addition to amplitude
Aspect has some differences, it is difficult to observe fault signature.Therefore, it is necessary to find a kind of feature of the data-driven side of study automatically
Method realizes the distributed expression per a kind of fault mode feature.
In order to realize the automated characterization study of fault data, select 360000 points as pretreated original in sampled point
Beginning data, 400 data points of every section of interception.Each fault mode training set chooses 240000 (600 × 400) a data points, surveys
Examination collection chooses 120000 (300 × 400) a data points, i.e. each fault mode has 600 training samples and 300 tests respectively
Sample.Therefore, the size for inputting picture is 20 × 20.
Fig. 5 gives case one petrochina mining site water filling plunger pump trouble feature learning result.Fig. 5 (a) is minimum entropy solution
The feature learning of the convolutional neural networks model of convolution enhancing is as a result, Fig. 5 (b) is using individual convolutional neural networks model
Feature learning by rotated three dimensional scatter plot as a result, can comprehensively be observed.It can be seen from the figure that all kinds of fault modes
Class between spacing clearly, illustrate that convolutional neural networks can effectively learn the feature to every a kind of fault mode；Single class
Spacing is compact in the class of fault mode, illustrates that convolutional neural networks have good robustness for the identification of each category feature.From
And obtain, after enhancing by minimum entropy deconvolution, convolutional neural networks learn the fault signature of oilfield plunger pump
Ability is remarkably reinforced.
Fig. 6 show one petrochina mining site water filling plunger pump trouble classification confusion matrix figure of case.(a), (b) institute in Fig. 6
The classification results for showing the respectively convolutional neural networks and individual convolutional neural networks of the enhancing of minimum entropy deconvolution, can see
Go out, the former classification Average Accuracy is 99.14%, and the latter is only 77.71%.As shown in Fig. 6 (a), minimum entropy deconvolution increases
Strong convolutional neural networks are 94% for the recognition accuracy of fault mode 1, and fault mode 1 is mistaken for 2 He of fault mode
The probability of fault mode 3 is 1% and 3%；100% can reach to the recognition accuracy of other fault modes, and individually roll up
The result of product neural network is without so ideal.As shown in Fig. 6 (b), individual convolutional neural networks are to fault mode 1 and failure
The recognition accuracy 100% of pattern 4；Fault mode 2 is 85%, and the probability that fault mode 2 is mistaken for fault mode 5 is
15%；Fault mode 3 is 68%, and it is 20% and 12% that fault mode 3, which is mistaken for fault mode 5 and the probability of fault mode 7,；
Fault mode 5 is 71%, and it is 11% and 18% that fault mode 5, which is mistaken for fault mode 2 and the probability of fault mode 3,；Failure
Pattern 6 is 51%, and the probability that fault mode 6 is mistaken for fault mode 3, fault mode 5 and fault mode 7 is 4%, 12% and
33%；Fault mode 7 is 69%, and the probability that fault mode 7 is mistaken for fault mode 2, fault mode 3 and fault mode 6 is
5%, 10% and 17%.Embodiment case two：
Take the fault data of certain Oil Field operation pump group, water filling plunger pump model：3ZS-5/35, fluid end is using straight
General formula is pressed plate valve integral pump head, and pumped (conveying) medium is sewage, diameter of plunger 30mm, pump speed 320r/min, and stroke is
110mm.The fault data of 6 kinds of oil-field flooding plunger pumps is taken in present case, fault type is respectively：(1) shape is worked normally
State is predefined as fault mode 1；(2) motor side bearing inner race is worn, and is predefined as fault mode 2；(3) outside motor end bearing
Circle abrasion, is predefined as fault mode 3；(4) crankcase bearing bush abrasion is predefined as fault mode 4；(5) liquid feed valve loosens, in advance
It is defined as fault mode 5；(6) belt pulley deflection is predefined as fault mode 6.Sample frequency 25.6k Hz.
