CN113743297A - Storage tank dome displacement data restoration method and device based on deep learning - Google Patents
Storage tank dome displacement data restoration method and device based on deep learning Download PDFInfo
- Publication number
- CN113743297A CN113743297A CN202111030338.8A CN202111030338A CN113743297A CN 113743297 A CN113743297 A CN 113743297A CN 202111030338 A CN202111030338 A CN 202111030338A CN 113743297 A CN113743297 A CN 113743297A
- Authority
- CN
- China
- Prior art keywords
- displacement data
- data
- measuring point
- 1dcnn
- storage tank
- 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
- 238000006073 displacement reaction Methods 0.000 title claims abstract description 163
- 238000000034 method Methods 0.000 title claims abstract description 35
- 238000013135 deep learning Methods 0.000 title claims abstract description 21
- 230000006870 function Effects 0.000 claims abstract description 59
- 238000000354 decomposition reaction Methods 0.000 claims abstract description 37
- 238000012545 processing Methods 0.000 claims abstract description 9
- 230000008439 repair process Effects 0.000 claims abstract description 8
- 238000012549 training Methods 0.000 claims abstract description 6
- 238000004364 calculation method Methods 0.000 claims description 14
- 238000012544 monitoring process Methods 0.000 claims description 13
- 210000002569 neuron Anatomy 0.000 claims description 12
- 238000004458 analytical method Methods 0.000 claims description 11
- 239000011159 matrix material Substances 0.000 claims description 10
- 230000008569 process Effects 0.000 claims description 10
- 238000013507 mapping Methods 0.000 claims description 9
- 238000001514 detection method Methods 0.000 claims description 6
- 238000012935 Averaging Methods 0.000 claims description 4
- 230000015654 memory Effects 0.000 claims description 4
- 230000001537 neural effect Effects 0.000 claims description 4
- QBPFLULOKWLNNW-UHFFFAOYSA-N chrysazin Chemical compound O=C1C2=CC=CC(O)=C2C(=O)C2=C1C=CC=C2O QBPFLULOKWLNNW-UHFFFAOYSA-N 0.000 claims description 3
- 238000001228 spectrum Methods 0.000 claims description 3
- 239000013598 vector Substances 0.000 claims description 3
- 238000013480 data collection Methods 0.000 claims description 2
- 238000005259 measurement Methods 0.000 claims 4
- 230000007774 longterm Effects 0.000 abstract description 4
- 239000000306 component Substances 0.000 description 43
- 238000010586 diagram Methods 0.000 description 6
- 238000013527 convolutional neural network Methods 0.000 description 5
- 238000012216 screening Methods 0.000 description 4
- 230000004913 activation Effects 0.000 description 3
- 230000000694 effects Effects 0.000 description 3
- 238000012986 modification Methods 0.000 description 2
- 230000004048 modification Effects 0.000 description 2
- 230000004044 response Effects 0.000 description 2
- 230000002159 abnormal effect Effects 0.000 description 1
- 230000001133 acceleration Effects 0.000 description 1
- 238000013473 artificial intelligence Methods 0.000 description 1
- 238000013528 artificial neural network Methods 0.000 description 1
- 238000002474 experimental method Methods 0.000 description 1
- 230000007787 long-term memory Effects 0.000 description 1
- 238000010801 machine learning Methods 0.000 description 1
- 239000012533 medium component Substances 0.000 description 1
- 238000003062 neural network model Methods 0.000 description 1
- 238000011176 pooling Methods 0.000 description 1
- 230000001737 promoting effect Effects 0.000 description 1
- 238000005067 remediation Methods 0.000 description 1
- 238000000926 separation method Methods 0.000 description 1
- 230000006403 short-term memory Effects 0.000 description 1
- 238000007873 sieving Methods 0.000 description 1
- 238000006467 substitution reaction Methods 0.000 description 1
- 238000012800 visualization Methods 0.000 description 1
Images
Classifications
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F2218/00—Aspects of pattern recognition specially adapted for signal processing
- G06F2218/02—Preprocessing
- G06F2218/04—Denoising
-
- G—PHYSICS
- G01—MEASURING; TESTING
- G01B—MEASURING LENGTH, THICKNESS OR SIMILAR LINEAR DIMENSIONS; MEASURING ANGLES; MEASURING AREAS; MEASURING IRREGULARITIES OF SURFACES OR CONTOURS
- G01B21/00—Measuring arrangements or details thereof, where the measuring technique is not covered by the other groups of this subclass, unspecified or not relevant
- G01B21/02—Measuring arrangements or details thereof, where the measuring technique is not covered by the other groups of this subclass, unspecified or not relevant for measuring length, width, or thickness
-
- 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
- G06N3/044—Recurrent networks, e.g. Hopfield networks
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N3/00—Computing arrangements based on biological models
- G06N3/02—Neural networks
- G06N3/04—Architecture, e.g. interconnection topology
- G06N3/045—Combinations of networks
-
- 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
Landscapes
- Engineering & Computer Science (AREA)
- Physics & Mathematics (AREA)
- Theoretical Computer Science (AREA)
- General Physics & Mathematics (AREA)
- Artificial Intelligence (AREA)
- General Engineering & Computer Science (AREA)
- General Health & Medical Sciences (AREA)
- Biomedical Technology (AREA)
- Data Mining & Analysis (AREA)
- Evolutionary Computation (AREA)
- Biophysics (AREA)
- Molecular Biology (AREA)
- Computing Systems (AREA)
- Computational Linguistics (AREA)
- Life Sciences & Earth Sciences (AREA)
- Mathematical Physics (AREA)
- Software Systems (AREA)
- Health & Medical Sciences (AREA)
- Computer Vision & Pattern Recognition (AREA)
- Signal Processing (AREA)
- Testing Or Calibration Of Command Recording Devices (AREA)
Abstract
The invention relates to a storage tank dome displacement data restoration method based on deep learning, belonging to the technical field of displacement data restoration, and the restoration method comprises the following steps: obtaining historical displacement data of the measuring points to be repaired and the surrounding associated measuring points, decomposing the historical displacement data into a plurality of intrinsic mode function components through a set empirical mode decomposition algorithm, training the 1DCNN-LSTM model to obtain an EEMD-1DCNN-LSTM model, predicting missing displacement data through the EEMD-1DCNN-LSTM model, and completing data repair. According to the method, the EEMD, the 1DCNN and the LSTM are combined into a new model, the method is very suitable for processing complex long-term time sequence dynamic information with spatial correlation, can greatly improve the prediction precision, and is very suitable for repairing CNG storage tank dome missing displacement data.
