Comparative learning-based multivariate time series data enhancement method

Abstract

The invention discloses a multivariate time sequence data enhancement method based on comparative learning, which comprises the following steps of: acquiring an original time sequence, a business topology and a covariable, carrying out peak valley and accumulation and mutation detection, and constructing atlas data; compliance verification is carried out on the enhancement operator; generating parameters by using a Mama three-domain converter under a graph mask, and projecting and rejecting sampling to obtain an enhanced view; performing multi-granularity InfoNCE on the view, and superposing image consistency, prediction consistency and circulation consistency loss; performing multi-step quantile prediction and uncertainty estimation by using an improved Chronos model, and selecting an enhancement type and intensity according to an upper confidence limit; and generating an enhanced proof list, carrying out distillation deployment, carrying out on-line monitoring on drifting, triggering degradation and rollback by a threshold value, and carrying out retraining. The method improves representation generalization and robustness, ensures compliance auditing, and is suitable for scenes of industrial Internet of Things, financial risk control and the like.

Description

Technical Field

[0001] This invention relates to the fields of time series data mining and machine learning technology, and in particular to a multivariate time series data augmentation method based on contrastive learning. Background Technology

[0002] With the rapid growth in demand for multivariate time-series self-supervised representation and robust modeling in scenarios such as Industrial Internet of Things (IIoT), energy load, financial risk control, and AIOps, research on contrastive learning and data augmentation has attracted attention. Existing methods often generate views using general operators such as jitter, time warp, interpolation resampling, frequency band perturbation, or cross-channel mixing, and are trained on InfoNCE-type targets before being validated offline using conventional prediction models. However, these methods generally suffer from the following problems in practical applications: Inconsistent sampling rates across multiple time sources and uncontrollable cross-channel time differences, coupled with the common use of global interpolation and fixed slicing that ignores business topology and covariates, lead to event boundary and semantic mismatches. Enhancement operators lack Markov equivalence preservation at the causal graph level, and time warping and amplitude affine transformations easily alter the skeleton or V-structure, incorporating "pseudo-invariance." Frequency domain perturbations are mostly limited to energy or bandwidth control, making it difficult to stably maintain the dominant frequency and harmonic ratio, disrupting periodic / mechanistic patterns. Physical invariants and business conservation (dimension, consistency, monotonic / conservation constraints) lack system verification and parameter projection, resulting in high compliance risks. View generation is primarily single-domain or serial, lacking coordination between the time domain, frequency domain, and event domain, as well as graph mask gating, making parameter strength dependent on experience and difficult to adapt. Contrastive learning is mostly at the single-granularity sample level, lacking event-level, fragment-level, and... Figure 1 The joint constraints such as consistency, forward prediction consistency, and cyclic consistency represent insufficient robustness; the enhancement optimization relies on human experience or single-round indicators and lacks a UCB strategy driven by quantile prediction and uncertainty assessment; the training-deployment link lacks "enhancement proof sheet" and online drift monitoring, making it difficult to trace.

[0003] Therefore, how to provide a multivariate time-series data augmentation method based on contrastive learning is a problem that urgently needs to be solved by those skilled in the art. Summary of the Invention

[0004] One objective of this invention is to propose a multivariate time-series data augmentation method based on contrastive learning. This invention integrates graph construction, causal graph verification, physical and spectral constraints, multi-domain augmentation generation, contrastive learning representation, improved Chronos model time-series prediction evaluation and policy feedback mechanism. It describes in detail the entire process of generating high-quality augmented views under the selection of augmentation operators, intensity control and compliance screening. It has the advantages of controllable augmentation process, closed-loop evaluation feedback, and strong adaptability to heterogeneous time-series data.

