CN114511767A - Quick state prediction method for timing diagram data - Google Patents
Quick state prediction method for timing diagram data Download PDFInfo
- Publication number
- CN114511767A CN114511767A CN202210129022.2A CN202210129022A CN114511767A CN 114511767 A CN114511767 A CN 114511767A CN 202210129022 A CN202210129022 A CN 202210129022A CN 114511767 A CN114511767 A CN 114511767A
- Authority
- CN
- China
- Prior art keywords
- data
- time
- features
- dimension
- enc
- Prior art date
- Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
- Granted
Links
- 238000000034 method Methods 0.000 title claims abstract description 28
- 238000010586 diagram Methods 0.000 title claims abstract description 15
- 230000009466 transformation Effects 0.000 claims abstract description 26
- 238000013507 mapping Methods 0.000 claims abstract description 10
- 238000012423 maintenance Methods 0.000 claims abstract description 6
- 238000007781 pre-processing Methods 0.000 claims abstract description 4
- 238000004364 calculation method Methods 0.000 claims description 20
- 239000011159 matrix material Substances 0.000 claims description 17
- 230000007246 mechanism Effects 0.000 claims description 13
- 239000013598 vector Substances 0.000 claims description 10
- 238000004821 distillation Methods 0.000 claims description 9
- 230000004927 fusion Effects 0.000 claims description 7
- 230000002123 temporal effect Effects 0.000 claims description 7
- 238000013528 artificial neural network Methods 0.000 claims description 4
- 230000014759 maintenance of location Effects 0.000 claims description 4
- 238000012545 processing Methods 0.000 claims description 3
- 230000009467 reduction Effects 0.000 claims description 2
- 230000008859 change Effects 0.000 abstract description 5
- 230000009286 beneficial effect Effects 0.000 abstract description 3
- 238000000605 extraction Methods 0.000 description 4
- 230000008569 process Effects 0.000 description 3
- 238000013459 approach Methods 0.000 description 2
- 238000013527 convolutional neural network Methods 0.000 description 2
- 230000007774 longterm Effects 0.000 description 2
- 238000011160 research Methods 0.000 description 2
- 230000002159 abnormal effect Effects 0.000 description 1
- 238000006243 chemical reaction Methods 0.000 description 1
- 238000004590 computer program Methods 0.000 description 1
- 238000001514 detection method Methods 0.000 description 1
- 230000000694 effects Effects 0.000 description 1
- 238000011156 evaluation Methods 0.000 description 1
- 238000003058 natural language processing Methods 0.000 description 1
- 230000003287 optical effect Effects 0.000 description 1
- 230000000737 periodic effect Effects 0.000 description 1
- 230000003595 spectral effect Effects 0.000 description 1
- 239000000126 substance Substances 0.000 description 1
- 230000002459 sustained effect Effects 0.000 description 1
- 238000012549 training Methods 0.000 description 1
- 230000001131 transforming effect Effects 0.000 description 1
Images
Classifications
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F18/00—Pattern recognition
- G06F18/20—Analysing
- G06F18/25—Fusion techniques
- G06F18/253—Fusion techniques of extracted features
-
- 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)
- Theoretical Computer Science (AREA)
- Physics & Mathematics (AREA)
- Data Mining & Analysis (AREA)
- Life Sciences & Earth Sciences (AREA)
- Artificial Intelligence (AREA)
- General Physics & Mathematics (AREA)
- General Engineering & Computer Science (AREA)
- Evolutionary Computation (AREA)
- Molecular Biology (AREA)
- Computational Linguistics (AREA)
- Software Systems (AREA)
- Mathematical Physics (AREA)
- Health & Medical Sciences (AREA)
- Biomedical Technology (AREA)
- Biophysics (AREA)
- Computing Systems (AREA)
- General Health & Medical Sciences (AREA)
- Bioinformatics & Cheminformatics (AREA)
- Evolutionary Biology (AREA)
- Bioinformatics & Computational Biology (AREA)
- Computer Vision & Pattern Recognition (AREA)
- Complex Calculations (AREA)
- Compression Or Coding Systems Of Tv Signals (AREA)
Abstract
The invention discloses a quick state prediction method facing to timing diagram data, which comprises the following steps: step 1, data preprocessing: converting the data format; step 2, feature transformation: mapping data information, data state information and time information observed by each sensor to a high-dimensional space; and 3, feature learning of the encoder end: fusing all the transformed features in the step 2 to generate features of a time dimension and a space dimension, fusing the features of the time dimension and the space dimension together, and transmitting the fused features into a decoder part; and 4, feature transformation at the decoder end: performing feature transformation by adopting a generating mode to generate finally learned features; and 5, predicting data and states. The invention considers the state characteristics of the data, and switches in from the angle of data maintenance time to reveal the change of the data state, thereby reducing the redundancy degree of the data to a certain extent and being beneficial to improving the prediction accuracy of the data.