Original signal waveform is as shown in Figure 4：(a) fault mode 1；(b) fault mode 2；(c) fault mode 3；(d) failure
Pattern 4；(e) fault mode 5；(f) fault mode 6.As seen from the figure, per a kind of fault-signal in addition to having some poor in terms of amplitude
It is different, it is difficult to observe fault signature.Therefore, it is necessary to find a kind of feature Auto-learning Method of data-driven, realize each
The distributed expression of class fault mode feature.
In order to realize the automated characterization study of fault data, select 200000 points as pretreated original in sampled point
Beginning data, 400 data points of every section of interception.Each fault mode training set chooses 120000 (300 × 400) a data points, surveys
Examination collection chooses 80000 (200 × 400) a data points, i.e. each fault mode has 300 training samples and 100 tests respectively
Sample.Therefore, the size for inputting picture is 20 × 20.
Fig. 8 gives case one petrochina mining site water filling plunger pump trouble feature learning result.Fig. 8 (a) is minimum entropy solution
The feature learning of the convolutional neural networks model of convolution enhancing is as a result, Fig. 8 (b) is using individual convolutional neural networks model
Feature learning by rotated three dimensional scatter plot as a result, can comprehensively be observed.It can be seen from the figure that all kinds of fault modes
Class between spacing clearly, illustrate that convolutional neural networks can effectively learn the feature to every a kind of fault mode；Single class
Spacing is compact in the class of fault mode, illustrates that convolutional neural networks have good robustness for the identification of each category feature.By
After this by minimum entropy deconvolution as it can be seen that enhance, convolutional neural networks learn the fault signature of oilfield plunger pump
Ability is remarkably reinforced.
Fig. 9 show one petrochina mining site water filling plunger pump trouble classification confusion matrix figure of case.(a), (b) institute in Fig. 9
The classification results for showing the respectively convolutional neural networks and individual convolutional neural networks of the enhancing of minimum entropy deconvolution, can see
Go out, the former classification Average Accuracy is 97.83%, and the latter is only 85.67%.As shown in Fig. 9 (a), minimum entropy deconvolution increases
Strong convolutional neural networks are 96% for the recognition accuracy of fault mode 1, and fault mode 1 is mistaken for 3 He of fault mode
The probability of fault mode 4 is 3% and 1%；The recognition accuracy of fault mode 3 is 94%, and fault mode 3 is mistaken for failure mould
Formula 1, fault mode 2, fault mode 4 and fault mode 6 probability be 2%, 1%, 1% and 2%；The identification of fault mode 6 is accurate
True rate is 97%, and the probability that fault mode 6 is mistaken for fault mode 2 is 3%；It is equal to the recognition accuracy of other fault modes
100% is can reach, and individually the result of convolutional neural networks is without so ideal.As shown in Fig. 6 (b), individual convolution god
Through network to the recognition accuracy 100% of fault mode 1 and fault mode 4；Fault mode 2 is 93%, and fault mode 2 is misjudged
Probability for fault mode 3 and fault mode 5 is 2% and 6%；Fault mode 3 is 60%, and fault mode 3 is mistaken for failure
Pattern 1, fault mode 2, fault mode 4, fault mode 5 and fault mode 6 probability be 2%, 12%, 4%, 15% and 8%；
Fault mode 5 be 77%, fault mode 5 be mistaken for fault mode 2, fault mode 3 and fault mode 6 probability be 7%,
9% and 6%；Fault mode 6 is 84%, and fault mode 6 is mistaken for fault mode 1, fault mode 2, fault mode 3 and failure
The probability of pattern 5 is 3%, 1%, 6% and 6%.
One of ordinary skill in the art will appreciate that implement the method for the above embodiments be can be with
Relevant hardware is instructed to complete by program, the program can be stored in a computer read/write memory medium,
The storage medium, such as ROM/RAM, disk, CD.
The above disclosure is only the preferred embodiments of the present invention, cannot limit the right model of the present invention with this certainly
It encloses, therefore equivalent changes made in accordance with the claims of the present invention, is still within the scope of the present invention.
