CN112925822A - Time series classification method, system, medium and device based on multi-representation learning - Google Patents

Time series classification method, system, medium and device based on multi-representation learning Download PDF

Info

Publication number
CN112925822A
CN112925822A CN202110180044.7A CN202110180044A CN112925822A CN 112925822 A CN112925822 A CN 112925822A CN 202110180044 A CN202110180044 A CN 202110180044A CN 112925822 A CN112925822 A CN 112925822A
Authority
CN
China
Prior art keywords
time
sequence
classification
time series
characterization
Prior art date
Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
Granted
Application number
CN202110180044.7A
Other languages
Chinese (zh)
Other versions
CN112925822B (en
Inventor
王少鲲
胡宇鹏
李学庆
曲磊钢
李振
展鹏
Current Assignee (The listed assignees may be inaccurate. Google has not performed a legal analysis and makes no representation or warranty as to the accuracy of the list.)
Shandong University
Original Assignee
Shandong University
Priority date (The priority date is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the date listed.)
Filing date
Publication date
Application filed by Shandong University filed Critical Shandong University
Priority to CN202110180044.7A priority Critical patent/CN112925822B/en
Publication of CN112925822A publication Critical patent/CN112925822A/en
Application granted granted Critical
Publication of CN112925822B publication Critical patent/CN112925822B/en
Active legal-status Critical Current
Anticipated expiration legal-status Critical

Links

Images

Classifications

    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06F—ELECTRIC DIGITAL DATA PROCESSING
    • G06F16/00—Information retrieval; Database structures therefor; File system structures therefor
    • G06F16/20—Information retrieval; Database structures therefor; File system structures therefor of structured data, e.g. relational data
    • G06F16/24—Querying
    • G06F16/245—Query processing
    • G06F16/2458—Special types of queries, e.g. statistical queries, fuzzy queries or distributed queries
    • G06F16/2474—Sequence data queries, e.g. querying versioned data
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06F—ELECTRIC DIGITAL DATA PROCESSING
    • G06F16/00—Information retrieval; Database structures therefor; File system structures therefor
    • G06F16/20—Information retrieval; Database structures therefor; File system structures therefor of structured data, e.g. relational data
    • G06F16/26—Visual data mining; Browsing structured data
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06F—ELECTRIC DIGITAL DATA PROCESSING
    • G06F16/00—Information retrieval; Database structures therefor; File system structures therefor
    • G06F16/20—Information retrieval; Database structures therefor; File system structures therefor of structured data, e.g. relational data
    • G06F16/28—Databases characterised by their database models, e.g. relational or object models
    • G06F16/284—Relational databases
    • G06F16/285—Clustering or classification
    • G06F16/287—Visualization; Browsing
    • 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)
  • Databases & Information Systems (AREA)
  • Data Mining & Analysis (AREA)
  • General Engineering & Computer Science (AREA)
  • General Physics & Mathematics (AREA)
  • Software Systems (AREA)
  • Computational Linguistics (AREA)
  • Mathematical Physics (AREA)
  • Artificial Intelligence (AREA)
  • Health & Medical Sciences (AREA)
  • Life Sciences & Earth Sciences (AREA)
  • Biomedical Technology (AREA)
  • Biophysics (AREA)
  • Evolutionary Computation (AREA)
  • General Health & Medical Sciences (AREA)
  • Molecular Biology (AREA)
  • Computing Systems (AREA)
  • Fuzzy Systems (AREA)
  • Probability & Statistics with Applications (AREA)
  • Image Analysis (AREA)

Abstract

The invention relates to a time series classification method, a system, a medium and a device based on multi-feature learning, which comprises the following steps: (1) performing multi-feature coding on a given time sequence based on different time sequence characterization strategies; (2) the representation fusion and enhancement are realized by utilizing a residual error network and a bidirectional long-time and short-time memory network; (3) and finishing classification by utilizing a multi-layer perceptron network and realizing classification interpretability by utilizing an attention mechanism. According to the method, a multi-channel time sequence representation learning model is constructed, and time sequence features can be comprehensively understood based on various representation strategies. The method is based on the representation fusion model of the residual error network and the bidirectional long-time and short-time memory network, can effectively fuse the multi-view representations and realize representation enhancement, thereby effectively improving the classification precision. The invention can effectively identify the important time sequence characteristics of the time sequence based on an attention mechanism, namely, can provide the interpretability basis of the classification result, namely, the classification interpretability is realized.