Description
Technical Field
The invention relates to the technical field of time sequence prediction, in particular to a quick state prediction method for time sequence diagram data.
Background
Timing data is ubiquitous at present, and is exemplified by traffic timing data, and prediction thereof is an important component for constructing intelligent traffic. Only if the traffic time sequence data is predicted more reliably, reasonable planning can be made in advance according to factors such as future traffic flow conditions and the like. In recent years, more and more research has focused on the study of such timing diagram data: lai et al solved the prediction problem of multivariate Time series using LSTNet (Long and Short Time-series Network), which integrates a CNN model and an RNN model, where CNN is used to extract Short-term dependencies of data, and RNN is used to discover Long-term patterns and trends of data, improving the accuracy of data prediction compared to a single model, but it ignores the spatial relationship existing between data. Zheng et al incorporates knowledge of Graph structures into temporal prediction, and proposes a GMAN (Graph Multi-Attention Network) model for long-term traffic information prediction. The STE (spatial temporal embedding) respectively uses GCN to obtain spatial features, uses an attention mechanism to obtain time features, and then fuses the spatial features and the time features, so that the extraction of space-time features is realized; however, when extracting the spatial features, the used adjacency matrix is fixed, and obviously, in the traffic map, the state between data points is changed. Also having the same problem are the STGCN (Spatial-Temporal graphic relational Networks) model proposed by Yu et al, the ASTGCN (extension-Based Spatial-Temporal graphic relational Networks) model proposed by Guo et al, and the like. The Graph WaveNet model proposed by Wu et al solves the above limitations, and uses a dynamically generated approach to construct the adjacency matrix of the Graph, so that the generated spatial relationship can be optimized by the model itself.
With the success of the transform model in the field of natural language processing, there is now some research on applying it to the field of time series prediction, such as the ASTGNN (Attention based Spatial-Temporal Graph Neural Network) proposed by Guo et al, which uses the transform model and uses it for extracting the time features of data, and incorporates the GCN part therein for extracting the Spatial features. However, in the transform model, a large number of matrix operations are involved, and the requirements on hardware devices such as a memory are high. Zhou et al propose an Informer model for optimizing the number of matrix calculations and reducing the calculation complexity, aiming at the problems of large calculation amount and large memory consumption of a Transformer model, but do not consider the spatial relationship between data when applied to the prediction of time series data.
In the above-mentioned methods, they focus on predicting the data value at a future time, which is often reliable when predicting a shorter time period in the future. However, the prediction range is expanded, that is, the accuracy of the prediction result is reduced as the prediction time increases in the long-time series prediction. Moreover, this prediction method only focuses on the characteristics of the data at a certain time, but ignores the status characteristics of the data, such as how long the data will last.
Disclosure of Invention
The present invention is directed to a method for predicting a state of a timing chart data, which is capable of predicting the timing chart data and predicting a corresponding holding time to reveal a corresponding state.
The purpose of the invention is realized by the following technical scheme:
a quick state prediction method facing to timing diagram data comprises the following steps:
and 3, feature learning of the encoder end: fusing all the transformed features in the step 2, transmitting the fused features into an encoder module based on a multi-head attention mechanism, and learning data features and data state features on a time dimension; transmitting the original data before feature transformation to a dynamic graph neural network module, learning data features on a space dimension, fusing the features of a time dimension and the space dimension together, and transmitting the fused data to a decoder part;
and 4, feature transformation at the decoder end: performing characteristic transformation in a generating mode; firstly, generating characteristics with time dimension and space dimension; then, carrying out multi-head attention calculation together with the features generated in the step 3 to generate finally learned features;
said step 4 comprises in particular the sub-steps of,
step 401: obtaining the fusion characteristics Z of the time characteristics, the data characteristics and the data state characteristics through the transformation of the step 2Fuse_dec＝ZTime_dec+ZData_dec+ZDur_dec；
Step 402: fusing the features ZFuse_decAs input, the characteristic representation Z in the time dimension is obtained by passing the decoder part based on the multi-head Attention mechanism through a MaskProbSparse Self-Attention moduleOut_dec_temporal；
Step 403: using Dynamic GCN to perform feature learning on an original data sequence to obtain data features on spatial dimensions, and then using 1D convolution to perform dimension transformation to obtain ZOut_dec_spatio；
Step 404: merging the temporal and spatial features of the decoder portion to obtain ZOut_dec＝ZOut_dec_temporal+ZOut_dec_spatio. Then, after FullAttention, Z is addedOut_decAs Q, ZOut_encAs K, V, to obtain a final characteristic output ZOut；
And 5, predicting data and states: and respectively transmitting the final generated characteristics of the decoder end into two different full-connection layers to predict data and states.