[0005] A multivariate time-series data augmentation method based on contrastive learning according to an embodiment of the present invention includes the following steps: The original time series, business topology and covariates are collected, and the events are segmented by peak and valley detection, cumulative sum and mutation detection. The graph data is constructed by combining differentiable causal discovery, time-delay Granger and short-time Fourier transform. Based on the graph data, compliance checks are performed on the preset enhancement operator to ensure that the causal graph Markov equivalence is preserved, the physical invariant constraints are satisfied, and the main frequency and harmonic ratio remain unchanged, and compliance space data is output. Within the range of operators and intensities defined by the compliant spatial data, a three-domain cooperative transformer based on Mamba is invoked to jointly generate enhancement parameters in the time domain, frequency domain, and event domain under graph mask gating, and to perform compliant projection and rejection sampling to produce enhanced view data. Multi-granularity InfoNCE comparative learning is performed on the enhanced view data, and the data is overlaid. Figure 1 The joint loss of consistent prediction, consistent forward prediction, and consistent cycle is used to obtain characterization data; Using enhanced view data as the evaluation object, the improved Chronos model is called to perform multi-step quantile prediction and uncertainty estimation. Combined with the upper confidence bound strategy, the enhancement type and intensity are determined, and enhancement strategy data is generated. Based on the representation data and enhancement strategy data, an enhanced proof is generated and knowledge distillation deployment is completed. Online monitoring is conducted to detect drift and compliance threshold triggers for degradation, rollback, and retraining.

[0006] Optionally, the process of collecting original time series, service topology, and covariates, performing event segmentation through peak-valley detection and cumulative sum-abrupt detection, and constructing spectral data by combining differentiable causal discovery, time-delay Granger calculus, and short-time Fourier transform specifically includes: The original time series is a multivariate time series. A unified time axis is established with the smallest time granularity. The gaps in the original time series with different sampling rates are filled by interpolation according to the linear ratio of adjacent sampling points, and missing markers are written at the filled positions. The business topology is a data structure of objects and connections corresponding to the original time series, including entity identifiers, entity types, connection edges between entities and edge attributes. Based on the entity identifiers, the channels and sensor sources in the original time series are mapped and verified to form an entity-channel correspondence table. The covariates are exogenous variables that affect the original time series and originate from external systems or public data sources. The covariates are synchronized with the original time series on the time axis. Gaps are filled by linear interpolation of adjacent sampling points and missing markers are added. The original time series is subjected to joint event segmentation by peak and valley detection and cumulative and abrupt change detection. Peak and valley detection finds local maxima and minima within a sliding window and calculates historical noise levels by segmentation according to covariates to set trigger thresholds. Cumulative and abrupt change detection progressively accumulates the difference from the baseline and compares it with the trigger threshold. If the difference exceeds the trigger threshold, it is marked as an abrupt change point. Continuous abrupt change points are paired with adjacent peaks and valleys to determine the start and end of the event and generate the event segmentation results. Based on the event segmentation results, differentiable causal discovery is performed to obtain the initial results of directed dependencies between variables. The initial results are verified by the time-delay Granger test and the main lag order is added. At the same time, covariates are added as exogenous nodes, and a causal graph containing directed edges and lag information is output. Construct a topology graph containing nodes, edges, and edge attributes based on the entity-channel correspondence table; The original time series is segmented based on the event segmentation results, and after windowing, a short-time Fourier transform is performed on each segment, which are then aggregated to form a spectrogram. The causal graph, topological graph, and spectrogram, along with the event segmentation results, are combined into spectral data.

