CN115859058A - UPS (uninterrupted Power supply) fault prediction method and system based on width learning network - Google Patents
UPS (uninterrupted Power supply) fault prediction method and system based on width learning network Download PDFInfo
- Publication number
- CN115859058A CN115859058A CN202310166830.0A CN202310166830A CN115859058A CN 115859058 A CN115859058 A CN 115859058A CN 202310166830 A CN202310166830 A CN 202310166830A CN 115859058 A CN115859058 A CN 115859058A
- Authority
- CN
- China
- Prior art keywords
- ups
- data
- value
- width learning
- algorithm
- 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
- 238000000034 method Methods 0.000 title claims abstract description 78
- 238000004422 calculation algorithm Methods 0.000 claims abstract description 78
- 230000002068 genetic effect Effects 0.000 claims abstract description 50
- 238000003745 diagnosis Methods 0.000 claims abstract description 19
- 230000035772 mutation Effects 0.000 claims abstract description 17
- 238000012545 processing Methods 0.000 claims abstract description 11
- 238000012544 monitoring process Methods 0.000 claims abstract description 9
- 238000002922 simulated annealing Methods 0.000 claims abstract description 9
- 238000010845 search algorithm Methods 0.000 claims abstract description 6
- 239000011159 matrix material Substances 0.000 claims description 44
- 230000006870 function Effects 0.000 claims description 27
- 230000008569 process Effects 0.000 claims description 19
- 239000013598 vector Substances 0.000 claims description 18
- 210000000349 chromosome Anatomy 0.000 claims description 13
- 238000010606 normalization Methods 0.000 claims description 13
- 238000012549 training Methods 0.000 claims description 13
- 238000013528 artificial neural network Methods 0.000 claims description 12
- 238000000354 decomposition reaction Methods 0.000 claims description 12
- 238000005457 optimization Methods 0.000 claims description 12
- 108090000623 proteins and genes Proteins 0.000 claims description 7
- 238000000605 extraction Methods 0.000 claims description 6
- 238000013507 mapping Methods 0.000 claims description 6
- 238000012360 testing method Methods 0.000 claims description 6
- 230000005856 abnormality Effects 0.000 claims description 4
- 230000003190 augmentative effect Effects 0.000 claims description 4
- 210000004027 cell Anatomy 0.000 claims description 4
- 238000007781 pre-processing Methods 0.000 claims description 4
- 230000003416 augmentation Effects 0.000 claims description 3
- 238000004364 calculation method Methods 0.000 claims description 3
- 239000000203 mixture Substances 0.000 claims description 3
- 238000012216 screening Methods 0.000 claims description 3
- 238000007619 statistical method Methods 0.000 claims description 3
- 238000002759 z-score normalization Methods 0.000 claims description 3
- 239000000284 extract Substances 0.000 claims description 2
- 230000002708 enhancing effect Effects 0.000 claims 1
- 238000012423 maintenance Methods 0.000 abstract description 15
- 230000007547 defect Effects 0.000 abstract description 3
- 238000010586 diagram Methods 0.000 description 14
- 230000008901 benefit Effects 0.000 description 5
- 230000008859 change Effects 0.000 description 5
- 230000032683 aging Effects 0.000 description 4
- 238000013473 artificial intelligence Methods 0.000 description 4
- 238000005516 engineering process Methods 0.000 description 4
- 230000002159 abnormal effect Effects 0.000 description 3
- 238000004590 computer program Methods 0.000 description 3
- 230000000694 effects Effects 0.000 description 2
- 230000036541 health Effects 0.000 description 2
- 238000013021 overheating Methods 0.000 description 2
- 230000000737 periodic effect Effects 0.000 description 2
- 230000009467 reduction Effects 0.000 description 2
- 230000005634 sigma model Effects 0.000 description 2
- 230000009466 transformation Effects 0.000 description 2
- WHXSMMKQMYFTQS-UHFFFAOYSA-N Lithium Chemical compound [Li] WHXSMMKQMYFTQS-UHFFFAOYSA-N 0.000 description 1
- 230000003213 activating effect Effects 0.000 description 1
- 230000001154 acute effect Effects 0.000 description 1
- 230000005540 biological transmission Effects 0.000 description 1
- 238000006243 chemical reaction Methods 0.000 description 1
- 230000002759 chromosomal effect Effects 0.000 description 1
- 238000003066 decision tree Methods 0.000 description 1
- 238000013135 deep learning Methods 0.000 description 1
- 238000009795 derivation Methods 0.000 description 1
- 230000006866 deterioration Effects 0.000 description 1
- 238000004146 energy storage Methods 0.000 description 1
- 230000007613 environmental effect Effects 0.000 description 1
- 238000011156 evaluation Methods 0.000 description 1
- 229910052739 hydrogen Inorganic materials 0.000 description 1
- 239000001257 hydrogen Substances 0.000 description 1
- 125000004435 hydrogen atom Chemical class [H]* 0.000 description 1
- 230000006872 improvement Effects 0.000 description 1
- 238000007689 inspection Methods 0.000 description 1
- 229910052744 lithium Inorganic materials 0.000 description 1
- 238000010801 machine learning Methods 0.000 description 1
- 238000005259 measurement Methods 0.000 description 1
- 238000012986 modification Methods 0.000 description 1
- 230000004048 modification Effects 0.000 description 1
- 230000003287 optical effect Effects 0.000 description 1
- 239000002245 particle Substances 0.000 description 1
- 230000002028 premature Effects 0.000 description 1
- 230000002035 prolonged effect Effects 0.000 description 1
- 238000011160 research Methods 0.000 description 1
- 239000003381 stabilizer Substances 0.000 description 1
- 230000000087 stabilizing effect Effects 0.000 description 1
- 230000003068 static effect Effects 0.000 description 1
- 238000012706 support-vector machine Methods 0.000 description 1
- 238000000844 transformation Methods 0.000 description 1
- 238000012795 verification Methods 0.000 description 1
Images
Classifications
-
- Y—GENERAL TAGGING OF NEW TECHNOLOGICAL DEVELOPMENTS; GENERAL TAGGING OF CROSS-SECTIONAL TECHNOLOGIES SPANNING OVER SEVERAL SECTIONS OF THE IPC; TECHNICAL SUBJECTS COVERED BY FORMER USPC CROSS-REFERENCE ART COLLECTIONS [XRACs] AND DIGESTS
- Y04—INFORMATION OR COMMUNICATION TECHNOLOGIES HAVING AN IMPACT ON OTHER TECHNOLOGY AREAS
- Y04S—SYSTEMS INTEGRATING TECHNOLOGIES RELATED TO POWER NETWORK OPERATION, COMMUNICATION OR INFORMATION TECHNOLOGIES FOR IMPROVING THE ELECTRICAL POWER GENERATION, TRANSMISSION, DISTRIBUTION, MANAGEMENT OR USAGE, i.e. SMART GRIDS
- Y04S10/00—Systems supporting electrical power generation, transmission or distribution
- Y04S10/50—Systems or methods supporting the power network operation or management, involving a certain degree of interaction with the load-side end user applications
Landscapes
- Supply And Distribution Of Alternating Current (AREA)
- Power Sources (AREA)
- Management, Administration, Business Operations System, And Electronic Commerce (AREA)
Abstract
The application discloses a UPS (uninterrupted power supply) fault prediction method and system based on a width learning network, wherein the method comprises the following steps: historical data of the UPS battery in different states are obtained; determining the value ranges of the parameters of the feature nodes, the enhanced nodes and the feature window number in the width learning network; introducing a simulated annealing algorithm, and improving mutation operation in the genetic algorithm to obtain an improved genetic algorithm; optimizing the width learning parameters through an improved genetic algorithm; determining the threshold range of the normal battery of the UPS according to a grid search algorithm and a K-sigma model; and (4) comprehensively using GA + BLS + K-sigma to obtain a final prediction and diagnosis model, and monitoring the state of the UPS power supply battery in real time. The method and the device solve the defects in the conventional operation and maintenance processing mode of the UPS battery, and the failure prediction method can give consideration to efficiency and accuracy, reduce the maintenance cost and the accidental shutdown time of equipment, and meet the high-reliability requirement of UPS failure prediction.
Description
Technical Field
The invention belongs to the technical field of fault prediction and health management, and particularly relates to a UPS fault prediction method and system based on a width learning network.
