CN108898120A - Water cooler method for diagnosing faults based on hybrid neural networks - Google Patents
Water cooler method for diagnosing faults based on hybrid neural networks Download PDFInfo
- Publication number
- CN108898120A CN108898120A CN201810724449.0A CN201810724449A CN108898120A CN 108898120 A CN108898120 A CN 108898120A CN 201810724449 A CN201810724449 A CN 201810724449A CN 108898120 A CN108898120 A CN 108898120A
- Authority
- CN
- China
- Prior art keywords
- network
- rbf
- water cooler
- output
- data
- 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
- G06F2218/00—Aspects of pattern recognition specially adapted for signal processing
- G06F2218/02—Preprocessing
- G06F2218/04—Denoising
- G06F2218/06—Denoising by applying a scale-space analysis, e.g. using wavelet analysis
-
- 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
- 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
Abstract
The invention discloses the water cooler method for diagnosing faults based on hybrid neural networks, water cooler is difficult to make Accurate Prediction in real time for single BP neural network, and BP neural network itself existing defects, this method combination RBF neural can approach the characteristic of arbitrary function, propose the hybrid production style RBF-BP based on wavelet transformation and carry out fault diagnosis to water cooler.By RBF network and the parallel connection of BP network as a neural network, the referred to as hidden layer of RBF-BP, the algorithm has the advantages that RBF network and BP network simultaneously, its learning process convergence rate is quickly, and it is avoided that the problem of training process is easily trapped into local minimum, the fault diagnosis that can be effectively applied to water cooler improves the performance of fault diagnosis.
Description
Technical field
The present invention relates to a kind of fault diagnosis technology fields of data-driven, and in particular to one kind is based on hybrid neural networks
Water cooler method for diagnosing faults.
Background technique
With flourishing for social economy, Heating,Ventilating and Air Conditioning has become the indispensable important equipment of skyscraper.Cold water
Unit is most important energy consumption equipment in Heating,Ventilating and Air Conditioning, and energy consumption accounts for about the 70% of total energy consumption, so being examined by the failure of cold water group
It is disconnected, it finds in time and solves failure to heating ventilation air-conditioning system reliability service and energy saving be of great significance.
Water cooler is substantially largely the random system of non-linear, coupling, parameter time-varying, the course of work.For cold
Water dispenser group fault diagnosis has had a more complete characteristic parameter selection, including chilled water supply water temperature, condenser inflow temperature, cold
Condenser leaving water temperature, the evaporator temperature difference, the condenser temperature difference, supercooling temperature, refrigerator delivery temperature and compression case bottom temperature etc..
The practical fault diagnosis that carries out can select part parameter as fault signature variable according to specific operating condition, and fault diagnosis
Method determines the high efficiency and accuracy of water cooler fault diagnosis result to a greater extent.
All method for diagnosing faults can be divided into the method based on signal processing, the method based on analytic modell analytical model and base
In the method for knowledge.Since the mathematical model for carrying out fault diagnosis in field of heating ventilation air conditioning is extremely complex, and Knowledge based engineering side
Method has the characteristic for not needing mathematical models, this will have a good application prospect.
What is largely utilized currently based on the water cooler fault diagnosis of knowledge is neural network, and the overwhelming majority is using single
The method that one neural network and other technologies such as wavelet decomposition, fuzzy system, genetic algorithm etc. combine, it is the most frequently used among these
Be exactly BP neural network, these only certain point from feature-extraction analysis and network reference services problem make up BP net
The defect of network itself is difficult two that solve the problems, such as neural network simultaneously.
Summary of the invention
To solve the problems in background technique, complex nonlinear and BP nerve net for water cooler process data
Network itself pace of learning is slow and the deficiencies of easily falling into local optimum, and the present invention provides solve the above problems based on composite nerve
The water cooler method for diagnosing faults of network.
The present invention is achieved through the following technical solutions:
Water cooler method for diagnosing faults based on hybrid neural networks, includes the following steps:
Historical data under step 1, acquisition process of refrigerastion be normal and the operating condition that breaks down, and data are pre-processed,
The pretreatment includes missing values processing and outlier processing;
Step 2 removes the noise in data using wavelet transformation;
Historical data after wavelet transformation is divided into trained input data and training output data by step 3, after processing
Input variable of the fault signature variable as neural network, output variable of the water cooler operating condition variable as neural network;
Step 4 establishes RBF-BP hybrid production style, constantly trains network using inputoutput data, and optimize
Weight threshold parameter, until network convergence, obtains fault diagnosis model;
Step 5 carries out real-time fault diagnosis using the successful RBF-BP hybrid neural networks of training.