Description

Time series classification method, system, medium and device based on multi-representation learning
Technical Field
The invention belongs to the technical field of time series data mining, and particularly relates to a time series classification method, a time series classification system, a time series classification medium and time series classification equipment based on multi-representation learning.
Background
With the rapid development of technologies such as internet, cloud computing, big data processing and the like, the technology has entered the big data era of everything interconnection. Among the massive heterogeneous data, there is a time-related time series data, which is widely distributed in almost all application fields of the real world, and is referred to as "time series" (for short). The time series can not only reflect specific data characteristics at a certain moment, but also reveal continuous change along with time, change trend and potential knowledge of data entities. Time series often has large data characteristics such as 'mass', 'high dimension', 'continuous generation', etc., and is very challenging to research, so that time series data mining is also considered as one of ten challenging problems in the field of data mining in this century.
In the time series data mining study, the Time Series Classification (TSC) problem, which focuses on: how to correctly classify massive time series into predefined data categories. The research results of time series classification are widely applied to the fields of industry, medical treatment, government affairs, finance and the like, such as abnormality detection based on the time series of the internet of things, disease diagnosis based on the time series of the electrocardiogram and the like, and thus have received wide attention from the academic and industrial fields.
Currently, the time series classification research is still in the starting stage, and the following two challenges mainly exist:
(1) how to improve the classification accuracy of the TSCs. The time series often has data characteristics of 'mass, high dimension, continuity' and the like, and how to accurately classify the time series with diversity is very challenging. The limitations of the existing TSC method are: on one hand, a time sequence feature extraction strategy is combined with various traditional classification algorithms, and then a voting mechanism is adopted to improve the classification precision to a certain extent, while the potential contribution of deep learning to the feature learning is ignored; on the other hand, the conventional TSC method can only simply use a single deep learning model (a convolutional neural network and a bidirectional long-and-short time memory network) to perform universal characterization learning on the time sequence and finish classification, and lacks adaptability understanding and characterization on various time sequence characteristics, so that the classification accuracy may not be effectively improved.
(2) How to enhance the class interpretability of the TSC. In addition to obtaining accurate classification results, researchers also desire a model with some explanatory power over the classification results. For example: when diagnosing the running state of the relevant equipment based on the time series, besides obtaining the classification result of "normal" or "abnormal", the researcher also hopes that the classification model can effectively display the classification basis in a certain data visualization mode, namely, the researcher is told that the classification model is a specific classification result according to what kind of data characteristics. Therefore, providing interpretable basis for classification results is also challenging.
Disclosure of Invention
Aiming at the defects of the prior art, the invention provides a time series classification method, a system, a medium and equipment based on multi-representation learning, and the time series classification is completed by constructing a multi-representation learning network model. Therefore, the invention firstly provides a time sequence characteristic coding network, and generates corresponding time sequence representation based on different time sequence representation strategies; secondly, the invention provides a representation fusion model based on a residual error network and a bidirectional long-time and short-time memory network, and the time sequence representation can be comprehensively understood and enhanced in a multi-view mode; finally, the invention completes time sequence classification by utilizing a multilayer perceptron network and carries out interpretable data visualization on the classification result by utilizing an attention mechanism.
Interpretation of terms:
1. a Piecewise Linear Representation (PLR) strategy, where PLR is a time series Representation strategy, and specifically, PLR can segment an original time series based on a series of time series points and use straight line segments between the time series points, that is, perform data Representation and dimension reduction on the original sequence in a linear manner.
2. The method comprises the steps of (a) a step of approximately aggregating (PAA) strategy, wherein the PAA is a time series characterization strategy, specifically, the PAA first performs an equally dividing operation on a given time series to obtain a series of subsequences, then calculates the mean value of each subsequence respectively, and performs data characterization on the original time series by using the obtained mean value sequence, and the characterization mode also realizes a data dimension reduction operation.
3. Element-wise multiplication (element-wise multiplication), element-wise multiplication, which multiplies elements at corresponding positions of two sequences while maintaining the total sequence length.
4. The bidirectional long-short-term memory network (BilSTM) is a deep learning network model which can capture time sequence characteristics in a time sequence based on different time units so as to realize time sequence data representation.
5. Residual network (ResidualNetwork, ResNet) residual network is a deep learning network model. The conventional Convolutional Neural Network (CNN) may have a phenomenon of gradient explosion or gradient disappearance as the number of network layers increases, and may cause a decrease in learning performance. Compared with a CNN model, the ResNet model adopts a series of residual modules, and a correlation channel is established between input and output of different network layers, so that network convergence is accelerated by learning the residual in the different network layers, and the data characterization capability can be improved by continuously deepening the number of model layers.
6. Multilayer perceptron network, MLP, a deep learning network model. A plurality of intermediate layers (hidden layers) are arranged between an input layer and an output layer of the model, full connection operation is utilized between the layers, data mapping and transformation operation is carried out, and corresponding data representation is obtained at the output layer.