Background
The UPS is an energy storage device that uses an inverter as a main component and provides a constant voltage and constant frequency uninterruptible power supply, and is mainly used to provide uninterrupted power supply for a single computer, a computer network system, or other power electronic devices. When the mains supply input is normal, the UPS supplies the mains supply to the load for use after stabilizing the voltage of the mains supply, and the UPS is an alternating current mains supply voltage stabilizer and also charges a built-in battery; when the commercial power is interrupted (power failure in accident), the UPS immediately supplies 220V alternating current to the load by the method of inversion conversion, so that the load can maintain normal work and the software and hardware of the load can be protected from being damaged.
The main manifestations of UPS power failure are: three important indexes of voltage, current and temperature deviate from the normal range, and phenomena of overcharge, overdischarge, overheating, aging and the like occur. The main modes of operation and maintenance of the conventional UPS power supply are periodic equipment inspection, periodic maintenance and after-repair, which can bring about excessive maintenance due to safety margin on one hand and cause acute deterioration of battery performance due to under-consideration of dynamic operation performance, resulting in unexpected consequences on the other hand. Therefore, some artificial intelligence methods based on data driving are used by scientific research personnel for monitoring the performance of the battery, predicting the fault of the battery and evaluating the health state of the battery in real time so as to realize the optional maintenance of the UPS battery, reduce the maintenance cost and reduce the accidental shutdown time of equipment. Such as: and fault diagnosis is carried out on the UPS storage battery by utilizing multi-model particle filter algorithm diagnosis, and machine learning algorithms such as a support vector machine, a deep neural network, a decision tree and the like are utilized. The algorithms are complex in structure, and time is consumed when a large amount of data is processed; in addition to the particularity of UPS application, the requirement of high reliability in prediction and evaluation is needed, and an artificial intelligence algorithm which can give consideration to both efficiency and accuracy is urgently needed to be found.
Compared with a deep neural network, the width learning emphasizes the width, the data features are processed in parallel, and when huge data needs to be processed, the defects that the most popular deep neural network at present is long in calculation time, complex in structure, high in accuracy and the like are well overcome. A satisfactory result is obtained on the basis of solving the problems of classification and fault diagnosis, but the selection of the number of the feature nodes, the enhanced nodes and the feature windows in the width learning network does not have a universal mature standard, and the selection can only be carried out by depending on manual experience, so that the accuracy of the width learning network cannot be guaranteed.
Disclosure of Invention
The invention aims to provide a UPS (uninterrupted power supply) fault prediction system based on a width learning network, which aims to solve the problems in the prior art in the background technology.
In order to achieve the purpose, the invention adopts the following technical scheme:
a UPS failure prediction method based on a width learning network comprises the following steps:
performing operations such as data preprocessing, normalization, wavelet packet characteristic extraction and the like on historical data of the UPS battery in different states, and dividing the processed data into a training set and a test set;
determining the value ranges of characteristic nodes, enhanced nodes and characteristic window number parameters in the width learning network;
introducing a simulated annealing algorithm, and improving the mutation operation in the genetic algorithm to obtain an improved genetic algorithm;
optimizing the width learning parameters through an improved genetic algorithm, and outputting width learning parameter values corresponding to the optimal solution;
determining the threshold range of the normal battery of the UPS according to a grid search algorithm and a K-sigma model;
and (4) comprehensively using GA + BLS + K-sigma to obtain a final prediction and diagnosis model, and monitoring the state of the UPS power supply battery in real time.
On the other hand, the invention provides a UPS failure prediction system based on a width learning network, comprising: a data processing module for expressing the pattern vector of the original feature space of high dimensionality with the new pattern vector of the feature space of low dimensionality, thereby finding the most representative and most effective feature method; the width learning system module is used for obtaining a network model according to the weight value of the neural network of which the inverse matrix obtained by the method is equivalent to; the genetic algorithm module is used for optimizing parameters of the network model so as to optimize an objective function; the threshold early warning module is established based on K-sigma of a grid search method, the value of a single data item can be compared with an early warning value according to the data of each battery cell of the UPS, such as current, voltage, temperature and the like collected in real time, and when the current, the voltage and the temperature exceed the standard, abnormality is found in time and the UPS is positioned for fault warning.
The method comprises a wavelet packet decomposition algorithm, wherein the wavelet packet decomposition algorithm is used for extracting high-frequency components in signals for screening, can effectively filter high-frequency noise of load signals, and extracts residual signal characteristics, and comprises the following steps: carrying out normalization processing on the UPS vertical data signals; performing wavelet packet 3 layer decomposition on the processed data to obtain each frequency band component, and eliminating a noise component; and reconstructing a data signal by using the wavelet packet coefficient, and then extracting the energy characteristic of the wavelet packet.
The wavelet coefficients are:
(ii) a The wavelet reconstruction formula is as follows:
。
the obtained inverse matrix is equivalent to the weight of the neural network, and the steps are as follows: firstly, the mapping from input data to characteristic nodes is established, and firstly, the input nodes are transposed to form a matrix
Z-score normalization is performed to ensure that input data is normalized to between 0 and 1; then is paired with>
Is augmented by adding a row 1 to the end of the training set to become->
(ii) a A feature node is generated for each window.
The generating of the feature node comprises the following steps: generating a random weight matrix we, we being (f + 1) xN 1 A dimensional random weight matrix with values in a Gaussian distribution; putting We into We { i }, wherein i represents the iteration quantity and the iteration number is N 2 ；
A 1 =H 1 Xwe; that is, the random convolution and bias of the weight value are performed once on the feature of each sample to obtain a new feature, and for each sample, the new feature can be expressed as:
;
to A 1 Carrying out normalization; to A 1 Carrying out sparse representation; solving the optimization problem in the sparse process by adopting a Lasso method, and newly generating a random feature vector A 1 Dimension of s × N 1 (ii) a Training set H after augmentation 1 Dimension s x (f + 1) in order to find a sparse matrix W such that H 1 ×W=A 1 Solving by:
node T for finally generating characteristics of a window 1 :T 1 =normal(H 1 ×W)；
For N 2 Each of the feature windows generating N 1 Each node is an s-dimensional eigenvector, and for the whole network, the characteristic node matrix y is one dimension of s x (N) 2 ×N 1 ) Of the matrix of (a).
The characteristic node enhancement comprises the steps of firstly standardizing and amplifying a characteristic node matrix y to obtain H 2 Different from the characteristic nodes, the coefficient matrix of the enhanced node is not a random matrix but a random matrix subjected to orthogonal normalization.
Encoding possible solutions, each individual representing a solution of the genetic algorithm; then, the processes of selection, crossing, mutation and the like of chromosome genes in the biological evolution are simulated, and new filial generations are continuously generated; and judging whether the new solution is good or bad through the fitness function until the optimal solution is generated.
The genetic algorithm further comprises the following steps: and (3) encoding: because three parameters of width learning are optimized, real number coding and the number N of characteristic windows of width learning are adopted 1 Number of characteristic nodes N 2 Number of enhanced nodes N 3 Respectively have a value range of [ a, b]，[c,d]，[e,f]Since real numbers are encoded, the length of the chromosome is N =3; the genetic composition of one of the chromosomes:
N={N i ,N j ,N k }，N i ∈[a,b]，N j ∈[c,d]，N k ∈[e,f](ii) a Selecting: calculating the value of the objective function corresponding to the feasible solution after encoding, and arranging the objective function values from large to small; and (3) crossing: performing the following cross operation on the selected individuals as parents to generate better-performing offspring;
the father generation is set as:
；
parent generation crossing to generate child generation process:
wherein i =1,2,3, γ ∈ [0,1], α ∈ [0,1], α + β =1;
two children were generated:
if it is
Out of bounds, then the child is generated with the following arithmetic intersection:
mutation: variation optimization is carried out through the idea of a simulated annealing algorithm, so that the method can be used for searching in the whole range at first, taking values near the optimal solution, and then obtaining the global optimal solution locally, and the specific process is as follows:
first, the parent generated by selection, crossover:
then a perturbation ω is given to the parent to generate a new individual:
omega = [ -1,1] continuous uniform random number, 0.95 is a change factor, and T is the iteration number of the algorithm;
and (3) comparing the fitness of the child and the parent:
if it is
；
Accepting the child as a new solution;
otherwise, the child is accepted with the following probability:
p（f(n)-f(N)）=exp((f(n)-f(N))/T)；
calculating the fitness of the objective function: and calculating the fitness value of the new generation of individuals by taking the output accuracy of the width learning network as an objective function.
