CN108197648A - A kind of Fault Diagnosis Method of Hydro-generating Unit and system based on LSTM deep learning models - Google Patents

A kind of Fault Diagnosis Method of Hydro-generating Unit and system based on LSTM deep learning models Download PDF

Info

Publication number
CN108197648A
CN108197648A CN201711463863.2A CN201711463863A CN108197648A CN 108197648 A CN108197648 A CN 108197648A CN 201711463863 A CN201711463863 A CN 201711463863A CN 108197648 A CN108197648 A CN 108197648A
Authority
CN
China
Prior art keywords
layers
lstm
input
output
deep learning
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
Application number
CN201711463863.2A
Other languages
Chinese (zh)
Other versions
CN108197648B (en
Inventor
李超顺
王若恒
涂文奇
陈昊
陈新彪
Current Assignee (The listed assignees may be inaccurate. Google has not performed a legal analysis and makes no representation or warranty as to the accuracy of the list.)
Huazhong University of Science and Technology
Original Assignee
Huazhong University of Science and Technology
Priority date (The priority date 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 date listed.)
Filing date
Publication date
Application filed by Huazhong University of Science and Technology filed Critical Huazhong University of Science and Technology
Priority to CN201711463863.2A priority Critical patent/CN108197648B/en
Publication of CN108197648A publication Critical patent/CN108197648A/en
Application granted granted Critical
Publication of CN108197648B publication Critical patent/CN108197648B/en
Active legal-status Critical Current
Anticipated expiration legal-status Critical

Links

Classifications

    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06F—ELECTRIC DIGITAL DATA PROCESSING
    • G06F18/00—Pattern recognition
    • G06F18/20—Analysing
    • G06F18/21—Design or setup of recognition systems or techniques; Extraction of features in feature space; Blind source separation
    • G06F18/214—Generating training patterns; Bootstrap methods, e.g. bagging or boosting
    • G—PHYSICS
    • G01—MEASURING; TESTING
    • G01R—MEASURING ELECTRIC VARIABLES; MEASURING MAGNETIC VARIABLES
    • G01R31/00—Arrangements for testing electric properties; Arrangements for locating electric faults; Arrangements for electrical testing characterised by what is being tested not provided for elsewhere
    • G01R31/34—Testing dynamo-electric machines
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06F—ELECTRIC DIGITAL DATA PROCESSING
    • G06F18/00—Pattern recognition
    • G06F18/20—Analysing
    • G06F18/24—Classification techniques
    • G06F18/241—Classification techniques relating to the classification model, e.g. parametric or non-parametric approaches
    • G06F18/2415—Classification techniques relating to the classification model, e.g. parametric or non-parametric approaches based on parametric or probabilistic models, e.g. based on likelihood ratio or false acceptance rate versus a false rejection rate
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
    • G06N3/00—Computing arrangements based on biological models
    • G06N3/02—Neural networks
    • G06N3/04—Architecture, e.g. interconnection topology
    • G06N3/048—Activation functions
    • 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 a kind of Fault Diagnosis Method of Hydro-generating Unit and system based on LSTM deep learning models, this method includes：The sample sequence of N number of unlike signal channel of Hydropower Unit is obtained, carrying out VMD to each time series decomposes to obtain K IMF component；It builds corresponding training set and treats diagnosis collection；LSTM models are built to the training set of each IMF component, feature extraction is carried out to each IMF component by two layers of LSTM layers；K LSTM layers of output of same signal path are connected to one Dense layers；Tagsort is carried out to multiple Dense layers output by Softmax layers；Deep learning neural network model is trained by RMSProp gradient descent algorithms, trained model is treated diagnosis collection diagnoses.The relatively good SNR estimation and compensation effects of variation mode decomposition VMD are combined the processing advantage of time series with shot and long term memory network LSTM by the present invention, effectively raise the accuracy of Approach for Hydroelectric Generating Unit Fault Diagnosis.