Further, the data information, the data state information and the time information are all in a data form of a data sequence.
Further, the step 2 comprises the following sub-steps:
step 201: transmitting data information into TCN module, extracting data characteristic ZData_enc；
Step 202: transmitting the maintaining time sequence into a TCN module, and extracting the characteristic Z of the data stateDur_enc；
Step 203: coding the time sequence information to obtain the characteristic Z of the timeTime_enc。
Furthermore, the data state information is a maintenance time sequence, and the state information reflects the rule of data change to a certain extent.
Further, the step 3 specifically includes the following sub-steps:
step 301: combining the characteristics in the step 2 to obtain a fusion characteristic Z containing time information, data characteristics and data state characteristicsFuse_enc＝ZTime_enc+ZData_enc+ZDur_enc。
Step 302: fusing the features ZFuse_encAs input, pass into the encoder portion based on the multi-headed attention mechanism; reducing the dimension of the characteristic sequence to 1/2 of the original dimension for the purpose of distillation; by reducing the dimension, on one hand, the requirements on hardware devices such as a memory are reduced, and on the other hand, the calculation efficiency is further accelerated.
Step 303: through L1A sub ProbSparse Self-orientation module and a MaxPool distillation module to obtain the output Z of the encoderOut_enc_temporal；
Step 304: using dynamic GCN to perform feature learning on an original data sequence to obtain data features on a space dimension, and then using 1D convolution and MaxPool to perform dimension transformation to obtain ZOut_enc_spatio；
Step 305: fusing the features of the time dimension and the space dimension together to obtain ZOut_enc＝ZOut_enc_temporal+ZOut_enc_spatioAnd passed into the decoder section.
Further, the processing procedure of the encoder part based on the multi-head attention mechanism is as follows: firstly, multi-head self-Attention calculation is carried out through a ProbSpareSelf-Attention module, and the time complexity is calculated from O (n) by using an optimized matrix calculation mode2) Reduction to O (nlog (n)); and then, reducing the dimension of the characteristic sequence to 1/2 of the original dimension through MaxPool convolution operation for the purpose of distillation.
Further, the step 5 specifically includes the following sub-steps:
step 501: the final characteristic vector Z obtained in the step 4 is processedOutInput FC1, output predicted timing graph data
Step 502: the final characteristic vector Z obtained in the step 4 is processedOutTransmitting the predicted retention time corresponding to the timing diagram data to FC2
The invention has the beneficial effects that: the invention considers the state characteristics of the data, switches in from the angle of data maintenance time to reveal the change of the data state, and fuses the change as information into the characteristics, so that the data characteristics are richer on one hand; on the other hand, the redundancy degree of the data is reduced to a certain degree, the prediction accuracy of the data is improved, and the method can also be applied to downstream tasks such as data completion, abnormal detection and the like.
Drawings
In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the drawings used in the description of the embodiments or the prior art will be briefly described below, it is obvious that the drawings in the following description are only some embodiments of the present invention, and for those skilled in the art, other drawings can be obtained according to the structures shown in the drawings without creative efforts.
FIG. 1 is a flow chart of the method of the present invention.
FIG. 2 is a schematic diagram of a TCN module.
Fig. 3 is a forward process of the model.
FIG. 4 is a diagram of a model framework of the present invention.
FIG. 5 is a schematic diagram of a detailed list of each layer of the model.
Detailed Description
It should be understood that the specific embodiments described herein are merely illustrative of the invention and are not intended to limit the invention.