When the value of the single data item is compared with the early warning value, the calculation mode is as follows:
firstly, the average value mu of the historical normal signal is calculated by a statistical method formula M The standard deviation sigma is calculated based on the current and voltage of data drive, and the temperature threshold value can be set by a K-sigma criterion;
;/>
the normal value ranges of the current, voltage and temperature signals are[μ M -kσ,μ M +kσ]；
Determining the value range of k in the k-sigma by the following steps:
calculating the mean value mu of the historical normal signal data m Standard deviation σ;
calculating the distance from each fault data to the normal data:
wherein ,χi For fault data, χ m Normal data;
respectively find d (x) i ,χ m ) Maximum and minimum values of (a):
d max =max{d(χ i ,χ m )}，
d min =min{d(χ i ,χ m )}；
respectively find d max Maximum data x of imax and dmin Minimum data x of imin ：
x imax =max{d max }，
x imin =min{d min }；
Determining the maximum value k of k max And the minimum value k min ：
x imin =μ m +k min σ，
x imax =μ m +k max σ；
To obtain the value range of k
；
The value of k is optimally determined using a grid search method.
The technical effects and advantages are as follows:
(1) The invention provides a UPS failure prediction method and a system based on width learning, wherein an artificial intelligence model can give consideration to efficiency and accuracy, realize the optional maintenance of a UPS battery, reduce the maintenance cost and the accidental shutdown time of equipment, and meet the high-reliability requirement of UPS failure prediction.
(2) The technical scheme provided by the invention screens the high-frequency components in the signals which can be extracted by decomposing the UPS storage battery signals by utilizing the wavelet packet, can effectively filter the high-frequency noise of the load signals, and can better represent various fault types of the UPS storage battery.
(3) According to the technical scheme provided by the invention, the width learning network parameters are optimized by improving the genetic algorithm, so that the problem of low accuracy caused by the fact that parameter values are determined by manual experience can be avoided, and the width learning diagnosis accuracy is improved.
(4) The technical scheme provided by the invention determines the alarm threshold range of various fault types through a grid search method and a K-sigma model, obtains high predicted reliability by using optimal deviation, finds various faults of the UPS battery to the maximum extent in time, achieves accurate early warning and ensures normal operation of equipment.
Drawings
FIG. 1 is a schematic diagram illustrating a wavelet packet decomposition according to an embodiment of the present invention;
FIG. 2 is a flow chart of data processing in an embodiment of the present invention;
FIG. 3 is a schematic diagram of width learning in an embodiment of the present invention;
FIG. 4 is a flow chart of improved genetic optimization breadth learning in an embodiment of the present invention;
FIG. 5 is a flowchart of a K-sigma model threshold early warning method based on a grid search method according to an embodiment of the present invention;
FIG. 6 is a flow chart of a UPS real-time failure prediction model of a breadth learning network in an embodiment of the present invention;
FIG. 7 is a time domain waveform diagram of decomposition of each node by three layers of wavelet packets in the embodiment of the present invention;
FIG. 8 is a graph of convergence curves for an embodiment of the present invention;
FIG. 9 is a diagram of the normal charging voltage Q-Q in an embodiment of the present invention;
FIG. 10 is a diagram illustrating real-time monitoring of a UPS power supply in an embodiment of the present invention;
fig. 11 is a flowchart of a method for monitoring a real-time status of a battery of a UPS power supply according to an embodiment of the present invention.
Detailed Description
The technical solutions in the embodiments of the present invention will be clearly and completely described below with reference to the drawings in the embodiments of the present invention, and it is obvious that the described embodiments are only a part of the embodiments of the present invention, and not all of the embodiments. The specific embodiments described herein are merely illustrative of the invention and do not delimit the invention. All other embodiments, which can be obtained by a person skilled in the art without making any creative effort based on the embodiments in the present invention, belong to the protection scope of the present invention.
The embodiment of the invention provides a UPS (uninterrupted power supply) fault prediction method based on a width learning network, which comprises the following steps as shown in FIG. 11:
performing operations such as data preprocessing, normalization, wavelet packet characteristic extraction and the like on historical data of the UPS battery in different states, and dividing the processed data into a training set and a test set;
determining the value ranges of the characteristic nodes, the enhanced nodes and the characteristic window number parameters in the width learning network;
introducing a simulated annealing algorithm, and improving the mutation operation in the genetic algorithm to obtain an improved genetic algorithm;
optimizing the width learning parameters through an improved genetic algorithm, and outputting width learning parameter values corresponding to the optimal solution;
determining the threshold range of the normal battery of the UPS according to a grid search algorithm and a K-sigma model;
and (4) comprehensively using the GA + BLS + K-sigma to obtain a final prediction and diagnosis model, and monitoring the state of the UPS battery in real time.
The embodiment of the invention provides a UPS (uninterrupted power supply) fault prediction system based on a width learning network, which comprises the following steps: a data processing module for expressing the pattern vector of the original feature space of high dimensionality with the new pattern vector of the feature space of low dimensionality, thereby finding the most representative and most effective feature method; the width learning system module is used for obtaining a network model according to the weight value of the neural network of which the inverse matrix obtained by the method is equivalent to; the genetic algorithm module is used for optimizing parameters of the network model so as to optimize an objective function; and the threshold early warning module is established based on K-sigma of a grid search method, can compare the value of a single data item with an early warning value according to the real-time collected data of each single battery of the UPS, such as current, voltage, temperature and the like, and timely finds abnormality and positions the UPS to perform fault warning when the current, the voltage and the temperature exceed the standard.
In this embodiment, the specific embodiment mode is as follows:
data processing, based on time-frequency domain feature extraction, the feature extraction refers to a method of finding the most representative and effective features by expressing a pattern vector of a high-dimensional original feature space with a new pattern vector of a low-dimensional feature space through a series of transformations.
When the UPS battery fails, different failure types may cause the amplitude of each frequency range of the signal to change, resulting in a change in energy in the corresponding frequency band. Therefore, the characteristics of each frequency band can be used as the basis for fault diagnosis.
The wavelet packet decomposition algorithm is utilized to analyze the general time sequence data signals, different frequency components of the signals can be effectively extracted, the obtained signal components have higher stationarity and periodicity, and the signal characteristics are more obvious. In addition, the high-frequency component with high volatility has low prediction precision, and the wavelet packet decomposition can be used for extracting the high-frequency component in the signal for screening, so that the high-frequency noise of the load signal can be effectively filtered, and the residual signal characteristic can be extracted. A three-layer wavelet packet decomposition is shown in fig. 1.
Wavelet coefficient of
Wavelet reconstruction formula is
The data processing flow is shown in fig. 2.
The working principle of width learning is as follows: the breadth learning system is a newly developed alternative scheme of a deep learning structure, the network can effectively avoid the problem of hyper-parameter redundancy in a multilayer network, and the system is a system based on a random vector function link neural network.
Different from the original RVFLNN, the enhancement nodes are directly established, the BLS maps the characteristics to obtain a series of mapping nodes, and then the corresponding enhancement nodes are obtained through an algorithm. The structure of the width learning system is shown in fig. 3.
X represents the input data matrix, then the matrix is operated on, and
expressed as the ith mapping feature, namely Z i ，W ei Are representative of suitable random weights.
Z i ≡[Z 1 ,…,Z i ]；
Z i Denoted as the ith mapped feature group. Similarly, the jth enhanced node, H j Is denoted as ζ j （Z i W hj +β hj ) All jth enhanced nodes are represented as:
H j ≡[H 1 ,…,H j ]；
in practice, the size of i, j is chosen empirically by a human. In addition to
The two functions may be different. For j ≠ r, ζ j ，ζ r The two functions may be different. To be non-trivial, the ith random feature>
And jth random feature
ζ j The subscripts of (a) will be omitted hereinafter.
X represents an input matrix, and N samples are contained in the X, and each sample has M-dimensional characteristics. Y is an output matrix belonging to
R N×C For n feature maps, each map yields k nodes. Can be shown by the following formula:
W ei ，β ei is a randomly generated matrix with all feature nodes denoted as Z n ≡[Z 1 ,…,Z n ]Denote the mth enhanced node group as H m ≡ζ(Z n W hm +β hm ) Thus the input a to the width system can be expressed as:
A=[Z n |H m ]；
the width system can thus be expressed as the following equation:
therefore, the following steps are carried out:
wherein Wm Is the link weight of the width structure, which can be computed from a ridge regression approximation.