Description

A kind of Fault Diagnosis Method of Hydro-generating Unit and system based on LSTM deep learning models
The invention belongs to Approach for Hydroelectric Generating Unit Fault Diagnosis technical fields, and LSTM deep learnings are based on more particularly, to one kind The Fault Diagnosis Method of Hydro-generating Unit and system of model.
Background technology
For a long time, periodic plan repair system has played important function, but this repair for the normal operation in power station System is not met by growing plant maintenance demand, and the cost of repair is still sufficiently expensive.And in social demand and Under the drive of science and technology, the fast development of fault diagnosis technology and system carries out reality for Hydropower Unit according to its operating status When monitoring provide the possibility of realization, therefore just seem to the research of Approach for Hydroelectric Generating Unit Fault Diagnosis technology and very must with maintenance It will.
With the development of machine learning, pattern-recognition, signal processing, artificial intelligence etc., based on data analysis and processing Diagnostic method starts to be developed, and conventional method is combined by this diagnostic method with intelligent method, by the big of sensor Amount data carry out processing analysis to extract fault signature.But it is opposite due to these traditional method for diagnosing faults based on data The model structure of shallow-layer is faced in the case of more and more huger with detection data such as dimension disaster, to complex nonlinear Uncertain plant learning ability it is limited wait the problems such as, conventional method seem when in face of modern complex device troubleshooting issue performance not Foot, therefore there is an urgent need to a kind of new methods to adapt to the variation of diagnostic requirements.
In view of the above problems, the new methods deep learning shot and long term memory network of the invention by artificial intelligence field (Long Short Term Memory, are abbreviated as：LSTM) with variation mode decomposition (Variational Mode Decomposition is abbreviated as：VMD it) is combined, realizes a kind of method for diagnosing faults and system for Hydropower Unit.The party Method, with respect to the network structure of conventional model deeper, is more had using deep learning model come the mass data to Hydropower Unit The feature extraction of effect, and be combined with the SNR estimation and compensation effect of VMD, in a new manner by deep learning theoretical origin to water The fault diagnosis of motor group, to reach relatively better Approach for Hydroelectric Generating Unit Fault Diagnosis effect.
Invention content
The purpose of the present invention is to provide a kind of Fault Diagnosis Method of Hydro-generating Unit based on LSTM deep learning models with System builds multi-level fuzzy judgment using the shot and long term memory network model and variation mode decomposition algorithm in deep learning field, To find a kind of more effective new method in Approach for Hydroelectric Generating Unit Fault Diagnosis field.
To achieve these goals, the technical solution adopted in the present invention is：
In a first aspect, the present invention provides a kind of Fault Diagnosis Method of Hydro-generating Unit based on LSTM deep learning models, Method includes：
Step 1：The sample sequence of N number of unlike signal channel of Hydropower Unit is obtained, each time series is become Divide mode decomposition, obtain decompositing the K IMF component come.
Step 2：Each IMF component is normalized, and builds corresponding training set and treats diagnosis collection.
Step 3：Shot and long term memory network model is built to the training set of each IMF component, is passed through at least two layers LSTM layers to carry out feature extraction to each IMF component.
Step 4：The LSTM layers output of K IMF component of same signal path is connected to one Dense layers, then will be N number of The Dense layers output of signal path is connected to Softmax layers to carry out last classification, and pass through cross entropy loss function and event Barrier label obtains error to be used to train.
Step 5：Deep learning neural network model is trained by RMSProp gradient descent algorithms, will be trained Model treat diagnosis collection diagnosed, obtain final diagnostic result.
Further, the premise that VMD is decomposed is one variational problem of construction, it is assumed that each ' mode ' is with center The finite bandwidth of frequency, variational problem can be described as seeking K IMF components uk(t), make the estimation bandwidth of each mode The sum of minimum, constraints be that each mode sum is original input signal, variational problem construction process is as follows：
(1) it is converted by Hilbert, obtains the analytic signal of each mode function, it is therefore an objective to the unilateral of it can be obtained Frequency spectrum, specific transformation is as follows, and wherein δ (t) is pulse signal function, uk(t) it is IMF components, * is convolutional calculation symbol, and j is represented Imaginary unit.Wherein,
(2) one is added in the analytic signal of each mode estimate centre frequency againFormula is as follows, wherein ωkCentered on Frequency, thus can be by the spectrum modulation of each mode to corresponding Base Band.Wherein,
(3) square L2 norms of above-mentioned demodulated signal gradient are calculated, estimate each mode signals bandwidth, construction makes total mould The variational problem of state signal bandwidth minimum, variational problem represent as follows, and wherein f is original signal f (t),Represent to the time into Row derivation, t are the time.Wherein,
Further, the algorithm of variation mode decomposition is by introducing penalty factor α and Lagrange multiplier operator λ (t) The Lagrangian formulation being expanded is as follows, and wherein f (t) is original signal.Wherein,
It is asked for after the Lagrangian formulation that extension is constructed to variational problem using alternating direction Multiplier Algorithm ADMM IMF components uk(t) it is as follows：
Step 1.1, initializationAnd n；
Step 1.2 performs cycle n=n+1；
Step 1.3, updateOnly calculate frequency domain ω>0 part, whereinFor the Fourier transformation of f (t), uk (ω) is signal uk(t) Fourier transformation, ωkCentered on frequency,More new formula is as follows：
Step 1.4, center frequency ωkRenewal equation it is as follows：
Step 1.5, update Lagrange multiplier operator λ, λ newer are：
Wherein τ is iteration coefficient；
Step 1.6 repeats step 1.2~step 1.5, until meeting stop condition, thus obtains final KStop Only condition is：
Step 1.7, by a K for the condition that meetsReal part is taken again by inversefouriertransform to be calculated last K IMF components uk(t).Specific formula is as follows, and wherein ifft () represents Fourier inversion,Expression takes real part.
It further, can be by selecting a part of IMF components gone out that same signal path is decomposed as input Classification output is carried out, analyzes influence size of the IMF components of various combination to classification results accuracy, so as to select IMF points Measure uk(t) several IMF components for being affected in result are used for feature extraction, are reached more with reducing unnecessary calculation amount High efficiency and accuracy.
Further, the time series of each IMF component is handled by shot and long term memory network LSTM, sequence Input of the value at each moment as a LSTM neuron is arranged, the forward calculation of neuron includes：
The forward calculation for forgeing door is ft=sigmoid (θf·[ht-1,xt]+bf), wherein ftTo forget the output valve of door, θfTo forget door input weight, bfTo forget door input biasing.
The forward calculation of input gate is it=sigmoid (θi·[ht-1,xt]+bi), wherein itFor the output valve of input gate, θiFor input gate input weight, biTo forget door input biasing.
The forward calculation of out gate is ot=sigmoid (θo·[ht-1,xt]+bo), wherein otFor the output valve of input gate, θoFor input gate input weight, boIt inputs and biases for input gate.
Input gate itThe input value to addition inside mnemon Cell of control is calculated as WhereinFor input gate itThe input value of control, θcFor input weight, bcIt is biased for input.
The update of the memory content of mnemon Cell is calculated asWherein ct-1It is upper one The value of the mnemon at a moment.
It finally exports to next layer of hidden layer output valve htBe calculated as ht=ot⊙tanh(ct), it is controlled by out gate Mnemon processed is exported.Wherein ht-1Hidden layer for the t-1 moment exports, xtFor input value, during the t of a corresponding IMF component The value at quarter.
Further, calculating used sigmoid activation primitive formula before is：
Used tanh activation primitive formula are：
Further, it is connected to behind LSTM layers using full articulamentum Dense layers, LSTM layers of output is defeated as its Incoming vector carries out further feature extraction, and Dense layers of activation primitive selection ReLu reduces gradient disperse problem, counts Calculation formula is Dense (x)=Re Lu (θ x+b), and wherein θ is weight, and b is biasing, and x is input, and K LSTM layers corresponding is defeated Go out the output vector of composition.