7. adam optimizer function, adam (adaptive motion estimation) optimizer, an adaptive moment estimation optimization function, which is a common optimization function in deep learning. After the loss function of the model is defined, the optimization function is needed to solve the parameter set in the model, so as to achieve the minimum loss. The optimization function dynamically adjusts the learning rate of the model parameters by utilizing the first moment estimation and the second moment estimation of the gradient, and defines the range of the learning rate by controlling the learning speed, thereby obtaining the stable model parameters.
The invention aims to provide a time series classification method based on multi-feature learning.
In order to achieve the purpose, the technical scheme adopted by the invention is as follows:
a time series classification method based on multi-feature learning is characterized by comprising the following steps: inputting the time sequence to be classified into a trained multi-representation learning network model to realize the classification of the time sequence, and specifically comprising the following steps:
(1) performing multi-feature coding on a given time sequence based on different time sequence characterization strategies;
(2) the representation fusion and enhancement are realized by utilizing a residual error network and a bidirectional long-time and short-time memory network;
(3) and finishing classification by utilizing a multi-layer perceptron network and realizing classification interpretability by utilizing an attention mechanism.
A time series is a set of data series generated based on a time sequence. Stock index futures data in financial markets, for example, which represents a certain trend of data change (rise and fall) in chronological order; sensor data collected in chronological order in industrial manufacturing; electrocardiographic data of patients acquired in time series in the medical industry may be referred to as time series data (in accordance with the international nomenclature for TimeSeries) for short. Time series is a common big data type, widely spread in almost all application fields of the real world. The time series can not only reflect specific data characteristics at a certain moment, but also reveal continuous change along with time, change trend and potential knowledge of data entities. The time series often has the characteristics of large data such as 'mass', 'high dimension', 'continuous generation', and the like, and the research on the time series is very challenging.
Preferably, in step (1), the implementation steps of the multi-feature coding include:
A. segment linear characterization (PLR) strategy based on ith time series T of time series data setiCarrying out characteristic coding, wherein the time sequence data set comprises N time sequences, i is more than or equal to 1 and is less than or equal to N, and obtaining a first coding sequence
Figure BDA0002941915050000031
B. PAA (PAA) strategy versus time series T based on segmented aggregation ApproximationiPerforming characteristic coding to obtain a second coding sequence
Figure BDA0002941915050000032
C. Using one-dimensional time sequence convolution operation to obtain the first code sequence in step A
Figure BDA0002941915050000033
The second coding sequence obtained in step B
Figure BDA0002941915050000034
And time series TiAnd performing data characterization to obtain a basic characterization sequence.
Further preferably, the specific implementation process of step C is:
performing two-layer one-dimensional convolutional neural network (1 DConvolulanalNetwork, 1DCNN) on the first coding sequence obtained in the step A
Figure BDA0002941915050000035
The second coding sequence obtained in step B
Figure BDA0002941915050000036
And time series TiPerforming one-dimensional time sequence convolution operation, wherein the size of a convolution kernel of each layer is 3 multiplied by 1, and time sequence coding representation can be completed according to the characteristics of the current time and the context time thereof; as shown in formula (I), formula (II) and formula (III):
Figure BDA0002941915050000037
Figure BDA0002941915050000041
Figure BDA0002941915050000042
in the formula (I), the formula (II) and the formula (III),
Figure BDA0002941915050000043
to represent
Figure BDA0002941915050000044
Data representation after two layers of one-dimensional convolution neural networks,
Figure BDA0002941915050000045
represents TiData representation after two layers of one-dimensional convolution neural networks,
Figure BDA0002941915050000046
to represent
Figure BDA0002941915050000047
And (4) data representation after two layers of one-dimensional convolution neural networks.
Preferably, in step (2), the specific implementation steps for characterizing fusion and enhancement include:
D. performing element-wise multiplication (element-wise multiplication) on the basic characterization sequence obtained in the step C, and performing characterization fusion;
E. d, respectively inputting the characterization sequences obtained in the step D after the characterizations are fused into a residual error network and a bidirectional long-time and short-time memory network, and deeply understanding the sequence characteristics; and merging the characterization sequences processed by the residual error network and the bidirectional long-time and short-time memory network again, thereby realizing characterization fusion and enhancement.
Further preferably, in step D, the element-level multiplication operation is represented by formula (IV) and formula (v):
Figure BDA0002941915050000048
Figure BDA0002941915050000049
in the formulae (IV) and (V), tanh represents an activation function,
Figure BDA00029419150500000410
to represent
Figure BDA00029419150500000411
And
Figure BDA00029419150500000412
the enhanced characterization after the fusion is carried out,
Figure BDA00029419150500000413
to represent
Figure BDA00029419150500000414
And
Figure BDA00029419150500000415
enhanced characterization after fusion.
Further preferably, in the step E, the specific implementation steps include:
a. for the above two characterization sequences
Figure BDA00029419150500000416
And
Figure BDA00029419150500000417
performing connection operation, inputting the connection operation into a bidirectional long-time memory network to obtain a characterization sequence
Figure BDA00029419150500000418
As shown in formula (VI):
Figure BDA00029419150500000419
in the formula (VI), the BilSTM () represents that the fusion representation is input into a bidirectional long-time and short-time memory network, the important time sequence characteristics in the fusion representation are fully understood, and the enhanced representation is obtained
Figure BDA00029419150500000420
b. Will be provided with
Figure BDA00029419150500000421
Inputting the residual error network to obtain a characterization sequence
Figure BDA00029419150500000422
c. Outputting the sequence of the step a and the step b
Figure BDA00029419150500000423
Performing characterization fusion based on the connection operation, wherein the characterization fusion is shown as a formula (VII):
Figure BDA00029419150500000424
in the formula (VII), the reaction mixture is,
Figure BDA00029419150500000425
represents the fused enhanced signature sequence, Concat () represents the two sequences
Figure BDA00029419150500000426
And performing connection operation to finish the characterization fusion.