The following are symbolic explanations:
generating a characteristic node: the kernel of the width learning is to find the pseudo-inverse of the characteristic node and the enhanced node to the target value. In the width learning system, the characteristic nodes and the enhanced nodes correspond to the input of the neural network, and the obtained inverse matrix is equivalent to the weight of the neural network. Firstly, the mapping from input data to characteristic nodes is established, and firstly, the input nodes are transposed to form a matrix
Z-score normalization is performed to ensure that the input data is normalized to between 0 and 1. Then, the product is processedIs paired and/or matched>
Is augmented by adding a row 1 to the end of the training set to become->
The aim is that the bias term can be added directly through matrix operation when generating the characteristic node. Then start generating feature nodes for each window:
(1) Generating a random weight matrix we, we being one (f + 1) xN 1 A dimensional random weight matrix with values in a Gaussian distribution;
(2) Putting We into We { i }, wherein i represents the iteration quantity and the iteration number is N 2 ；
（3）A 1 =H 1 Xwe; that is, the random convolution and bias of the weight value are performed once on the feature of each sample to obtain a new feature, and for each sample, the new feature can be expressed as:
(4) To A 1 Normalization is carried out;
(5) To A 1 Carrying out sparse representation; and solving the optimization problem in the sparse process by adopting a Lasso method. Newly generated random feature vector A 1 Dimension of s × N 1 Training set H after augmentation 1 The dimension is s × (f + 1). The objective is to find a sparse matrix W such that H 1 ×W=A 1 . The solution is given by:
(6) Node T for finally generating characteristics of a window 1 :T 1 =normal(H 1 ×W)；
For N 2 Each of the feature windows generating N 1 And each node is an s-dimensional feature vector. For the entire network, the characteristic node matrix y is a matrix of dimensions sx (N) 2 ×N 1 ) Moment ofAnd (5) arraying.
And generating an enhanced node, wherein the characteristic node is linear, so that the enhanced node is introduced to increase the nonlinearity of the system.
(1) Like the characteristic nodes, firstly, the characteristic node matrix y is standardized and augmented to obtain H 2 . Different from the characteristic nodes, the coefficient matrix of the enhanced node is not a random matrix but a random matrix subjected to orthogonal normalization. Suppose that
(N 2 ×N 1 )>N 3 ) Then the coefficient matrix wh of the enhanced node can be expressed as (N) 2 ×N 1 )×N 3 The dimensions are orthogonal normalized random matrices. The method aims to map the characteristic nodes to a high-dimensional subspace through nonlinearity, so that the expression capacity of the network is stronger.
(2) Activating the enhanced node:
;
s is the scaling of the enhanced node;
(3) Input T of the final generation network 3。
Final input T of the network 3 Is composed of
Feature dimension of each sample is N 1 ×N 2 +N 3
(4) And (5) solving the pseudo-inverse.
；
The algorithm is as follows: width learning
Inputting: training sample X, number of characteristic nodes n and number of enhanced nodes m
And (3) outputting: w
For i=0;i≤n
e random value W ei ，β ei ;
Computing
;
end
Setting a feature mapping group Z n =[Z 1 ,…,Z n ];
For j=1;j≤m
Random value W hj ，β hj ；
Calculate H j =[ζ(Z n W hj +β hj )];
end
Setting enhanced node group H m =[H 1 ,…,H m ];
Is provided with
And calculate->
General formula (4-5)
Is provided with
。
The improved genetic algorithm is the establishing process of the width learning network model, wherein the characteristic nodes are reasonably arranged, the nodes are enhanced, the values of the characteristic window parameters are directly related to the performance of the whole network, and the improved genetic algorithm is utilized to optimize the network model parameters, so that the target function (the accuracy of the network model is highest) is optimal.
The genetic algorithm is a method for searching an optimal solution by simulating a natural evolution process, and has the main advantages that a structural object is directly operated, the optimal objective function value is used as search information, and complex operations such as derivation and the like are avoided; by adopting a probabilistic optimization method, the optimized search space can be automatically acquired and guided without a determined rule, and the search direction is adaptively adjusted; the system has expandability and is convenient to be mixed with other technologies for use.
The genetic algorithm mainly comprises the steps of coding possible solutions, enabling each individual to represent a solution of the genetic algorithm, simulating processes of selection, crossing, variation and the like of chromosome genes in biological evolution, continuously generating new filial generations, and judging the quality of the new solutions through a fitness function until the optimal solutions are generated.
The coding modes include binary coding, real number coding and the like, and the algorithm can be more efficient only when the coding modes correspond to correct environments. Otherwise, the problems of premature convergence, low algorithm efficiency and the like easily occur.
The key steps of the genetic algorithm of the invention are as follows:
(1) And (3) encoding: since the three parameters of the width learning are optimized and decoding is needed when binary coding is adopted, which leads to the reduction of the efficiency of the algorithm, real number coding is adopted. Number of width learning feature windows N 1 Number of characteristic nodes N 2 Number of enhanced nodes N 3 Respectively have a value range of [ a, b]，[c,d]，[e,f]. Since real numbers are encoded, the length of the chromosome is N =3,
the genetic composition of one of the chromosomes: n = { N i ,N j ,N k }；
N i ∈[a,b]，N j ∈[c,d]，N k ∈[e,f]。
(2) Selecting: and calculating the value of the objective function corresponding to the feasible solution after encoding, arranging the objective function values from large to small, and leaving the feasible solution with high fitness in a roulette selection mode to eliminate the feasible solution with low fitness.
(3) And (3) crossing: the selected individuals were used as parents to perform the following crossover operations to produce better performing offspring.
The father generation is set as:
；
parent generation crossing to generate child generation process:
i=1,2,3 γ∈[0,1],α∈[0,1],α+β=1；
two children were generated:
if it is
Out of bounds, then the child is generated with the following arithmetic intersection:
(4) Mutation: the diversity of population genes is a precondition for ensuring that a genetic algorithm finds a global optimal solution. In the evolution process, because the selection of eliminated individuals can lead to the reduction of gene diversity, and mutation operation can make up for the defect of insufficient genes, the mutation operation of the genetic algorithm is necessary. The traditional mutation operation adopts a random mutation optimization mode which easily causes 'precocity' to cause local convergence, so that the efficiency of the algorithm is not high. In order to improve the global optimization capability of the genetic algorithm and the convergence rate of the algorithm, a simulated annealing algorithm is introduced to generate a mutation operator in genetic algorithm operation operators.
Variation optimization is carried out through the idea of a simulated annealing algorithm, so that the whole range is searched at first, values are taken near the optimal solution, and then the global optimal solution is obtained locally. The specific process is as follows:
first, the parent generated by selection, crossover:
then a perturbation ω is given to the parent to generate a new individual:
note: omega = [ -1,1] continuous uniform random number, 0.95 is a change factor, and T is the iteration number of the algorithm;
and (3) comparing the fitness of the child and the parent:
if it is
；
Accepting the offspring as a new solution;
otherwise, the child is accepted with the following probability:
p（f(n)-f(N)）=exp((f(n)-f(N))/T)；
calculating the fitness of the objective function: and (3) calculating the fitness value of the new generation of individuals by taking the output accuracy of the width learning network as an objective function to see whether a loop termination condition is met, and otherwise, repeating the steps (1), (2), (3), (4) and (5). Until a termination condition is met. The process of optimizing BLS parameters of the BLS model using the improved genetic algorithm is shown in fig. 4.
The K-sigma model threshold early warning method based on the grid search method comprises the following steps:
the UPS equipment threshold value early warning model can compare the value of a single data item with an early warning value according to the current, voltage, temperature and other UPS single battery data collected in real time, and when the current, the voltage and the temperature exceed the standard, abnormity is found in time and the UPS is positioned for fault warning. For the collected current, voltage and temperature values, whether the current, voltage and temperature are abnormal or not needs to be determined by means of threshold judgment and the like. And calculating according to the normal current, voltage and temperature data signals in the historical data of the UPS battery. Firstly, the average value mu of the historical normal signal is calculated by a statistical method formula M Standard deviation σ. The temperature threshold setting may be calculated by the K-sigma criterion based on the current, voltage of the data drive.