The calculation formula of ReLu is：
Further, by Softmax layers Dense layers of feature vector is carried out tagsort obtain probability output to Amount, Softmax recurrence are the graders that depth network supervised learning part is widely used in current depth Learning Studies.Make For a Nonlinear Classifier, it is combined with the part of unsupervised learning in depth network, tends to reach very high Classification accuracy, specific formula are：
Wherein x be Softmax layers input to Amount is in a model Dense layer of output, and θ is softmax layers of weight matrix, and element p in the vector of the left side (y=n | x；θ) The probability that the value for representing output y is n, finally forms a probability output vector.
Further, Dropout technologies has been used in the LSTM multilayer deep learning neural network models of structure Prevent the over-fitting of neural network model.Dropout is that a certain proportion of implicit node is allowed not work at random in training pattern, this A little idle nodes can will not carry out right value update in this time training, and it is next time trained when can in proportion select again not The node of work, this method can effectively reduce over-fitting and reduce the error rate of classification.
Since sigmoid etc. swashs when further, in order to overcome General loss function such as mean square deviation loss function derivation The gradient disperse problem that the characteristic of function living is brought, the present invention use loss letter of the cross entropy loss function as entire model Number, loss function are by the way that desired output and reality output are carried out a kind of parameter of an optimization being calculated, tool Body formula is：
Wherein L (f (x(i)；θ),y(i)) for loss function, m is number of samples, y(i)For the desired output of i-th of sample, f (x(i)；θ) the reality output for i-th of sample, corresponding Softmax layers practical probability output vector.
Further, gradient optimal method when gradient declines training uses RMSProp algorithms, RMSProp algorithm needles There is relatively common gradient descent algorithm to have relatively better training effect to Recognition with Recurrent Neural Network.RMSProp in this model Algorithm is by combining cross entropy loss function L (f (x(i)；θ),y(i)) gradient derivation is carried out to each weight θ, being obtained makes loss The weight changing value that function becomes smaller, while gradient cumulative amount r and attenuation coefficient ρ are introduced, accelerate convergence rate to a certain extent And suboptimization is avoided, specific gradient updating step is as follows：
Step 5.1, the size for calculating gradient g, L (f (x(i)；θ),y(i)) it is cross entropy loss function,It represents to weight θ Derivation.
Wherein,
Step 5.2, calculating gradient cumulative amount r, wherein ρ are attenuation coefficient.Wherein, r ← ρ r+ (1- ρ) g ⊙ g.
Step 5.3, calculating weight updated value Δ θ, η are learning rate, and δ (usually takes 10 for a small constant-6), for small Numerical stability when number removes.Wherein,
Step 5.4, the value that weight θ is updated by Δ θ.Wherein, θ ← θ+Δ θ.
Second aspect, the present invention also provides a kind of Approach for Hydroelectric Generating Unit Fault Diagnosis systems based on LSTM deep learning models System, system include：
IMF component acquiring units, for obtaining the sample sequence of N number of unlike signal channel of Hydropower Unit, to each Time series carries out variation mode decomposition, obtains K IMF component；
It training set and treats diagnosis collection acquiring unit, for each IMF component to be normalized, and builds corresponding Training set and treat diagnosis collection；
Feature extraction unit, for building shot and long term memory network model to the training set of each IMF component, by extremely LSTM layers two layers few to carry out feature extraction to each intrinsic mode function；
Training module, for K LSTM layers of output of same signal path to be connected to one Dense layers, then will be N number of The Dense layers output of signal path is connected to Softmax layers to classify, and pass through cross entropy loss function and faulty tag Error is obtained to be used to train；
Diagnostic module is trained deep learning neural network model for passing through RMSProp gradient descent algorithms, will Trained model is treated diagnosis collection and is diagnosed, and obtains final diagnostic result.
LSTM neural networks in variation mode decomposition VMD algorithms and deep learning are combined to form one kind by the present invention More accurate efficient algorithm, to improve the accuracy of current Approach for Hydroelectric Generating Unit Fault Diagnosis, also further improves the dimension of hydroelectric power plant Repair efficiency.
Further, larger noise element is had in the signal path that the present invention is obtained for Hydropower Unit in-site measurement The problem of, vibration signal is decomposed using variation mode decomposition VMD to obtain the effective waveform component after SNR estimation and compensation Extract feature again.(Recurrent Neural Networks, are abbreviated as Recognition with Recurrent Neural Network simultaneously：RNN) this feedback-type Neural network and its variant shot and long term memory network LSTM, gating cycle unit GRU etc. have the data of time series More efficient processing capacity.Because there is gradient disperse when propagating training along time reversal in RNN neural networks, from And the information of long period span can not be used effectively, therefore selection uses the shot and long term in deep learning in the present invention Memory network LSTM carries out feature extraction to the vector sequence after VMD SNR estimation and compensations.
Description of the drawings
Fig. 1 is the Approach for Hydroelectric Generating Unit Fault Diagnosis mould provided in an embodiment of the present invention based on shot and long term memory network deep learning Type figure；
Fig. 2 is the calculating expanded view of Recognition with Recurrent Neural Network provided in an embodiment of the present invention；
Fig. 3 is the memory internal unit of shot and long term memory network provided in an embodiment of the present invention and the calculating figure of door；
Fig. 4 is a kind of Approach for Hydroelectric Generating Unit Fault Diagnosis system based on LSTM deep learning models provided in an embodiment of the present invention Organization Chart.
Specific embodiment
In order to make the purpose , technical scheme and advantage of the present invention be clearer, below in conjunction with attached drawing to the present invention into Row is further described.It should be appreciated that specific steps described herein are only used to explain the present invention, it is not used to limit The present invention.
Embodiment 1：
A kind of Fault Diagnosis Method of Hydro-generating Unit based on LSTM deep learning models that the embodiment of the present invention proposes, reference Diagnostic model figure as shown in Figure 1, specific implementation step are as follows：
Step 1：Obtain the sampled value of the vibration signal of N number of unlike signal channel of Hydropower Unit.Each signal is led to The vibration signal in road carries out variation mode decomposition, obtains decompositing the K IMF component come.
Wherein, the premise that VMD is decomposed is one variational problem of construction, it is assumed that each ' mode ' has centre frequency Finite bandwidth, variational problem can be described as seeking K IMF components uk(t), make the sum of estimation bandwidth of each mode most Small, constraints is that each mode sum is original input signal, and variational problem construction process is as follows：
(1) it is converted by Hilbert, obtains the analytic signal of each mode function, it is therefore an objective to the unilateral of it can be obtained Frequency spectrum, specific transformation is as follows, and wherein δ (t) is pulse signal function, uk(t) it is IMF components, * is convolutional calculation symbol, and j is represented Imaginary unit.
(2) one is added in the analytic signal of each mode estimate centre frequency againFormula is as follows, wherein ωkFor in Frequency of heart, thus can be by the spectrum modulation of each mode to corresponding Base Band.
(3) square L2 norms of above-mentioned demodulated signal gradient are calculated, estimate each mode signals bandwidth, construction makes total mould The variational problem of state signal bandwidth minimum, variational problem represent as follows, and wherein f is original signal f (t),Represent to the time into Row derivation, t are the time.
Further, the algorithm of variation mode decomposition is by introducing penalty factor α and Lagrange multiplier operator λ (t) The Lagrangian formulation being expanded is as follows, and wherein f (t) is original signal.
It is asked for after the Lagrangian formulation that extension is constructed to variational problem using alternating direction Multiplier Algorithm ADMM IMF components uk(t) it is as follows：