The technical solutions in the embodiments of the present invention will be clearly and completely described below with reference to the drawings in the embodiments of the present invention, and it is obvious that the described embodiments are only a part of the embodiments of the present invention, and not all of the embodiments. All other embodiments, which can be derived by a person skilled in the art from the embodiments given herein without making any creative effort, shall fall within the protection scope of the present invention.
step 1: data preprocessing: the data format is converted, and the data reported by the sensor is converted from the format of (time point, data value) to the format of (data value, maintenance time) according to the time sequence.
For a particular sensor XiIn other words, the data form is originally
We merge data segments with the same data to form a new format of
On one hand, the conversion mode can reduce the redundancy of data and save the memory resource, and under the worst condition, the data length is equal to the length of the original data sequence; on the other hand, using a sustained time sequence of data
The change of the data state is revealed, and the duration time sequence of the data is used as an additional input for enriching the content of the data characteristics, which is beneficial to improving the prediction accuracy of the data.
Step 2: and (3) feature transformation: for data information observed by each sensor, mapping the data information into a high-dimensional space by using a TCN module; mapping the data state information of each sensor to a high-dimensional space in a TCN convolution mode; and (3) carrying out characteristic coding on the time information of each sensor according to the date, and mapping to a high-dimensional space in a linear transformation mode.
Step 201: transmitting the data value sequence into a TCN module, and extracting the characteristic Z of the data sequenceData_enc；
Because the data dimension of the sensors is small, the TCN module is used here to map the features of each sensor into a high-dimensional space, facilitating feature learning by the encoder module.
Step 202: transmitting the maintaining time sequence into a TCN module, and extracting the state characteristic Z of the dataDur_enc(ii) a TCN is a feature extraction module composed of a hole convolution and a residual structure, as shown in FIG. 2.
The hole convolution kernel is described using the formula:
wherein x ∈ RnRepresents a sequence of 1D, s is the sequence for performing the convolution operation, D is the expansion factor, k is the number of filters, and f represents the filter.
Compared with the common convolution operation, the method does not leak future information and cause the condition of state information leakage when extracting the features.
Step 203: coding the time sequence information to obtain a characteristic vector Z of timeTime_enc。
For time-series data, there may be a periodic regularity between data, and the sequentiality between data is certain. However, in the multi-head attention mechanism, the data is insensitive to the sequence information of the data, the position information between the data is not enough to be embodied in a relative position coding mode, and the characteristic of the periodicity of the data cannot be well extracted. Therefore, time information of the data is encoded to reveal order information between the data from a global perspective.
The pseudo code of the whole model is shown in FIG. 3, and the details of the parameters of each layer of the model are shown in FIG. 5.
And step 3: feature learning at the encoder side: fusing the features in the step 2, transmitting the fused features into an encoder module based on a multi-head attention mechanism, and learning data features and data state features on a time dimension; and transmitting the original data before feature transformation to a dynamic graph neural network module, learning data features on a spatial dimension, fusing the features of the temporal dimension and the spatial dimension, and transmitting the fused features to a decoder part.
Step 301: combining the characteristics in the step 2 to obtain a fusion characteristic Z containing time information, data characteristics and data state characteristicsFuse_enc＝ZTime_enc+ZData_enc+ZDur_enc。
Step 302: fusing the features ZFuse_encAs input, an encoder section based on a multi-head attention mechanism is introduced. Firstly, multi-head Self-Attention calculation is carried out through a ProbSparse Self-Attention module, the part uses an optimized matrix calculation mode, and the time complexity is reducedFrom O (n)2) The calculation efficiency of the model is accelerated by reducing the calculation rate to O (nlog (n)); and then, the aim of distillation is achieved by MaxPool convolution operation, and the dimension of the characteristic sequence is reduced to 1/2 of the original dimension. By reducing the dimension, on one hand, the requirements on hardware devices such as a memory are reduced, and on the other hand, the calculation efficiency is further accelerated.
When calculating the Attention of a plurality of heads, the method of ProbSparse Self-Attention is adopted, and the formula is expressed as follows:
wherein the content of the first and second substances,
and d represents the dimension corresponding to Q and K.
