CN109993281A - A kind of causality method for digging based on deep learning - Google Patents
A kind of causality method for digging based on deep learning Download PDFInfo
- Publication number
- CN109993281A CN109993281A CN201910242406.3A CN201910242406A CN109993281A CN 109993281 A CN109993281 A CN 109993281A CN 201910242406 A CN201910242406 A CN 201910242406A CN 109993281 A CN109993281 A CN 109993281A
- Authority
- CN
- China
- Prior art keywords
- data
- model
- time series
- alternative features
- causality
- Prior art date
- Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
- Pending
Links
Classifications
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- 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
- G06N3/049—Temporal neural networks, e.g. delay elements, oscillating neurons or pulsed inputs
-
- 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)
- Theoretical Computer Science (AREA)
- Physics & Mathematics (AREA)
- General Health & Medical Sciences (AREA)
- Computing Systems (AREA)
- Biomedical Technology (AREA)
- Biophysics (AREA)
- Computational Linguistics (AREA)
- Data Mining & Analysis (AREA)
- Evolutionary Computation (AREA)
- Life Sciences & Earth Sciences (AREA)
- Molecular Biology (AREA)
- Artificial Intelligence (AREA)
- General Engineering & Computer Science (AREA)
- General Physics & Mathematics (AREA)
- Mathematical Physics (AREA)
- Software Systems (AREA)
- Health & Medical Sciences (AREA)
- Management, Administration, Business Operations System, And Electronic Commerce (AREA)
- Feedback Control In General (AREA)
Abstract
The invention discloses a kind of causality method for digging based on deep learning, data are pre-processed first with technologies such as missing data supplement, data normalization, one-hot codings, it is then based on the single argument time series forecasting that Keras deep learning frame carries out target signature using LSTM, adjustment model structure and a series of hyper parameters obtain optimal models, record R of the model on test set2Then score is predicted all alternative features using the model, its R on test set is obtained2Score subtracts each other the Granger causality score as alternative features and target signature with the two scores, so far then show that the number of a quantification is used to describe the Granger causality between alternative features and target signature.This method is suitable in the analysis of Influential Factors problem of other times sequence.In conclusion the Granger causality method for digging proposed by the present invention based on deep learning has towards mass data, excavates to more depth, application field widely advantage.
Description
Technical field
The invention belongs to data mining technology fields, more particularly to are based on depth learning technology from Multivariate Time Series
Granger causality is excavated in data.
Background technique
Time series data is the data point set being observed using constant as time interval, and key property has
Two, one is to have correlation between observation dependent on the time, secondly for other than having lifting trend, most times
Sequence also has seasonal trend, i.e., has specific variation in specific time window.As data carrying cost drops significantly
Low, the plenty of time sequence data generated in actual life is also completely recorded, and this data are in finance, commodity valence
The fields such as lattice, traffic are generally existing.Such as in air quality field, atmospheric visibility decline in recent years, air quality are deteriorated, compared with
More cities are often accompanied by haze weather, all produce biggish negative effect to people's lives and work, therefore air quality is asked
Topic receives government and the highest attention of people.For the Glan of visibility and meteorologic factor (sendible temperature, temperature, air pressure etc.)
Outstanding causality excavates hiding relationship and some boisterous occurrence laws that can reveal that between these features, thus
Certain theories integration can be provided for air contaminant treatment.Therefore the Granger causality excavated between multivariable is very
With practical significance.
In causality excavation applications, Granger Causality is the common method of causality in hunting time sequence data.
As shown in Figure 1, it is assumed that have a time series X.It is the sampled point { x by different time1, x2, x3..., xnCollectively form
, now using the future of the past prediction X of X, for example use x1~xn-jPredict xn-j+1~xn, a mistake is generated during prediction
Poor δ1.It is the same with X, it is assumed that having time sequence Y, it is shaped like X, by { y1, y2, y3..., ynCollectively constitute, recycle X and Y common
Past prediction X future, such as with { x1~xn-j|y1~yn-jRemove prediction xn-j+1~xn, the process of prediction generate one it is new
Error delta2.If δ1Greater than δ2, that is to say, that the associated prediction error of X and Y is less than the prediction error of X itself, in such case
Under, claim Y and X to have Granger causality.