Description
Technical Field
The invention belongs to the technical field of storage tank dome displacement data restoration, and relates to a storage tank dome displacement data restoration method and device based on deep learning.
Background
The adoption of the displacement sensor has great significance for evaluating the dynamic response of the LNG storage tank structure. However, in the shaking table experiment, some displacement sensors may fail or be abnormal, so that data is lost, and the data is difficult to recover.
The existing method for predicting the structural displacement of the LNG storage tank based on an artificial intelligence method is mainly divided into two methods. One method is a 'shallow layer' machine learning method, acceleration sensing data has high nonlinearity and non-Gaussian, and a 'shallow layer' model has certain limitation on long-term prediction of displacement response, cannot process massive monitoring data and has low accuracy. The other method is a traditional deep neural network model and has the characteristics of universality, high efficiency and the like, but the accuracy needs to be further improved. Therefore, the existing prediction methods cannot be used to repair the displacement data of the sensor.
Disclosure of Invention
In view of the above, the present invention provides a method and an apparatus for repairing displacement data of a storage tank dome based on deep learning of an EEMD-1DCNN-LSTM model.
In order to achieve the purpose, the invention provides the following technical scheme:
a storage tank dome displacement data restoration method based on deep learning is used for restoration when LNG storage tank dome displacement data are missing, and comprises the following steps:
step S1, taking a measuring point with LNG storage tank dome displacement data missing as a measuring point to be repaired, selecting a plurality of measuring points around the measuring point to be repaired as associated measuring points, acquiring historical displacement data of the measuring point to be repaired in a certain period before the data missing period and historical displacement data of the associated measuring points in a corresponding period, and decomposing the historical displacement data of the measuring points into a plurality of intrinsic mode function components respectively through a set empirical mode decomposition algorithm;
step S2, taking the intrinsic mode function component obtained through decomposition as an input feature of a 1DCNN-LSTM model, extracting a spatial correlation feature of the displacement of the associated measuring point and the displacement of the measuring point to be repaired through the 1DCNN model, and then sending the extracted spatial correlation feature to the LSTM model to obtain a dependency characteristic in time;
step S3, defining a loss function of the 1DCNN-LSTM model, and ending the training when the value of the loss function of the 1DCNN-LSTM model converges to a fixed value and keeps unchanged to obtain an EEMD-1DCNN-LSTM model;
step S4, obtaining displacement data of the measuring point to be repaired before the data missing period and displacement data of the associated measuring point before the data missing period and the data missing period, inputting the displacement data into an EEMD-1DCNN-LSTM model to predict the displacement data of the measuring point to be repaired at the data missing period; and taking the predicted displacement data as displacement data of the measuring point to be repaired in the data missing period to finish the repair of the missing data.
Further, the ensemble empirical mode decomposition algorithm is implemented by the following steps:
step S11, selecting the processing times m of the original signal;
s12, selecting m random white noises with different amplitudes, and combining the original signal with each white noise respectively to obtain m new signals;
step S13, performing empirical mode decomposition on the m new signals respectively to obtain a series of intrinsic mode function components;
and step S14, respectively averaging the intrinsic mode function components of the corresponding modes to obtain a set empirical mode decomposition result.
Further, the scales of m white noises are uniformly distributed, and the energy of the white noises is also uniformly distributed on the frequency spectrum.
Further, the time sequence of the displacement data of one measuring point forms one-dimensional data; decomposing the one-dimensional data into a plurality of IMF sequences through EEMD to form two-dimensional data; and after data of a plurality of associated measuring points are obtained, three-dimensional data are formed and are used as an input feature mapping group of the 1DCNN model.
Further, the architecture of the single neural unit of the LSTM includes an input gate, a forgetting gate, an output gate, and a memory unit, and is used for implementing input and output of information, and the operation process is as follows:
Γi＝σ(Wi,xxt+Wi,hht-1+bi)
Γf＝σ(Wf,xxt+Wf,hht-1+bf)
Γo＝σ(Wo,xxt+Wo,hht-1+bo)
ht＝Γo*tanh(Ct)
wherein, Wi,x、Wi,h、Wf,x、Wf,h、Wo,x、Wo,h、Wc,x、Wc,hRepresenting a weight matrix; bi、bf、bc、boRepresenting a bias matrix; x is the number oftAn input feature representing time t; c. Ct-1Representing neurons before updating; c. CtRepresenting the updated neuron; h ist-1An output characteristic representing the time (t-1); h istAn output characteristic representing time t; gamma-shapediRepresenting an input gate; gamma-shapedfIndicating a forgetting gate; gamma-shapedoAn output gate is shown;
is a candidate neuron; sigma is a Sigmoid function; tan h is the hyperbolic tangent function.
Further, the loss function l (x, y) of the 1DCNN-LSTM model is defined as:
where N represents the number of samples, xiDenotes the actual value of the i-th sample, yiRepresenting the predicted value of the ith sample.