Preferably, in step (3), the classification is completed by using a multi-layer perceptron network, and the interpretability of the classification is realized by using an attention mechanism, and the specific implementation steps include:
F. performing representation learning again on the fusion sequence obtained after representation fusion and enhancement in the step (2) by utilizing a multilayer perceptron network;
G. utilizing a Softmax mechanism to obtain probability distribution of the current sequence to different categories, and realizing time sequence classification;
H. and acquiring attention scores of the classification model to different parts of the current sequence based on an attention mechanism, and realizing interpretability of the classification result.
More preferably, step F means: will be provided with
Figure BDA0002941915050000051
Inputting the multi-layer perceptron network to obtain a characterization sequence
Figure BDA0002941915050000052
More preferably, step G means:
using Softmax mechanism, for
Figure BDA0002941915050000053
Carrying out normalization operation to obtain the ith time sequence TiProbability P (L | T) of belonging to different data classesi) As shown in formula (VIII):
Figure BDA0002941915050000054
in the formula (VIII), L represents the number of data classes, TiRefers to the ith time sequence, and the softmax () function carries out normalization operation and will
Figure BDA0002941915050000055
Is mapped to [0,1 ] for the total sequence number in (2)]Within the range, T is obtainediProbability distributions belonging to different categories.
According to the calculated probability distribution, T is calculatediAnd dividing the data into data categories with the highest probability to realize time series classification.
Further preferably, the loss function K (θ) is represented by the formula (ix):
Figure BDA0002941915050000056
in the formula (IX), N represents the total number of time series in the time series data set, KiRepresents TiA 0-1 vector corresponding to the L data classes;
all parameters of a Multi-representation Learning network Model (MLN) are trained by using an adam optimizer function.
Further preferably, the parameters in the multi-representation learning network model are solved by using an optimization function, wherein the optimization function is an adam optimizer function in Tensorflow.
A multi-feature learning based time series classification system, comprising:
a multi-feature information encoding module that performs feature encoding on a given time series based on a Piecewise Aggregation Approximation (PAA) policy and a Piecewise Linear Representation (PLR) policy;
the multi-feature fusion module is used for effectively fusing multi-feature sequences based on a residual error network and a bidirectional long-time and short-time memory network to realize feature fusion and enhancement;
and the classification module completes data classification and performs data visualization on classification results by utilizing a multilayer perceptron network and an attention mechanism, and provides a classification interpretability basis.
It is a second object of the present invention to provide a computer-readable storage medium.
In order to achieve the purpose, the invention adopts the following technical scheme:
a computer-readable storage medium having stored therein a plurality of instructions adapted to be loaded by a processor of a terminal device and to execute a method of multi-feature learning based time series classification.
A third object of the present invention is to provide a terminal device.
In order to achieve the purpose, the invention adopts the following technical scheme:
a terminal device comprising a processor and a computer readable storage medium, the processor being configured to implement instructions; a computer readable storage medium stores instructions adapted to be loaded by a processor and to perform said one multi-feature learning based time series classification.
Compared with the prior art, the invention has the beneficial effects that:
1. according to the method, a multi-channel time sequence representation learning model is constructed, and time sequence features can be comprehensively understood based on various representation strategies.
2. The method is based on the representation fusion model of the residual error network and the bidirectional long-time and short-time memory network, can effectively fuse the multi-view representations and realize representation enhancement, thereby effectively improving the classification precision.
3. The invention can effectively identify the important time sequence characteristics of the time sequence based on an attention mechanism, namely, can provide the interpretability basis of the classification result, namely, the classification interpretability is realized.
Drawings
FIG. 1 is a schematic flow chart of a time series classification method based on multi-feature learning according to the present invention;
FIG. 2(a) is a schematic illustration of an attention visualization of the present invention on an Adiac data set;
FIG. 2(b) is a schematic view of an attention visualization on the DistalPhalanxtW dataset according to the present invention;
FIG. 2(c) is a schematic illustration of the attention visualization on Coffee dataset according to the present invention;
fig. 2(d) is a schematic view of the attention visualization on the Herring data set according to the present invention.
Detailed Description
The invention is further described with reference to the drawings and examples.
It is to be understood that the following detailed description is exemplary and is intended to provide further explanation of the invention as claimed. Unless defined otherwise, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this invention belongs.
It is noted that the terminology used herein is for the purpose of describing particular embodiments only and is not intended to be limiting of exemplary embodiments according to the invention. As used herein, the singular forms "a", "an" and "the" are intended to include the plural forms as well, and it should be understood that when the terms "comprises" and/or "comprising" are used in this specification, they specify the presence of stated features, steps, operations, devices, components, and/or combinations thereof, unless the context clearly indicates otherwise.
Example 1
A time series classification method based on Multi-representation Learning, as shown in fig. 1, implements time series classification by constructing a Multi-representation Learning network (MLN) model. Firstly, comprehensively understanding the sequence characteristics by using a plurality of characterization strategies through the MLN model; secondly, effectively fusing multiple characteristics based on a residual error network and a bidirectional long-time and short-time memory network to realize characteristic enhancement; and finally, realizing time series classification by using a multi-layer perceptron network, and giving an interpretable basis of a classification result based on an attention mechanism. The method specifically comprises the following steps:
(1) performing multi-feature coding on a given time sequence based on different time sequence characterization strategies;
