CN-121980309-A - Self-adaptive enhancement dynamic diagram comparison learning multivariate time sequence classification method

Abstract

The invention discloses a multi-element time sequence classification method for self-adaptive enhancement dynamic graph contrast learning, which is oriented to multi-element time sequence classification tasks under a few-labeling condition and comprises the following steps of 1) self-adaptive enhancement strategy, 2) dynamic graph contrast learning and 3) two-stage combined training. The method combines trend-period decomposition and random substitution-based statistical significance test to judge the time sequence dependence strength in the preprocessing stage, generates an enhanced recommendation strategy for matching data characteristics, introduces a dynamic diagram contrast characterization learning and two-stage training mechanism, and improves classification performance and model stability while fully utilizing unlabeled data.

Inventors

• WANG HUIJIAO
• WU ZINING
• YE FANG
• PENG ZIYAO

Assignees

• 桂林电子科技大学

Dates

Publication Date

20260505

Application Date

20260119

Claims (1)

1. 1. A self-adaptive enhancement dynamic diagram contrast learning multi-element time sequence classification method is characterized by comprising the following steps: 1) Before contrast learning training, generating a data set-level enhancement recommendation set based on a trend-periodic structure of a training set, and carrying out self-adaptive removal on an enhancement operator which possibly destroys semantic consistency by combining time sequence dependency significance to obtain a final enhancement operator pool for constructing a subsequent weak/strong disturbance view, wherein the method comprises the following steps of: Given training set Wherein the first The univariate time sequence corresponding to each variable dimension is recorded as For a pair of STL decomposition is carried out to obtain a trend term, a period term and a residual term, as shown in a formula (1): (1), Wherein +represents at each point in time The above-mentioned item-by-item additions, As a component of the trend the composition, As a component of the period of time, Based on samples in training set 、 、 The statistical features of the data set are used for constructing a data set level structure image for representing trend intensity, periodic features and noise level of the data set, and Top- - Individual enhancement operators form enhanced recommendation sets Randomly extracting from training set Samples of each sequence The Lag-1 autocorrelation statistic is calculated as shown in formula (2): (2), Wherein the method comprises the steps of Is the pearson correlation coefficient, pair Proceeding with Sub-random permutation Construction of zero distribution And calculates a normalized saliency score as shown in equation (3): (3), Wherein the method comprises the steps of Is that Is set in the standard deviation of (2), The maximum significance of the extracted sample in the variable dimension is obtained, and the average value is obtained to obtain the data set level discriminant as shown in a formula (4): (4), setting a discrimination threshold When (when) When the time is determined to be a strong time sequence dependent data set, the recommendation set is selected In the method, enhancement operators which possibly destroy time sequence or key dependency relationship are removed in a self-adaptive manner to obtain a filtered recommendation set When (1) When the time sequence dependency strength of the data set is judged to be lower than the threshold value, the data set belongs to the data set which is not strongly time sequence dependent, and the recommendation set is unchanged Dividing the enhancement operator into a weak enhancement operator pool and a strong enhancement operator pool, wherein the weak enhancement operator pool initial set is set as The initial set of the strong enhancement pool is Recommendation set after screening Merging and updating to obtain a final operator pool And And uniformly representing any enhancement operator as parameter transformation Wherein A specific enhancement operator is represented as such, In order to enhance the parameters of the device, Sampling the enhancement operators in a uniform random manner from the weak enhancement operator pool and the strong enhancement operator pool, respectively, for each input sample during the training phase, as shown in equation (5): (5), Wherein, the Expressed in a collection Uniform sampling on the same, weak enhancement operator based on random sampling Strong enhancement operator For the original sample And respectively applying weak/strong disturbance, and respectively obtaining weak/strong disturbance views as shown in a formula (6): (6), Wherein the method comprises the steps of And Semantically consistent representation, weakly enhanced view is The strong enhancement view is ; 2) Dynamic graph contrast learning based on the weakly enhanced view obtained in step 1) And strong enhancement view Order-making Will be The input feature extraction network obtains a structured time sequence representation, adopts a time sequence comparison objective function to carry out unsupervised constraint, and specifically comprises the following steps: 2.1 Multi-scale time sequence feature extraction and time sequence dependency modeling, namely extracting time sequence features in a time dimension by adopting multi-scale expansion convolution, and setting an expansion rate set as For each expansion rate The outputs of the branches are spliced in the channel dimension, and the multi-scale characteristic representation is obtained as shown in a formula (7): (7), Wherein the method comprises the steps of Features that the branches with different expansion convolution rates are spliced according to channels, In order for the time convolution kernel to be of a size, To output the feature dimension for the