Claims (8)
1. a kind of oilfield water filling plunger pump trouble diagnostic method, it is characterised in that including：
S1：The fault mode of all kinds of known fault types in predefined oilfield water filling plunger pump, and acquire each failure classes
The one-dimensional original vibration signal of type；
S2:Minimum entropy deconvolution is done to original vibration signal to be filtered, and two-dimensional matrix is established to filtered signal, as
The training sample of model；
S3:Convolution sum down-sampling operation is carried out to two-dimensional matrix using convolutional neural networks, completely reservation table levies signal characteristic
Part, reduce the higher-dimension interference component in signal, solve the automatic problem concerning study of feature of fault-signal, obtain known fault mould
The convolutional neural networks diagnostic model based on the enhancing of minimum entropy deconvolution of formula；
S4:Minimum entropy deconvolution filtering is carried out to the fault-signal that needs monitor, establishes two-dimensional matrix, constructs test sample, it is defeated
Enter known diagnosis model, finally determines fault type.
2. oilfield water filling plunger pump trouble diagnostic method according to claim 1, it is characterised in that：The step
S1 acquires the one-dimensional original vibration signal of each known fault type by acceleration transducer.
3. oilfield water filling plunger pump trouble diagnostic method according to claim 1, it is characterised in that：The step
S2 includes：
(1) select the kurtosis of output signal y as the object function of filtering：
(2) relationship between y and f is constructed, can be summarized as：
In formula, x (n) (n=1,2 ..., N) is input signal x by the list entries after data sampling, f (l) (l=1,
2 ..., L) be filter first of weight coefficient, the functional relation of output sequence and filter weight can be expressed as：
(3) object function for maximizing f (l), can obtain：
(4) it is constantly iteratively solved by f (l), update filter weight f：
(5) the optimum filter weight f obtained by formula (9), and formula (6) is substituted into, the output of impact enhancing is calculated
Signal y, and build two-dimensional matrix using it.
4. oilfield water filling plunger pump trouble diagnostic method according to claim 3, it is characterised in that：By step S3 institutes
Input picture of the two-dimensional matrix as convolutional neural networks is built, the convolution operation in the step S3 includes：
The characteristic dimension of picture and the consistent convolution operation of original image after holding convolution are：After a convolution operation
Picture size be：
In formula (10), inputmaplengthIndicate the length of the input picture of convolutional layer, inputmapwidthIndicate the defeated of convolutional layer
Enter the width of picture；stridelengthIndicate the length of convolution kernel moving step length, stridewidthIndicate convolution kernel moving step length
Width；outputmaplengthIndicate the output picture length after a convolution operation, size inputmaplength/
stridelength, and round up；outputmapwidthIndicate the output picture width after a convolution operation, size
For inputmapwidth/stridewidth, and round up.
5. the oilfield water filling plunger pump trouble diagnostic method according to claim 3, it is characterised in that：It will step
Input picture of the two-dimensional matrix as convolutional neural networks constructed by rapid S3, the convolution operation in the step S3 include：
The consistent convolution operation of the characteristic dimension and original image that need not keep the picture after convolution is：If by primary
Picture size after convolution operation is：
outputmapwidth=[(inputmapwidth-filterwidth)+1/stridewidth]
In formula (11), filter indicates convolution kernel, filterlengthIndicate the length of convolution kernel, filterwidthIndicate convolution kernel
Width.
6. the oilfield water filling plunger pump trouble diagnostic method according to claim 4 or 5, it is characterised in that：
Use maximum value or use average value in the step S3 carry out down-sampling operation, to reduce the size of picture.
7. the oilfield water filling plunger pump trouble diagnostic method according to claim 6, it is characterised in that：It is rolling up
After product operation and down-sampling operation are completed, one in convolutional neural networks is randomly deleted by the Dropout of full articulamentum
A little hidden neurons, and keep input and output neuron number constant, full articulamentum connects into the output picture of its last layer
One one-dimensional vector, penalty values are calculated by loss function, to be updated to the continuous iteration of network, finally obtain convolutional Neural
The ideal weighted value W of network and biasing b, that is, the convolutional neural networks for obtaining the minimum entropy deconvolution enhancing of known fault pattern are examined
Disconnected model,
The loss function is the function evaluated with actual result for model output result, is selected as mean square error letter
Number：
In formula (12), z is desired output；F (wq+b) is the reality output of model, and q is the input of neuron, and w is weighted value,
B is biasing, and f () is activation primitive.