(2) the representation fusion and enhancement are realized by utilizing a residual error network and a bidirectional long-time and short-time memory network;
(3) and finishing classification by utilizing a multi-layer perceptron network and realizing classification interpretability by utilizing an attention mechanism.
A time series is a set of data series generated based on a time sequence. Stock index futures data in financial markets, for example, which represents a certain trend of data change (rise and fall) in chronological order; sensor data collected in chronological order in industrial manufacturing; electrocardiographic data of patients acquired in time series in the medical industry may be referred to as time series data (in accordance with the international nomenclature for TimeSeries) for short. Time series is a common big data type, widely spread in almost all application fields of the real world. The time series can not only reflect specific data characteristics at a certain moment, but also reveal continuous change along with time, change trend and potential knowledge of data entities. The time series often has the characteristics of large data such as 'mass', 'high dimension', 'continuous generation', and the like, and the research on the time series is very challenging.
Example 2
The method for classifying time series based on multi-feature learning according to embodiment 1 is characterized in that:
as shown in fig. 1, in step (1), the specific implementation steps of the multi-feature coding include:
A. segment linear characterization (PLR) strategy based on ith time series T of time series data setiCarrying out characteristic coding, wherein the time sequence data set comprises N time sequences, i is more than or equal to 1 and is less than or equal to N, and obtaining a first coding sequence
Figure BDA0002941915050000071
B. PAA (PAA) strategy versus time series T based on segmented aggregation ApproximationiPerforming characteristic coding to obtain a second coding sequence
Figure BDA0002941915050000072
C. Using one-dimensional time sequence convolution operation to obtain the first code sequence in step A
Figure BDA0002941915050000073
The second coding sequence obtained in step B
Figure BDA0002941915050000074
And time series TiPerforming data characterization to obtain basic characterization sequence
Figure BDA0002941915050000075
The concrete implementation process of the step C is as follows:
performing two-layer one-dimensional convolutional neural network (1 DConvolulanalNetwork, 1DCNN) on the first coding sequence obtained in the step A
Figure BDA0002941915050000081
The second coding sequence obtained in step B
Figure BDA0002941915050000082
And time series TiPerforming one-dimensional time sequence convolution operation, wherein the size of a convolution kernel of each layer is 3 multiplied by 1, and time sequence coding representation can be completed according to the characteristics of the current time and the context time thereof; as shown in formula (I), formula (II) and formula (III):
Figure BDA0002941915050000083
Figure BDA0002941915050000084
Figure BDA0002941915050000085
in the formula (I), the formula (II) and the formula (III),
Figure BDA0002941915050000086
to represent
Figure BDA0002941915050000087
Data representation after two layers of one-dimensional convolution neural networks,
Figure BDA0002941915050000088
represents TiData representation after two layers of one-dimensional convolution neural networks,
Figure BDA0002941915050000089
to represent
Figure BDA00029419150500000810
And (4) data representation after two layers of one-dimensional convolution neural networks.
Example 3
The method for classifying time series based on multi-feature learning according to embodiment 1 is characterized in that:
in the step (2), the specific implementation steps of the characterization fusion and the enhancement comprise:
D. performing element-wise multiplication (element-wise multiplication) on the basic characterization sequence obtained in the step C, and performing characterization fusion;
E. d, respectively inputting the characterization sequences obtained in the step D after the characterizations are fused into a residual error network and a bidirectional long-time and short-time memory network, and deeply understanding the sequence characteristics; and merging the characterization sequences processed by the residual error network and the bidirectional long-time and short-time memory network again, thereby realizing characterization fusion and enhancement.
In step D, element-level multiplication is shown as formula (IV) and formula (v):
Figure BDA00029419150500000811
Figure BDA00029419150500000812
in the formulae (IV) and (V), tanh represents an activation function,
Figure BDA00029419150500000813
to represent
Figure BDA00029419150500000814
And
Figure BDA00029419150500000815
the enhanced characterization after the fusion is carried out,
Figure BDA00029419150500000816
to represent
Figure BDA00029419150500000817
And
Figure BDA00029419150500000818
enhanced characterization after fusion.
In step E, the concrete implementation steps include:
a. for the above two characterization sequences
Figure BDA00029419150500000819
And
Figure BDA00029419150500000820
performing connection operation, inputting the connection operation into a bidirectional long-time memory network to obtain a characterization sequence
Figure BDA00029419150500000821
As shown in formula (VI):
Figure BDA00029419150500000822
in the formula (VI), LSTM () represents that the fusion representation is input into a bidirectional long-time and short-time memory network, the important time sequence characteristics in the fusion representation are fully understood, and the enhanced representation is obtained
Figure BDA0002941915050000091
b. Will be provided with
Figure BDA0002941915050000092
Inputting the residual error network to obtain a characterization sequence
Figure BDA0002941915050000093
c. Outputting the sequence of the step a and the step b
Figure BDA0002941915050000094
Performing characterization fusion based on the connection operation, wherein the characterization fusion is shown as a formula (VII):
Figure BDA0002941915050000095
in the formula (VII), the reaction mixture is,
Figure BDA0002941915050000096
represents the fused enhanced signature sequence, Concat () represents the two sequences
Figure BDA0002941915050000097
And performing connection operation to finish the characterization fusion.
Example 4
The method for classifying time series based on multi-feature learning according to embodiment 1 is characterized in that:
in the step (3), classification is completed by utilizing a multilayer perceptron network, and classification interpretability is realized by utilizing an attention mechanism, and the specific implementation steps comprise:
F. performing representation learning again on the fusion sequence obtained after representation fusion and enhancement in the step (2) by utilizing a multilayer perceptron network;
G. utilizing a Softmax mechanism to obtain probability distribution of the current sequence to different categories, and realizing time sequence classification;
H. and acquiring attention scores of the classification model to different parts of the current sequence based on an attention mechanism, and realizing interpretability of the classification result.
In step F, the following steps are carried out: will be provided with