Normal value ranges for current, voltage and temperature signals are [ mu ] M -kσ,μ M +kσ]To ensure predictive and diagnostic accuracy, the k-sigma model k is scoped.
Firstly, multi-dimensional historical fault signal data X are collected n×5 ) (n rows, 5 columns);
wherein each column represents signal data (including historical normal signals) for one historical fault type of the UPS battery and the number of rows represents the dimension of the signal. Determining the value range of k in the k-sigma by the following steps:
(1) Calculating the mean value mu of the historical normal signal data M Standard deviation σ;
(2) Calculating the distance from each fault data to the normal data;
note: chi shape i For fault data, χ m Normal data;
(3) Respectively find out d (x) i ,χ m ) Maximum and minimum values of (a):
d max =max{d(χ i ,χ m ) }，
d min =min{d(χ i ,χ m ) }；
(4) Respectively find d max Maximum data x of imax and dmin Minimum data x of imin ：
x imax =max{d max }，
x imin =min{d min }；
(5) Determining the maximum value k of k max And the minimum value k min ：
x imin =μ m +k min σ，
x imax =μ m +k max σ；
(6) To obtain the value range of k
。
The value of k is optimally determined using a grid search method. The grid searching method comprises the following steps: (1) Determination by historical normal signals
μ M Sigma, (2) determining the value range of k, wherein the step length of k is 1, (3) testing the corresponding value of each k by using fault data for calculationAccuracy, (4) termination condition k ≧ k max And (5) outputting the corresponding K value when the accuracy of each threshold range is highest.
When the actual measured value is in the range, the current, voltage, temperature and battery of the UPS are normal, and when the actual real-time current or voltage value is not in the range, the corresponding current, voltage, temperature and battery are abnormal, and corresponding treatment is needed. The flow of the K-sigma model threshold early warning method based on the grid search method is shown in fig. 5.
Aiming at the improvement of the genetic algorithm and the threshold alarm model, the flow of the UPS battery real-time fault prediction model based on the width learning network is shown in FIG. 6.
Specific steps of the proposed scheme of the present invention are introduced as follows for the flowchart:
step 1: collecting historical data of the UPS storage battery in the fault types of overcharge, overdischarge, aging, overheat and normal states respectively;
step 2: and carrying out operations such as data preprocessing, normalization, wavelet packet feature extraction and the like on the collected fault historical data, and dividing the processed data into a training set and a test set.
And step 3: and setting parameters of the GA, and determining the value ranges of three parameters, namely a width learning system characteristic node, an enhanced node and a characteristic window according to a large number of relevant documents.
And 4, step 4: and carrying out real number coding on three parameter variables needing optimization.
And 5: initializing the population, calculating the fitness value of the individual and carrying out selection operation on the fitness value.
Step 6: and crossing the parents of the children, and generating new children with better performance based on the variation of the simulated annealing algorithm.
And 7: and (4) calculating the fitness value of the new filial generation, if the fitness value does not meet the termination condition, circulating the steps 5-7 until the condition is met, and outputting the specific optimal values of the three parameters corresponding to the maximum fitness.
And 8: and taking the normal historical data of the UPS as the input of the K-sigma model, and determining the K value which enables the model accuracy to be highest by utilizing a grid search method, thereby determining the optimal threshold early warning range of the normal battery of the UPS.
And step 9: the UPS real-time data are collected, whether faults occur or not is judged through the K-sigma model, if faults are early warned, the BLS model outputs the predicted fault types, the state of the UPS storage battery is accurately and efficiently detected in real time, reliable power supply is guaranteed, and the service life of a battery pack is prolonged. The above is the detailed technical scheme of the invention.
According to the UPS failure prediction model provided by the embodiment, the efficiency and the accuracy can be considered, the optional maintenance of the UPS battery is realized, the maintenance cost is reduced, the accidental shutdown time of equipment is reduced, and the requirement of high reliability of the UPS failure prediction is met. The high-frequency components in the signals which can be extracted by decomposing the UPS storage battery signals by using the wavelet packets are screened, so that the high-frequency noise of the load signals can be effectively filtered, and various fault types of the UPS storage battery can be better represented. By improving the genetic algorithm to optimize the parameters of the width learning network, the problem of low accuracy caused by the fact that the parameter value is determined by manual experience can be avoided, and the accuracy of width learning diagnosis is improved. The threshold value range of various fault type alarms is determined through a grid search method and a K-sigma model, the predicted high reliability is obtained through the optimal deviation, various faults of the UPS battery are found out to the maximum extent in time, accurate early warning is achieved, and normal operation of equipment is guaranteed.
In some embodiments, the UPS battery mainly includes two types, i.e., a lead storage battery and a lithium battery, which has the great advantages of higher density, longer life, smaller volume, relatively light weight, environmental friendliness, etc., but is expensive, and once the battery is damaged, the maintenance period is long, which is the most fatal disadvantage that the battery may be burned or even exploded when the battery is subjected to too high current or an impact. The UPS batteries on the market are mainly lead storage batteries, so the implementation of the present invention is exemplified by lead storage batteries.
In this embodiment, for example, the charging voltage data signal of the UPS battery cell is collected, and an industrial-grade UPS battery is adopted, where the voltage of a single battery is U =12v, the voltage of a single battery is U =384V, the total voltage of the single battery is U =384V, and the battery capacity of the single battery is C =120AH. Acquiring historical data of the UPS battery with the fault types of overcharge, overdischarge, overheat and aging faults and normal voltage, normalizing the data, denoising by wavelet packet transformation, extracting features and acquiring feature vectors corresponding to the fault types.
Normalization with 0-1:
performing wavelet packet three-layer decomposition on the voltage signal to obtain an original signal and a time domain waveform diagram of each node, as shown in fig. 7, wherein (3,0) is a time domain waveform diagram of a node one; (3,1) is the time domain waveform for node two; (3,2) is the time domain waveform diagram for node three; (3,3) is the time domain waveform diagram for node four; (3,4) is the time domain waveform diagram for node five; (3,5) is the time domain waveform for node six; (3,6) is the time domain waveform diagram for node seven; (3,7) is a time domain waveform diagram for node eight.
As can be seen from fig. 7, the frequency components of the net voltage data are mainly concentrated in the low frequency portion, the first 2 frequency band signals are concentrated in the main components of the net voltage signal, the signal fluctuation in the high frequency component is large, but the amplitude is small, the ratio in the whole input signal is very small, and the signal noise is large. Therefore, high-frequency components are removed, only signals of [3.0] and [3.1] nodes are reserved, and signal reconstruction is carried out by using the following formula.
The wavelet coefficients are:
the wavelet reconstruction formula is as follows:
obtaining a voltage signal characteristic vector x of each fault type of the UPS battery Feature vector . In the embodiment, 500 groups of samples are collected, the front 450 groups of samples are divided into a training set, and the rear 50 groups of samples are divided into a testing set.
Determining the value range of the width learning parameter: as shown in the above, through reference of a large number of documents and cross experimental verification, the regularization parameter C = 2^ -30, the enhanced node scaling factor s =0.8, the characteristic node m of the width learning network, the enhanced node N and the characteristic window number N are respectively determined to be [2 ], [2 ] 20 and [ 50 ]. Improved genetic algorithms.
The embodiment combines some advantages and disadvantages of the genetic algorithm and combines the practical problem of width learning parameter optimization to improve the genetic algorithm so as to realize the purpose of the invention.
And (3) encoding: for the width learning network feature node m, the enhanced node N and the feature window number N,
the parameters are real encoded, wherein the chromosomal genome of the population of individuals becomes:
N={N i ,N j ,N k }；
note: n is a radical of hydrogen i ∈[2 20] ，N j ∈[2 20]，N k ∈[50 500]。
Selecting: according to the individual genes of the coding population, the fitness (the accuracy rate of the width learning output) of the individuals is calculated, the individuals are arranged from large to small, and a roulette strategy is used for selection operation.
And (3) crossing: generating a random number for the chromosome left after the selection operation, if the random number is greater than or equal to the crossover probability P c Performing the following crossover operation, otherwise, reselecting a pair of chromosomes from the rest chromosomes to continue generating a random number, and determining whether the crossover operation can be performed, in this embodiment, the crossover probability P c =0.8. In the chromosome crossing process, both the crossing starting point position and the crossing length are randomly generated so as to improve the randomness of the population and avoid the algorithm from falling into local convergence.