Step 1.1, initializationAnd n；
Step 1.2 performs cycle n=n+1；
Step 1.3, updateOnly calculate frequency domain ω>0 part, calculation formula is as follows whereinFor in Fu of f (t) Leaf transformation,For signal uk(t) Fourier transformation, ωkCentered on frequency；
Step 1.4, center frequency ωkRenewal equation it is as follows：
Step 1.5, update Lagrange multiplier operator λ, λ newer are as follows, and wherein τ is iteration coefficient：
Step 1.6 repeats step 1.2~step 1.5, until meeting stop condition, thus obtains final KStop Only condition is：
Step 1.7, by a K for the condition that meetsReal part is taken again by inversefouriertransform to be calculated last K IMF components uk(t), formula is as follows, and wherein ifft () represents Fourier inversion,Expression takes real part.
Each IMF component is normalized in step 2, and builds corresponding training set and treat diagnosis collection.Normalizing During change to same channel important progress normalization on the whole, while the input sequence in training set as a sample The length of row can take with the sample sequence of the time span of the signal comprising lowest frequency with guarantee to all useful informations into Row extraction, while sample information is divided into multiple periods as sample, it divides training set using certain proportion and waits to diagnose Collection.
Step 3 builds shot and long term memory network model to the training set of each IMF component, passes through two layers of LSTM layers To carry out feature extraction to each IMF component.Wherein first layer LSTM networks output sequence in a manner of multi output, the sequence are used Feature extraction is carried out again in being supplied to LSTM layers of the second layer.LSTM layers of the second layer exports, then the LSTM with same channel to be single Layer output merges, and the feature vector that the LSTM layers output of K obtained IMF component merges is used to connect entirely as Dense The input of layer.Simultaneously two layers LSTM layers take Dropout values as 0.2 to prevent over-fitting.
The forward calculation process of wherein LSTM is as shown in figure 3, as follows, the wherein h that specifically calculates stept-1For t-1 moment hidden layers Output, xtFor input value, the value of the t moment of a corresponding IMF component.
The calculating for forgeing door is as follows, wherein ftTo forget the output valve of door, θfTo forget door input weight, bfTo forget door Input biasing.
ft=sigmoid (θf·[ht-1,xt]+bf) (10)
The calculating of input gate is as follows, wherein itFor the output valve of input gate, θiFor input gate input weight, biTo forget door Input biasing.
it=sigmoid (θi·[ht-1,xt]+bi) (11)
The calculating of out gate is as follows, wherein otFor the output valve of input gate, θoFor input gate input weight, boTo forget door Input biasing.
ot=sigmoid (θo·[ht-1,xt]+bo) (12)
Input gate itThe input value calculating to addition inside mnemon Cell of control is as follows, whereinFor input gate it The input value of control, θcFor input weight, bcIt is biased for input.
As follows, wherein c is calculated to the update of the memory content of mnemon Cellt-1Mnemon for a upper moment Value.
Hidden layer output valve htCalculating it is as follows.
ht=ot⊙tanh(ct) (15)
Step 4, same signal path K IMF component LSTM layers output be connected to one Dense layers, then will be N number of The Dense layers output of signal path is connected to Softmax layers to carry out last classification, and pass through intersection
Error is obtained with faulty tag to be used to train.
Wherein Dense layers is used to carry out further feature extraction to the feature of LSTM layers of extraction, is used further to next layer The classified calculating of Softmax.To prevent over-fitting, and calculation amount is reduced, it is 0.5 that Dense layers, which take Dropout values,.Dense layers Calculation formula is as follows, and wherein θ is weight, and x is input, and b is biasing, and ReLu is activation primitive：
Dense (x)=ReLu (θ x+b) (16)
The calculation formula of ReLu is as follows：
Output is normalized by Softmax layers in Dense layers of output vector, and Softmax layers of output are corresponding each The probability of failure, each of the vector of output represent probability size of this corresponding for 1 Hydropower Unit failure, pass through probability Size is that can determine whether the corresponding failure cause of input most probable of signal path at this time.The calculation formula of Softmax is as follows, wherein X is Softmax layer of input vector, is in a model Dense layers of output, and θ is softmax layers of weight matrix, the left side to In amount element p (y=n | x；The probability that the value for θ) representing output y is n, finally forms a probability output vector.
Due to the characteristic of the activation primitives such as sigmoid during in order to overcome General loss function such as mean square deviation loss function derivation The gradient disperse problem brought, corresponding error loss function are constructed by cross entropy to realize.Loss function be pass through by Desired output carries out a kind of parameter of an optimization being calculated with reality output, and the calculating of cross entropy is as follows, wherein L (f(x(i)；θ),y(i)) it is loss function, m is number of samples, is the desired output of i-th of sample, is the reality of i-th of sample Output, corresponding Softmax layers practical probability output vector.
Derivation of the cross entropy when reversely asking for gradient is as follows, represents to inputWeight θjCarry out derived function：
Step 5 is trained deep learning neural network model by RMSProp gradient descent algorithms, will train Model treat diagnosis collection diagnosed, obtain final diagnostic result.RMSProp algorithms have relatively for Recognition with Recurrent Neural Network Common gradient descent algorithm has relatively better training effect, by combining cross entropy loss function L (f (x(i)；θ),y(i)) Gradient derivation is carried out to each weight θ and biasing b, the weight changing value that loss function is made to become smaller is obtained, while introduce gradient and tire out R and attenuation coefficient ρ is measured, accelerate convergence rate to a certain extent and avoids suboptimization.RMSprop algorithm iteration processes It is as follows：
Step 5.1, the size for calculating gradient g, L (f (x(i)；θ),y(i)) it is cross entropy loss function,It represents to weight θ Derivation.
Step 5.2, calculating gradient cumulative amount r, wherein ρ are attenuation coefficient.
r←ρr+(1-ρ)g⊙g (22)
Step 5.3, calculating weight updated value Δ θ, η are learning rate, and δ (usually takes 10 for a small constant-6), for small Numerical stability when number removes.
Step 5.4, the value that weight θ is updated by Δ θ.
θ←θ+Δθ (24)
It is trained the deep learning network model after optimization to can be used to treat diagnosis collection progress fault diagnosis, specially The vibration signal sequence of each signal path of certain a kind of failure is carried out to be input to network after VMD decomposes renormalization processing In model, obtained final probability vector is analyzed, if who the probability value in the vector of the corresponding failure is most Greatly and closest to 1, show that the network diagnosis likelihood of failure is maximum, the diagnostic result for representing the failure is correct.
Embodiment 2：
The embodiment of the present invention additionally provides a kind of Approach for Hydroelectric Generating Unit Fault Diagnosis system based on LSTM deep learning models, such as Shown in Fig. 4, system includes：
IMF component acquiring units, for obtaining the sample sequence of N number of unlike signal channel of Hydropower Unit, to each Time series carries out variation mode decomposition, obtains K IMF component；
It training set and treats diagnosis collection acquiring unit, for each IMF component to be normalized, and builds corresponding Training set and treat diagnosis collection；
Feature extraction unit, for building shot and long term memory network model to the training set of each IMF component, by extremely LSTM layers two layers few to carry out feature extraction to each intrinsic mode function；
Training module, for K LSTM layers of output of same signal path to be connected to one Dense layers, then will be N number of The Dense layers output of signal path is connected to Softmax layers to classify, and pass through cross entropy loss function and faulty tag Error is obtained to be used to train；
Diagnostic module is trained deep learning neural network model for passing through RMSProp gradient descent algorithms, will Trained model is treated diagnosis collection and is diagnosed, and obtains final diagnostic result.
As it will be easily appreciated by one skilled in the art that the foregoing is merely the specific implementation step of the present invention, not limiting The system present invention, all any modification, equivalent and improvement made all within the spirits and principles of the present invention etc., should be included in Within protection scope of the present invention.