It is very widely used transform domain denoising method by wavelet transformation, can be kept away using wavelet transformation removal noise
Exempt to be lost with Fourier transformation removal noise bring signal, therefore noise, reduction number can be effectively removed using wavelet transformation
Useful signal in;Wavelet transformation be mainly characterized by gradually carrying out signal (function) by flexible shift operations it is multiple dimensioned
Refinement, is finally reached high frequency treatment time subdivision, and frequency is segmented at low frequency, can adapt to the requirement of time frequency signal analysis automatically, utilizes
To treated, historical data denoises wavelet transformation, and decomposition obtains eigenvectors matrix, extracts water cooler fault signature
Vector.In fault diagnosis field, can be rolled over using wavelet transformation removal noise to avoid with Fourier transformation denoising bring signal
Damage.RBF network has many advantages, such as that study is fast, can be avoided and fall into local optimum, therefore the present invention is mutual by two kinds of single networks
It is bonded a network, enhances the stability of fault diagnosis network, improves the convergence rate of network.
Further, the method that the processing of missing values described in step 1 uses data interpolation, the method choosing of the data interpolation
Use Newton interpolating method;
If the sampling time sequence of historical data is { t1,t2,…,tn, suitable interpolating function f is established using known point
(t), unknown point is by corresponding points xiFind out approximate functional value f (ti) replace;
Known n sampled point is to (t1,x1), (t2,x2) ..., (tn,xn) all scale quotient formula:
The above difference coefficient formula of simultaneous establishes following interpolation polynomial f (t):
F (t)=f (t1)+(t-t1)f[t2, t1]+(t-t1)(t-t2)f[t3, t2, t1]+
(t-t1)(t-t2)(t-t3)f[t4, t3, t2, t1]+……+
(t-t1)(t-t2)……(t-tn-1)f[tn, tn-1... ..., t2, t1]+
(t-t1)(t-t2)……(t-tn)f[tn, tn-1... ..., t1, t]
=P (t)+R (t)
Wherein:
P (t)=f (t1)+(t-t1)f[t2, t1]+(t-t1)(t-t2)f[t3, t2, t1]+
(t-t1)(t-t2)(t-t3)f[t4, t3, t2, t1]+……+
(t-t1)(t-t2)……(t-tn-1)f[tn, tn-1... ..., t2, t1]
R (t)=(t-t1)(t-t2)……(t-tn)f[tn, tn-1... ..., t1, t]
P (t) is Newton interpolation approximating function, and R (t) is error function, and the corresponding point t of the functional value of missing is substituted into interpolation
The approximation f (t) that multinomial is lacked.
Further, outlier processing described in step 1 is considered as missing values, at the method handled using missing values
Reason.
Further, described in step 2 using wavelet transformation removal data in noise be using small echo basic function and
The wavelet decomposition number of plies decomposes signals and associated noises, then threshold value quantizing processing is carried out to high-frequency signal, by low frequency signal and place
Signal is reconstructed in high-frequency signal after reason;
The basic function of small echo is:
Wherein:A is scale factor;B is time shift method;a-1/2Guarantee the different values to a, in stretching for wavelet function
Keep energy equal in journey;
The frequency domain presentation of wavelet function is as follows:
Further, in RBF-BP hybrid production style described in step 4, RBF neural uses Gaussian function
For radial basis function, expression is:
In formula:X is that m ties up training input vector;ckFor the center of k-th of radial basis function, with X dimension having the same;σk
For the radial sound stage width degree of k-th of hidden neuron;||X-c k||2Indicate X~ckBetween euclideam norm, with | | X-c k |
|2Increase, Φk(X) 0 can gradually be decayed to;
If the number of RBF network hidden neuron is K, then the output of radial basis function network is:
Wherein:W0 is deviation;wk(k=1,2 ... ..., K) is weight of the hidden layer to output layer;
If the number of BP neural network hidden neuron is P, then the output of BP network is:
In formula:X is that m ties up training input vector;w* k(k=1,2 ..., P) it is weight of the BP network hidden layer to output layer;b
Then remember that the output of RBF-BP hybrid neural networks is for the threshold value of BP neuron:
Be m peacekeeping target output T for given training sample X it is n dimension, defines output error:
ε=ti-y;
Using training data to the training of RBF-BP internet off-line, the weight and threshold value of network are optimized;By to institute
There is training sample to carry out the mean square error that neural network output is calculated, so that it is determined that the fitness of each individual;
If error function is:
Wherein:For the target output of v-th of output node of training sample u;For v-th of the output of training sample u
The reality output of node;
By several iterative calculation, RBF-BP neural network will reach global minima.