A deep learning-based storage tank dome displacement data restoration device, comprising:
the displacement data acquisition module is used for acquiring displacement data of each measuring point of the dome of the LNG storage tank in real time and transmitting the displacement data to the calculation analysis module;
the calculation analysis module is used for monitoring whether displacement data of each measuring point of a dome of the LNG storage tank are missing or not in real time, when the displacement data of one measuring point is monitored to be missing, the measuring point with the missing displacement data is used as a measuring point to be repaired, a plurality of measuring points are selected around the measuring point to be repaired to be used as associated measuring points, the displacement data of the measuring point to be repaired and the displacement data of the associated measuring points are input into an EEMD-1DCNN-LSTM model, the missing displacement data are predicted, the missing displacement data are supplemented by the predicted displacement data, and data repair is completed; and
and the output module is used for outputting the displacement data of each measuring point acquired by the displacement data acquisition module and the repaired displacement data of the measuring point to be repaired.
Further, the calculation analysis module comprises a data reading unit, a monitoring unit, an EEMD-1DCNN-LSTM model and a storage unit;
the data reading unit is used for reading displacement data of each measuring point of the LNG storage tank dome, which is acquired by the displacement data acquisition module;
the monitoring unit is used for monitoring whether displacement data of each measuring point of the dome of the LNG storage tank is lost or not in real time;
the EEMD-1DCNN-LSTM model comprises a set empirical mode decomposition unit and a 1DCNN-LSTM model, wherein the set empirical mode decomposition unit is used for decomposing displacement data of the measuring points to be repaired and each associated measuring point into a plurality of intrinsic mode function components through a set empirical mode decomposition algorithm, and vectors formed by the intrinsic mode function components are used as input characteristics of the 1DCNN-LSTM model; the 1DCNN-LSTM model is used for predicting displacement data of the missing of the measuring point to be repaired according to input characteristics;
the storage unit is used for storing displacement data of each measuring point of the LNG storage tank dome acquired by the displacement data acquisition module and displacement data predicted by the EEMD-1DCNN-LSTM model.
The device further comprises a detection unit and an alarm unit, wherein the detection unit is used for designating a measuring point with normal displacement data as a measuring point to be repaired, comparing the displacement data of the measuring point predicted by the EEMD-1DCNN-LSTM model with the real displacement data of the measuring point collected by the displacement data collection module, and when the difference value between the displacement data and the measuring point exceeds a preset threshold value, enabling the alarm unit to output an alarm signal to prompt that the predicted data has overlarge offset.
In the invention, EEMD, 1DCNN and LSTM are combined into a new model, wherein EEMD can decompose complex nonlinear displacement data into linear combinations of eigenmode functions with limited frequencies from high to low, and each decomposed IMF component comprises local characteristic signals of different time scales of an original signal; the 1DCNN model has the characteristics of local connection, weight sharing and the like, and can reserve and extract the spatial correlation characteristics among IMFs; the LSTM model can fully excavate the nonlinear relation between variables and adaptively sense the characteristic information of the upper time sequence and the lower time sequence; therefore, the combined new model is very suitable for processing the complex long-term time sequence dynamic information with space correlation, can greatly improve the prediction precision, and is very suitable for repairing the missing displacement data of the dome of the CNG storage tank. In addition, the EEMD algorithm and the 1DCNN-LSTM model have low requirements on hardware and low implementation cost.
Drawings
For the purposes of promoting a better understanding of the objects, aspects and advantages of the invention, reference will now be made to the following detailed description taken in conjunction with the accompanying drawings in which:
FIG. 1 is a flow chart of a preferred embodiment of the method for deep learning based tank dome displacement data remediation of the present invention;
FIG. 2 is a schematic diagram of obtaining data of a measuring point to be repaired and a related measuring point for prediction;
FIG. 3 is a schematic diagram of a decomposition of data by the EEMD algorithm;
FIG. 4 is a schematic structural diagram of the 1DCNN-LSTM model;
FIG. 5 is a schematic diagram of a one-dimensional convolutional neural network computation process;
FIG. 6 is a schematic diagram of the architecture of a single neuron architecture of the LSTM;
FIG. 7 is a block diagram of the storage tank dome displacement data restoration device based on deep learning according to a preferred embodiment of the present invention.
Detailed Description
The embodiments of the present invention are described below with reference to specific embodiments, and other advantages and effects of the present invention will be easily understood by those skilled in the art from the disclosure of the present specification. The invention is capable of other and different embodiments and of being practiced or of being carried out in various ways, and its several details are capable of modification in various respects, all without departing from the spirit and scope of the present invention. It should be noted that the drawings provided in the following embodiments are only for illustrating the basic idea of the present invention in a schematic way, and the features in the following embodiments and examples may be combined with each other without conflict.
As shown in FIG. 1, a preferred embodiment of the deep learning-based storage tank dome displacement data restoration method of the present invention comprises the following steps:
and S1, taking the measuring points with the missing LNG storage tank dome displacement data as measuring points to be repaired, selecting a plurality of measuring points around the measuring points to be repaired as associated measuring points, and acquiring historical displacement data of the measuring points to be repaired in a certain period before the data missing period and historical displacement data of the associated measuring points in a corresponding period. As shown in fig. 2, assuming that the displacement data from time step T to time step (T +1) is missing from the point to be repaired, the displacement data of the point to be repaired K time steps before time step T and the displacement data of the associated point from time step (T-K) to time step (T +1) are obtained.
Then, the historical displacement data of the above-mentioned measuring points are respectively decomposed into a plurality of IMF (Intrinsic Mode Functions) components by an EEMD (ensemble empirical Mode Decomposition) algorithm. The IMF component satisfies the following two conditions:
(1) in the whole time range of the function, the number of the extreme points is equal to or different from the number of the zero crossing points by 1;
(2) at any time point, the mean of the upper envelope and the lower envelope is 0.