Current existing multivariate time series Causality Analysis is concentrated mainly on qualitative Granger Causality between variable
Relational checking, some researchs based on Granger causality also only come fitting data and provide knot on the basis of linear regression
Fruit, this method can not accomplish nonlinear data complicated in actual life the prediction of high-accuracy, have very big office
It is sex-limited.With the development of artificial intelligence technology, the application of deep learning is also more universal, and shot and long term remembers (Long Short-
Term Memory, LSTM) it is a kind of time recurrent neural network (RNN), it is suitably used among processing and predicted time sequence
Every the critical event very long with delay, proposed earliest by Hochreiter & Schmidhuber in 1997, by numerous
Experts and scholars refine and promote, now because performance has been widely used in remarkably in the tasks such as time series forecasting.
The purpose of design of LSTM is very clear: avoiding long-rang dependence problem.For LSTM, long-time " remembeing " information
It is a kind of behavior of default, rather than is difficult to the thing learnt.The characteristics of LSTM is exactly that each layer is added to other than RNN structure
Valve node.Valve has 3 classes: forgeing valve (forget gate), inputs valve (input gate) and output valve
(output gate).These valves can open or close, for by the memory state (shape of network before of judgment models network
State) in the result of this layer output whether reach threshold value to be added in the calculating of this current layer.Valve node utilizes sigmoid
Function is calculated the memory state of network as input；By valve output and current layer if output result reaches threshold value
The output result is forgotten by calculated result multiplication as next layer of input if not reaching threshold value.Each layer includes
The weight of valve node can all update in the training process of model backpropagation each time.When valve open when, before mould
The training result of type will be associated with current model and calculate, and the just no longer shadow of the calculated result before when valve is closed
Ring current calculating.Therefore, influence of the early stage sequence to final result can be realized by the switch of control valve.
Summary of the invention
The technical problem to be solved by the present invention is to provide a kind of Granger causality excavation side based on deep learning
Method pre-processes data first with technologies such as missing data supplement, data normalization, one-hot codings, is then based on
Keras deep learning frame carries out the single argument time series forecasting of target signature using LSTM, adjusts model structure and a series of super
Parameter obtains optimal models, records R of the model on test set2Score, then using the model for all alternative features
It is predicted, obtains its R on test set2Score subtracts each other the lattice as alternative features and target signature with the two scores
Blue outstanding person's causality score, so far then it can be concluded that the number of a quantification is for describing between alternative features and target signature
Granger causality.
Multidimensional time-series in the larger situation of data-oriented magnitude of the present invention are proposed and are excavated based on depth learning technology
Wherein between target signature and other alternative features (one or more dimensions) Granger causality method.Data are carried out first
Pretreatment, observes data type, distribution and the deletion condition of its each feature, category feature is needed to carry out One-hot
Coding or LabelEncoder are standardized, and for the missing values in data, can attempt to utilize mode or median
To carry out completion.Keras deep learning frame is then based on to utilize LSTM to carry out the single argument time series forecasting of target signature, is adjusted
Integral mould structure and a series of hyper parameters obtain optimal models, record R of the model on test set2Score, then utilizing should
Model predicts all alternative features, obtains its R on test set2Score is subtracted each other with the two scores as standby
The Granger causality score of feature and target signature is selected, so far then it can be concluded that the number of a quantification is standby for describing
Select the Granger causality between feature and target signature.