Figure BDA00029419150500000913
Inputting the multi-layer perceptron network to obtain a characterization sequence
Figure BDA0002941915050000098
In the step G, the following steps are carried out: using Softmax mechanism, for
Figure BDA0002941915050000099
Carrying out normalization operation to obtain the ith time sequence TiProbability P (L | T) of belonging to different data classesi) As shown in formula (VIII):
Figure BDA00029419150500000910
in the formula (VIII), L represents the number of data classes, TiMeans that the ith time sequence, 1o2tmax () function is normalized and will
Figure BDA00029419150500000911
Is mapped to [0,1 ] for the total sequence number in (2)]Within the range, T is obtainediProbability distributions belonging to different categories.
According to the calculated probability distribution, T is calculatediAnd dividing the data into data categories with the highest probability to realize time series classification.
The loss function K (θ) is shown by equation (IX):
Figure BDA00029419150500000912
in the formula (IX), N represents the total number of time series in the time series data set, KiRepresents TiA 0-1 vector corresponding to the L data classes;
all parameters of a Multi-representation Learning network Model (MLN) are trained by using an adam optimizer function.
And solving parameters in the multi-representation learning network model by using an optimization function, wherein the optimization function is an adam optimizer function in Tensflow.
Table 1 is a comparative model introduction table of the present invention;
TABLE 1
Figure BDA0002941915050000101
Table 2 is a classification accuracy comparison table of the present invention;
TABLE 2
Figure BDA0002941915050000102
The classification precision is compared with the classification precision of the internationally leading similar model, and the result of the table 2 shows that the classification precision of the multi-feature learning network model has obvious superiority.
The multi-feature learning network model can utilize an attention mechanism to identify important features in a time series. The attention visualization of four reference data sets is selected, and fig. 2(a) is a schematic view of the attention visualization of the method on an Adiac data set; FIG. 2(b) is a schematic view of the attention visualization on the DistalPhalanxtW data set by the method of the present embodiment; FIG. 2(c) is a schematic view of the attention visualization of the method of the present embodiment on the Coffee data set; FIG. 2(d) is a schematic view of the attention visualization of the method of the present embodiment on the Herring data set; in fig. 2(a) to 2(d), the abscissa represents the time index, and the ordinate represents the time series value. The darker the color in the time series, the higher the attention score. The time series with the same data category and the sequence segments with high attention scores are relatively concentrated. Thus, the model MLN gives a more definite class interpretability.
Example 6
A multi-feature learning based time series classification system, comprising:
a multi-feature information encoding module that performs feature encoding on a given time series based on a Piecewise Aggregation Approximation (PAA) policy and a Piecewise Linear Representation (PLR) policy;
the multi-feature fusion module is used for effectively fusing multi-feature sequences based on a residual error network and a bidirectional long-time and short-time memory network to realize feature fusion and enhancement;
and the classification module completes data classification and performs data visualization on classification results by utilizing a multilayer perceptron network and an attention mechanism, and provides a classification interpretability basis.
Example 7
A computer-readable storage medium having stored therein a plurality of instructions adapted to be loaded by a processor of a terminal device and to perform the multi-token learning based time series classification method according to any one of embodiments 1-6.
Example 8
A terminal device comprising a processor and a computer readable storage medium, the processor being configured to implement instructions; a computer readable storage medium for storing a plurality of instructions adapted to be loaded by a processor and to perform multi-token learning based time series classification as described in any of embodiments 1-6.
As will be appreciated by one skilled in the art, embodiments of the present invention may be provided as a method, system, or computer program product. Accordingly, the present invention may take the form of an entirely hardware embodiment, an entirely software embodiment or an embodiment combining software and hardware aspects. Furthermore, the present invention may take the form of a computer program product embodied on one or more computer-usable storage media (including, but not limited to, disk storage, CD-ROM, optical storage, and the like) having computer-usable program code embodied therein.
The present invention is described with reference to flowchart illustrations and/or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of the invention. It will be understood that each flow and/or block of the flow diagrams and/or block diagrams, and combinations of flows and/or blocks in the flow diagrams and/or block diagrams, can be implemented by computer program instructions. These computer program instructions may be provided to a processor of a general purpose computer, special purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, create means for implementing the functions specified in the flowchart flow or flows and/or block diagram block or blocks.
These computer program instructions may also be stored in a computer-readable memory that can direct a computer or other programmable data processing apparatus to function in a particular manner, such that the instructions stored in the computer-readable memory produce an article of manufacture including instruction means which implement the function specified in the flowchart flow or flows and/or block diagram block or blocks.
These computer program instructions may also be loaded onto a computer or other programmable data processing apparatus to cause a series of operational steps to be performed on the computer or other programmable apparatus to produce a computer implemented process such that the instructions which execute on the computer or other programmable apparatus provide steps for implementing the functions specified in the flowchart flow or flows and/or block diagram block or blocks.
The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention, and various modifications and changes may be made by those skilled in the art. Any modification, equivalent replacement, or improvement made within the spirit and principle of the present invention should be included in the protection scope of the present invention.
Although the embodiments of the present invention have been described with reference to the accompanying drawings, it is not intended to limit the scope of the present invention, and it should be understood by those skilled in the art that various modifications and variations can be made without inventive efforts by those skilled in the art based on the technical solution of the present invention.