As shown in fig. 3, the ensemble empirical mode decomposition algorithm is implemented by the following steps:
step S11, selecting the processing times m of the original signal;
s12, selecting m random white noises with different amplitudes, and combining the original signal with each white noise to obtain m new signals n1(t)，n2(t)，……，nj(t)，……，nm(t); the scales of m white noises are uniformly distributed, and the energy of the white noises is also uniformly distributed on a frequency spectrum.
Step S13, performing EMD (Empirical Mode Decomposition, ensemble Empirical Mode Decomposition) on each of the m new signals to obtain a series of IMF components. In the following, n is defined as1(t) performing EMD as an example to explain the specific process of the EMD algorithm:
step S131, adding n1(t) signal to be decomposed x (t) as EMD.
And S132, screening the signal x (t) to be decomposed. The screening process is to subtract the average envelope function of the signal to be decomposed to obtain a new function; the method specifically comprises the following steps: finding out all maximum value points of the signal x (t) to be decomposed, and fitting an upper envelope line of the signal x (t) to be decomposed by using a cubic spline function; finding out all minimum value points of the signal x (t) to be decomposed, and fitting a cubic spline function into a lower envelope line of the signal x (t) to be decomposed; calculating the mean value of the upper envelope and the lower envelope to obtain a first mean envelope function m1(t); subtracting the first average envelope function m from the signal x (t) to be decomposed1(t) obtaining a first intermediate component function d1,1(t)。
Step S133, determining the intermediate component function d1,1(t) whether or not IMF component is satisfiedIf satisfied, d1,1(t) is denoted as the first IMF component IMF1(t) of the signal to be decomposed; if not, continue to step S12 for d1,1(t) performing screening until the intermediate component function satisfies the condition of the IMF component. Assuming a medium component function d obtained after K screening1,k(t) if the IMF component is satisfied, d is1,k(t) is noted as the first IMF component of the signal to be decomposed, for n1(t) the first IMF component of which is denoted as a1,1。
Step S134, subtracting the first IMF component from the signal x (t) to be decomposed to obtain a first residual component function r1(t); the first residual component function r1(t) continuing the decomposition (i.e. the separation of the IMF component from the signal by iterative sieving) according to steps S12 and S13 to obtain a second IMF component, for n1(t) the second IMF component of which is denoted as a2,1(ii) a Using a first residual component function r1(t) subtracting the second IMF component to obtain a second residual component function r2(t) of (d). Continuing to apply the second residual component function r according to steps S12 and S132(t) carrying out decomposition; suppose that after N decompositions, the Nth residual component function r is obtainedn(t) is a monotonic function, the decomposition is stopped and the residual component function r is usedn(t) as a residual amount RES. As shown in FIG. 2, at this time, the signal n is applied1(t) is decomposed into n IMF components (i.e., a)1,1、a2,1、……、ai,1、……、aN,1) And a residual amount RES.
According to the above method, n is2(t) decomposition by EMD into i.e. a1,2、a2,2、……、ai,2、……、aN,2；
……；
N is to bej(t) decomposition by EMD into i.e. a1,j、a2,j、……、ai,j、……、aN,j；
……；
N is to bem(t) decomposition by EMD into i.e. a1,m、a2,m、……、ai,m、……、aN,m。
Step S14, respectively averaging IMF components of corresponding modes to obtain a1、a2、……、ai、……、aNAnd obtaining a final IMF component, namely a set empirical mode decomposition result. The formula for averaging is:
compared with Fourier transform and wavelet decomposition, the EMD does not need to set a basis function and has self-adaptability, so that the EMD has a wider application range. After decomposing the signal x (t) to be decomposed, the first IMF component comprises the component with the smallest time scale (the highest frequency) in the signal x (t) to be decomposed, the corresponding frequency component is gradually reduced along with the increase of the order of the IMF component, and r isn(t) (i.e., the residual amount RES in the present embodiment) has the lowest frequency component. According to the convergence condition of EMD decomposition, the residue r obtained by decompositionnWhen (t) is a monotonic function, the time period will be longer than the recording length of the signal, so that the residue r can be reducedn(t) as a trend term for the signal to be decomposed x (t).
Step S2, decomposing the obtained eigenmode function component (namely a)1、a2、……、ai、……、aN) As the input characteristics of the 1DCNN-LSTM model, extracting the spatial correlation characteristics of the displacement of the associated measuring point and the displacement of the measuring point to be repaired through the 1DCNN (one-dimensional convolutional neural network) model, and then sending the extracted spatial correlation characteristics to an LSTM (long-short term memory; long and short term memory network) model, and obtaining the dependency characteristics in time. As shown in FIG. 4, the 1DCNN-LSTM model is formed by splicing the 1DCNN model and the LSTM model.
CNN is widely used in the field of image processing, and generally, an image is three-dimensional data, that is, X · Y · Z. The time series of the displacement data of one sensor (i.e. measuring point) is 1D, and is decomposed into a plurality of IMF sequences through EEMD, so that the 2D data is changed, and a plurality of sensors are processed at the same time, so that one dimension is added, and finally the data can be regarded as 3D data processing, and the data is processed in a mode similar to image data from the surface. The 1DCNN includes convolutional and pooling layers, and its operating principle is as follows.
The convolutional layer functions to extract the features of a local region, and different convolutional kernels correspond to different feature extractors. The neurons of the convolutional layer are one-dimensional structures as are the fully connected networks. Since the convolutional network is mainly applied to image processing, and an image has a two-dimensional structure, in order to more fully utilize local information of the image, neurons are generally organized into a three-dimensional neural layer having a size of M × width N × depth D, which is considered to be formed by D feature maps of two-dimensional structures having a size of M × N.