To achieve the goals above, the invention adopts the following technical scheme: in order to construct a LSTM model to instruction
Practice collection to be trained, first have to pre-process data, including converts supervised learning problem, steady for time series problem
Time series, data normalization etc..It, can be by using the observation from the last one moment for a time series problem
Feature X as input and the observation at current time realize conversion as output Y, since it is desired that conversion is one group of time
Sequence data has clearly one-to-one input/output relation so can not be combined into and think real supervised learning like that, especially
In data set when most starting or is last, two positions always have a position that can not be focused to find out corresponding relationship in training, in order to
Such problems is solved, input feature vector can be set to 0 when most starting, and its corresponding output is exactly the first of time series
A element.For jiggly time series, stationary sequence can be converted it by the method that first-order difference is handled.Its
It is secondary, be standardized before data input can very effective promotion convergence rate and effect, especially if when activation primitive
Sigmoid or tanh, the maximum section of gradient are near 0, when input value is very big or very little, sigmoid or
With regard to substantially flat, that is, when carry out gradient decline and optimize, gradient can tend to 0, and cause for the variation of person tanh
Optimal speed is very slow.And tanh function when the default activation function of LSTM, its output area is in [- 1,1], while this is the time
The preferred range of sequence data.Therefore MinMaxScaler class can be used data set is transformed into the range of [- 1,1].It is complete
LSTM model can be then constructed at the pretreatment of data, ready data set is decomposed into training set and test set first, so
Training set and test set are decomposed into afterwards and output and input variable, and by input remodeling at 3D format expected from LSTM, i.e. [sample
Example, time step, feature].Then LSTM model is defined and is fitted, 50 neurons are defined in the first hidden layer, in output layer
The middle target signature for defining a neuron for prediction, input shape will be a time steps with all alternative features,
The efficient Adam version of mean absolute error (MAE) loss function and stochastic gradient descent is used simultaneously, which will be suitable for
50 t raining periods that batch size is 72 come finally by validation_data () parameter is arranged in fit () function
Model training and test loss during track training, in end of run, training and prediction loss are all drawn and come out, and such as scheme
Shown in 2.Here on test set, in order to preferably portray Granger causality power, R is used2To measure the effect of prediction
Quality, formula are as follows:
WhereinIt is the time series predicted, yiThe true value for the time series observed is represented,When indicating this
Between sequence mean value, therefore, R2It can be construed to change the explained variance of prediction model, when the value more approaches 1, illustrate model
Prediction effect is better, otherwise then illustrates that the predictive ability of the model is negative if negative value, uses R2It can more formally define
Granger causality.By only being predicted twice with target signature and using all alternative features, predicted twice to calculate
R on test set2Difference, claim its Granger causality score, to quantify between alternative features and target signature
Granger causality is strong and weak.
By being tested based on above-mentioned technical description, the causality of last available alternative features and target signature
Score, the value the big, illustrates that the Granger causality of alternative features and target signature is stronger, on the contrary then illustrate that relationship is weaker
Even without Granger causality.
A kind of Granger causality method for digging based on deep learning the following steps are included:
Step 1 obtains time related sequence data, and cleans to these data.
Step 2, for shortage of data, it is unsmooth situations such as the data after cleaning are pre-processed.
Step 3 carries out model buildings for pretreated data, constructs training set and test set, parameter optimization obtain pre-
Survey the optimal solution of model.
Step 4 predicts the only data of target signature using trained model and containing the data of alternative features respectively,
The Granger causality score of alternative features and target signature is obtained using training result twice.
Preferably, step 2 specifically includes the following steps:
Step 2.1, for the missing values in data, carry out completion using mode or median；
Step 2.2 need to carry out One-hot coding for category feature or LabelEncoder is standardized；
Step 2.3 is directed to jiggly time series, converts it into steady sequence by the method that first-order difference is handled
Column；
Time series problem is converted supervised learning problem by step 2.4；
Preferably, step 3 specifically includes the following steps:
Step 3.1, according to data volume size cutting training set and test set；
Step 3.2 defines LSTM model, and 50 neurons are defined in the first hidden layer, and a mind is defined in output layer
Target signature through member for prediction, input shape will be a time steps with all alternative features；
Step 3.3 defines loss function, uses the height of mean absolute error (MAE) loss function and stochastic gradient descent
Adam version is imitated, which will be suitable for 50 t raining periods of the batch size for 72, then trains LSTM model, constantly adjusts
Hyper parameter obtains model optimal solution；
Preferably, step 4 specifically includes the following steps:
Step 4.1, using above-mentioned model, respectively when containing only target signature and two kinds comprising all alternative features
Predicted target values；
Step 4.2, the R predicted twice2Value calculates the R predicted twice2The difference of value, can be obtained alternative features and mesh
Mark the Granger causality score of feature.