layer Modeling long-distance time sequence dependence acquisition of input lightweight multi-head attention module ; 2.2 Group graph construction and intra-group propagation update-partitioning the time dimension into Group when Cannot be covered by Zero filling is carried out on the time dimension during integer division to obtain Satisfies the following conditions And make each group length as First, the The time slices corresponding to the groups are shown in formula (8): (8), Obtaining the time sequence data slice of the group as according to the formula (8) The set of high-dimensional feature slices is In intra-group graph structure learning, for each group index Introducing a learnable source/target dual embedding And (3) with Generating intra-group variable association scoring matrix And is opposite to Performing sparsification to preserve strong-correlation connection to obtain intra-group adjacency matrix For a pair of Symmetric normalization is carried out to obtain The update phase is propagated within the group, As a node characteristic input, at Performing feature propagation and linear transformation under constraint to obtain update features in the group Wherein In order for the parameters to be able to be learned, And then, introducing a coordinate graph attention mechanism to jointly enhance the variable dimension and the time dimension to obtain And finally, splicing all groups of outputs along the time dimension to obtain a whole section of structural characteristics as shown in a formula (9): (9), 2.3 Time sequence contrast objective function construction, namely average aggregation is carried out on node dimensions, and any view is obtained First, the The representation of the individual time steps is defined as shown in equation (10): (10), Each time step vector obtained by the formula (10) Arranged along a time dimension, a sequence representation matrix is constructed Random sampling time anchor point Wherein To predict the number of steps, in weak enhancement sequence The previous segment is used as context input, and the input time sequence is from the attention context modeling network The self-attention aggregation history information is adopted, and the obtained context expression is shown in a formula (11): (11), Followed by each prediction step for weak enhancement sequences Using linear pre-measuring heads Generating future representation predictions Directly obtaining a strongly enhanced view at a future time according to equation (10) True characteristic representation of (2) Taking the target as a prediction target, adopting a negative sample structure in the batch to construct contrast constraint, and setting the batch size as The similarity matrix definition is shown in equation (12): (12), Wherein the method comprises the steps of For the vector inner product, the timing contrast penalty definition is as shown in equation (13): (13), Adopting time sequence contrast loss, forming positive sample pairs by the representations of the same sample aligned by time offset under the weak enhancement view and the strong enhancement view, forming negative sample pairs by the corresponding representations of other samples in the same batch, and performing contrast optimization to restrict the time sequence consistency and class discriminant of characterization under the disturbance view; 3) Two-stage combined training, namely constructing a non-labeling sample set under the condition of less labeling And labeling sample sets The method adopts a two-stage combined training process of 'non-labeling pretraining-labeling fine tuning', wherein the non-labeling training stage is that Performing self-supervision training on each sample Generating weak enhancement views by adopting enhancement strategies determined in a mode of step 1) respectively And strong enhancement view Respectively obtaining weak enhancement representations after respectively carrying out the dynamic diagram modeling flow of the step 2) And strong enhancement representation Will be And Input contrast learning process to calculate time sequence contrast loss And according to contrast loss For model parameters Optimization is performed in which A learning parameter set of the learning module is compared with the feature extraction and dynamic graph modeling flow; In the labeling fine tuning stage, model parameters obtained by pre-training are used as initialization parameters, and a sample set is labeled Training classification tasks, and generating weak enhancement views of the samples based on enhancement strategies And strong enhancement view For any input sequence Model output belongs to The predictive posterior probability of a class is noted as During fine tuning training, the weak enhancement view is displayed Input model to obtain class prediction probability The cross entropy is used to obtain the classification loss as shown in equation (14): (14), At the same time, according to the weak enhancement view And strong enhancement view Respectively corresponding to weak enhancement representations And strong enhancement representation Calculated loss of timing contrast Introducing a fine tuning objective function as a regularization term to obtain a joint optimization objective, as shown in formula (15): (15), Wherein the method comprises the steps of For comparing the regular intensity coefficients, the method is used for adjusting the influence of comparison constraint on the process of less labeling and fine adjustment, after model training is completed, an enhanced view is not constructed any more in a test stage, and an original sequence is directly processed Inputting the feature extraction and dynamic graph modeling flow, and outputting classification results through classification branches, wherein the prediction category is as follows Wherein For the category number, the predictive label and the training set real label And comparing and calculating the classification accuracy.