Feature maps (Feature maps) are features extracted by convolution, each of which may be a type of extracted Feature. To improve the representation capability of the convolutional network, multiple different feature maps may be used at each layer to better represent the features.
Without loss of generality, assume the structure of a convolutional layer as follows:
(1) inputting a feature mapping group:
is a three-dimensional Tensor (Tensor); wherein, M represents the number of time steps included in the displacement data of one sensor, N represents the number of IMF components obtained by one-time EEMD decomposition of the displacement data of the sensor, and D represents the number of the sensors. Each Slice (Slice) matrix
D is more than or equal to 1 and less than or equal to D.
(2) Outputting a characteristic mapping group:
is a three-dimensional tensor in which each slice matrix
Is an output characteristic mapping, and P is more than or equal to 1 and less than or equal to P.
(3) And (3) convolution kernel:
is a four-dimensional tensor; where U represents the number of rows of the convolution kernel and V represents the number of columns of the convolution kernel, for example: UxV may take a value of 3 x 5. Each slice matrix
D is more than or equal to 1 and less than or equal to D; p is more than or equal to 1 and less than or equal to P.
To calculate an output feature map YpUsing a convolution kernel Wp,1、Wp,2、……、Wp,DSeparately mapping X to input features1、X2、……、XDConvolution is carried out, then the convolution results are added, and a scalar offset b is added to obtain the net input Z of the convolution layerpHere, the Net input means a Net activity value (Net activity) that has not undergone a nonlinear Activation function.
Obtaining output characteristic mapping Y after nonlinear activation functionp。
Yp＝f(Zp)
Wherein
A three-dimensional convolution kernel; bpRepresenting a bias matrix; f () is a nonlinear activation function, typically a ReLU function. The calculation process is shown in fig. 5, where the dashed boxes represent convolution kernels. If P feature maps are desired to be output by the convolutional layer, the above calculation process may be repeated P times to obtain P output feature maps Y1、Y1、……、YP. At the input of
Output is as
Each output signature map requires D filters and one offset in the convolutional layer. Assuming that the size of each filter is U × V, then a total of P × D × (U × V) + P parameters are required.
From the calculation process, the 1DCNN model has the characteristics of local connection, weight sharing and the like, and can reserve and extract the spatial correlation characteristics among IMFs.
Output feature mapping Z of CNNpIs sent to the corresponding LSTM. As shown in fig. 6, the architecture of the LSTM single neural unit includes an input gate, a forgetting gate, an output gate, and a memory unit, and is used to implement input and output of information, and the operation process is as follows:
Γf＝σ(Wf,xxt+Wf,hht-1+bf)
Γo＝σ(Wo,xxt+Wo,hht-1+bo)
ht＝Γo*tanh(Ct)
wherein, Wi,x、Wi,h、Wf,x、Wf,h、Wo,x、Wo,h、Wc,x、Wc,hRepresenting a weight matrix; bi、bf、bc、boRepresenting a bias matrix; x is the number oftRepresenting input features at time t, i.e. CNN pairsOutput feature mapping Y according to time of dayp；ct-1Representing neurons before updating; c. CtRepresenting the updated neuron; h ist-1An output characteristic representing the time (t-1) (i.e., the last time); h istAn output characteristic representing time t (i.e., the current time); gamma-shapediRepresenting an input gate; gamma-shapedfIndicating a forgetting gate; gamma-shapedoAn output gate is shown;
is a candidate neuron; sigma is a Sigmoid function; tan h is the hyperbolic tangent function.
The training algorithm of the LSTM neural network is specifically as follows: firstly, calculating the output value of each LSTM unit in the forward direction; then, reversely calculating an error term of each LSTM unit, and calculating the gradient of each weight by using the corresponding error term; finally, the weights are updated by a gradient descent algorithm. The LSTM model can fully mine the nonlinear relation among variables, adaptively sense the characteristic information of upper and lower time sequences, and is very suitable for processing complex long-term time sequence dynamic information.
And step S3, defining a loss function of the 1DCNN-LSTM model, and ending the training when the value of the loss function of the 1DCNN-LSTM model converges to a fixed value and keeps unchanged to obtain the EEMD-1DCNN-LSTM model.
Wherein the loss function l (x, y) of the 1DCNN-LSTM model can be defined as:
where N represents the number of samples, xiRepresenting the actual value (i.e. true value), y, of the ith sampleiRepresenting the predicted value of the ith sample.
And when the value of the loss function converges to a fixed value and keeps unchanged, considering the parameters of the 1DCNN-LSTM model at the moment as the optimal model parameters, and stopping the model training.
Step S4, obtaining displacement data of the measuring point to be repaired before the data missing period and displacement data of the associated measuring point before the data missing period and the data missing period, inputting the displacement data into an EEMD-1DCNN-LSTM model to predict the displacement data of the measuring point to be repaired at the data missing period; and taking the predicted displacement data as displacement data of the measuring point to be repaired in the data missing period to finish the repair of the missing data.
The invention also provides a storage tank dome displacement data restoration device based on deep learning, as shown in fig. 7, a preferred embodiment of the storage tank dome displacement data restoration device based on deep learning of the invention comprises a displacement data acquisition module, a calculation analysis module and an output module.
The displacement data acquisition module is used for acquiring displacement data of each measuring point of the dome of the LNG storage tank in real time and transmitting the displacement data to the calculation analysis module;
the calculation analysis module is used for monitoring whether displacement data of each measuring point of a dome of the LNG storage tank are missing or not in real time, when the situation that the displacement data of one measuring point is missing is monitored, the measuring point with the missing displacement data is used as a measuring point to be repaired, a plurality of measuring points are selected around the measuring point to be repaired to be used as associated measuring points, the displacement data of the measuring point to be repaired and the displacement data of the associated measuring points are input into an EEMD-1DCNN-LSTM model, the missing displacement data are predicted, the missing displacement data are supplemented through the predicted displacement data, and data repair is completed.