Opposite with the prior art, the present invention has following clear superiority:
The present invention is based on depth learning technologies to excavate the Granger causality between time series, first with missing number
Data are pre-processed according to technologies such as supplement, data normalization, one-hot codings, are then based on Keras deep learning frame benefit
The single argument time series forecasting of target signature is carried out with LSTM, adjusts model structure and a series of hyper parameters obtain optimal models, remember
Record R of the model on test set2Then score is predicted all alternative features using the model, obtain it and surveying
R on examination collection2Score subtracts each other the Granger causality score as alternative features and target signature with the two scores, until
This is then it can be concluded that the number of a quantification is used to describe the Granger causality between alternative features and target signature.Base
In deep learning prediction model can more accurate fit time sequence data, the accuracy rate of prediction compared to line linear regression and
Conventional machines learning method is higher, therefore the present invention can excavate the pass of the Granger Causality between time series to a deeper level
System, and the numerical value of a quantification is provided to measure the power of Granger causality.In addition, when this method is suitable for other
Between sequence analysis of Influential Factors problem on.In conclusion the Granger causality proposed in this paper based on deep learning is dug
Pick method has towards mass data, excavates to more depth, application field widely advantage.
Detailed description of the invention:
Fig. 1 is the general flow figure of Granger causality test；
Fig. 2 is the loss variation diagram of model training process of the present invention；
Fig. 3 is that the present invention relates to the flow charts of method；
Specific embodiment
Yi Xiajiehejutishishili,Bing Canzhaofutu,Dui Benfamingjinyibuxiangxishuoming.
Hardware device used in the present invention has PC machine 1,1080 video card 2；
As shown in figure 3, the present invention provides a kind of Granger causality method for digging based on deep learning, specifically include
Following steps:
Step 1, time related sequence data are obtained, and these data are cleaned.
Step 2, for shortage of data, it is unsmooth situations such as data are pre-processed.
Step 2.1, for the missing values in data, completion is carried out using mode or median.
Step 2.2, One-hot coding need to be carried out for category feature or LabelEncoder is standardized.
Step 2.3, for jiggly time series, steady sequence is converted it by the method that first-order difference is handled
Column.
Step 2.4, supervised learning problem is converted by time series problem.
Step 3, model buildings, construct training set and test set, parameter optimization obtain the optimal solution of prediction model.
Step 3.1, according to data volume size cutting training set and test set.
Step 3.2, LSTM model is defined, 50 neurons are defined in the first hidden layer, a mind is defined in output layer
Target signature through member for prediction, input shape will be a time steps with all alternative features, while using average
The efficient Adam version of absolute error (MAE) loss function and stochastic gradient descent, it is 72 which, which will be suitable for batch size,
50 t raining periods.
Step 3.3, loss function is defined, training LSTM model, constantly adjustment hyper parameter obtain model optimal solution.
Step 4, the only data of target signature are predicted respectively using trained model and containing the data of alternative features,
The Granger causality score of alternative features and target signature is obtained using training result twice；
Step 4.1, using above-mentioned model, respectively when containing only target signature and two kinds comprising all alternative features
Predicted target values.
Step 4.2, the R predicted twice2Value, calculates their difference, alternative features and target signature can be obtained
Granger causality score.