The present invention has the advantage that and beneficial effect:
1, the present invention is very widely used transform domain denoising method by wavelet transformation, is made an uproar using wavelet transformation removal
Sound can be lost to avoid with Fourier transformation removal noise bring signal, therefore can be effectively removed and be made an uproar using wavelet transformation
Sound, the useful signal in restoring data;Wavelet transformation is mainly characterized by through flexible shift operations to signal (function) gradually
Multi-scale refinement is carried out, high frequency treatment time subdivision is finally reached, frequency is segmented at low frequency, can adapt to time frequency signal analysis automatically
It is required that treated, historical data is denoised using wavelet transformation, decomposition obtains eigenvectors matrix, extracts water cooler
Fault feature vector;In fault diagnosis field, can be brought using wavelet transformation removal noise to avoid with Fourier transformation denoising
Signal lose;RBF network has many advantages, such as that study is fast, can be avoided and fall into local optimum, therefore the present invention is single by two kinds
Network, which be combined with each other, constitutes a network, enhances the stability of fault diagnosis network, improves the convergence rate of network;
2, the present invention uses RBF network hidden layer as a neural network in parallel with BP network, which is provided simultaneously with
The advantages of two kinds of networks, the training speed of network is fast, and avoids and fall into lacking for local minimum using single BP network
It falls into;The hybrid neural networks are applied in the fault diagnosis of water cooler, the efficiency of fault diagnosis and accurate can be greatly improved
Degree, also functions to the energy conservation of heating ventilation air-conditioning system very big effect.
Detailed description of the invention
Attached drawing described herein is used to provide to further understand the embodiment of the present invention, constitutes one of the application
Point, do not constitute the restriction to the embodiment of the present invention.In the accompanying drawings:
Attached drawing 1 is water cooler Troubleshooting Flowchart of the present invention.
Attached drawing 2 is 3 layers of decomposition tree structure diagram of small echo.
Attached drawing 3 is RBF-BP hybrid neural networks schematic diagram.
Attached drawing 4 is RBF-BP hybrid neural networks system construction drawing.
Attached drawing 5 (a) is diagnosis effect figure of the present invention to the 1 condenser fouling of water cooler failure.
Attached drawing 5 (b) is diagnosis regression accuracy figure of the present invention to the 1 condenser fouling of water cooler failure.
Attached drawing 6 (a) is the diagnosis effect figure that the present invention has on-condensible gas to water cooler failure 2.
Attached drawing 6 (b) is the diagnosis regression accuracy figure that the present invention has on-condensible gas to water cooler failure 2.
Attached drawing 7 (a) is the present invention to the excessive diagnosis effect figure of 3 refrigerant of water cooler failure.
Attached drawing 7 (b) is the present invention to the excessive diagnosis regression accuracy figure of 3 refrigerant of water cooler failure.
Attached drawing 8 (a) is diagnosis effect figure of the present invention to the leakage of 4 refrigerant of water cooler failure.
Attached drawing 8 (b) is diagnosis regression accuracy figure of the present invention to the leakage of 4 refrigerant of water cooler failure.
Attached drawing 9 (a) is the present invention to the excessive diagnosis effect figure of 5 lubricating oil of water cooler failure.
Attached drawing 9 (b) is the present invention to the excessive diagnosis regression accuracy figure of 5 lubricating oil of water cooler failure.
Specific embodiment
To make the objectives, technical solutions, and advantages of the present invention clearer, below with reference to embodiment and attached drawing, to this
Invention is described in further detail, and exemplary embodiment of the invention and its explanation for explaining only the invention, are not made
For limitation of the invention.
Embodiment 1
As shown in Figure 1, the water cooler method for diagnosing faults based on hybrid neural networks, includes the following steps:
Historical data under step 1, acquisition process of refrigerastion be normal and the operating condition that breaks down, and data are pre-processed,
The pretreatment includes missing values processing and outlier processing;
Step 2 removes the noise in data using wavelet transformation;
Historical data after wavelet transformation is divided into trained input data and training output data by step 3, after processing
Input variable of the fault signature variable as neural network, output variable of the water cooler operating condition variable as neural network;
Step 4 establishes RBF-BP hybrid production style, constantly trains network using inputoutput data, and optimize
Weight threshold parameter, until network convergence, obtains fault diagnosis model;
Step 5 carries out real-time fault diagnosis using the successful RBF-BP hybrid neural networks of training.
When implementation, the method that the processing of missing values described in step 1 uses data interpolation, the method selection of the data interpolation
Newton interpolating method;
If the sampling time sequence of historical data is { t1,t2,…,tn, suitable interpolating function f is established using known point
(t), unknown point is by corresponding points xiFind out approximate functional value f (ti) replace;
Known n sampled point is to (t1,x1), (t2,x2) ..., (tn,xn) all scale quotient formula:
The above difference coefficient formula of simultaneous establishes following interpolation polynomial f (t):
F (t)=f (t1)+(t-t1)f[t2, t1]+(t-t1)(t-t2)f[t3, t2, t1]+
(t-t1)(t-t2)(t-t3)f[t4, t3, t2, t1]+……+
(t-t1)(t-t2)……(t-tn-1)f[tn, tn-1... ..., t2, t1]+
(t-t1)(t-t2)……(t-tn)f[tn, tn-1... ..., t1, t]
=P (t)+R (t)
Wherein:
P (t)=f (t1)+(t-t1)f[t2, t1]+(t-t1)(t-t2)f[t3, t2, t1]+
(t-t1)(t-t2)(t-t3)f[t4, t3, t2, t1]+……+
(t-t1)(t-t2)……(t-tn-1)f[tn, tn-1... ..., t2, t1]
R (t)=(t-t1)(t-t2)……(t-tn)f[tn, tn-1... ..., t1, t]
P (t) is Newton interpolation approximating function, and R (t) is error function, and the corresponding point t of the functional value of missing is substituted into interpolation
The approximation f (t) that multinomial is lacked.