The reason that the sparse form of self-attention can be used is that after calculating the self-attention score using Q and K, it is found that the effect on V follows a long tail distribution, i.e. only a small fraction of the attention score plays a more important role. Therefore, the similarity between Q and K (measured here using KL divergence) is calculated, and the vector of top K (where K is a parameter set to K ═ ln L _ Q, and L _ Q is the number of vectors in Q) is taken as the top K (where K is a parameter) vector
This approach can reduce computational complexity without affecting attention results, from O (n)2) The calculation efficiency of the model is accelerated by reducing the calculation rate to O (nlog (n));
then, the dimension of the data is reduced to half of the input by using the operation of distillation (here, MaxPool convolution) to reduce the requirement on the memory. Meanwhile, the data dimension of the lower network layer is reduced, and the operation speed is accelerated.
Step 303: through L1A sub ProbSparse Self-orientation module and a MaxPool distillation module to obtain the output Z of the encoderOut_enc_temporal。
Step 304: using dynamic GCN to perform feature learning on an original data sequence to obtain data features on a space dimension, and then using 1D convolution and MaxPool to perform dimension transformation to obtain ZOut_enc_spatio。
Since the spatial relationship between the data should be dynamically changing in time series. Specifically, at different time points, the adjacency matrix formed between data should be different, so when initializing the adjacency matrix, we use the full-connection matrix to represent the relationship between data. And the relation between the data at each time is changed and is represented by using a dynamically generated correlation strength matrix S.
At this time, we mark the hidden representation of data at different time as Ei(i∈[1,t]) So the correlation strength matrix S is represented as:
sjkthe (0 ≦ j, k ≦ N) matrix represents the strength of association between the j point and the k point.
The spectral domain based GCN is then used to update the data representation E at each timei。
Ei＝GCN(Ei)＝σ(A⊙SiEiW)
Wherein: eiFor a hidden layer representation of the incoming data,
Siis the association strength matrix of each node in the learned graph. A denotes the original adjacency matrix. A, Si∈RN*N. An indication of a corresponding position multiplication. W is a matrix of weights that is,
finally, obtaining the spatial characteristics at each time, and forming Z by transforming dimensionsOut_enc_spatioAnd (6) outputting.
Step 305: fusing the features of the time dimension and the space dimension together to obtain ZOut_enc＝ZOut_enc_temporal+ZOut_enc_spatioAnd passed into the decoder section.
And 4, step 4: feature transformation at decoder side: and performing feature transformation by adopting a generating mode. Firstly, generating features with time dimension and space dimension similar to step 3; then, the multi-head attention calculation is performed together with the features generated in step 3, and finally learned features are generated.
Step 401: fusion characteristic Z of information, data characteristic and data state characteristic between acquisitionsFuse_dec＝ZTime_dec+ZData_dec+ZDur_dec。
Since the predictive value is obtained by using the generating equation in the decoder, the decoder should have a part of the input information, i.e. the fusion characteristic Z similar to step 301, as in the encoderFuse_dec＝ZTime_dec+ZData_dec+ZDur_dec。
Step 402: fusing the features ZFuse_decAs input, the characteristic representation Z in the time dimension is acquired by passing the decoder part based on the multi-head Attention mechanism through a Mask ProbSparse Self-Attention moduleOut_dec_temporal。
Different from the encoder part, the encoder part is training data, and the input time series data can be used for feature extraction. However, in the encoder part, data in the prediction time period needs to be generated, so that the time sequence is strictly observed, and future information cannot be leaked. Therefore, after the Mask ProbSparse Self-Attention module is used to calculate the Attention score of the head, the upper triangular part of the matrix is set to negative infinity, and then after the calculation of the softmax layer, the influence caused by future information can be ignored.
Step 403: using Dynamic GCN to perform feature learning on an original data sequence to obtain data features on spatial dimensions, and then performing dimension transformation through 1D convolution to obtain ZOut_dec_spatio。
Characteristic information Z of decoder transmitted by encoderOut_encIncluding features in both the spatial and temporal dimensions, and thus in ZOut_encBefore performing the multi-head attention calculation, it should be ensured that the features of the decoder portion also contain information in the spatial dimension and the time dimension, and therefore, the input data is subjected to the extraction of the spatial features here.