Claims (10)

1. a kind of Fault Diagnosis Method of Hydro-generating Unit based on LSTM deep learning models, which is characterized in that method includes：
The sample sequence of N number of unlike signal channel of Hydropower Unit is obtained, variation mode decomposition is carried out to each time series, Obtain K IMF component；
Each IMF component is normalized, and builds corresponding training set and treats diagnosis collection；
Shot and long term memory network model is built to the training set of each IMF component, by least two layers of LSTM layers come to every One intrinsic mode function carries out feature extraction；
K LSTM layers of output of same signal path are connected to one Dense layers, then the Dense layers of N number of signal path is defeated Go out to be connected to Softmax layers to classify, and error is obtained to be used to train with faulty tag by cross entropy loss function；
Deep learning neural network model is trained by RMSProp gradient descent algorithms, trained model is treated Diagnosis collection is diagnosed, and obtains final diagnostic result.
2. the Fault Diagnosis Method of Hydro-generating Unit according to claim 1 based on LSTM deep learning models, feature exist In, it is described that variation mode decomposition is carried out to each time series, K IMF component is obtained, is specifically included：
Make the sum of the estimation bandwidth of each mode minimum, constraints is that each mode sum is original input signal, is become Divide problem construction process as follows：
Pass throughHilbert transformation is carried out, obtains the analytic signal of each mode function, wherein δ (t) is Pulse signal function, uk(t) it is IMF components, * is convolutional calculation symbol, and j represents imaginary unit；
Pass through formulaOne, which is added in, to the analytic signal of each mode estimates centre frequencyWherein ωkCentered on frequency so that the spectrum modulation of each mode to corresponding Base Band；
Pass through formula againIt calculates above-mentioned Square L2 norms of demodulated signal gradient, estimate each mode signals bandwidth, by the sum of bandwidth of mode signals of estimation, Construction makes the variation Solve problems of total bandwidth minimum, for subsequent variation mode decomposition；Wherein f (t) is original signal,Table Show and derivation is carried out to the time, t is the time.
3. the Fault Diagnosis Method of Hydro-generating Unit according to claim 2 based on LSTM deep learning models, feature exist In described by the sum of bandwidth of mode signals of estimation, construction makes the variation Solve problems of total bandwidth minimum, for subsequent Variation mode decomposition, specifically includes：
The algorithm of variation mode decomposition is bright by introducing the glug that penalty factor α and Lagrange multiplier operator λ (t) are expanded Day, expression formula was completed, and wherein Lagrangian formulation is as follows：
Wherein f (t) is original signal；
After the Lagrangian formulation that extension is constructed to variational problem IMF points are asked for using alternating direction Multiplier Algorithm ADMM Measure uk(t) it is as follows：
Step 1.1, initializationAnd n；
Step 1.2 performs cycle n=n+1；
Step 1.3, updateOnly calculate frequency domain ω>0 part, whereinFor the Fourier transformation of original signal f (t),For signal uk(t) Fourier transformation, ωkCentered on frequency；
Step 1.4, center frequency ωkRenewal equation it is as follows：
Step 1.5, update Lagrange multiplier operator λ, λ newer are as follows, and wherein τ is iteration coefficient：
Step 1.6 repeats step 1.2~step 1.5, until meeting stop conditionBy This obtained K
Step 1.7, the K that stop condition will be metBy inversefouriertransform, realistic portion is a last K is calculated again IMF components uk(t), calculating formula isWherein ifft () represents Fourier inversion,Table Show and take real part.
4. the Fault Diagnosis Method of Hydro-generating Unit based on LSTM deep learning models according to Claims 2 or 3, feature It is, by the way that a part of IMF components gone out that same signal path is decomposed is selected to carry out classification output as input, analysis is not With influence size of the IMF components combined to classification results accuracy, IMF components u is selectedk(t) it is affected in result IMF components be used for feature extraction.
5. according to any Fault Diagnosis Method of Hydro-generating Unit based on LSTM deep learning models of claim 1-3, Be characterized in that, the time series of each IMF component handled by shot and long term memory network LSTM, sequence each when Input of the value at quarter as a LSTM neuron, the forward calculation of neuron include：
The forward direction of door is forgotten according to ft=sigmoid (θf·[ht-1,xt]+bf) calculate, wherein ftTo forget the output valve of door, θf To forget door input weight, bfTo forget door input biasing；
The forward direction of input gate is according to it=sigmoid (θi·[ht-1,xt]+bi) calculate, wherein itFor the output valve of input gate, θi For input gate input weight, biTo forget door input biasing；
The forward direction of out gate is according to ot=sigmoid (θo·[ht-1,xt]+bo) calculate, wherein otFor the output valve of input gate, θo For input gate input weight, boIt inputs and biases for input gate；
Input gate itControl to the input value added inside mnemon Cell according toMeter It calculates, whereinFor input gate itThe input value of control, θcFor input weight, bcIt is biased for input；
To the update of the memory content of mnemon Cell according toIt calculates, wherein ct-1It is upper one The value of the mnemon at moment；
It exports to next layer of hidden layer output valve htAccording to ht=ot⊙tanh(ct) calculate, and control memory single by out gate Member is exported；Wherein ht-1Hidden layer for the t-1 moment exports, xtFor input value, the value of the t moment of a corresponding IMF component.
6. according to any Fault Diagnosis Method of Hydro-generating Unit based on LSTM deep learning models of claim 1-3, Be characterized in that, be connected to behind LSTM layers using full articulamentum Dense layers, using LSTM layers of output as its input vector come into Row feature extraction obtains feature vector；Dense layers of activation primitive selection ReLu, the calculation formula of Dense for Dense (x)= ReLu (θ x+b), wherein θ are weight, and b is biasing, and x is input, the output vector of output composition K LSTM layers corresponding；
The calculation formula of ReLu is：
7. the Fault Diagnosis Method of Hydro-generating Unit according to claim 6 based on LSTM deep learning models, feature exist In carrying out tagsort to Dense layers of feature vector by Softmax layers and obtain probability output vector, specific formula is such as Under：
Wherein x is Softmax layer of input vector, is in a model Dense layers of output, and θ is softmax layers of weight square Battle array, in the vector of the left side element p (y=n | x；The probability that the value for θ) representing output y is n, corresponding multiple probability value compositions Softmax layers of probability output vector.
8. the Fault Diagnosis Method of Hydro-generating Unit according to claim 1 based on LSTM deep learning models, feature exist In using loss function of the cross entropy loss function as entire model, loss function is by the way that desired output and reality is defeated Go out and carry out parameter, specific formula is：
Wherein L (f (x(i)；θ),y(i)) for loss function, m is number of samples, y(i)For the desired output of i-th of sample, f (x(i)； θ) the reality output for i-th of sample, corresponding Softmax layers practical probability output vector.
9. the Fault Diagnosis Method of Hydro-generating Unit according to claim 1 based on LSTM deep learning models, feature exist In RMSProp algorithms are by combining cross entropy loss function L (f (x(i)；θ),y(i)) gradient derivation is carried out to each weight θ, it asks The weight changing value that loss function of sening as an envoy to becomes smaller, while introduce gradient cumulative amount r and attenuation coefficient ρ, specific gradient updating step It is as follows：
Step 5.1, basisCalculate the size of gradient g, L (f (x(i)；θ),y(i)) it is cross entropy Loss function,It represents to weight θ derivations；
Step 5.2 calculates gradient cumulative amount r according to r ← ρ r+ (1- ρ) g ⊙ g, and wherein ρ is attenuation coefficient；
Step 5.3, basisIt is learning rate to calculate weight updated value Δ θ, η, and δ is (logical for one small constant Often take 10-6), numerical stability during for being removed by decimal；
Step 5.4, according to θ ← θ+Δ θ, the value of update weight θ.
10. a kind of Approach for Hydroelectric Generating Unit Fault Diagnosis system based on LSTM deep learning models, which is characterized in that system includes：
IMF component acquiring units, for obtaining the sample sequence of N number of unlike signal channel of Hydropower Unit, to each time Sequence carries out variation mode decomposition, obtains K IMF component；
It training set and treats diagnosis collection acquiring unit, for each IMF component to be normalized, and builds corresponding instruction Practice collection and treat diagnosis collection；
Feature extraction unit for building shot and long term memory network model to the training set of each IMF component, passes through at least two The LSTM layers of layer to carry out feature extraction to each intrinsic mode function；
Training module, for K LSTM layers of same signal path output to be connected to one Dense layers, then by N number of signal The Dense layers output of channel is connected to Softmax layers to classify, and pass through cross entropy loss function and obtained with faulty tag Error is used to training；
Diagnostic module is trained deep learning neural network model for passing through RMSProp gradient descent algorithms, by training Good model is treated diagnosis collection and is diagnosed, and obtains final diagnostic result.
CN201711463863.2A 2017-12-28 2017-12-28 Hydroelectric generating set fault diagnosis method and system based on LSTM deep learning model Active CN108197648B (en)

Priority Applications (1)

Application Number Priority Date Filing Date Title
CN201711463863.2A CN108197648B (en) 2017-12-28 2017-12-28 Hydroelectric generating set fault diagnosis method and system based on LSTM deep learning model

Applications Claiming Priority (1)

Application Number Priority Date Filing Date Title
CN201711463863.2A CN108197648B (en) 2017-12-28 2017-12-28 Hydroelectric generating set fault diagnosis method and system based on LSTM deep learning model

Publications (2)

Publication Number Publication Date
CN108197648A true CN108197648A (en) 2018-06-22
CN108197648B CN108197648B (en) 2020-06-05

Family

ID=62585562

Family Applications (1)

Application Number Title Priority Date Filing Date
CN201711463863.2A Active CN108197648B (en) 2017-12-28 2017-12-28 Hydroelectric generating set fault diagnosis method and system based on LSTM deep learning model

Country Status (1)

Country Link
CN (1) CN108197648B (en)