Mutation: the traditional mutation operation adopts a random mutation optimizing mode which easily causes 'precocity' to cause local convergence, so that the efficiency of the algorithm is not high m =0.3，T（Number of iterations) =50.
The specific process is as follows:
first, the parent generated by selection, crossover:
then a perturbation ω is given to the parent to generate a new individual:
note: ω = [ -1,1] continuous uniform random number, 0.95 is the variation factor, and T is the number of algorithm iterations.
And (3) comparing the fitness of the child and the parent:
if it is
，
The child is accepted as the new solution.
Otherwise, the child is accepted with the following probability:
p（f(n)-f(N)）=exp((f(n)-f(N))/T)；
and calculating the fitness value of the new generation of population, judging whether a termination condition (N is more than or equal to 50) is met, and if the termination condition is not met, outputting the values of the three parameters corresponding to the maximum target function. The output in this example is m =12,n =56.
The results of the improved genetic algorithm and the conventional algorithm based on matlab in this example are shown in fig. 8. From the results, the traditional genetic algorithm has a phenomenon of 'precocity' to a certain extent, so that local convergence is caused, and the diagnosis accuracy of the breadth learning network is influenced.
Calculating a K-sigma model threshold based on a grid search method:
checking whether the voltage data signal conforms to normal distribution: a Q-Q plot of the collected historical normal charging voltage data is shown in fig. 9.
From the Q-Q diagram, positiveThe normal voltage data exists at a partial point not in the vicinity of a straight line (normal distribution), so that the data signal does not conform to the normal distribution. In order to ensure accuracy, an optimal threshold range is set, and a K-sigma model is introduced. Determining the range of K as [2 7 ] by cross-counting fault signal data]. And optimizing the K by adopting a grid search algorithm to avoid human errors. μ of normal data M σ is determined by the following equation:
obtaining: mu.s M =13.6681, σ =0.1055, and the optimum K =4 is determined by the grid search method. So the range of the normal signal of the charging voltage is [13.2461,14.0901 ]]When the actual measurement value is in the range, the UPS battery is normal; and when the actual real-time voltage value is not in the range, the corresponding battery is abnormal, and finally, the BLS model outputs the predicted fault type.
By using historical data of overcharge, overdischarge, overheating, aging faults and normal states, a width learning training diagnosis model is obtained through an improved genetic algorithm, a corresponding optimal threshold value interval is obtained through a grid search algorithm and a K sigma model, and finally a satisfactory prediction and diagnosis model is obtained. Results of three diagnosis modes, namely a width learning model, a width learning + improved genetic algorithm and a width learning + improved genetic algorithm + K sigma model are compared, and the results are shown in the following table.
The result shows that the width learning combined with the improved genetic algorithm has more time consumption than the simple width learning, but the accuracy is greatly improved, which indicates that the parameter of the width learning determined by the artificial experience is unreliable, the maximum advantage of the width learning on the fault diagnosis cannot be exerted, and the result may be caused by artificial errors and is not ideal. On the basis, the accuracy of the model can be further improved by combining the K-sigma model, and a satisfactory effect is obtained on the diagnosis of the battery fault of the UPS.
And detecting the state of the UPS storage battery in real time on line.
In this embodiment, the width learning, the improved genetic algorithm, and the K-sigma final diagnosis model are applied to the UPS power supply battery with the overcharge fault, and the results are shown in fig. 10 by online monitoring the real-time status of the UPS power supply before and after the fault for 15 days.
The result shows that the dotted line is the threshold range determined by the final width learning diagnosis model according to the normal charging voltage data, when the equipment breaks down, the accurate prediction of the fault can be realized in advance through the real-time monitoring model, the early warning and the accurate maintenance are realized in advance, and the normal and efficient operation of the equipment is ensured.
In this embodiment, an electronic device is provided, comprising a memory in which a computer program is stored and a processor configured to run the computer program to perform the method in the above embodiments.
The programs described above may be executed on a processor or may be stored in memory (or referred to as computer-readable media), which may include non-transitory and non-transitory, removable and non-removable media, and may be implemented in any method or technology for storage of information. The information may be computer readable instructions, data structures, modules of a program, or other data. Examples of computer storage media include, but are not limited to, phase change memory (PRAM), static Random Access Memory (SRAM), dynamic Random Access Memory (DRAM), other types of Random Access Memory (RAM), read Only Memory (ROM), electrically Erasable Programmable Read Only Memory (EEPROM), flash memory or other memory technology, compact disc read only memory (CD-ROM), digital Versatile Disks (DVD) or other optical storage, magnetic cassettes, magnetic tape, magnetic disk storage or other magnetic storage devices, or any other non-transmission medium, which can be used to store information that can be accessed by a computing device. As defined herein, a computer readable medium does not include a transitory computer readable medium such as a modulated data signal and a carrier wave.
These computer programs may also be loaded onto a computer or other programmable data processing apparatus to cause a series of operational steps to be performed on the computer or other programmable apparatus to produce a computer implemented process such that the instructions which execute on the computer or other programmable apparatus provide steps for implementing the functions specified in the flowchart flow or flows and/or block diagram block or blocks, and corresponding steps may be implemented by different modules.
Such an apparatus or system is provided in this embodiment. The system is called a UPS failure prediction system based on a width learning network, and comprises: a data processing module for expressing the pattern vector of the original feature space of high dimensionality with the new pattern vector of the feature space of low dimensionality, thereby finding the most representative and most effective feature method; the width learning system module is used for obtaining a network model according to the weight value of the neural network of which the inverse matrix obtained by the method is equivalent to; the genetic algorithm module is used for optimizing parameters of the network model so as to optimize an objective function; the threshold early warning module is established based on K-sigma of a grid search method, the value of a single data item can be compared with an early warning value according to the data of each battery cell of the UPS, such as current, voltage, temperature and the like collected in real time, and when the current, the voltage and the temperature exceed the standard, abnormality is found in time and the UPS is positioned for fault warning.
The system or the apparatus is used for implementing the functions of the method in the foregoing embodiments, and each module in the system or the apparatus corresponds to each step in the method, which has been described in the method and is not described herein again.
According to the UPS failure prediction method and the UPS failure prediction system, the artificial intelligence model can give consideration to efficiency and accuracy, the optional maintenance of the UPS battery is achieved, the maintenance cost is reduced, the accidental shutdown time of equipment is reduced, and the requirement of high reliability of UPS failure prediction is met. The high-frequency components in the signals which can be extracted by decomposing the UPS storage battery signals by using the wavelet packets are screened, so that the high-frequency noise of the load signals can be effectively filtered, and various fault types of the UPS storage battery can be better characterized. By optimizing the width learning network parameters through improving the genetic algorithm, the problem of low accuracy caused by the fact that parameter values are determined by means of manual experience can be solved, and the width learning diagnosis accuracy is improved. The threshold value range of various fault type alarms is determined through a grid search method and a K-sigma model, the predicted high reliability is obtained through the optimal deviation, various faults of the UPS battery are found out to the maximum extent in time, accurate early warning is achieved, and normal operation of equipment is guaranteed.
Finally, it should be noted that: although the present invention has been described in detail with reference to the foregoing embodiments, it will be apparent to those skilled in the art that modifications may be made to the embodiments or portions thereof without departing from the spirit and scope of the invention.
Claims (10)
1. A UPS failure prediction method based on a width learning network is characterized by comprising the following steps:
performing operations such as data preprocessing, normalization, wavelet packet characteristic extraction and the like on historical data of the UPS battery in different states, and dividing the processed data into a training set and a test set;
determining the value ranges of the characteristic nodes, the enhanced nodes and the characteristic window number parameters in the width learning network;
introducing a simulated annealing algorithm, and improving the mutation operation in the genetic algorithm to obtain an improved genetic algorithm;
optimizing the width learning parameters through an improved genetic algorithm, and outputting width learning parameter values corresponding to the optimal solution;
determining the threshold range of the normal battery of the UPS according to a grid search algorithm and a K-sigma model;
and (4) comprehensively using the GA + BLS + K-sigma to obtain a final prediction and diagnosis model, and monitoring the state of the UPS battery in real time.