Step 404: merging the temporal and spatial features of the decoder portion to obtain ZOut_dec＝ZOut_dec_temporal+ZOut_dec_spatio. Then, Z is subjected to Full extensionOut_decAs Q, ZOut_encAs K, V, to obtain a final characteristic output ZOut。
And 5: the output of the decoder is connected to two full connection layers (FC), and the timing chart data and the corresponding holding time are respectively predicted. An output part of Decoder as in fig. 4;
step 501: the final characteristic vector Z obtained in the step 4 is processedOutInput FC1, output predicted timing graph data
Step 502: the final characteristic vector Z obtained in the step 4 is processedOutTransmitting the predicted retention time corresponding to the timing diagram data to FC2
MSE was used as an evaluation index:
wherein | X | represents the total number of data in the segment, | T | represents the total number of data retention time in the segment,
representing predicted data, x, t representing originalThe data of (1).
It should be noted that, for simplicity of description, the above-mentioned embodiments of the method are described as a series of acts or combinations, but those skilled in the art should understand that the present application is not limited by the order of acts described, as some steps may be performed in other orders or simultaneously according to the present application. Further, those skilled in the art should also appreciate that the embodiments described in the specification are preferred embodiments and that the acts and elements referred to are not necessarily required in this application.
In the above embodiments, the descriptions of the respective embodiments have respective emphasis, and for parts that are not described in detail in a certain embodiment, reference may be made to related descriptions of other embodiments.
It will be understood by those skilled in the art that all or part of the processes of the methods of the embodiments described above can be implemented by a computer program, which can be stored in a computer-readable storage medium, and when executed, can include the processes of the embodiments of the methods described above. The storage medium may be a magnetic disk, an optical disk, a ROM, a RAM, etc.
The above disclosure is only for the purpose of illustrating the preferred embodiments of the present invention, and it is therefore to be understood that the invention is not limited by the scope of the appended claims.
Claims (7)
1. A method for fast state prediction oriented to timing diagram data is characterized by comprising the following steps:
step 1, data preprocessing: converting a data format, namely converting the data reported by the sensor from a time point + data value format into a data value + maintenance time format according to a time sequence;
step 2, feature transformation: for data information observed by each sensor, mapping the data information into a high-dimensional space by using a TCN module; mapping the data state information of each sensor to a high-dimensional space in a TCN convolution mode; for the time information of each sensor, carrying out feature coding according to date, and mapping to a high-dimensional space in a linear transformation mode;
and 3, feature learning of the encoder end: fusing all the transformed features in the step 2, transmitting the fused features into an encoder module based on a multi-head attention mechanism, and learning data features and data state features on a time dimension; transmitting the original data before feature transformation to a dynamic graph neural network module, learning data features on a space dimension, fusing the features of a time dimension and the space dimension together, and transmitting the fused data to a decoder part;
and 4, feature transformation at the decoder end: performing characteristic transformation in a generating mode; firstly, generating characteristics with time dimension and space dimension; then, carrying out multi-head attention calculation together with the features generated in the step 3 to generate finally learned features;
said step 4 comprises in particular the sub-steps of,
step 401: obtaining the fusion characteristics Z of the time characteristics, the data characteristics and the data state characteristics through the transformation of the step 2Fuse_dec＝ZTime_dec+ZData_dec+ZDur_dec；
Step 402: fusing the features ZFuse_decAs input, the characteristic representation Z in the time dimension is acquired by passing the decoder part based on the multi-head Attention mechanism through a Mask ProbSparse Self-Attention moduleOut_dec_temporal；
Step 403: using Dynamic GCN to perform feature learning on an original data sequence to obtain data features on spatial dimensions, and then obtaining Z through dimension transformation of 1D convolutionOut_dec_spatio；
Step 404: merging the temporal and spatial features of the decoder portion to obtain ZOut_dec＝ZOut_dec_temporal+ZOut_dec_spatio(ii) a Then, Z is subjected to Full extensionOut_decAs Q, ZOut_encAs K, V, to obtain a final characteristic output ZOut；
And 5, predicting data and states: and respectively transmitting the final generated characteristics of the decoder end into two different full-connection layers to predict data and states.
2. The method as claimed in claim 1, wherein the data information, the data status information and the time information are in the form of data of a data sequence.