Cited By (52)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN109035226A (en) * 2018-07-12 2018-12-18 武汉精测电子集团股份有限公司 Mura defects detection method based on LSTM model
CN109129574A (en) * 2018-11-08 2019-01-04 山东大学 Service robot kinematic system cloud fault diagnosis system and method
CN109143851A (en) * 2018-07-11 2019-01-04 佛山科学技术学院 The method of the identification of multiple labeling failure deep learning and its result intelligent expression
CN109242212A (en) * 2018-10-16 2019-01-18 中国矿业大学 A kind of wind-powered electricity generation prediction technique based on change Mode Decomposition and length memory network
CN109276255A (en) * 2018-11-27 2019-01-29 平安科技（深圳）有限公司 A kind of limb tremor detection method and device
CN109276244A (en) * 2018-09-03 2019-01-29 南京理工大学 The recognition methods that age-care based on brain wave information is intended to
CN109408552A (en) * 2018-08-08 2019-03-01 南京航空航天大学 The monitoring of the civil aircraft system failure and recognition methods based on LSTM-AE deep learning frame
CN109443752A (en) * 2018-10-17 2019-03-08 北京信息科技大学 A kind of Gearbox vibration signal noise reduction and method for diagnosing faults based on VMD
CN109460005A (en) * 2018-11-01 2019-03-12 上海理工大学 Dynamic industrial process method for diagnosing faults based on GRU deep neural network
CN109555566A (en) * 2018-12-20 2019-04-02 西安交通大学 A kind of turbine rotor method for diagnosing faults based on LSTM
CN109583346A (en) * 2018-11-21 2019-04-05 齐鲁工业大学 EEG feature extraction and classifying identification method based on LSTM-FC
CN109657604A (en) * 2018-12-17 2019-04-19 中国人民解放军战略支援部队信息工程大学 Satellite width phase signals identification demodulation method and device based on Recognition with Recurrent Neural Network
CN109738776A (en) * 2019-01-02 2019-05-10 华南理工大学 Fan converter open-circuit fault recognition methods based on LSTM
CN109781399A (en) * 2019-02-27 2019-05-21 天津大学 A kind of air cooling refrigeration unit sensor fault diagnosis method neural network based
CN109814523A (en) * 2018-12-04 2019-05-28 合肥工业大学 Method for diagnosing faults based on CNN-LSTM deep learning method and more attribute time series datas
CN109931678A (en) * 2019-03-13 2019-06-25 中国计量大学 Air-conditioning fault diagnosis method based on deep learning LSTM
CN109948833A (en) * 2019-02-25 2019-06-28 华中科技大学 A kind of Hydropower Unit degradation trend prediction technique based on shot and long term memory network
CN110046663A (en) * 2019-04-16 2019-07-23 重庆大学 A kind of complex electromechanical systems fault critical state discrimination method
CN110069886A (en) * 2019-05-05 2019-07-30 四川大学 The identification of cable initial failure and classification method based on VMD and CNN
CN110153802A (en) * 2019-07-04 2019-08-23 西南交通大学 A kind of cutting-tool wear state discrimination method based on convolutional neural networks and long Memory Neural Networks conjunctive model in short-term
CN110232395A (en) * 2019-03-01 2019-09-13 国网河南省电力公司电力科学研究院 A kind of fault diagnosis method of electric power system based on failure Chinese text
CN110263846A (en) * 2019-06-18 2019-09-20 华北电力大学 The method for diagnosing faults for being excavated and being learnt based on fault data depth
CN110261773A (en) * 2019-07-01 2019-09-20 南京航空航天大学 A kind of aerogenerator failure symptom extracting method and system
CN110297479A (en) * 2019-05-13 2019-10-01 国网浙江省电力有限公司紧水滩水力发电厂 A kind of Fault Diagnosis Method of Hydro-generating Unit based on the fusion of convolutional neural networks information
CN110309603A (en) * 2019-07-05 2019-10-08 华北电力大学（保定） A kind of short-term wind speed forecasting method and system based on wind speed characteristics
CN110441065A (en) * 2019-07-04 2019-11-12 杭州华电江东热电有限公司 Gas turbine online test method and device based on LSTM
CN110458281A (en) * 2019-08-02 2019-11-15 中科新松有限公司 The deeply study rotation speed prediction technique and system of ping-pong robot
CN110531266A (en) * 2019-10-14 2019-12-03 国网山东省电力公司电力科学研究院 A kind of synchronous machinery excitation winding interturn short-circuit fault early warning method
CN110609229A (en) * 2019-09-17 2019-12-24 电子科技大学 Wind driven generator blade imbalance fault detection method based on deep learning
CN110736968A (en) * 2019-10-16 2020-01-31 清华大学 Radar abnormal state diagnosis method based on deep learning
CN110858289A (en) * 2018-08-24 2020-03-03 河南工业大学 Grain moisture detection method and system based on deep learning
CN110865625A (en) * 2018-08-28 2020-03-06 中国科学院沈阳自动化研究所 Process data anomaly detection method based on time series
CN110909782A (en) * 2019-11-15 2020-03-24 湘潭大学 Method for diagnosing machine tool spindle fault based on multi-feature combined deep learning
CN110991471A (en) * 2019-08-01 2020-04-10 南京航空航天大学 Fault diagnosis method for high-speed train traction system
CN111008203A (en) * 2018-11-20 2020-04-14 中国石油化工股份有限公司 Method, device and terminal for diagnosing misalignment positioning of oil well metering instrument
CN111175852A (en) * 2019-12-27 2020-05-19 中国电子科技集团公司第十四研究所 Airport fog forecast early warning method based on long-time memory algorithm
CN111222467A (en) * 2020-01-08 2020-06-02 燕山大学 Wind driven generator component fault classification detection method
CN111273623A (en) * 2020-02-25 2020-06-12 电子科技大学 Fault diagnosis method based on Stacked LSTM
CN111428788A (en) * 2020-03-24 2020-07-17 西安交通大学 Deep learning-based multi-fault diagnosis method and system for steam turbine generator set rotor
CN111458640A (en) * 2020-04-22 2020-07-28 西安交通大学 Three-phase current-based alternating current asynchronous motor rotor broken bar fault diagnosis method
CN112101116A (en) * 2020-08-17 2020-12-18 北京无线电计量测试研究所 Analog circuit fault diagnosis method based on deep learning
CN112101244A (en) * 2020-09-18 2020-12-18 无锡科晟光子科技有限公司 Method for predicting optical fiber vibration signal danger level based on deep learning
CN112146879A (en) * 2020-08-21 2020-12-29 江苏大学 Rolling bearing fault intelligent diagnosis method and system
CN112488171A (en) * 2020-11-25 2021-03-12 江南大学 Variable mode decomposition SRU fault prediction method
CN113392771A (en) * 2021-06-16 2021-09-14 天津职业技术师范大学(中国职业培训指导教师进修中心) Plant growth state diagnosis method, system and device based on deep learning
CN113485102A (en) * 2021-06-22 2021-10-08 西安法士特汽车传动有限公司 Method for identifying vehicle running condition based on long-term and short-term memory neural network
CN113609932A (en) * 2021-07-21 2021-11-05 广州大学 Fault detection method, system, device and medium based on long-term and short-term memory network
CN113777485A (en) * 2021-09-07 2021-12-10 江南大学 Permanent magnet synchronous motor open-circuit fault diagnosis method based on value function error
CN113822329A (en) * 2021-08-10 2021-12-21 国网新源控股有限公司 Hydro-power generating unit main shaft throw signal processing method and device
CN115221769A (en) * 2021-04-15 2022-10-21 广州中国科学院先进技术研究所 Fault prediction method, system, electronic equipment and storage medium
CN116340830A (en) * 2023-05-19 2023-06-27 山东通维信息工程有限公司 Highway electromechanical system fault classification method based on deep memory model
CN117557304A (en) * 2024-01-11 2024-02-13 国网浙江省电力有限公司 Electric quantity and electricity price level fusion prediction method based on modal decomposition and neural network

Citations (3)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN106596116A (en) * 2016-11-29 2017-04-26 西安理工大学 Vibration fault diagnosis method of wind generating set
CN107044349A (en) * 2017-04-17 2017-08-15 清华大学深圳研究生院 A kind of breakdown in the motor Forecasting Methodology
CN107315892A (en) * 2017-08-10 2017-11-03 北京交通大学 A kind of Method for Bearing Fault Diagnosis based on extreme learning machine

Patent Citations (3)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN106596116A (en) * 2016-11-29 2017-04-26 西安理工大学 Vibration fault diagnosis method of wind generating set
CN107044349A (en) * 2017-04-17 2017-08-15 清华大学深圳研究生院 A kind of breakdown in the motor Forecasting Methodology
CN107315892A (en) * 2017-08-10 2017-11-03 北京交通大学 A kind of Method for Bearing Fault Diagnosis based on extreme learning machine

Non-Patent Citations (6)