[0007] Optionally, based on the spectral data, compliance checks are performed on the preset enhancement operator to ensure causal graph Markov equivalence preservation, physical invariant constraint satisfaction, and invariance of dominant frequency and harmonic ratio, outputting compliance space data, specifically including: Based on the spectral data, preset enhancement operators are determined, and candidate parameter ranges and step granularity are set for each type of enhancement operator; The preset enhancement operators are checked for causal graph Markov equivalence preservation. Based on the directed edge relations and conditional independence relations of the causal graph, the enhancement is limited to the channel set that is adjacent or in the same Markov blanket of the causal graph. Adding or deleting paths that would change the conditional independence relations is prohibited. Cross-channel mixing is only allowed between channel pairs that are directly connected in the causal graph. Time warp and amplitude affine are subject to consistent rules between parent and child channels to avoid changes in directional relations. Those that do not meet the requirements are judged as non-compliant. The physical invariant constraints of the preset enhancement operators are checked. Based on the entity connections and static constraints in the topology diagram, it is checked whether the entity relationships of each channel are maintained after enhancement, and whether the dimensional and conservation relationships given by the business invariant list are destroyed. Parameters that touch the constraint boundary are truncated according to the boundary value and the reason for truncation is recorded. Candidates that violate the constraints are eliminated. The preset enhancement operator is checked to ensure that the main frequency and harmonic ratio remain unchanged. Based on the main frequency position, main peak bandwidth and harmonic ratio of each order in the spectrum at the event segment level, the frequency domain energy adjustment is limited to the allowed frequency band. It is prohibited to introduce new main peaks or move recorded main peaks beyond the tolerance range. The allowed frequency bands for bandpass resampling are given by the set of frequency bands marked on the spectrum. Candidates that exceed the allowed frequency bands are eliminated. Finally, the list of enhanced operators obtained after screening, the corresponding upper and lower limits of parameters, and the list of allowed parameters are combined and output as compliant spatial data.

[0008] Optionally, within the operator and intensity range defined by the compliant spatial data, a Mamba-based three-domain cooperative transformer is invoked to jointly generate enhancement parameters in the time, frequency, and event domains under graph mask gating, and compliant projection and rejection sampling are performed to produce enhanced view data. Specifically, this includes: A three-domain cooperative transformer based on Mamba is established, wherein the three-domain cooperative transformer is a parameterized data transformation module implemented using the selective state-space network Mamba; Within the upper and lower limits and step granularity of the parameters given in the compliant spatial data, candidate parameter sequences are generated for time warp, amplitude affine, phase perturbation, bandpass resampling, controlled cross-channel mixing and missing replay, and graph mask gating is generated. The graph mask gating is to mark channel pairs as allowed or prohibited based on the causal graph adjacency relationship. Enhancements are implemented in a fixed order: within the allowed event segments, time warp is applied first, followed by amplitude affine projection; compliant projection is performed on out-of-bounds values, whereby values ​​exceeding the upper or lower limit are replaced with the corresponding boundary values ​​and the replacement position is recorded; phase perturbation and bandpass resampling are applied sequentially within the allowed frequency bands; candidates not in the allowed frequency bands are directly discarded; and operations involving cross-channel effects are masked and prohibited using a map mask; controlled cross-channel mixing is applied within the allowed cross-channel mapping and event segment range; and non-adjacent channel pairs in the causal graph are not processed. Rejection sampling is implemented, which involves verifying the consistency of the results after enhancement. Samples that have event boundaries exceeding the limits, cross-channel relationships and map masks conflicting, or main peaks moving out of the allowed frequency band are marked as rejected and removed. Enhancement results are synthesized in the order of "time domain → frequency domain → event domain", and enhancement view data is generated and written with operator identifier, final parameter value, active channel, event segment and frequency band number.

[0009] Optionally, the step involves performing multi-granularity InfoNCE comparative learning on the enhanced view data and overlaying it. Figure 1 The joint loss of consistent prediction, prospective prediction, and cyclic consistency yields characterization data, specifically including: Load the enhanced view data and the corresponding original time series, form a training batch according to the sample identifier and event segment number, input each sequence into the shared time series encoder to obtain the intermediate representation, and then map it into a normalized vector through the projection head; Using the original time series and augmented view data of the same sample and the same time span as positive sample pairs, and different samples or different event segments within the same batch as negative sample pairs, multi-granularity InfoNCE contrastive learning is implemented for the augmented view data. The multi-granularity includes sample level pairing of whole sequences, event level pairing by event segment number, and segment level pairing by sliding window. Cosine similarity is calculated for each granularity and temperature scaling is applied. The contrastive losses of sample level, event level, and segment level are accumulated according to preset weights. Superimposed on comparative learning Figure 1 To achieve joint loss, based on the causal graph, the induced relationship results are constructed from the channel correlation statistics of the encoder output, and compared with the adjacent relationships in the causal graph item by item. The deviation is included in the penalty term. The look-ahead prediction consistent joint loss is superimposed, and a look-ahead head is set. The look-ahead head is a lightweight prediction module that outputs the statistics and interval values ​​of the next time period for the original time series and the enhanced view data respectively, and compares the differences under the same statistical caliber, and includes the differences in the penalty term. The system employs a superimposed cyclic consistent joint loss and sets up an inverse mapping head, which is a lightweight restoration module. The original morphological sequence is input and output as the augmented view data. The difference between the original morphological sequence and the original temporal sequence at the corresponding time point is included in the penalty term. In one forward calculation, the contrastive loss and the joint loss are accumulated to form the total loss. In the backward update, only the parameters of the temporal encoder, the projection head, the look-ahead head, and the inverse mapping head are updated until convergence. After training is completed, the representation data is output.