2. A UPS failure prediction system based on a width learning network is characterized by comprising:
a data processing module for expressing the pattern vector of the original feature space of high dimensionality with the new pattern vector of the feature space of low dimensionality, thereby finding the most representative and most effective feature method;
the width learning system module is used for obtaining a network model according to the weight value of the neural network of which the inverse matrix obtained by the method is equivalent to;
the genetic algorithm module is used for optimizing parameters of the network model so as to optimize an objective function;
the threshold early warning module is established based on K-sigma of a grid search method, the value of a single data item can be compared with an early warning value according to the data of each battery cell of the UPS, such as current, voltage, temperature and the like collected in real time, and when the current, the voltage and the temperature exceed the standard, abnormality is found in time and the UPS is positioned for fault warning.
3. The breadth learning network-based UPS failure prediction system of claim 2, wherein: the method comprises a wavelet packet decomposition algorithm, wherein the wavelet packet decomposition algorithm is used for extracting high-frequency components in signals for screening, can effectively filter high-frequency noise of load signals, and extracts residual signal characteristics, and comprises the following steps:
carrying out normalization processing on the UPS vertical data signals;
performing wavelet packet 3-layer decomposition on the processed data to obtain each frequency band component, and eliminating noise components;
and reconstructing a data signal by using the wavelet packet coefficient, and then extracting the energy characteristic of the wavelet packet.
4. The UPS failure prediction system based on the breadth learning network of claim 3, wherein: the wavelet coefficients are:
；
the wavelet reconstruction formula is as follows:
。
5. the breadth learning network-based UPS failure prediction system of claim 2, wherein: the obtained inverse matrix is equivalent to the weight of the neural network, and the steps are as follows:
firstly, the mapping from input data to characteristic nodes is established, and firstly, the input nodes are transposed to form a matrix
Z-score normalization is performed to ensure that input data is normalized to between 0 and 1;
then to
Is augmented by adding a row 1 to the end of the training set to become->
；
A feature node is generated for each window.
6. The UPS failure prediction system based on the width learning network as claimed in claim 5, wherein: generating a characteristic node, comprising the following steps:
generating a random weight matrix we, we being (f + 1) xN 1 A dimensional random weight matrix with values in a Gaussian distribution;
putting We into We { i }, wherein i represents the iteration quantity and the iteration number is N 2 ；
A 1 ＝H 1 Xwe; i.e. for each sampleThe features are subjected to random convolution and bias of weights once to obtain new features, and for each sample, the new features can be expressed as:
;
to A 1 Carrying out normalization;
to A 1 Carrying out sparse representation; solving the optimization problem in the sparse process by adopting a Lasso method, and newly generating a random feature vector A 1 Dimension of s × N 1 (ii) a Training set H after augmentation 1 Dimension s x (f + 1) in order to find a sparse matrix W such that H 1 ×W=A 1 Solving by:
;
node T for finally generating characteristics of a window 1 :T 1 =normal(H 1 ×W)；
For N 2 Each generating N 1 Each node is an s-dimensional eigenvector, and for the whole network, the characteristic node matrix y is one dimension of s x (N) 2 ×N 1 ) Of the matrix of (a).
7. The UPS failure prediction system based on the width learning network as claimed in claim 6, wherein: enhancing the characteristic nodes, including firstly standardizing and amplifying the characteristic node matrix y to obtain H 2 Different from the characteristic nodes, the coefficient matrix of the enhanced node is not a random matrix but a random matrix subjected to orthogonal normalization.
8. The breadth learning network-based UPS failure prediction system of claim 7, wherein: the steps of the genetic algorithm are as follows:
encoding possible solutions, each individual representing a solution of a genetic algorithm;
then, the processes of selection, crossing, mutation and the like of chromosome genes in the biological evolution are simulated, and new filial generations are continuously generated;
and judging whether the new solution is good or bad through the fitness function until the optimal solution is generated.
9. The breadth learning network-based UPS failure prediction system of claim 8, wherein: the genetic algorithm further comprises the following steps:
and (3) encoding: because three parameters of width learning are optimized, real number coding is adopted, and the number N of characteristic windows of width learning is adopted 1 Number of feature nodes N 2 Number of enhanced nodes N 3 Has a value range of [ a, b ] respectively]，[c,d]，[e,f]Since it is real number encoded, the length of the chromosome is N =3;
the genetic composition of one of the chromosomes: n = { N i ,N j ,N k }，N i ∈[a,b]，N j ∈[c,d]，N k ∈[e,f]；
Selecting: calculating the value of the objective function corresponding to the feasible solution after encoding, and arranging the objective function values from large to small;
and (3) crossing: performing the following cross operation on the selected individuals as parents to generate better-performing offspring;
let father generation be:
；
parent generation crossing to generate child generation process:
wherein ,
；
two children were generated:
if it is
Out of bounds, the child is generated with the following arithmetic intersection: />
Mutation: variation optimization is carried out through the idea of a simulated annealing algorithm, so that the method can be used for searching in the whole range at first, taking values near the optimal solution, and then obtaining the global optimal solution locally, and the specific process is as follows:
first, the parent generated by selection, crossover:
then a perturbation ω is given to the parent to generate a new individual:
and (3) comparing the fitness of the child and the parent:
if it is
；
Accepting the child as a new solution
Otherwise, the child is accepted with the following probability:
p（f(n)-f(N)）=exp((f(n)-f(N))/T)；
calculating the fitness of the objective function: and calculating the fitness value of the new generation of individuals by taking the output accuracy of the width learning network as an objective function.
10. The width learning network-based UPS failure prediction system of claim 9, wherein: when the value of a single data item is compared with the early warning value, the calculation mode is as follows:
firstly, the average value mu of the historical normal signal is calculated by a statistical method formula m The standard deviation sigma is calculated based on the current and voltage of data drive, and the temperature threshold value can be set by a K-sigma criterion;
normal value ranges for current, voltage and temperature signals are [ mu ] M -kσ,μ M +kσ]；
Determining the value range of k in the k-sigma by the following steps:
calculating the mean value mu of the historical normal signal data m Standard deviation σ;
calculating the distance from each fault data to the normal data:
wherein ,χi For fault data, x m Normal data;
respectively find d (x) i ,χ m ) Maximum and minimum values of (a):
d max =max{d(χ i ,χ m ) }，
d min =min{d(χ i ,χ m ) }；
respectively find d max Maximum data x of imax and dmin Minimum data x of imin ：
x imax =max{d max }，
x imin =min{d min }；
Determining the maximum value k of k max And the minimum value k min ：
x imin =μ m +k min σ，
x imax =μ m +k max σ；
Obtaining the value range k of k
[k min ，k max ]；
The value of k is optimally determined using a grid search method.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202310166830.0A CN115859058B (en) | 2023-02-27 | 2023-02-27 | UPS fault prediction method and system based on width learning network |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202310166830.0A CN115859058B (en) | 2023-02-27 | 2023-02-27 | UPS fault prediction method and system based on width learning network |
Publications (2)
| Publication Number | Publication Date |
|---|---|
| CN115859058A true CN115859058A (en) | 2023-03-28 |
| CN115859058B CN115859058B (en) | 2023-05-30 |
Family
ID=85659015
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN202310166830.0A Active CN115859058B (en) | 2023-02-27 | 2023-02-27 | UPS fault prediction method and system based on width learning network |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN115859058B (en) |