Outlier processing described in step 1 is considered as missing values, is handled using the method that missing values are handled.
Using the basic function and wavelet decomposition layer that the noise in wavelet transformation removal data is using small echo described in step 2
Several pairs of signals and associated noises decompose, and then carry out threshold value quantizing processing to high-frequency signal, by low frequency signal and treated high frequency
Signal is reconstructed in signal;
The basic function of small echo is:
Wherein:A is scale factor;B is time shift method;a-1/2Guarantee the different values to a, in stretching for wavelet function
Keep energy equal in journey;
The frequency domain presentation of wavelet function is as follows:
In RBF-BP hybrid production style described in step 4, RBF neural is radial base letter using Gaussian function
Number, expression are:
In formula:X is that m ties up training input vector;ckFor the center of k-th of radial basis function, with X dimension having the same;σk
For the radial sound stage width degree of k-th of hidden neuron;||X-c k||2Indicate X~ckBetween euclideam norm, with | | X-c k |
|2Increase, Φk(X) 0 can gradually be decayed to;
If the number of RBF network hidden neuron is K, then the output of radial basis function network is:
Wherein:W0 is deviation;wk(k=1,2 ... ..., K) is weight of the hidden layer to output layer;
If the number of BP neural network hidden neuron is P, then the output of BP network is:
In formula:X is that m ties up training input vector;w* k(k=1,2 ..., P) it is weight of the BP network hidden layer to output layer;b
Then remember that the output of RBF-BP hybrid neural networks is for the threshold value of BP neuron:
For given training sample X (m dimension) and target output T (n dimension), output error is defined:
ε=ti-y;
Using training data to the training of RBF-BP internet off-line, the weight and threshold value of network are optimized;By to institute
There is training sample to carry out the mean square error that neural network output is calculated, so that it is determined that the fitness of each individual;
If error function is:
Wherein:For the target output of v-th of output node of training sample u;For v-th of the output of training sample u
The reality output of node;
By several iterative calculation, RBF-BP neural network will reach global minima.
Data acquisition intervals are 5min, are got parms 36, are that sensor directly measures, including temperature, pressure, valve
Position, flow, compressor horsepower and electric current, water cooler alarm etc..This experiment filters out 8 main shadows from 36 initial data
The feature for ringing core coolant circulation, as the characteristic parameter of fault diagnosis, monitoring process operation conditions.
The Fault characteristic parameters of water cooler are the input data of network, as shown in table 1.
The Fault characteristic parameters of 1 water cooler of table
Serial number | Fault characteristic parameters |
1 | Chilled water supply water temperature |
2 | Condenser inflow temperature |
3 | Leaving condenser water temperature |
4 | The evaporator temperature difference |
5 | The condenser temperature difference |
6 | Supercooling temperature |
7 | Refrigerant discharge temperature |
8 | Compression case bottom temperature |
When water cooler is run, save historical data using sensor measurement data, and by database, when using data
Directly called from database.The feelings for shortage of data and data exception occur are possible to during measuring or saving data
Condition, in response to this, we should carry out pretreatment operation to data first.It is different using rejecting here for abnormal data
Constant value, and handle as missing values.
Missing values are handled using data interpolation method, select Newton interpolating method here, with inheritedness and are easy to change
The characteristics of node is substantially exactly the functional value that missing values are obtained by construction interpolation polynomial.
Original signal is influenced by various complicated factors, so that general in acquisition signal all contain a large amount of noise, is covered
The characteristic information of signal is covered.Fault characteristic parameters are chosen, is decomposed using 3 layers of db3 wavelet basis function and is used as Wavelet Denoising Method.It is small
3 layers of wave basic function decomposition tree construction are as shown in Fig. 2, to extract useful original signal.Will carry out wavelet transformation after
Input data of the data as neural network has more accurate diagnosis effect to training network.Training data is divided into defeated
Enter data and output data, wherein input of the water cooler Fault characteristic parameters as neural network, operating condition is nerve
The output of network, main fault mode is as shown in table 2,
The parameters such as suitable weight and threshold value are set, RBF-BP hybrid production style is established, as shown in figure 3, passing through
RBF network hidden layer as hybrid neural networks in parallel with BP network, BP e-learning speed is made up using RBF network
Slowly, the deficiencies of easily falling into local minimum, and by the data training neural network after processing and denoising, to obtain convergent failure
Diagnostic model.The output of RBF-BP hybrid neural networks is:
It is examined using the real time fail that the established RBF-BP hybrid neural networks based on wavelet transformation carry out water cooler
Disconnected, Fig. 4 is RBF-BP hybrid neural networks system construction drawing.Wherein δ is that the RBF-BPNN based on wavelet transformation exports Y and is controlled
The residual error of object reality output Y*.When no fault occurs, system residual epsilon<δ will regard residual signals as noise processed at this time;
When system jam, i.e. residual epsilon>δ cannot be eliminated simply with the mode of de-noising, at this time can according to the threshold value of setting
Judge that failure has occurred in controlled device.