3. The method for fast predicting the status of timing diagram data according to claim 1, wherein the step 2 comprises the following sub-steps:
step 201: transmitting data information into TCN module, extracting data characteristic ZData_enc；
Step 202: transmitting the maintaining time sequence into a TCN module, and extracting the characteristic Z of the data stateDur_enc；
Step 203: coding the time sequence information to obtain the characteristic Z of the timeTime_enc。
4. The method as claimed in claim 1, wherein the data state information is a time sequence.
5. The method as claimed in claim 1, wherein the step 3 comprises the following sub-steps:
step 301: combining the characteristics in the step 2 to obtain a fusion characteristic Z containing time information, data characteristics and data state characteristicsFuse_enc＝ZTime_enc+ZData_enc+ZDur_enc；
Step 302: fusing the features ZFuse_encAs input, the characteristic sequence is transmitted to an encoder part based on a multi-head attention mechanism for processing so as to achieve the purpose of distillation, and the dimensionality of the characteristic sequence is reduced to 1/2 of the original dimensionality;
step 303: through L1Second ProbSparse Self-Attention module and MaxPhol distillation module to obtain output Z of encoderOut_enc_temporal；
Step 304: using Dynamic GCN to perform feature learning on an original data sequence to obtain data features on a space dimension, and then using 1D convolution and MaxPool transformation dimension to obtain ZOut_enc_spatio；
Step 305: fusing the features of the time dimension and the space dimension together to obtain ZOut_enc＝ZOut_enc_temporal+ZOut_enc_spatioAnd passed into the decoder section.
6. The method of claim 1, wherein the processing procedure of the encoder portion based on the multi-head attention mechanism is as follows: firstly, multi-head Self-Attention calculation is carried out through a ProbSparse Self-Attention module, and the time complexity is calculated from O (n) by using an optimized matrix calculation mode2) Reduction to O (nlog (n)); and then, the dimension of the characteristic sequence is reduced to 1/2 of the original dimension through MaxPool convolution operation for achieving the purpose of distillation.
7. The method as claimed in claim 1, wherein the step 5 comprises the following sub-steps:
step 501: the final characteristic vector Z obtained in the step 4 is processedOutInput FC1, output predicted timing graph data
Step 502: the final characteristic vector Z obtained in the step 4 is processedOutTransmitting the predicted retention time corresponding to the timing diagram data to FC2
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202210129022.2A CN114511767B (en) | 2022-02-11 | 2022-02-11 | Rapid state prediction method for time sequence diagram data |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202210129022.2A CN114511767B (en) | 2022-02-11 | 2022-02-11 | Rapid state prediction method for time sequence diagram data |
Publications (2)
| Publication Number | Publication Date |
|---|---|
| CN114511767A true CN114511767A (en) | 2022-05-17 |
| CN114511767B CN114511767B (en) | 2023-12-01 |
Family
ID=81552158
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN202210129022.2A Active CN114511767B (en) | 2022-02-11 | 2022-02-11 | Rapid state prediction method for time sequence diagram data |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN114511767B (en) |
Cited By (1)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN117726195A (en) * | 2024-02-07 | 2024-03-19 | 创意信息技术股份有限公司 | City management event quantity change prediction method, device, equipment and storage medium |
Citations (7)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN111027610A (en) * | 2019-12-03 | 2020-04-17 | 腾讯科技（深圳）有限公司 | Image feature fusion method, apparatus, and medium |
| CN112330158A (en) * | 2020-11-06 | 2021-02-05 | 北京建筑大学 | Method for identifying traffic index time sequence based on autoregressive differential moving average-convolution neural network |
| CN112365095A (en) * | 2020-12-03 | 2021-02-12 | 浙江汉德瑞智能科技有限公司 | Flight delay analysis and prediction method based on weather and flow control influence |
| CN112836287A (en) * | 2020-11-10 | 2021-05-25 | 华北电力大学 | Neural network-based electric vehicle resource flexibility prediction method |
| CN113159414A (en) * | 2021-04-19 | 2021-07-23 | 华南理工大学 | Traffic speed prediction method based on timing diagram neural network |
| US20210295999A1 (en) * | 2020-03-18 | 2021-09-23 | Hitachi, Ltd. | Patient state prediction apparatus, patient state prediction method, and patient state prediction program |