Claims (10)

1. A time series classification method based on multi-feature learning is characterized by comprising the following steps: inputting the time sequence to be classified into a trained multi-representation learning network model to realize the classification of the time sequence, and specifically comprising the following steps:
(1) performing multi-feature coding on a given time sequence based on different time sequence characterization strategies;
(2) the representation fusion and enhancement are realized by utilizing a residual error network and a bidirectional long-time and short-time memory network;
(3) and finishing classification by utilizing a multi-layer perceptron network and realizing classification interpretability by utilizing an attention mechanism.
2. The method for classifying time series based on multi-feature learning as claimed in claim 1, wherein in step (1), the multi-feature coding is implemented by:
A. ith time sequence T of time sequence data set based on piecewise linear characterization strategyiCarrying out characteristic coding, wherein the time sequence data set comprises N time sequences, i is more than or equal to 1 and is less than or equal to N, and obtaining a first coding sequence
Figure FDA0002941915040000011
B. Approximation strategy for time series T based on segmentation and aggregationiPerforming characteristic coding to obtain a second coding sequence
Figure FDA0002941915040000012
C. Using one-dimensional time sequence convolution operation to obtain the first code sequence in step A
Figure FDA0002941915040000013
The second coding sequence obtained in step B
Figure FDA0002941915040000014
And time series TiAnd performing data characterization to obtain a basic characterization sequence.
3. The method for classifying time series based on multi-feature learning as claimed in claim 2, wherein in step (1), step C is implemented as follows:
the first coding sequence obtained in the step A is subjected to two layers of one-dimensional convolution neural networks
Figure FDA0002941915040000015
Obtained in step BSecond coding sequence
Figure FDA0002941915040000016
And time series TiPerforming one-dimensional time sequence convolution operation to finish time sequence coding representation; as shown in formula (I), formula (II) and formula (III):
Figure FDA0002941915040000017
Figure FDA0002941915040000018
Figure FDA0002941915040000019
in the formula (I), the formula (II) and the formula (III),
Figure FDA00029419150400000110
to represent
Figure FDA00029419150400000111
Data representation after two layers of one-dimensional convolution neural networks,
Figure FDA00029419150400000112
represents TiData representation after two layers of one-dimensional convolution neural networks,
Figure FDA00029419150400000113
to represent
Figure FDA00029419150400000114
And (4) data representation after two layers of one-dimensional convolution neural networks.
4. The method for classifying time series based on multi-feature learning as claimed in claim 1, wherein in step (1), in step (2), the specific implementation steps of feature fusion and enhancement comprise:
D. c, carrying out element-level multiplication operation on the basic characterization sequence obtained in the step C, and carrying out characterization fusion;
E. d, respectively inputting the characterization sequences obtained in the step D after the characterizations are fused into a residual error network and a bidirectional long-time and short-time memory network, and deeply understanding the sequence characteristics; and merging the characterization sequences processed by the residual error network and the bidirectional long-time and short-time memory network again, thereby realizing characterization fusion and enhancement.
5. The method for classifying time series based on multi-feature learning as claimed in claim 1, wherein in step (1), in step D, element-level multiplication is shown as formula (IV) and formula (v):
Figure FDA0002941915040000021
Figure FDA0002941915040000022
in the formulae (IV) and (V), tanh represents an activation function,
Figure FDA0002941915040000023
to represent
Figure FDA0002941915040000024
And
Figure FDA0002941915040000025
the enhanced characterization after the fusion is carried out,
Figure FDA0002941915040000026
to represent
Figure FDA0002941915040000027
And
Figure FDA0002941915040000028
enhanced characterization after fusion.
6. The method for classifying time series based on multi-feature learning as claimed in claim 1, wherein in step (1), step E, the specific implementation steps comprise:
a. for the above two characterization sequences
Figure FDA0002941915040000029
And
Figure FDA00029419150400000210
performing connection operation, inputting the connection operation into a bidirectional long-time memory network to obtain a characterization sequence
Figure FDA00029419150400000211
As shown in formula (VI):
Figure FDA00029419150400000212
in the formula (VI), the BilSTM () represents that the fusion representation is input into a bidirectional long-time and short-time memory network, the important time sequence characteristics in the fusion representation are fully understood, and the enhanced representation is obtained
Figure FDA00029419150400000213
b. Will be provided with
Figure FDA00029419150400000214
Inputting the residual error network to obtain a characterization sequence
Figure FDA00029419150400000215
c. Step a and step bStep b output sequence
Figure FDA00029419150400000216
Performing characterization fusion based on the connection operation, wherein the characterization fusion is shown as a formula (VII):
Figure FDA00029419150400000217
in the formula (VII), the reaction mixture is,
Figure FDA00029419150400000218
represents the fused enhanced signature sequence, Concat () represents the two sequences
Figure FDA00029419150400000219
And performing connection operation to finish the characterization fusion.
7. The method for classifying time series based on multi-feature learning as claimed in claim 1, wherein in step (1), in step (3), classification is completed by using a multi-layer perceptron network, and classification interpretability is realized by using an attention mechanism, and the specific implementation steps comprise:
F. performing representation learning again on the fusion sequence obtained after representation fusion and enhancement in the step (2) by utilizing a multilayer perceptron network;
G. utilizing a Softmax mechanism to obtain probability distribution of the current sequence to different categories, and realizing time sequence classification;
H. based on an attention mechanism, acquiring attention scores of the classification model to different parts of the current sequence to realize interpretability of the classification result;
more preferably, step F means: will be provided with
Figure FDA0002941915040000031
Inputting the multi-layer perceptron network to obtain a characterization sequence
Figure FDA0002941915040000032
More preferably, step G means:
using Softmax mechanism, for
Figure FDA0002941915040000033
Carrying out normalization operation to obtain the ith time sequence TiProbability P (L | T) of belonging to different data classesi) As shown in formula (VIII):
Figure FDA0002941915040000034
in the formula (VIII), L represents the number of data classes, TiRefers to the ith time sequence, and the softmax () function carries out normalization operation and will
Figure FDA0002941915040000035
Is mapped to [0,1 ] for the total sequence number in (2)]Within the range, T is obtainediProbability distributions belonging to different categories;
according to the calculated probability distribution, T is calculatediDividing the data into data categories with the highest probability to realize time series classification;
further preferably, the loss function K (θ) is represented by the formula (ix):
Figure FDA0002941915040000036
in the formula (IX), N represents the total number of time series in the time series data set, KiRepresents TiA 0-1 vector corresponding to the L data classes;
training all parameters of the multi-representation learning network model by using an adam optimizer function;
further preferably, the parameters in the multi-representation learning network model are solved by using an optimization function, wherein the optimization function is an adam optimizer function in Tensorflow.
8. A time series classification system based on multi-feature learning, which is used for implementing a time series classification method based on multi-feature learning as claimed in any one of claims 1-7, and comprises:
the multi-characteristic information coding module is used for coding the characteristics of a given time sequence based on a piecewise aggregation approximation strategy and a piecewise linear representation strategy;
the multi-feature fusion module is used for effectively fusing multi-feature sequences based on a residual error network and a bidirectional long-time and short-time memory network to realize feature fusion and enhancement;
and the classification module completes data classification and performs data visualization on classification results by utilizing a multilayer perceptron network and an attention mechanism, and provides a classification interpretability basis.
9. A computer-readable storage medium having stored thereon a plurality of instructions adapted to be loaded by a processor of a terminal device and to perform a method of multi-feature learning based time series classification according to any one of claims 1 to 7.
10. A terminal device comprising a processor and a computer readable storage medium, the processor being configured to implement instructions; a computer readable storage medium storing a plurality of instructions adapted to be loaded by a processor and to perform a multi-feature learning based time series classification as claimed in any one of claims 1 to 7.
CN202110180044.7A 2021-02-08 2021-02-08 Time series classification method, system, medium and device based on multi-representation learning Active CN112925822B (en)