* Cited by examiner, † Cited by third party
Title
MEI YUAN,YUTING WU AND LI LIN: "Fault diagnosis and remaining useful life estimation of aero engine using LSTM neural network", 《2016 IEEE/CSAA INTERNATIONAL CONFERENCE ON AIRCRAFT UTILITY SYSTEMS (AUS)》 *
NOBUYUKI YOSHIKAWA,NACIM BELKHIR AND SINJI SUZUKI: "Recurrent Neural Network-based Fault Detector for Aileron Failures of Aircraft", 《2017 11TH ASIAN CONTROL CONFERENCE (ASCC)》 *
付文龙: "水电机组振动信号分析与智能故障诊断方法研究", 《中国博士学位论文全文数据库 工程科技Ⅱ辑（月刊）》 *
张勋康，陈文献，杨洋，李涛涛: "基于ＶＭＤ分解和支持向量机的水电机组振动故障诊断", 《电网与清洁能源》 *
张建付，宋雨，李刚，王传洋，焦亚菲: "基于长短时记忆神经网络的风电机组滚动轴承故障诊断方法", 《计算机测量与控制》 *
杨洪柏，张宏利，蒋超，刘树林: "变分模态分解与神经网络结合的轴承故障诊断", 《机械设计与制造》 *

Cited By (76)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN109143851A (en) * 2018-07-11 2019-01-04 佛山科学技术学院 The method of the identification of multiple labeling failure deep learning and its result intelligent expression
CN109143851B (en) * 2018-07-11 2021-06-01 佛山科学技术学院 Method for recognizing multi-mark fault deep learning and intelligently expressing result thereof
CN109035226B (en) * 2018-07-12 2021-11-23 武汉精测电子集团股份有限公司 Mura defect detection method based on LSTM model
CN109035226A (en) * 2018-07-12 2018-12-18 武汉精测电子集团股份有限公司 Mura defects detection method based on LSTM model
CN109408552A (en) * 2018-08-08 2019-03-01 南京航空航天大学 The monitoring of the civil aircraft system failure and recognition methods based on LSTM-AE deep learning frame
CN109408552B (en) * 2018-08-08 2022-01-04 南京航空航天大学 Fault monitoring and identifying method for civil aircraft system based on LSTM-AE deep learning framework
CN110858289B (en) * 2018-08-24 2023-11-03 河南工业大学 Grain moisture detection method and system based on deep learning
CN110858289A (en) * 2018-08-24 2020-03-03 河南工业大学 Grain moisture detection method and system based on deep learning
CN110865625A (en) * 2018-08-28 2020-03-06 中国科学院沈阳自动化研究所 Process data anomaly detection method based on time series
CN109276244A (en) * 2018-09-03 2019-01-29 南京理工大学 The recognition methods that age-care based on brain wave information is intended to
CN109242212A (en) * 2018-10-16 2019-01-18 中国矿业大学 A kind of wind-powered electricity generation prediction technique based on change Mode Decomposition and length memory network
CN109443752A (en) * 2018-10-17 2019-03-08 北京信息科技大学 A kind of Gearbox vibration signal noise reduction and method for diagnosing faults based on VMD
CN109460005A (en) * 2018-11-01 2019-03-12 上海理工大学 Dynamic industrial process method for diagnosing faults based on GRU deep neural network
CN109129574A (en) * 2018-11-08 2019-01-04 山东大学 Service robot kinematic system cloud fault diagnosis system and method
CN111008203A (en) * 2018-11-20 2020-04-14 中国石油化工股份有限公司 Method, device and terminal for diagnosing misalignment positioning of oil well metering instrument
CN109583346A (en) * 2018-11-21 2019-04-05 齐鲁工业大学 EEG feature extraction and classifying identification method based on LSTM-FC
CN109276255A (en) * 2018-11-27 2019-01-29 平安科技（深圳）有限公司 A kind of limb tremor detection method and device
CN109276255B (en) * 2018-11-27 2023-02-28 平安科技（深圳）有限公司 Method and device for detecting tremor of limbs
CN109814523B (en) * 2018-12-04 2020-08-28 合肥工业大学 CNN-LSTM deep learning method and multi-attribute time sequence data-based fault diagnosis method
CN109814523A (en) * 2018-12-04 2019-05-28 合肥工业大学 Method for diagnosing faults based on CNN-LSTM deep learning method and more attribute time series datas
CN109657604B (en) * 2018-12-17 2021-02-12 中国人民解放军战略支援部队信息工程大学 Satellite amplitude-phase signal identification and demodulation method and device based on cyclic neural network
CN109657604A (en) * 2018-12-17 2019-04-19 中国人民解放军战略支援部队信息工程大学 Satellite width phase signals identification demodulation method and device based on Recognition with Recurrent Neural Network
CN109555566A (en) * 2018-12-20 2019-04-02 西安交通大学 A kind of turbine rotor method for diagnosing faults based on LSTM
CN109738776A (en) * 2019-01-02 2019-05-10 华南理工大学 Fan converter open-circuit fault recognition methods based on LSTM
CN109948833A (en) * 2019-02-25 2019-06-28 华中科技大学 A kind of Hydropower Unit degradation trend prediction technique based on shot and long term memory network
CN109781399A (en) * 2019-02-27 2019-05-21 天津大学 A kind of air cooling refrigeration unit sensor fault diagnosis method neural network based
CN110232395B (en) * 2019-03-01 2023-01-03 国网河南省电力公司电力科学研究院 Power system fault diagnosis method based on fault Chinese text
CN110232395A (en) * 2019-03-01 2019-09-13 国网河南省电力公司电力科学研究院 A kind of fault diagnosis method of electric power system based on failure Chinese text
CN109931678B (en) * 2019-03-13 2020-09-25 中国计量大学 Air conditioner fault diagnosis method based on deep learning LSTM
CN109931678A (en) * 2019-03-13 2019-06-25 中国计量大学 Air-conditioning fault diagnosis method based on deep learning LSTM
CN110046663A (en) * 2019-04-16 2019-07-23 重庆大学 A kind of complex electromechanical systems fault critical state discrimination method
CN110069886A (en) * 2019-05-05 2019-07-30 四川大学 The identification of cable initial failure and classification method based on VMD and CNN
CN110297479A (en) * 2019-05-13 2019-10-01 国网浙江省电力有限公司紧水滩水力发电厂 A kind of Fault Diagnosis Method of Hydro-generating Unit based on the fusion of convolutional neural networks information
CN110263846A (en) * 2019-06-18 2019-09-20 华北电力大学 The method for diagnosing faults for being excavated and being learnt based on fault data depth
CN110261773B (en) * 2019-07-01 2021-04-13 南京航空航天大学 Aviation generator fault symptom extraction method and system
CN110261773A (en) * 2019-07-01 2019-09-20 南京航空航天大学 A kind of aerogenerator failure symptom extracting method and system
CN110153802B (en) * 2019-07-04 2020-05-01 西南交通大学 Tool wear state identification method based on convolution neural network and long-term and short-term memory neural network combined model
CN110441065B (en) * 2019-07-04 2022-02-08 杭州华电江东热电有限公司 Gas turbine on-line detection method and device based on LSTM
CN110153802A (en) * 2019-07-04 2019-08-23 西南交通大学 A kind of cutting-tool wear state discrimination method based on convolutional neural networks and long Memory Neural Networks conjunctive model in short-term
CN110441065A (en) * 2019-07-04 2019-11-12 杭州华电江东热电有限公司 Gas turbine online test method and device based on LSTM
CN110309603A (en) * 2019-07-05 2019-10-08 华北电力大学（保定） A kind of short-term wind speed forecasting method and system based on wind speed characteristics
CN110309603B (en) * 2019-07-05 2023-10-24 华北电力大学（保定） Short-term wind speed prediction method and system based on wind speed characteristics
CN110991471B (en) * 2019-08-01 2023-07-11 南京航空航天大学 Fault diagnosis method for high-speed train traction system
CN110991471A (en) * 2019-08-01 2020-04-10 南京航空航天大学 Fault diagnosis method for high-speed train traction system
CN110458281A (en) * 2019-08-02 2019-11-15 中科新松有限公司 The deeply study rotation speed prediction technique and system of ping-pong robot
CN110609229A (en) * 2019-09-17 2019-12-24 电子科技大学 Wind driven generator blade imbalance fault detection method based on deep learning
CN110531266A (en) * 2019-10-14 2019-12-03 国网山东省电力公司电力科学研究院 A kind of synchronous machinery excitation winding interturn short-circuit fault early warning method
CN110531266B (en) * 2019-10-14 2021-08-24 国网山东省电力公司电力科学研究院 Synchronous motor excitation winding turn-to-turn short circuit fault early warning method
CN110736968A (en) * 2019-10-16 2020-01-31 清华大学 Radar abnormal state diagnosis method based on deep learning
CN110909782A (en) * 2019-11-15 2020-03-24 湘潭大学 Method for diagnosing machine tool spindle fault based on multi-feature combined deep learning
CN111175852A (en) * 2019-12-27 2020-05-19 中国电子科技集团公司第十四研究所 Airport fog forecast early warning method based on long-time memory algorithm
CN111222467B (en) * 2020-01-08 2022-03-25 燕山大学 Wind driven generator component fault classification detection method
CN111222467A (en) * 2020-01-08 2020-06-02 燕山大学 Wind driven generator component fault classification detection method
CN111273623A (en) * 2020-02-25 2020-06-12 电子科技大学 Fault diagnosis method based on Stacked LSTM
CN111273623B (en) * 2020-02-25 2022-10-18 电子科技大学 Fault diagnosis method based on Stacked LSTM
CN111428788A (en) * 2020-03-24 2020-07-17 西安交通大学 Deep learning-based multi-fault diagnosis method and system for steam turbine generator set rotor
CN111458640A (en) * 2020-04-22 2020-07-28 西安交通大学 Three-phase current-based alternating current asynchronous motor rotor broken bar fault diagnosis method
CN111458640B (en) * 2020-04-22 2021-05-28 西安交通大学 Three-phase current-based alternating current asynchronous motor rotor broken bar fault diagnosis method
CN112101116A (en) * 2020-08-17 2020-12-18 北京无线电计量测试研究所 Analog circuit fault diagnosis method based on deep learning
CN112146879A (en) * 2020-08-21 2020-12-29 江苏大学 Rolling bearing fault intelligent diagnosis method and system
CN112101244A (en) * 2020-09-18 2020-12-18 无锡科晟光子科技有限公司 Method for predicting optical fiber vibration signal danger level based on deep learning
CN112488171A (en) * 2020-11-25 2021-03-12 江南大学 Variable mode decomposition SRU fault prediction method
CN115221769A (en) * 2021-04-15 2022-10-21 广州中国科学院先进技术研究所 Fault prediction method, system, electronic equipment and storage medium
CN115221769B (en) * 2021-04-15 2023-10-10 广州中国科学院先进技术研究所 Fault prediction method, system, electronic equipment and storage medium
CN113392771A (en) * 2021-06-16 2021-09-14 天津职业技术师范大学(中国职业培训指导教师进修中心) Plant growth state diagnosis method, system and device based on deep learning
CN113485102A (en) * 2021-06-22 2021-10-08 西安法士特汽车传动有限公司 Method for identifying vehicle running condition based on long-term and short-term memory neural network
CN113609932A (en) * 2021-07-21 2021-11-05 广州大学 Fault detection method, system, device and medium based on long-term and short-term memory network
CN113609932B (en) * 2021-07-21 2022-09-16 广州大学 Fault detection method, system, device and medium based on long-term and short-term memory network
CN113822329A (en) * 2021-08-10 2021-12-21 国网新源控股有限公司 Hydro-power generating unit main shaft throw signal processing method and device
CN113822329B (en) * 2021-08-10 2023-11-03 国网新源控股有限公司 Method and device for processing main shaft swing degree signal of hydroelectric generating set
CN113777485B (en) * 2021-09-07 2022-05-17 江南大学 Permanent magnet synchronous motor open-circuit fault diagnosis method based on value function error
CN113777485A (en) * 2021-09-07 2021-12-10 江南大学 Permanent magnet synchronous motor open-circuit fault diagnosis method based on value function error
CN116340830A (en) * 2023-05-19 2023-06-27 山东通维信息工程有限公司 Highway electromechanical system fault classification method based on deep memory model
CN116340830B (en) * 2023-05-19 2023-08-18 山东通维信息工程有限公司 Highway electromechanical system fault classification method based on deep memory model
CN117557304A (en) * 2024-01-11 2024-02-13 国网浙江省电力有限公司 Electric quantity and electricity price level fusion prediction method based on modal decomposition and neural network
CN117557304B (en) * 2024-01-11 2024-03-29 国网浙江省电力有限公司 Electric quantity and electricity price level fusion prediction method based on modal decomposition and neural network