The calculation analysis module comprises a data reading unit, a monitoring unit, an EEMD-1DCNN-LSTM model and a storage unit.
The data reading unit is used for reading displacement data of each measuring point of the LNG storage tank dome, which is acquired by the displacement data acquisition module; preferably, a module comprising a GPS data acquisition unit and/or a Beidou positioning data acquisition unit is adopted.
The monitoring unit is used for monitoring whether displacement data of each measuring point of the dome of the LNG storage tank is lost or not in real time;
the EEMD-1DCNN-LSTM model comprises a set empirical mode decomposition unit and a 1DCNN-LSTM model, wherein the set empirical mode decomposition unit is used for decomposing displacement data of the measuring points to be repaired and each associated measuring point into a plurality of intrinsic mode function components through a set empirical mode decomposition algorithm, and vectors formed by the intrinsic mode function components are used as input characteristics of the 1DCNN-LSTM model; the 1DCNN-LSTM model is used for predicting displacement data of the missing measuring point to be repaired according to input characteristics.
The storage unit is used for storing displacement data of each measuring point of the LNG storage tank dome acquired by the displacement data acquisition module and displacement data predicted by the EEMD-1DCNN-LSTM model.
The output module is used for outputting the displacement data of each measuring point acquired by the displacement data acquisition module and the repaired displacement data of the measuring point to be repaired. The predictive data output module preferably employs a visualization module, such as a display, to visually output the historical data and the predictive data.
In order to detect whether the prediction deviation of the model is overlarge, the model further comprises a detection unit and an alarm unit, wherein the detection unit is used for appointing a measuring point with normal displacement data as a measuring point to be repaired, the displacement data of the measuring point is predicted through the EEMD-1DCNN-LSTM model, the predicted displacement data is compared with the real displacement data of the measuring point, which is acquired by the displacement data acquisition module, and when the difference value between the predicted displacement data and the real displacement data exceeds a preset threshold value, the alarm unit outputs an alarm signal to prompt that the predicted data is overlarge in deviation. Therefore, the operator is reminded of needing to train the model again so as to improve the accuracy of the model; thereby alerting the operator that the model may need to be retrained to improve the prediction accuracy of the model. Of course, the operator does not handle the alarm and does not affect the model operation.
In this embodiment, the analysis module adopts the EEMD algorithm and the 1DCNN-LSTM model, and has low requirements on calculation and storage capabilities, low requirements on hardware, and low implementation cost.
Finally, the above embodiments are only intended to illustrate the technical solutions of the present invention and not to limit the present invention, and although the present invention has been described in detail with reference to the preferred embodiments, it will be understood by those skilled in the art that modifications or equivalent substitutions may be made on the technical solutions of the present invention without departing from the spirit and scope of the technical solutions, and all of them should be covered by the claims of the present invention.
Claims (9)
1. A storage tank dome displacement data restoration method based on deep learning is used for restoration when LNG storage tank dome displacement data are missing, and is characterized by comprising the following steps:
step S1, taking a measuring point with LNG storage tank dome displacement data missing as a measuring point to be repaired, selecting a plurality of measuring points around the measuring point to be repaired as associated measuring points, acquiring historical displacement data of the measuring point to be repaired in a certain period before the data missing period and historical displacement data of the associated measuring points in a corresponding period, and decomposing the historical displacement data of the measuring points into a plurality of intrinsic mode function components respectively through a set empirical mode decomposition algorithm;
step S2, taking the intrinsic mode function component obtained through decomposition as an input feature of a 1DCNN-LSTM model, extracting a spatial correlation feature of the displacement of the associated measuring point and the displacement of the measuring point to be repaired through the 1DCNN model, and then sending the extracted spatial correlation feature to the LSTM model to obtain a dependency characteristic in time;
step S3, defining a loss function of the 1DCNN-LSTM model, and ending the training when the value of the loss function of the 1DCNN-LSTM model converges to a fixed value and keeps unchanged to obtain an EEMD-1DCNN-LSTM model;
step S4, obtaining displacement data of the measuring point to be repaired before the data missing period and displacement data of the associated measuring point before the data missing period and the data missing period, inputting the displacement data into an EEMD-1DCNN-LSTM model to predict the displacement data of the measuring point to be repaired at the data missing period; and taking the predicted displacement data as displacement data of the measuring point to be repaired in the data missing period to finish the repair of the missing data.
2. The deep learning-based storage tank dome displacement data restoration method according to claim 1, wherein the ensemble empirical mode decomposition algorithm is implemented by the following steps:
step S11, selecting the processing times m of the original signal;
s12, selecting m random white noises with different amplitudes, and combining the original signal with each white noise respectively to obtain m new signals;
step S13, performing empirical mode decomposition on the m new signals respectively to obtain a series of intrinsic mode function components;
and step S14, respectively averaging the intrinsic mode function components of the corresponding modes to obtain a set empirical mode decomposition result.
3. The deep learning-based storage tank dome displacement data restoration method according to claim 2, wherein m scales of the white noise are uniformly distributed, and the energy of the white noise is also uniformly distributed on a frequency spectrum.
4. The deep learning-based storage tank dome displacement data restoration method according to claim 1, wherein the time series of displacement data of one measuring point forms one-dimensional data; decomposing the one-dimensional data into a plurality of IMF sequences through EEMD to form two-dimensional data; and after data of a plurality of associated measuring points are obtained, three-dimensional data are formed and are used as an input feature mapping group of the 1DCNN model.