8. the oilfield water filling plunger pump trouble diagnostic method according to claim 7, it is characterised in that：It is described
Step S4 be：To the one-dimensional original vibration signal of oilfield water filling plunger pump unknown failure type, minimum entropy uncoiling is carried out
Product is filtered, and establishes two-dimensional matrix by direct truncation according to filtered one-dimensional vibration signal, as test
Sample inputs the convolutional neural networks diagnostic model of the minimum entropy deconvolution enhancing of the known fault pattern, finally determines
Fault type.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201810724488.0A CN108757426B (en) | 2018-07-04 | 2018-07-04 | Fault diagnosis method for water injection plunger pump in petroleum mine field |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201810724488.0A CN108757426B (en) | 2018-07-04 | 2018-07-04 | Fault diagnosis method for water injection plunger pump in petroleum mine field |
Publications (2)
| Publication Number | Publication Date |
|---|---|
| CN108757426A true CN108757426A (en) | 2018-11-06 |
| CN108757426B CN108757426B (en) | 2021-03-30 |
Family
ID=63976014
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN201810724488.0A Active CN108757426B (en) | 2018-07-04 | 2018-07-04 | Fault diagnosis method for water injection plunger pump in petroleum mine field |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN108757426B (en) |
Cited By (4)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN109765053A (en) * | 2019-01-22 | 2019-05-17 | 中国人民解放军海军工程大学 | Utilize the Fault Diagnosis of Roller Bearings of convolutional neural networks and kurtosis index |
| CN110991424A (en) * | 2019-12-25 | 2020-04-10 | 安徽工业大学 | Fault diagnosis method based on minimum entropy deconvolution and stacking sparse self-encoder |
| CN112446389A (en) * | 2019-08-15 | 2021-03-05 | 北京国双科技有限公司 | Fault judgment method and device |
| CN113409213A (en) * | 2021-06-22 | 2021-09-17 | 中铁工程装备集团有限公司 | Plunger pump fault signal time-frequency graph noise reduction enhancement method and system |
Citations (4)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN105910827A (en) * | 2016-04-25 | 2016-08-31 | 东南大学 | Induction motor fault diagnosis method based on discriminant convolutional feature learning |
| CN106874957A (en) * | 2017-02-27 | 2017-06-20 | 苏州大学 | A kind of Fault Diagnosis of Roller Bearings |
| CN107526853A (en) * | 2016-06-22 | 2017-12-29 | 北京航空航天大学 | Rolling bearing fault mode identification method and device based on stacking convolutional network |
| CN108171318A (en) * | 2017-11-30 | 2018-06-15 | 河南大学 | One kind is based on the convolutional neural networks integrated approach of simulated annealing-Gaussian function |
-
2018
- 2018-07-04 CN CN201810724488.0A patent/CN108757426B/en active Active
Patent Citations (5)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN105910827A (en) * | 2016-04-25 | 2016-08-31 | 东南大学 | Induction motor fault diagnosis method based on discriminant convolutional feature learning |
| CN105910827B (en) * | 2016-04-25 | 2017-04-05 | 东南大学 | Induction machine method for diagnosing faults based on identification convolution feature learning |
| CN107526853A (en) * | 2016-06-22 | 2017-12-29 | 北京航空航天大学 | Rolling bearing fault mode identification method and device based on stacking convolutional network |
| CN106874957A (en) * | 2017-02-27 | 2017-06-20 | 苏州大学 | A kind of Fault Diagnosis of Roller Bearings |
| CN108171318A (en) * | 2017-11-30 | 2018-06-15 | 河南大学 | One kind is based on the convolutional neural networks integrated approach of simulated annealing-Gaussian function |
Cited By (5)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN109765053A (en) * | 2019-01-22 | 2019-05-17 | 中国人民解放军海军工程大学 | Utilize the Fault Diagnosis of Roller Bearings of convolutional neural networks and kurtosis index |
| CN112446389A (en) * | 2019-08-15 | 2021-03-05 | 北京国双科技有限公司 | Fault judgment method and device |
| CN110991424A (en) * | 2019-12-25 | 2020-04-10 | 安徽工业大学 | Fault diagnosis method based on minimum entropy deconvolution and stacking sparse self-encoder |