Above embodiments are only exemplary embodiment of the present invention, are not used in the limitation present invention, protection scope of the present invention
It is defined by the claims.Those skilled in the art can within the spirit and scope of the present invention make respectively the present invention
Kind modification or equivalent replacement, this modification or equivalent replacement also should be regarded as being within the scope of the present invention.
Claims (5)
1. a kind of causality method for digging based on deep learning, it is characterised in that: the following steps are included:
Step 1 obtains time related sequence data, and cleans to these data；
Step 2 pre-processes the data after cleaning for shortage of data, unsmooth situation；
Step 3 carries out model buildings for pretreated data, constructs training set and test set, and parameter optimization obtains prediction mould
The optimal solution of type；
Step 4 is predicted the only data of target signature respectively using trained model and containing the data of alternative features, utilized
Training result obtains the Granger causality score of alternative features and target signature twice.
2. a kind of causality method for digging based on deep learning according to claim 1, it is characterised in that: step 2
Specifically include following steps,
Step 2.1, for the missing values in data, carry out completion using mode or median；
Step 2.2 need to carry out One-hot coding for category feature or LabelEncoder is standardized；
Step 2.3 is directed to jiggly time series, converts it into stationary sequence by the method that first-order difference is handled；
Time series problem is converted supervised learning problem by step 2.4.
3. a kind of causality method for digging based on deep learning according to claim 1, it is characterised in that: step 3
Specifically includes the following steps:
Step 3.1, according to data volume size cutting training set and test set；
Step 3.2 defines LSTM model, and 50 neurons are defined in the first hidden layer, and a neuron is defined in output layer
For the target signature of prediction, inputting shape will be a time step with all alternative features；
Step 3.3 defines loss function, uses the efficient of mean absolute error (MAE) loss function and stochastic gradient descent
Adam version, the model will be suitable for 50 t raining periods of the batch size for 72, then train LSTM model, and constantly adjustment is super
Parameter obtains model optimal solution.
4. a kind of causality method for digging based on deep learning according to claim 1, it is characterised in that: step 4
Specifically includes the following steps:
Step 4.1, using above-mentioned model, predicted respectively when containing only target signature and two kinds comprising all alternative features
Target value；
Step 4.2, the R predicted twice2Value calculates the R predicted twice2The difference of value, can be obtained alternative features and target is special
The Granger causality score of sign.
5. a kind of causality method for digging based on deep learning according to claim 1, it is characterised in that:
First have to pre-process data, including by time series problem be converted into supervised learning problem, stationary time series,
Data normalization；It, can be as input by using the observation from the last one moment for a time series problem
The observation at feature X and current time realizes conversion as output Y, since it is desired that conversion is one group of time series data,
There is clearly one-to-one input/output relation so can not be combined into and think real supervised learning like that, especially in data set
When most starting or is last, two positions always have a position that can not be focused to find out corresponding relationship in training, such in order to solve
Input feature vector can be set to 0 when most starting by problem, and its corresponding output is exactly first element of time series；Needle
To jiggly time series, stationary sequence is converted it by the method that first-order difference is handled；Secondly, before data input
Be standardized can very effective promotion convergence rate and effect, especially if sigmoid or tanh when activation primitive,
The maximum section of its gradient is near 0, and the variation of sigmoid or tanh is with regard to base when input value very big perhaps very little
This is flat, that is, when carry out gradient decline and optimize, gradient can tend to 0, and cause optimal speed very slow；And