5. The deep learning-based storage tank dome displacement data restoration method according to claim 1, wherein the LSTM single neural unit architecture comprises an input gate, a forgetting gate, an output gate and a memory unit, and is used for realizing information input and output, and the operation process is as follows:
Γi＝σ(Wi,xxt+Wi,hht-1+bi)
Γf＝σ(Wf,xxt+Wf,hht-1+bf)
Γo＝σ(Wo,xxt+Wo,hht-1+bo)
ht＝Γo*tanh(Ct)
wherein, Wi,x、Wi,h、Wf,x、Wf,h、Wo,x、Wo,h、Wc,x、Wc,hRepresenting a weight matrix; bi、bf、bc、boRepresenting a bias matrix; x is the number oftAn input feature representing time t; c. Ct-1Representing neurons before updating; c. CtRepresenting the updated neuron; h ist-1An output characteristic representing the time (t-1); h istAn output characteristic representing time t; gamma-shapediRepresenting an input gate; gamma-shapedfIndicating a forgetting gate; gamma-shapedoAn output gate is shown;
is a candidate neuron; sigma is a Sigmoid function; tan h is the hyperbolic tangent function.
6. The deep learning-based storage tank dome displacement data restoration method according to claim 1, wherein the loss function l (x, y) of the 1DCNN-LSTM model is defined as:
where N represents the number of samples, xiDenotes the actual value of the i-th sample, yiRepresenting the predicted value of the ith sample.
7. A storage tank dome displacement data prosthetic devices based on deep learning, characterized in that includes:
the displacement data acquisition module is used for acquiring displacement data of each measuring point of the dome of the LNG storage tank in real time and transmitting the displacement data to the calculation analysis module;
the calculation analysis module is used for monitoring whether displacement data of each measuring point of a dome of the LNG storage tank are missing or not in real time, when the displacement data of one measuring point is monitored to be missing, the measuring point with the missing displacement data is used as a measuring point to be repaired, a plurality of measuring points are selected around the measuring point to be repaired to be used as associated measuring points, the displacement data of the measuring point to be repaired and the displacement data of the associated measuring points are input into an EEMD-1DCNN-LSTM model, the missing displacement data are predicted, the missing displacement data are supplemented by the predicted displacement data, and data repair is completed; and
and the output module is used for outputting the displacement data of each measuring point acquired by the displacement data acquisition module and the repaired displacement data of the measuring point to be repaired.
8. The deep learning-based storage tank dome displacement data restoration device according to claim 7, wherein the calculation analysis module comprises a data reading unit, a monitoring unit, an EEMD-1DCNN-LSTM model and a storage unit;
the data reading unit is used for reading displacement data of each measuring point of the LNG storage tank dome, which is acquired by the displacement data acquisition module;
the monitoring unit is used for monitoring whether displacement data of each measuring point of the dome of the LNG storage tank is lost or not in real time;
the EEMD-1DCNN-LSTM model comprises a set empirical mode decomposition unit and a 1DCNN-LSTM model, wherein the set empirical mode decomposition unit is used for decomposing displacement data of the measuring points to be repaired and each associated measuring point into a plurality of intrinsic mode function components through a set empirical mode decomposition algorithm, and vectors formed by the intrinsic mode function components are used as input characteristics of the 1DCNN-LSTM model; the 1DCNN-LSTM model is used for predicting displacement data of the missing of the measuring point to be repaired according to input characteristics;
the storage unit is used for storing displacement data of each measuring point of the LNG storage tank dome acquired by the displacement data acquisition module and displacement data predicted by the EEMD-1DCNN-LSTM model.
9. The deep learning-based storage tank dome displacement data restoration device according to claim 7, further comprising a detection unit and an alarm unit, wherein the detection unit is configured to designate a measurement point with normal displacement data as a measurement point to be restored, compare the displacement data of the measurement point predicted by the EEMD-1DCNN-LSTM model with the actual displacement data of the measurement point collected by the displacement data collection module, and when the difference between the two exceeds a preset threshold, enable the alarm unit to output an alarm signal to prompt that the predicted data is shifted too much.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202111030338.8A CN113743297B (en) | 2021-09-03 | 2021-09-03 | Storage tank dome displacement data restoration method and device based on deep learning |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202111030338.8A CN113743297B (en) | 2021-09-03 | 2021-09-03 | Storage tank dome displacement data restoration method and device based on deep learning |
Publications (2)
| Publication Number | Publication Date |
|---|---|
| CN113743297A true CN113743297A (en) | 2021-12-03 |
| CN113743297B CN113743297B (en) | 2024-08-06 |
Family
ID=78735217
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN202111030338.8A Active CN113743297B (en) | 2021-09-03 | 2021-09-03 | Storage tank dome displacement data restoration method and device based on deep learning |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN113743297B (en) |
Cited By (1)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN117556187A (en) * | 2023-11-10 | 2024-02-13 | 纬创软件(武汉)有限公司 | Cloud data restoration method and system based on deep learning and readable storage medium |
Citations (9)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| IL31478A0 (en) * | 1968-03-26 | 1969-03-27 | Monsanto Co | Bulky thermoplastic filamentary yarn with improved tufting performance |
| CN106126896A (en) * | 2016-06-20 | 2016-11-16 | 中国地质大学（武汉） | The mixed model wind speed forecasting method learnt based on empirical mode decomposition and the degree of depth and system |