| CN113722383A (en) * | 2021-09-13 | 2021-11-30 | 福韵数据服务有限公司 | Investigation device and method based on time sequence information |
-
2022
- 2022-02-11 CN CN202210129022.2A patent/CN114511767B/en active Active
Patent Citations (7)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN111027610A (en) * | 2019-12-03 | 2020-04-17 | 腾讯科技（深圳）有限公司 | Image feature fusion method, apparatus, and medium |
| US20210295999A1 (en) * | 2020-03-18 | 2021-09-23 | Hitachi, Ltd. | Patient state prediction apparatus, patient state prediction method, and patient state prediction program |
| CN112330158A (en) * | 2020-11-06 | 2021-02-05 | 北京建筑大学 | Method for identifying traffic index time sequence based on autoregressive differential moving average-convolution neural network |
| CN112836287A (en) * | 2020-11-10 | 2021-05-25 | 华北电力大学 | Neural network-based electric vehicle resource flexibility prediction method |
| CN112365095A (en) * | 2020-12-03 | 2021-02-12 | 浙江汉德瑞智能科技有限公司 | Flight delay analysis and prediction method based on weather and flow control influence |
| CN113159414A (en) * | 2021-04-19 | 2021-07-23 | 华南理工大学 | Traffic speed prediction method based on timing diagram neural network |
| CN113722383A (en) * | 2021-09-13 | 2021-11-30 | 福韵数据服务有限公司 | Investigation device and method based on time sequence information |
Non-Patent Citations (3)
| Title |
|---|
| SREEJA NAG等: "Prototyping operational autonomy for Space Traffic Management", ACTA ASTRONAUTICA, vol. 180, pages 489 - 506, XP086471340, DOI: 10.1016/j.actaastro.2020.11.056 * |
| 余沁: "电力总负荷的时间序列方法超短期预测", 中国优秀硕士学位论文全文数据库 工程科技Ⅱ辑, pages 042 - 603 * |
| 尹月华等: "基于GIS的北京市道路交通流特性研究", 公路工程, vol. 45, no. 04, pages 102 - 108 * |
Cited By (2)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN117726195A (en) * | 2024-02-07 | 2024-03-19 | 创意信息技术股份有限公司 | City management event quantity change prediction method, device, equipment and storage medium |
| CN117726195B (en) * | 2024-02-07 | 2024-05-07 | 创意信息技术股份有限公司 | City management event quantity change prediction method, device, equipment and storage medium |
Also Published As
| Publication number | Publication date |
|---|---|
| CN114511767B (en) | 2023-12-01 |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| US11941522B2 (en) | Address information feature extraction method based on deep neural network model | |
| CN110929092B (en) | Multi-event video description method based on dynamic attention mechanism | |
| AU2019204399B2 (en) | A neural dialog state tracker for spoken dialog systems using dynamic memory networks | |
| CN112597296B (en) | Abstract generation method based on plan mechanism and knowledge graph guidance | |
| CN111372123B (en) | Video time sequence segment extraction method based on local to global | |
| CN115273464A (en) | Traffic flow prediction method based on improved space-time Transformer | |
| CN115145551A (en) | Intelligent auxiliary system for machine learning application low-code development | |
| Yu et al. | Forecasting a short‐term wind speed using a deep belief network combined with a local predictor | |
| CN115841119B (en) | Emotion cause extraction method based on graph structure | |
| CN116341651A (en) | Entity recognition model training method and device, electronic equipment and storage medium | |
| CN114817773A (en) | Time sequence prediction system and method based on multi-stage decomposition and fusion | |
| CN114511767A (en) | Quick state prediction method for timing diagram data | |
| CN113569062A (en) | Knowledge graph completion method and system | |
| EP4231202A1 (en) | Apparatus and method of data processing | |
| He et al. | Distributional drift adaptation with temporal conditional variational autoencoder for multivariate time series forecasting | |
| CN116629362A (en) | Interpreteable time graph reasoning method based on path search | |
| WO2024012735A1 (en) | Training of a machine learning model for predictive maintenance tasks | |
| Miao et al. | Low‐latency transformer model for streaming automatic speech recognition | |
| CN115689639A (en) | Commercial advertisement click rate prediction method based on deep learning | |
| CN114547276A (en) | Three-channel diagram neural network-based session recommendation method | |
| CN115169285A (en) | Event extraction method and system based on graph analysis | |
| CN114780841A (en) | KPHAN-based sequence recommendation method | |
| Rauf et al. | BCE4ZSR: Bi-encoder empowered by teacher cross-encoder for zero-shot cold-start news recommendation | |
| CN115062115B (en) | Method for predicting response emotion type in dialogue strategy | |
| CN117131876A (en) | Text processing method, device, 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 | ||
| GR01 | Patent grant | ||
| GR01 | Patent grant |