Tanh function when the default activation function of LSTM, its output area is in [- 1,1], while this is the first choice of time series data
Range；Therefore MinMaxScaler class can be used data set is transformed into the range of [- 1,1]；Complete the pretreatment of data
LSTM model can be then constructed, ready data set is decomposed into training set and test set first, then by training set and survey
Examination collection, which is decomposed into, outputs and inputs variable, and by input remodeling at 3D format expected from LSTM, i.e., [sample, time step, feature]；
Then LSTM model is defined and is fitted, 50 neurons are defined in the first hidden layer, neuron is defined in output layer
For the target signature of prediction, inputting shape will be a time step with all alternative features, while use average absolute
The efficient Adam version of error MAE loss function and stochastic gradient descent, it is 50 of 72 which, which will be suitable for batch size,
T raining period, finally by fit () function be arranged validation_data () parameter carry out track training during model
Training and test loss, in end of run, training and prediction loss are all drawn and come out, as shown in Figure 2；Here in test set
On, in order to preferably portray Granger causality power, use R2To measure the effect quality of prediction, formula are as follows:
WhereinIt is the time series predicted, yiThe true value for the time series observed is represented,Indicate this time sequence
The mean value of column, therefore, R2It can be construed to change the explained variance of prediction model, when the value more approaches 1, illustrate the prediction of model
Effect is better, otherwise then illustrates that the predictive ability of the model is negative if negative value, uses R2More formally define Granger because
Fruit relationship；By only being predicted twice with target signature and using all alternative features, predicted twice to calculate in test set
On R2Difference, claim its Granger causality score, thus quantify Granger between alternative features and target signature because
Fruit relationship is strong and weak.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201910242406.3A CN109993281A (en) | 2019-03-28 | 2019-03-28 | A kind of causality method for digging based on deep learning |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201910242406.3A CN109993281A (en) | 2019-03-28 | 2019-03-28 | A kind of causality method for digging based on deep learning |
Publications (1)
| Publication Number | Publication Date |
|---|---|
| CN109993281A true CN109993281A (en) | 2019-07-09 |
Family
ID=67131761
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN201910242406.3A Pending CN109993281A (en) | 2019-03-28 | 2019-03-28 | A kind of causality method for digging based on deep learning |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN109993281A (en) |
Cited By (7)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN110731787A (en) * | 2019-09-26 | 2020-01-31 | 首都师范大学 | fatigue state causal network method based on multi-source data information |
| CN111026852A (en) * | 2019-11-28 | 2020-04-17 | 广东工业大学 | Financial event-oriented hybrid causal relationship discovery method |
| CN112101480A (en) * | 2020-09-27 | 2020-12-18 | 西安交通大学 | Multivariate clustering and fused time sequence combined prediction method |
| CN113300897A (en) * | 2021-06-16 | 2021-08-24 | 中移（杭州）信息技术有限公司 | Causal relationship identification method, terminal device and storage medium |
| CN113627663A (en) * | 2021-08-04 | 2021-11-09 | 浙江大学 | Dynamic cause and effect analysis method based on geographic time sequence in city |
| CN113866391A (en) * | 2021-09-29 | 2021-12-31 | 天津师范大学 | Deep learning model prediction factor interpretation method and application thereof in soil water content prediction |
| CN117688472A (en) * | 2023-12-13 | 2024-03-12 | 华东师范大学 | Unsupervised domain adaptive multivariate time sequence classification method based on causal structure |
-
2019
- 2019-03-28 CN CN201910242406.3A patent/CN109993281A/en active Pending
Cited By (14)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN110731787B (en) * | 2019-09-26 | 2022-07-22 | 首都师范大学 | Fatigue state causal network method based on multi-source data information |
| CN110731787A (en) * | 2019-09-26 | 2020-01-31 | 首都师范大学 | fatigue state causal network method based on multi-source data information |
| CN111026852B (en) * | 2019-11-28 | 2023-06-30 | 广东工业大学 | Financial event-oriented hybrid causal relationship discovery method |
| CN111026852A (en) * | 2019-11-28 | 2020-04-17 | 广东工业大学 | Financial event-oriented hybrid causal relationship discovery method |