| CN110059844A (en) * | 2019-02-01 | 2019-07-26 | 东华大学 | Energy storage device control method based on set empirical mode decomposition and LSTM |
| CN110852527A (en) * | 2019-11-20 | 2020-02-28 | 成都理工大学 | Reservoir physical property parameter prediction method combining deep learning |
| CN110909931A (en) * | 2019-11-20 | 2020-03-24 | 成都理工大学 | Logging curve prediction method based on modal decomposition reconstruction and depth LSTM-RNN model |
| WO2020151075A1 (en) * | 2019-01-23 | 2020-07-30 | 五邑大学 | Cnn-lstm deep learning model-based driver fatigue identification method |
| WO2020156348A1 (en) * | 2019-01-31 | 2020-08-06 | 青岛理工大学 | Structural damage identification method based on ensemble empirical mode decomposition and convolution neural network |
| KR102149495B1 (en) * | 2019-08-19 | 2020-08-28 | 고려대학교 산학협력단 | Optimization apparatus for training conditions of environmental prediction model and operating thereof |
| CN112149285A (en) * | 2020-08-31 | 2020-12-29 | 中国地质大学（武汉） | Landslide prediction method based on optimization parameter selection |
-
2021
- 2021-09-03 CN CN202111030338.8A patent/CN113743297B/en active Active
Patent Citations (9)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| IL31478A0 (en) * | 1968-03-26 | 1969-03-27 | Monsanto Co | Bulky thermoplastic filamentary yarn with improved tufting performance |
| CN106126896A (en) * | 2016-06-20 | 2016-11-16 | 中国地质大学（武汉） | The mixed model wind speed forecasting method learnt based on empirical mode decomposition and the degree of depth and system |
| WO2020151075A1 (en) * | 2019-01-23 | 2020-07-30 | 五邑大学 | Cnn-lstm deep learning model-based driver fatigue identification method |
| WO2020156348A1 (en) * | 2019-01-31 | 2020-08-06 | 青岛理工大学 | Structural damage identification method based on ensemble empirical mode decomposition and convolution neural network |
| CN110059844A (en) * | 2019-02-01 | 2019-07-26 | 东华大学 | Energy storage device control method based on set empirical mode decomposition and LSTM |
| KR102149495B1 (en) * | 2019-08-19 | 2020-08-28 | 고려대학교 산학협력단 | Optimization apparatus for training conditions of environmental prediction model and operating thereof |
| CN110852527A (en) * | 2019-11-20 | 2020-02-28 | 成都理工大学 | Reservoir physical property parameter prediction method combining deep learning |
| CN110909931A (en) * | 2019-11-20 | 2020-03-24 | 成都理工大学 | Logging curve prediction method based on modal decomposition reconstruction and depth LSTM-RNN model |
| CN112149285A (en) * | 2020-08-31 | 2020-12-29 | 中国地质大学（武汉） | Landslide prediction method based on optimization parameter selection |
Non-Patent Citations (2)
| Title |
|---|
| ZENGSHUN CHEN 等: "An Improved Method Based on EEMD-LSTM to Predict Missing Measured Data of Structural Sensors", APPL. SCI., 8 September 2022 (2022-09-08) * |
| 赵建鹏;周俊;: "基于长短时记忆网络的旋转机械状态预测研究", 噪声与振动控制, no. 04, 18 August 2017 (2017-08-18) * |
Cited By (2)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN117556187A (en) * | 2023-11-10 | 2024-02-13 | 纬创软件(武汉)有限公司 | Cloud data restoration method and system based on deep learning and readable storage medium |
| CN117556187B (en) * | 2023-11-10 | 2024-05-10 | 纬创软件(武汉)有限公司 | Cloud data restoration method and system based on deep learning and readable storage medium |
Also Published As
| Publication number | Publication date |
|---|---|
| CN113743297B (en) | 2024-08-06 |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN109947086B (en) | Mechanical fault migration diagnosis method and system based on counterstudy | |
| CN111832216B (en) | Rolling bearing residual service life prediction method based on EEMD-MCNN-GRU | |
| CN110361778B (en) | Seismic data reconstruction method based on generation countermeasure network | |
| EP2327019B1 (en) | Systems and methods for real time classification and performance monitoring of batch processes | |
| CN111767930A (en) | Method for detecting abnormal time series data of Internet of things and related equipment thereof | |
| US11954923B2 (en) | Method for rating a state of a three-dimensional test object, and corresponding rating system | |
| CN114862811B (en) | Defect detection method based on variation automatic encoder | |
| CN117540174B (en) | Building structure multi-source heterogeneous data intelligent analysis system and method based on neural network | |
| CN113988357A (en) | High-rise building wind-induced response prediction method and device based on deep learning | |
| EP3588441A1 (en) | Imagification of multivariate data sequences | |
| CN113743297A (en) | Storage tank dome displacement data restoration method and device based on deep learning | |
| DE112018004348T5 (en) | METHODS, DEVICES AND SYSTEMS FOR MONITORING DEVICES | |
| CN114549922B (en) | Method and device for identifying and evaluating dynamic characteristics of system by convolutional neural network | |
| CN113688773B (en) | Storage tank dome displacement data restoration method and device based on deep learning | |
| CN113688770A (en) | Long-term wind pressure missing data completion method and device for high-rise building | |
| Techane et al. | Rotating machinery prognostics and application of machine learning algorithms: Use of deep learning with similarity index measure for health status prediction | |
| Warner et al. | A computationally-efficient probabilistic approach to model-based damage diagnosis | |
| CN116680988A (en) | Porous medium permeability prediction method based on Transformer network | |
| CN116070985A (en) | Dangerous chemical vehicle loading and unloading process identification method | |
| CN113762203B (en) | Cross-domain self-adaptive SAR image classification method, device and equipment based on simulation data | |
| CN114298101A (en) | Dynamic weighing method and system | |
| Gerg et al. | A perceptual metric prior on deep latent space improves out-of-distribution synthetic aperture sonar image classification | |
| CN113688771B (en) | LNG storage tank acceleration response data supplementing method and device | |
| CN113688774B (en) | Advanced learning-based high-rise building wind induced response prediction and training method and device | |
| Gulgec | Machine Learning Based Sensing and Data Processing Strategies for Next Generation Structural Health Monitoring |
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 |