The present invention is verified by MATLAB modeling.Respectively 5 kinds of typical faults are carried out with the experiment of fault diagnosis, modeling
Training sample be 550 groups, test data be 150 groups, wherein preceding 50 groups of data be normal data, introduced since the 51st data
Failure, i.e., rear 100 groups of data are fault data.Obtained experimental result such as Fig. 5 to Fig. 9, curve are test value, and wherein * is indicated
Actual output, zero indicates the output of test.Output valve is that 0 expression water cooler operates normally, and output valve is 1 expression cooling-water machine
Group breaks down.
Can be seen that from 5 kinds of fault diagnosis effects to the rate of failing to report of normal data is 0%, it is seen that the present invention is to cold water
Unit has accurate testing result when working normally, and effectively reduces in process detection and fails to report generation.
Fig. 5 (a) -5 (b) is respectively the present invention to the diagnosis effect figure of 1 condenser fouling of water cooler failure and returns essence
Degree figure.It is respectively 99% and 90% that the training regression accuracy of the method for the present invention and survey, which set regression accuracy, it can be seen from the figure that originally
Inventive method can accurately diagnose the Condenser fouling fault of water cooler.
Fig. 6 (a) -6 (b) is respectively that the present invention has the diagnosis effect figure of on-condensible gas to water cooler failure 2 and returns essence
Degree figure.It is respectively 99% and 90% that the training regression accuracy of the method for the present invention and survey, which set regression accuracy, it can be seen from the figure that originally
What inventive method can accurately diagnose water cooler has on-condensible gas failure.
Fig. 7 (a) -7 (b) is respectively the present invention to the excessive diagnosis effect figure of 3 refrigerant of water cooler failure and returns essence
Degree figure.It is respectively 98% and 99% that the training regression accuracy of the method for the present invention and survey, which set regression accuracy, it can be seen from the figure that originally
Inventive method energy and its refrigerant excess failure for accurately diagnosing water cooler.
Fig. 8 (a) -8 (b) is respectively the present invention to the diagnosis effect figure of 4 refrigerant of water cooler failure leakage and returns essence
Degree figure.It is respectively 99% and 95% that the training regression accuracy of the method for the present invention and survey, which set regression accuracy, it can be seen from the figure that originally
Inventive method can accurately diagnose the refrigerant leakage failure of water cooler.
Fig. 9 (a) -9 (b) is respectively the present invention to the excessive diagnosis effect figure of 5 lubricating oil of water cooler failure and returns essence
Degree figure.It is respectively 98% and 92% that the training regression accuracy of the method for the present invention and survey, which set regression accuracy, it can be seen from the figure that originally
Inventive method can accurately diagnose the lubricating oil excess failure of water cooler.
In order to illustrate the detection effect of the method for the present invention vividerly, 5 kinds of fault data test effects are subjected to list and are said
It is bright, as shown in table 3:
The fault mode of 3 water cooler of table
Using RBF network hidden layer as a neural network in parallel with BP network, which has been provided simultaneously with two kinds of nets
The training speed of the advantages of network, network are fast, and avoid the defect that local minimum is fallen into using single BP network;This is mixed
Application of Neural Network is closed into the fault diagnosis of water cooler, the Efficiency and accuracy of fault diagnosis can be greatly improved, for warm
The energy conservation of logical air-conditioning system also functions to very big effect.
Above-described specific embodiment has carried out further the purpose of the present invention, technical scheme and beneficial effects
It is described in detail, it should be understood that being not intended to limit the present invention the foregoing is merely a specific embodiment of the invention
Protection scope, all within the spirits and principles of the present invention, any modification, equivalent substitution, improvement and etc. done should all include
Within protection scope of the present invention.
Claims (5)
1. the water cooler method for diagnosing faults based on hybrid neural networks, which is characterized in that include the following steps:
Historical data under step 1, acquisition process of refrigerastion be normal and the operating condition that breaks down, and data are pre-processed, it is described
Pretreatment includes missing values processing and outlier processing;
Step 2 removes the noise in data using wavelet transformation;
Step 3, the historical data after wavelet transformation is divided into trained input data and training output data, will treated therefore
Hinder input variable of the characteristic variable as neural network, output variable of the water cooler operating condition variable as neural network;
Step 4 establishes RBF-BP hybrid production style, constantly trains network using inputoutput data, and optimize weight
Threshold parameter, until network convergence, obtains fault diagnosis model;
Step 5 carries out real-time fault diagnosis using the successful RBF-BP hybrid neural networks of training.