Also Published As

Publication number Publication date
CN108197648B (en) 2020-06-05

Similar Documents

Publication Publication Date Title
CN108197648A (en) A kind of Fault Diagnosis Method of Hydro-generating Unit and system based on LSTM deep learning models
CN111830408B (en) Motor fault diagnosis system and method based on edge calculation and deep learning
CN109141847B (en) Aircraft system fault diagnosis method based on MSCNN deep learning
CN109146162B (en) A kind of probability wind speed forecasting method based on integrated Recognition with Recurrent Neural Network
CN109948833A (en) A kind of Hydropower Unit degradation trend prediction technique based on shot and long term memory network
CN109597401A (en) A kind of equipment fault diagnosis method based on data-driven
CN110007652A (en) A kind of Hydropower Unit degradation trend interval prediction method and system
CN108197743A (en) A kind of prediction model flexible measurement method based on deep learning
Di et al. Ensemble deep transfer learning driven by multisensor signals for the fault diagnosis of bevel-gear cross-operation conditions
CN114676742A (en) Power grid abnormal electricity utilization detection method based on attention mechanism and residual error network
CN111459144A (en) Airplane flight control system fault prediction method based on deep cycle neural network
CN110163075A (en) A kind of multi-information fusion method for diagnosing faults based on Weight Training
CN112784920B (en) Yun Bianduan coordinated rotating component reactance domain self-adaptive fault diagnosis method
CN107292446A (en) A kind of mixing wind speed forecasting method based on consideration component relevance wavelet decomposition
CN108875108A (en) A kind of rail vehicle circuit failure diagnosis method based on PCA and BRB
Yan et al. Deep learning technology for chiller faults diagnosis
CN108805206A (en) A kind of modified LSSVM method for building up for analog circuit fault classification
Tang et al. Prediction of bearing performance degradation with bottleneck feature based on LSTM network
CN111141879A (en) Deep learning air quality monitoring method, device and equipment
CN117056678B (en) Machine pump equipment operation fault diagnosis method and device based on small sample
CN114386452A (en) Method for detecting faults of sun wheel of nuclear power circulating water pump
CN106682312A (en) Industrial process soft-measurement modeling method of local weighing extreme learning machine model
CN112365009A (en) Secondary equipment abnormity diagnosis method based on deep learning network
Hu et al. Research on the fault identification method of oil pumping unit based on residual network
CN114841063A (en) Aero-engine residual life prediction method based on deep learning

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