| CN113409213A (en) * | 2021-06-22 | 2021-09-17 | 中铁工程装备集团有限公司 | Plunger pump fault signal time-frequency graph noise reduction enhancement method and system |
| CN113409213B (en) * | 2021-06-22 | 2023-11-14 | 中铁工程装备集团有限公司 | Method and system for enhancing noise reduction of time-frequency diagram of fault signal of plunger pump |
Also Published As
| Publication number | Publication date |
|---|---|
| CN108757426B (en) | 2021-03-30 |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN108757426A (en) | Oilfield water filling plunger pump trouble diagnostic method | |
| CN109272123B (en) | Sucker-rod pump working condition early warning method based on convolution-circulation neural network | |
| Wang et al. | A minimum entropy deconvolution-enhanced convolutional neural networks for fault diagnosis of axial piston pumps | |
| CN106769052B (en) | A kind of mechanical system rolling bearing intelligent failure diagnosis method based on clustering | |
| CN110163302B (en) | Indicator diagram identification method based on regularization attention convolution neural network | |
| CN104712542B (en) | A kind of reciprocating compressor sensitive features based on Internet of Things are extracted and method for diagnosing faults | |
| CN110029986B (en) | Beam-pumping unit working fluid level prediction method based on particle swarm extreme learning machine | |
| CN110500217B (en) | Diesel engine common rail system oil pump fault detection method based on measurable data characteristics | |
| CN109997025A (en) | Engine Gernral Check-up device and method | |
| CN110110768A (en) | Fault Diagnosis of Roller Bearings based on Concurrent Feature study and multi-categorizer | |
| CN110363337B (en) | Oil measuring method and system of oil pumping unit based on data driving | |
| CN108398249B (en) | Method and device for predicting residual life of sealing ring, engineering machine and server | |
| CN111042917B (en) | Common rail fuel injector weak fault diagnosis method based on GOA-MCKD and hierarchical discrete entropy | |
| DE102010025851A1 (en) | A method of training a rolling bearing state classification system and a rolling bearing state classification system and a rolling bearing state classification system | |
| CN109359873A (en) | One kind being based on PCA-T2Ball screw assembly, health evaluating method | |
| CN110728377B (en) | Intelligent fault diagnosis method and system for electromechanical equipment | |
| EP4042125A1 (en) | Method for leakage detection | |
| CN113988479A (en) | Pumping well multi-well dynamic liquid level depth prediction method based on dynamic and static information feature fusion neural network | |
| Wang et al. | A working condition diagnosis model of sucker rod pumping wells based on big data deep learning | |
| CN105781958A (en) | Fault self-diagnosis device for one-way valve of membrane pump and diagnosis method | |
| CN113029619A (en) | Underground scraper fault diagnosis method based on C4.5 decision tree algorithm | |
| CN116244596A (en) | Industrial time sequence data anomaly detection method based on TCN and attention mechanism | |
| Wei et al. | Cavitation state recognition for control valve using AlexNet-type neural networks with three-channel images transformed by time series | |
| CN115979660A (en) | Filter fault diagnosis method for internal combustion engine based on MPA optimization | |
| CN113505639B (en) | Rotary machine multi-parameter health state assessment method based on TPE-XGBoost |
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 | ||
| GR01 | Patent grant | ||
| GR01 | Patent grant | ||
| EE01 | Entry into force of recordation of patent licensing contract | ||
| EE01 | Entry into force of recordation of patent licensing contract |
Application publication date: 20181106 Assignee: ZHEJIANG TIANDE PUMPS Co.,Ltd. Assignor: Wenzhou University Contract record no.: X2021330000855 Denomination of invention: Fault diagnosis method of water injection column plug pump in petroleum field Granted publication date: 20210330 License type: Common License Record date: 20211229 |