2. the water cooler method for diagnosing faults according to claim 1 based on hybrid neural networks, which is characterized in that step
The method that the processing of missing values described in rapid 1 uses data interpolation, the method selection Newton interpolating method of the data interpolation;
If the sampling time sequence of historical data is { t1, t2..., tn, interpolating function f (t), unknown point are established using known point
By corresponding points xiFind out functional value f (ti) replace;
Known n sampled point is to (t1, x1), (t2, x2) ..., (tn, xn) all scale quotient formula:
……
The above difference coefficient formula of simultaneous establishes following interpolation polynomial f (t):
F (t)=f (t1)+(t-t1)f[t2, t1]+(t-t1)(t-t2)f[t3, t2, t1]+
(t-t1)(t-t2)(t-t3)f[t4, t3, t2, t1]+……+
(t-t1)(t-t2)……(t-tn-1)f[tn, tn-1... ..., t2, t1]+
(t-t1)(t-t2)……(t-tn)f[tn, tn-1... ..., t1, t]
=P (t)+R (t)
Wherein:
P (t)=f (t1)+(t-t1)f[t2, t1]+(t-t1)(t-t2)f[t3, t2, t1]+
(t-t1)(t-t2)(t-t3)f[t4, t3, t2, t1]+……+
(t-t1)(t-t2)……(t-tn-1)f[tn, tn-1... ..., t2, t1]
R (t)=(t-t1)(t-t2)……(t-tn)f[tn, tn-1... ..., t1, t]
P (t) is Newton interpolation approximating function, and R (t) is error function, and the corresponding point t of the functional value of missing is substituted into interpolation polynomial
The approximation f (t) that formula is lacked.
3. the water cooler method for diagnosing faults according to claim 2 based on hybrid neural networks, which is characterized in that step
Outlier processing described in rapid 1 is considered as missing values, is handled using the method that missing values are handled.
4. the water cooler method for diagnosing faults according to claim 1 based on hybrid neural networks, which is characterized in that step
Removing the noise in data using wavelet transformation described in rapid 2 is basic function and the wavelet decomposition number of plies using small echo to noisy letter
It number is decomposed, threshold value quantizing processing then is carried out to high-frequency signal, by low frequency signal and treated that high-frequency signal carries out weight
Structure signal;
The basic function of small echo is:
Wherein:A is scale factor;B is time shift method;a-1/2Guarantee the different values to a, in the telescopic process of wavelet function
Keep energy equal;
The frequency domain presentation of wavelet function is as follows:
5. the water cooler method for diagnosing faults according to claim 1 based on hybrid neural networks, which is characterized in that step
In RBF-BP hybrid production style described in rapid 4, RBF neural is radial basis function using Gaussian function, specific
Expression formula is:
In formula:X is that m ties up training input vector;ckFor the center of k-th of radial basis function, with X dimension having the same;σkIt is
The radial sound stage width degree of k hidden neuron;||X-ck||2Indicate X~ckBetween euclideam norm, with | | X-ck||2Increasing
Greatly, Φk(X) 0 can gradually be decayed to;
If the number of RBF network hidden neuron is K, then the output of radial basis function network is:
Wherein:w0For deviation;wk(k=1,2 ... ..., K) is weight of the hidden layer to output layer;
If the number of BP neural network hidden neuron is P, then the output of BP network is:
In formula:X is that m ties up training input vector;w* k(k=1,2 ... ..., P) is weight of the BP network hidden layer to output layer;B is BP
The threshold value of neuron then remembers that the output of RBF-BP hybrid neural networks is:
Be m peacekeeping target output T for given training sample X it is n dimension, defines output error:
ε=ti-y;
Using training data to the training of RBF-BP internet off-line, the weight and threshold value of network are optimized;By to all instructions
Practice sample and carry out the mean square error that neural network output is calculated, so that it is determined that the fitness of each individual;
If error function is:
Wherein:For the target output of v-th of output node of training sample u;For v-th output node of training sample u
Reality output;
By several iterative calculation, RBF-BP neural network will reach global minima.