Description

Self-adaptive enhancement dynamic diagram comparison learning multivariate time sequence classification method Technical Field The invention relates to a time sequence data analysis and pattern recognition technology, in particular to a time sequence classification technology oriented to a few labeling condition, and particularly relates to a semi-supervised time sequence classification technology which fuses an adaptive data enhancement strategy and a contrast learning mechanism and combines dynamic diagram structural modeling to perform characterization learning and classification decision, in particular to a self-adaptive enhancement dynamic diagram contrast learning multi-element time sequence classification method. Background The multivariate time series classification is used for learning the distinguishing characteristics from the multi-channel continuous observation data and outputting the classification result. In practical application, labeling samples are often high in acquisition cost and limited in quantity, so that a model is difficult to fully learn and stably judge and express under a few labeling conditions, meanwhile, the correlation among variables in a multi-element sequence has dynamic change characteristics, and if an effective modeling mechanism is lacking, the problems of insufficient key dependency and characterization and unstable complex pattern recognition are easy to occur. On the other hand, time sequence representation learning based on contrast learning generally relies on data enhancement to construct training views, but existing random or fixed enhancement strategies are easy to be mismatched with data structure characteristics, can destroy sample semantic consistency and cause performance fluctuation, and accordingly affect generalization capability and robustness of a model. To enhance variable dependent modeling capability, graph neural networks are introduced into the timing tasks and form "graph-timing" fusion modeling routes. For example GRAPHWAVENET learns the correlations among variables through an adaptive adjacency matrix and integrates time sequence modeling, a dynamic graph modeling method based on a neural ordinary differential equation (Neural ODE) is used for describing the evolution of the structure and the state under continuous time, and a graph-annotation-force network type method improves the effectiveness of node aggregation and feature propagation through attention weights. The above research shows that the composition mode that the variables are regarded as nodes and the dependence are regarded as edges is helpful for describing the internal association of the multi-element sequence, but under the condition of few labels, the structure learning and the characteristic characterization of the dynamic graph are easily affected by the insufficient supervision signals, and the stable description of the structure changing along with the time and the generalization capability of the cross samples are difficult to be considered. In the context of label scarcity, self-supervised learning (SSL) is becoming an important technological route, where contrast learning draws in pairs of positive and negative samples by constructing different enhanced views of the same sequence to learn migratable characterizations with unlabeled data. Typical hierarchical contrast learning frameworks (e.g., TS2 Vec) are capable of capturing timing features at different semantic levels, and graph contrast learning frameworks (e.g., GCA) further emphasize that critical structures and attributes should be preserved as much as possible during enhancement to maintain the effectiveness of contrast objectives. Meanwhile, relevant review and evaluation work indicates that the effect of contrast learning in time sequence/graph time sequence tasks is highly sensitive to a data enhancement strategy, namely, semantic information such as trends, periods or local forms and the like can be damaged by random or fixed enhancement under different data structures, so that views are inconsistent, contrast constraint is weakened, performance fluctuation is caused, robustness is reduced, and meanwhile, the enhancement strategy needs to be compatible with disturbance of a topology level and a feature level and is matched with structural characteristics of the data. Disclosure of Invention The invention aims to solve the problems that in the prior art, multivariate time sequence classification is insufficient in supervision information under a few labeling conditions, dependency relations among variables are difficult to effectively describe along with time change, and generalization and robustness are insufficient due to fixed or random enhancement of easily damaged sequence semantic consistency, and provides a self-adaptive enhancement dynamic graph contrast learning multivariate time sequence classification method. The method combines trend-period decomposition and random substitution-based statistical s