Priority Applications (1)

Application Number Priority Date Filing Date Title
CN202110180044.7A CN112925822B (en) 2021-02-08 2021-02-08 Time series classification method, system, medium and device based on multi-representation learning

Applications Claiming Priority (1)

Application Number Priority Date Filing Date Title
CN202110180044.7A CN112925822B (en) 2021-02-08 2021-02-08 Time series classification method, system, medium and device based on multi-representation learning

Publications (2)

Publication Number Publication Date
CN112925822A true CN112925822A (en) 2021-06-08
CN112925822B CN112925822B (en) 2022-09-27

Family

ID=76171417

Family Applications (1)

Application Number Title Priority Date Filing Date
CN202110180044.7A Active CN112925822B (en) 2021-02-08 2021-02-08 Time series classification method, system, medium and device based on multi-representation learning

Country Status (1)

Country Link
CN (1) CN112925822B (en)

Cited By (1)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN114015825A (en) * 2021-11-09 2022-02-08 上海交通大学 Method for monitoring abnormal state of blast furnace heat load based on attention mechanism

Citations (4)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN109359698A (en) * 2018-10-30 2019-02-19 清华大学 Leakage loss recognition methods based on long Memory Neural Networks model in short-term
CN110188637A (en) * 2019-05-17 2019-08-30 西安电子科技大学 A kind of Activity recognition technical method based on deep learning
CN111443165A (en) * 2020-03-27 2020-07-24 华中科技大学 Odor identification method based on gas sensor and deep learning
CN112183582A (en) * 2020-09-07 2021-01-05 中国海洋大学 Multi-feature fusion underwater target identification method

Patent Citations (4)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN109359698A (en) * 2018-10-30 2019-02-19 清华大学 Leakage loss recognition methods based on long Memory Neural Networks model in short-term
CN110188637A (en) * 2019-05-17 2019-08-30 西安电子科技大学 A kind of Activity recognition technical method based on deep learning
CN111443165A (en) * 2020-03-27 2020-07-24 华中科技大学 Odor identification method based on gas sensor and deep learning
CN112183582A (en) * 2020-09-07 2021-01-05 中国海洋大学 Multi-feature fusion underwater target identification method

Non-Patent Citations (3)

* Cited by examiner, † Cited by third party
Title
WEI LUO: "Multi-resolution Representation for Streaming Time Series Retrieval", 《WORLD SCIENTIFIC》 *
丁一明: "时间序列流数据相似性搜索的研究", 《中国优秀博硕士学位论文全文数据库(硕士)信息科技辑》 *
胡宇鹏: "时间序列数据挖掘中的特征表示与分类方法的研究", 《中国优秀博硕士学位论文全文数据库(博士)信息科技辑》 *

Cited By (2)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN114015825A (en) * 2021-11-09 2022-02-08 上海交通大学 Method for monitoring abnormal state of blast furnace heat load based on attention mechanism
CN114015825B (en) * 2021-11-09 2022-12-06 上海交通大学 Method for monitoring abnormal state of blast furnace heat load based on attention mechanism

Also Published As

Publication number Publication date
CN112925822B (en) 2022-09-27

Similar Documents

Publication Publication Date Title
CN107273845B (en) Facial expression recognition method based on confidence region and multi-feature weighted fusion
CN108229280A (en) Time domain motion detection method and system, electronic equipment, computer storage media
Lin et al. Effective recognition of control chart patterns in autocorrelated data using a support vector machine based approach
CN107146237B (en) Target tracking method based on online state learning and estimation
CN109389171B (en) Medical image classification method based on multi-granularity convolution noise reduction automatic encoder technology
Huang et al. End-to-end continuous emotion recognition from video using 3D ConvLSTM networks
Karnati et al. Flepnet: feature level ensemble parallel network for facial expression recognition
Li et al. Efficient two-step networks for temporal action segmentation
Chen et al. DOF: A demand-oriented framework for image denoising
CN112925822B (en) Time series classification method, system, medium and device based on multi-representation learning
Fouad et al. Hyper-parameter optimization of convolutional neural network based on particle swarm optimization algorithm
Drass et al. Semantic segmentation with deep learning: detection of cracks at the cut edge of glass
CN113705715B (en) Time sequence classification method based on LSTM and multi-scale FCN
CN114492631A (en) Spatial attention calculation method based on channel attention
Jiao et al. Non-local duplicate pooling network for salient object detection
Su et al. Deep reinforcement learning-based progressive sequence saliency discovery network for mitosis detection in time-lapse phase-contrast microscopy images
Zhukov et al. Evaluation of Explanation Methods of AI--CNNs in Image Classification Tasks with Reference-based and No-reference Metrics
Jia et al. Arank: Toward specific model pruning via advantage rank for multiple salient objects detection
CN108564167A (en) The recognition methods of abnormal data among a kind of data set
Pei et al. FGO-Net: Feature and Gaussian Optimization Network for visual saliency prediction
Kachole et al. Bimodal SegNet: Instance Segmentation Fusing Events and RGB Frames for Robotic Grasping
Yun et al. A multi-target detection algorithm for infrared image based on retinex and LeNet5 neural network
Wu et al. Diffusion-based network for unsupervised landmark detection
Ongkittikul et al. Covolutional Multinomial Logistic Regression for Face Recognition
Cao et al. Improved YOLOv3 model based on ResNeXt for target detection

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