Priority Applications (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN201810724449.0A CN108898120A (en) | 2018-07-04 | 2018-07-04 | Water cooler method for diagnosing faults based on hybrid neural networks |
Applications Claiming Priority (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN201810724449.0A CN108898120A (en) | 2018-07-04 | 2018-07-04 | Water cooler method for diagnosing faults based on hybrid neural networks |
Publications (1)
Publication Number | Publication Date |
---|---|
CN108898120A true CN108898120A (en) | 2018-11-27 |
Family
ID=64347844
Family Applications (1)
Application Number | Title | Priority Date | Filing Date |
---|---|---|---|
CN201810724449.0A Pending CN108898120A (en) | 2018-07-04 | 2018-07-04 | Water cooler method for diagnosing faults based on hybrid neural networks |
Country Status (1)
Country | Link |
---|---|
CN (1) | CN108898120A (en) |
Cited By (10)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN109977621A (en) * | 2019-04-30 | 2019-07-05 | 西南石油大学 | A kind of air-conditioning failure prediction method based on deep learning |
CN110333093A (en) * | 2019-07-19 | 2019-10-15 | 东莞市泰丰空调制冷设备有限公司 | A kind of water cooler test platform |
CN110779745A (en) * | 2019-10-12 | 2020-02-11 | 杭州安脉盛智能技术有限公司 | Heat exchanger early fault diagnosis method based on BP neural network |
CN111310781A (en) * | 2019-12-30 | 2020-06-19 | 华南理工大学 | Refrigeration equipment refrigerant leakage identification method based on PCA-ANN |
CN111504673A (en) * | 2020-03-27 | 2020-08-07 | 青岛海信日立空调系统有限公司 | Fault diagnosis method and system for water chilling unit and air conditioner |
CN111539516A (en) * | 2020-04-22 | 2020-08-14 | 谭雄向 | Power grid fault diagnosis system and method based on big data processing |
CN112364999A (en) * | 2020-10-19 | 2021-02-12 | 深圳市超算科技开发有限公司 | Training method and device for water chiller adjustment model and electronic equipment |
CN112395946A (en) * | 2020-10-20 | 2021-02-23 | 宋晓健 | Mud pulse signal denoising method based on wavelet transform-RBF neural network |
CN112976332A (en) * | 2020-12-24 | 2021-06-18 | 中山艾尚智同信息科技有限公司 | Neural network based method for predicting rheological property of ready-mixed concrete |
CN113761792A (en) * | 2021-08-03 | 2021-12-07 | 三峡大学 | Six-pulse rectifier fault diagnosis method based on neural network |
Citations (8)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN103208091A (en) * | 2013-04-25 | 2013-07-17 | 国家电网公司 | Electric larceny preventing method based on data mining of electric load management system |
CN104633829A (en) * | 2013-11-06 | 2015-05-20 | 上海思控电气设备有限公司 | Building cooling station energy-saving control device and method thereof |
CN104751260A (en) * | 2013-12-30 | 2015-07-01 | 通用电气公司 | System and method for identifying abnormal reading of meter |
CN105913450A (en) * | 2016-06-22 | 2016-08-31 | 武汉理工大学 | Tire rubber carbon black dispersity evaluation method and system based on neural network image processing |
CN106931595A (en) * | 2017-03-09 | 2017-07-07 | 南京福致通电气自动化有限公司 | A kind of central air-conditioning coefficient of performance on-line analysis |
CN106971027A (en) * | 2017-03-09 | 2017-07-21 | 西安建筑科技大学 | A kind of handpiece Water Chilling Units fault signature system of selection based on DR BN models |
CN107831866A (en) * | 2017-11-27 | 2018-03-23 | 国家计算机网络与信息安全管理中心 | A kind of water-cooling system state evaluating method, equipment and computer-readable recording medium |
CN107844799A (en) * | 2017-10-17 | 2018-03-27 | 西安建筑科技大学 | A kind of handpiece Water Chilling Units method for diagnosing faults of integrated SVM mechanism |
-
2018
- 2018-07-04 CN CN201810724449.0A patent/CN108898120A/en active Pending
Patent Citations (8)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN103208091A (en) * | 2013-04-25 | 2013-07-17 | 国家电网公司 | Electric larceny preventing method based on data mining of electric load management system |
CN104633829A (en) * | 2013-11-06 | 2015-05-20 | 上海思控电气设备有限公司 | Building cooling station energy-saving control device and method thereof |
CN104751260A (en) * | 2013-12-30 | 2015-07-01 | 通用电气公司 | System and method for identifying abnormal reading of meter |
CN105913450A (en) * | 2016-06-22 | 2016-08-31 | 武汉理工大学 | Tire rubber carbon black dispersity evaluation method and system based on neural network image processing |
CN106931595A (en) * | 2017-03-09 | 2017-07-07 | 南京福致通电气自动化有限公司 | A kind of central air-conditioning coefficient of performance on-line analysis |
CN106971027A (en) * | 2017-03-09 | 2017-07-21 | 西安建筑科技大学 | A kind of handpiece Water Chilling Units fault signature system of selection based on DR BN models |
CN107844799A (en) * | 2017-10-17 | 2018-03-27 | 西安建筑科技大学 | A kind of handpiece Water Chilling Units method for diagnosing faults of integrated SVM mechanism |
CN107831866A (en) * | 2017-11-27 | 2018-03-23 | 国家计算机网络与信息安全管理中心 | A kind of water-cooling system state evaluating method, equipment and computer-readable recording medium |
Non-Patent Citations (3)
Title |
---|
李仿华 等: ""基于遗传优化的RBF-BP网络的实时故障检测"", 《微型机与应用》 * |