| CN112101480A (en) * | 2020-09-27 | 2020-12-18 | 西安交通大学 | Multivariate clustering and fused time sequence combined prediction method |
| CN112101480B (en) * | 2020-09-27 | 2022-08-05 | 西安交通大学 | Multivariate clustering and fused time sequence combined prediction method |
| CN113300897A (en) * | 2021-06-16 | 2021-08-24 | 中移（杭州）信息技术有限公司 | Causal relationship identification method, terminal device and storage medium |
| CN113300897B (en) * | 2021-06-16 | 2022-10-18 | 中移（杭州）信息技术有限公司 | Causal relationship identification method, terminal device and storage medium |
| CN113627663A (en) * | 2021-08-04 | 2021-11-09 | 浙江大学 | Dynamic cause and effect analysis method based on geographic time sequence in city |
| CN113627663B (en) * | 2021-08-04 | 2023-11-10 | 浙江大学 | Dynamic causal analysis method based on geographic time sequence in city |
| CN113866391A (en) * | 2021-09-29 | 2021-12-31 | 天津师范大学 | Deep learning model prediction factor interpretation method and application thereof in soil water content prediction |
| CN113866391B (en) * | 2021-09-29 | 2024-03-08 | 天津师范大学 | Deep learning model prediction factor interpretation method and application thereof in soil water content prediction |
| CN117688472A (en) * | 2023-12-13 | 2024-03-12 | 华东师范大学 | Unsupervised domain adaptive multivariate time sequence classification method based on causal structure |
| CN117688472B (en) * | 2023-12-13 | 2024-05-24 | 华东师范大学 | Unsupervised domain adaptive multivariate time sequence classification method based on causal structure |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN109993281A (en) | A kind of causality method for digging based on deep learning | |
| CN110148285B (en) | Intelligent oil well parameter early warning system based on big data technology and early warning method thereof | |
| CN111639430B (en) | Natural gas pipeline leakage identification system driven by digital twinning | |
| CN101950382B (en) | Method for optimal maintenance decision-making of hydraulic equipment with risk control | |
| CN111985610B (en) | Oil pumping well pump efficiency prediction system and method based on time sequence data | |
| CN104636751A (en) | Crowd abnormity detection and positioning system and method based on time recurrent neural network | |
| CN106199174A (en) | Extruder energy consumption predicting abnormality method based on transfer learning | |
| CN114282443B (en) | Residual service life prediction method based on MLP-LSTM supervised joint model | |
| CN113554466A (en) | Short-term power consumption prediction model construction method, prediction method and device | |
| CN111160659B (en) | Power load prediction method considering temperature fuzzification | |
| CN112733440A (en) | Intelligent fault diagnosis method, system, storage medium and equipment for offshore oil-gas-water well | |
| CN116610816A (en) | Personnel portrait knowledge graph analysis method and system based on graph convolution neural network | |
| Wu | MOOC learning behavior analysis and teaching intelligent decision support method based on improved decision tree C4. 5 algorithm | |
| CN116911571A (en) | Mine operation and maintenance monitoring system | |
| CN115017513A (en) | Intelligent contract vulnerability detection method based on artificial intelligence | |
| CN115481726A (en) | Industrial robot complete machine health assessment method and system | |
| Gökçe et al. | Performance comparison of simple regression, random forest and XGBoost algorithms for forecasting electricity demand | |
| CN116720098A (en) | Abnormal behavior sensitive student behavior time sequence modeling and academic early warning method | |
| CN117077631A (en) | Knowledge graph-based engineering emergency plan generation method | |
| CN109635008A (en) | A kind of equipment fault detection method based on machine learning | |
| CN109034453A (en) | A kind of Short-Term Load Forecasting Method based on multiple labeling neural network | |
| CN118152579A (en) | Knowledge-driven multi-parameter electric pump well working condition intelligent diagnosis method | |
| CN115293249A (en) | Power system typical scene probability prediction method based on dynamic time sequence prediction | |
| Ban et al. | Physics‐Informed Gas Lifting Oil Well Modelling using Neural Ordinary Differential Equations | |
| CN118114812B (en) | Shale gas yield prediction method, computer equipment and storage medium |
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 | ||
| WD01 | Invention patent application deemed withdrawn after publication |
Application publication date: 20190709 |
|
| WD01 | Invention patent application deemed withdrawn after publication |