[0010] Optionally, the step of using enhanced view data as the evaluation object, calling an improved Chronos model to perform multi-step quantile prediction and uncertainty estimation, and combining an upper confidence bound strategy to determine the enhancement type and intensity, and generating enhancement strategy data, specifically includes: The improved Chronos model includes a discretized encoder, a language modeling decoder, and an inverse quantization demultiplexer; Using compliance space data as a candidate pool, evaluation combinations are generated according to enhancement type, intensity, allowed event segments and frequency bands. The original time series segments and enhanced view segments are segmented from the historical window corresponding to each evaluation combination and the covariates are aligned. The data is then input into a discretization encoder for event adaptive discretization. The distribution of each channel is statistically analyzed in segments according to the event segmentation results. Adaptive bucketing is established within the segments to map continuous values ​​to symbols. Conditional cue vectors are generated for covariates and injected with the symbols aligned with the symbol time axis. Values ​​that touch the compliance boundary are written into boundary symbols. Missing symbols are written into missing test locations and the length of continuous missing values ​​is recorded. The original time series symbols, enhanced view symbols and covariate cue vectors are output. The original time-series symbols, enhanced view symbols, and covariate cues are concatenated in time and fed into the language modeling decoder. The autoregressive method predicts the future symbol sequence. During training, the cross-entropy loss is used to supervise the next symbol. Unenhanced control samples and enhanced samples are simultaneously included in the same batch. After calculating the loss separately, the parameters are updated with equal weight. During the decoding stage, temperature sampling and kernel sampling are performed within the upper and lower limits of the parameters in the compliant space data to generate multiple future symbol trajectories. At the same time, non-compliant future symbol trajectories are pruned according to the allowed event segments and frequency bands, and the symbol trajectory data is output. The symbol trajectory data is input into the dequantization restorer, which gradually restores the real-value trajectory according to the reverse rule of the discretization mapping. At each time step, multiple real-value trajectories are sorted to obtain the quantile values ​​of the preset quantile set, and the quantile bandwidth and trajectory variance are calculated as uncertainty indicators. The quantile prediction is aligned with the actual observation of the corresponding time window, and the quantile loss is gradually accumulated according to the quantile point as the prediction cost. The amount of reduction of quantile loss relative to the unenhanced control sample is used as the prediction improvement amount, and the uncertainty indicators are recorded together as the evaluation result. The evaluation results employ an upper confidence bound strategy for decision-making, using the predicted improvement amount as the benefit item and superimposing quantile bandwidth and trajectory variance as risk penalties to obtain a single score. Within the scope of compliance space data, the enhancement type and intensity with the highest score are selected and organized into enhancement strategy data.