王婷婷: ""供水管网系统DMA分区流量数据聚类分析研究"", 《中国硕士学位论文全文数据库 工程科技Ⅱ辑 》 * |
石书彪 等: ""基于小波去噪和神经网络的冷水机组故障诊断"", 《制冷学报》 * |
Cited By (12)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN109977621A (en) * | 2019-04-30 | 2019-07-05 | 西南石油大学 | A kind of air-conditioning failure prediction method based on deep learning |
CN110333093A (en) * | 2019-07-19 | 2019-10-15 | 东莞市泰丰空调制冷设备有限公司 | A kind of water cooler test platform |
CN110779745A (en) * | 2019-10-12 | 2020-02-11 | 杭州安脉盛智能技术有限公司 | Heat exchanger early fault diagnosis method based on BP neural network |
CN110779745B (en) * | 2019-10-12 | 2021-07-06 | 杭州安脉盛智能技术有限公司 | Heat exchanger early fault diagnosis method based on BP neural network |
CN111310781A (en) * | 2019-12-30 | 2020-06-19 | 华南理工大学 | Refrigeration equipment refrigerant leakage identification method based on PCA-ANN |
CN111504673A (en) * | 2020-03-27 | 2020-08-07 | 青岛海信日立空调系统有限公司 | Fault diagnosis method and system for water chilling unit and air conditioner |
CN111539516A (en) * | 2020-04-22 | 2020-08-14 | 谭雄向 | Power grid fault diagnosis system and method based on big data processing |
CN112364999A (en) * | 2020-10-19 | 2021-02-12 | 深圳市超算科技开发有限公司 | Training method and device for water chiller adjustment model and electronic equipment |
CN112395946A (en) * | 2020-10-20 | 2021-02-23 | 宋晓健 | Mud pulse signal denoising method based on wavelet transform-RBF neural network |
CN112976332A (en) * | 2020-12-24 | 2021-06-18 | 中山艾尚智同信息科技有限公司 | Neural network based method for predicting rheological property of ready-mixed concrete |
CN113761792A (en) * | 2021-08-03 | 2021-12-07 | 三峡大学 | Six-pulse rectifier fault diagnosis method based on neural network |
CN113761792B (en) * | 2021-08-03 | 2024-01-19 | 三峡大学 | Neural network-based six-pulse rectifier fault diagnosis method |
Similar Documents
Publication | Publication Date | Title |
---|---|---|
CN108898120A (en) | Water cooler method for diagnosing faults based on hybrid neural networks | |
CN110555479B (en) | Fault feature learning and classifying method based on 1DCNN and GRU fusion | |
CN109446187B (en) | Method for monitoring health state of complex equipment based on attention mechanism and neural network | |
CN109522600B (en) | Complex equipment residual service life prediction method based on combined deep neural network | |
CN108334059B (en) | Particle swarm optimization-based BP neural network model fault diagnosis method | |
CN108931387B (en) | Fault diagnosis method based on multi-sensor signal analysis and capable of providing accurate diagnosis decision | |
CN112257530B (en) | Rolling bearing fault diagnosis method based on blind signal separation and support vector machine | |
CN107844799B (en) | Water chilling unit fault diagnosis method of integrated SVM (support vector machine) mechanism | |
Volponi et al. | The use of Kalman filter and neural network methodologies in gas turbine performance diagnostics: a comparative study | |
CN108804740B (en) | Long-distance pipeline pressure monitoring method based on integrated improved ICA-KRR algorithm | |
CN107992665B (en) | Online fault diagnosis and analysis method for alternating current filter of extra-high voltage converter station | |
CN101887479B (en) | Rapid diagnosis method for rotating stall of axial flow compressor based on determined learning theory | |
CN108363844B (en) | Method for predicting exhaust temperature in starting process of aircraft engine | |
CN114061957A (en) | Health assessment method for main bearing of diesel engine | |
CN112461543B (en) | Rotary machine fault diagnosis method based on multi-classification support vector data description | |
CN113538353B (en) | Five-phase asynchronous motor rolling bearing fault diagnosis method based on single-channel diagram data enhancement and migration training residual error network | |
CN110702438B (en) | Centrifugal water chilling unit and fresh air system combined fault diagnosis method | |
CN116720150A (en) | Mechanical refrigeration system fault diagnosis method and system | |
Liao et al. | The fault diagnosis for centrifugal compressor based on time series analysis with neutral network | |
CN113537010B (en) | Fifteen-phase asynchronous motor rolling bearing fault diagnosis method based on single-channel diagram data enhancement and migration training residual error network | |
CN107622251A (en) | A kind of aircraft fuel pump signal degradation feature extracting method and device | |
CN115014748A (en) | Fault diagnosis method for seed cotton sorting spray valve | |
CN113409213A (en) | Plunger pump fault signal time-frequency graph noise reduction enhancement method and system | |
Tian et al. | Fault diagnosis method for inter-shaft bearings based on information exergy and random forest | |
CN108984824B (en) | Method for detecting air leakage rate of annular cooling machine |
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 | ||
RJ01 | Rejection of invention patent application after publication |
Application publication date: 20181127 |
|
RJ01 | Rejection of invention patent application after publication |