Cited By (5)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN116455079A (en) * | 2023-05-04 | 2023-07-18 | 奥斯塔娜(常州)电子有限公司 | Big data-based electricity consumption integrated safety supervision system and method |
| CN116626514A (en) * | 2023-05-23 | 2023-08-22 | 浙江海得智慧能源有限公司 | Method, system and equipment for screening invalid battery cells |
| CN117332361A (en) * | 2023-12-01 | 2024-01-02 | 赫芝特电气技术（广州）有限公司 | Online early warning method and detection system for uninterruptible power supply |
| CN117992738A (en) * | 2024-04-03 | 2024-05-07 | 山东兰蝶节能环保科技有限公司 | Automatic electricity-saving monitoring method and system for electric energy of electricity-saving cabinet |
| CN117992738B (en) * | 2024-04-03 | 2024-06-11 | 山东兰蝶节能环保科技有限公司 | Automatic electricity-saving monitoring method and system for electric energy of electricity-saving cabinet |
Citations (12)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN103926526A (en) * | 2014-05-05 | 2014-07-16 | 重庆大学 | Analog circuit fault diagnosis method based on improved RBF neural network |
| CN108090658A (en) * | 2017-12-06 | 2018-05-29 | 河北工业大学 | Arc fault diagnostic method based on time domain charactreristic parameter fusion |
| CN108876054A (en) * | 2018-07-06 | 2018-11-23 | 国网河南省电力公司郑州供电公司 | Short-Term Load Forecasting Method based on improved adaptive GA-IAGA optimization extreme learning machine |
| US20190377625A1 (en) * | 2018-06-08 | 2019-12-12 | Microsoft Technology Licensing, Llc | Computing node failure and health prediction for cloud-based data center |
| CN111325308A (en) * | 2020-02-14 | 2020-06-23 | 集美大学 | Nonlinear system identification method |
| US20200285997A1 (en) * | 2019-03-04 | 2020-09-10 | Iocurrents, Inc. | Near real-time detection and classification of machine anomalies using machine learning and artificial intelligence |
| CN112804239A (en) * | 2021-01-22 | 2021-05-14 | 山东维平信息安全测评技术有限公司 | Traffic safety analysis modeling method and system |
| US20210311813A1 (en) * | 2020-07-20 | 2021-10-07 | Beijing Baidu Netcom Science And Technology Co., Ltd. | Fault prediction method, apparatus and storage medium |
| CN114818806A (en) * | 2022-04-25 | 2022-07-29 | 重庆大学 | Gearbox fault diagnosis method based on wavelet packet and depth self-encoder |
| CN114861778A (en) * | 2022-04-22 | 2022-08-05 | 哈尔滨理工大学 | Method for rapidly classifying rolling bearing states under different loads by improving width transfer learning |
| CN115062839A (en) * | 2022-06-13 | 2022-09-16 | 陕西省地震局 | Extreme seismic region intensity evaluation method and system, electronic device and readable storage medium |
| CN115293206A (en) * | 2022-08-09 | 2022-11-04 | 贵州大学 | Bearing fault diagnosis method based on sparrow algorithm optimization width learning system |
-
2023
- 2023-02-27 CN CN202310166830.0A patent/CN115859058B/en active Active
Patent Citations (12)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN103926526A (en) * | 2014-05-05 | 2014-07-16 | 重庆大学 | Analog circuit fault diagnosis method based on improved RBF neural network |
| CN108090658A (en) * | 2017-12-06 | 2018-05-29 | 河北工业大学 | Arc fault diagnostic method based on time domain charactreristic parameter fusion |
| US20190377625A1 (en) * | 2018-06-08 | 2019-12-12 | Microsoft Technology Licensing, Llc | Computing node failure and health prediction for cloud-based data center |
| CN108876054A (en) * | 2018-07-06 | 2018-11-23 | 国网河南省电力公司郑州供电公司 | Short-Term Load Forecasting Method based on improved adaptive GA-IAGA optimization extreme learning machine |
| US20200285997A1 (en) * | 2019-03-04 | 2020-09-10 | Iocurrents, Inc. | Near real-time detection and classification of machine anomalies using machine learning and artificial intelligence |
| CN111325308A (en) * | 2020-02-14 | 2020-06-23 | 集美大学 | Nonlinear system identification method |
| US20210311813A1 (en) * | 2020-07-20 | 2021-10-07 | Beijing Baidu Netcom Science And Technology Co., Ltd. | Fault prediction method, apparatus and storage medium |
| CN112804239A (en) * | 2021-01-22 | 2021-05-14 | 山东维平信息安全测评技术有限公司 | Traffic safety analysis modeling method and system |
| CN114861778A (en) * | 2022-04-22 | 2022-08-05 | 哈尔滨理工大学 | Method for rapidly classifying rolling bearing states under different loads by improving width transfer learning |
| CN114818806A (en) * | 2022-04-25 | 2022-07-29 | 重庆大学 | Gearbox fault diagnosis method based on wavelet packet and depth self-encoder |
| CN115062839A (en) * | 2022-06-13 | 2022-09-16 | 陕西省地震局 | Extreme seismic region intensity evaluation method and system, electronic device and readable storage medium |
| CN115293206A (en) * | 2022-08-09 | 2022-11-04 | 贵州大学 | Bearing fault diagnosis method based on sparrow algorithm optimization width learning system |
Non-Patent Citations (2)
| Title |
|---|
| M.M.ETTEFAGH等: "Bearing Fault Diagnosis using Hybrid Genetic Algorithm K-means Clustering" * |
| 张景景: "基于模糊神经网络的非线性系统故障诊断" * |
Cited By (7)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN116455079A (en) * | 2023-05-04 | 2023-07-18 | 奥斯塔娜(常州)电子有限公司 | Big data-based electricity consumption integrated safety supervision system and method |
| CN116455079B (en) * | 2023-05-04 | 2024-04-26 | 奥斯塔娜(常州)电子有限公司 | Big data-based electricity consumption integrated safety supervision system and method |
| CN116626514A (en) * | 2023-05-23 | 2023-08-22 | 浙江海得智慧能源有限公司 | Method, system and equipment for screening invalid battery cells |
| CN117332361A (en) * | 2023-12-01 | 2024-01-02 | 赫芝特电气技术（广州）有限公司 | Online early warning method and detection system for uninterruptible power supply |
| CN117332361B (en) * | 2023-12-01 | 2024-03-01 | 赫芝特电气技术（广州）有限公司 | Online early warning method and detection system for uninterruptible power supply |
| CN117992738A (en) * | 2024-04-03 | 2024-05-07 | 山东兰蝶节能环保科技有限公司 | Automatic electricity-saving monitoring method and system for electric energy of electricity-saving cabinet |
| CN117992738B (en) * | 2024-04-03 | 2024-06-11 | 山东兰蝶节能环保科技有限公司 | Automatic electricity-saving monitoring method and system for electric energy of electricity-saving cabinet |
Also Published As
| Publication number | Publication date |
|---|---|
| CN115859058B (en) | 2023-05-30 |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN115859058A (en) | UPS (uninterrupted Power supply) fault prediction method and system based on width learning network | |
| Ren et al. | A review of machine learning state-of-charge and state-of-health estimation algorithms for lithium-ion batteries | |
| Liu et al. | Sequence fault diagnosis for PEMFC water management subsystem using deep learning with t-SNE | |
| CN109800875A (en) | Chemical industry fault detection method based on particle group optimizing and noise reduction sparse coding machine | |
| CN112508442B (en) | Transient stability assessment method and system based on automatic and interpretable machine learning | |
| US20210334658A1 (en) | Method for performing clustering on power system operation modes based on sparse autoencoder | |
| CN115563563A (en) | Fault diagnosis method and device based on transformer oil chromatographic analysis | |
| CN114580934A (en) | Early warning method for food detection data risk based on unsupervised anomaly detection | |
| Lee et al. | An efficient prediction for heavy rain from big weather data using genetic algorithm | |
| CN115238980A (en) | Micro-grid transient power angle trajectory prediction method based on attention mechanism and space-time diagram convolutional neural network | |
| CN115018512A (en) | Electricity stealing detection method and device based on Transformer neural network | |
| CN112766618A (en) | Anomaly prediction method and device | |
| CN113361692B (en) | Lithium battery remaining life combined prediction method | |
| Li et al. | Degradation Pattern Recognition and Features Extrapolation for Battery Capacity Trajectory Prediction | |
| CN116611003A (en) | Transformer fault diagnosis method, device and medium | |
| CN112380763A (en) | System and method for analyzing reliability of in-pile component based on data mining | |
| Pietroń et al. | Fast and scalable neuroevolution deep learning architecture search for multivariate anomaly detection | |
| Wang | Analysis of bank credit risk evaluation model based on BP neural network | |
| Zhou et al. | A proactive failure tolerant mechanism for SSDs storage systems based on unsupervised learning | |
| CN114565051B (en) | Method for testing product classification model based on influence degree of neurons | |
| Kumar et al. | Software Fault Prediction using Wrapper based Feature Selection Approach employing Genetic Algorithm | |
| CN114492559A (en) | Power equipment fault diagnosis method based on data time-frequency domain modeling | |
| Liu et al. | Adaptive robustness evaluation for complex system prognostics and health management software platform | |
| CN116304846B (en) | CVT internal insulation abnormality online assessment method based on self-supervision learning | |
| CN117577981B (en) | Photovoltaic power generation energy storage control method and system |
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 |