[0011] Optionally, the step of generating enhanced proof certificates based on representation data and enhancement strategy data, completing knowledge distillation deployment, and online monitoring of drift and compliance threshold triggering degradation, rollback, and retraining specifically includes: Based on representational data and enhancement strategy data, construct enhanced proof single data; The enhanced proof data is used for archiving of enhancement schemes during the deployment process. During the deployment system operation phase, real-time data is continuously collected and compared with the representation content in the enhanced proof. If the representation result of the real-time data is detected to deviate from the corresponding representation content, and the deviation exceeds the set drift threshold, the current enhancement strategy is considered to be invalid. When the drift threshold is triggered, a compliance response mechanism is initiated based on the status of the enhancement strategy. The compliance response mechanism includes: performing a downgrade process, which involves pausing the current enhancement strategy and switching to a standard enhancement template that has been verified to be stable during multiple rounds of training; performing a rollback process, which involves switching the current data processing flow to a path without enhancement or the last effective enhancement scheme; and performing a reconstruction process, which uses the current drift data as input to regenerate new map data, compliance space data, and enhancement strategy data.

[0012] The beneficial effects of this invention are: This invention enhances the ability to analyze the structure and dependencies of original time-series data by constructing multi-source heterogeneous spectral data and comprehensively incorporating differentiable causal relationships, business topology, and spectral features. It establishes a compliant augmentation space by performing causal graph Markov equivalence preservation, physical invariant constraints, and frequency domain consistency checks on augmentation operators, effectively ensuring the rationality and credibility of augmentation operations. Using a Mamba-based three-domain cooperative transformer, multi-domain augmentation parameters are jointly generated under graph mask gating, achieving targeted and diverse augmentation views. Multi-granularity InfoNCE contrastive learning is introduced, and multiple structural consistency losses are superimposed, significantly enhancing the model's feature robustness under different augmentation views. An improved Chronos model is used for multi-step quantile prediction and uncertainty estimation, combined with an upper confidence bound strategy to select the optimal augmentation type and intensity, achieving quantitative evaluation and feedback optimization of augmentation strategies. A closed-loop monitoring mechanism is formed through augmentation proofs and knowledge distillation deployment, possessing continuous adaptive and dynamic evolution capabilities for augmentation strategies, significantly improving the effectiveness and practicality of multivariate time-series data augmentation. Attached Figure Description

[0013] The accompanying drawings are provided to further illustrate the invention and form part of the specification. They are used in conjunction with embodiments of the invention to explain the invention and do not constitute a limitation thereof. In the drawings: Figure 1 This is a flowchart of a multivariate time-series data augmentation method based on contrastive learning proposed in this invention; Figure 2 This is a schematic diagram of a multivariate time-series data augmentation method based on contrastive learning proposed in this invention; Figure 3 This is a framework diagram of the improved Chronos model in a multivariate time-series data augmentation method based on contrastive learning proposed in this invention. Detailed Implementation

[0014] The present invention will now be described in further detail with reference to the accompanying drawings. These drawings are simplified schematic diagrams, illustrating only the basic structure of the invention, and therefore only show the components relevant to the invention.

[0015] refer to Figure 1-3 A multivariate time-series data augmentation method based on contrastive learning includes the following steps: The original time series, business topology and covariates are collected, and the events are segmented by peak and valley detection, cumulative sum and mutation detection. The graph data is constructed by combining differentiable causal discovery, time-delay Granger and short-time Fourier transform. Based on the graph data, compliance checks are performed on the preset enhancement operator to ensure that the causal graph Markov equivalence is preserved, the physical invariant constraints are satisfied, and the main frequency and harmonic ratio remain unchanged, and compliance space data is output. Within the range of operators and intensities defined by the compliant spatial data, a three-domain cooperative transformer based on Mamba is invoked to jointly generate enhancement parameters in the time domain, frequency domain, and event domain under graph mask gating, and to perform compliant projection and rejection sampling to produce enhanced view data. Multi-granularity InfoNCE comparative learning is performed on the enhanced view data, and the data is overlaid. Figure 1 The joint loss of consistent prediction, consistent forward prediction, and consistent cycle is used to obtain characterization data; Using enhanced view data as the evaluation object, the improved Chronos model is called to perform multi-step quantile prediction and uncertainty estimation. Combined with the upper confidence bound strategy, the enhancement type and intensity are determined, and enhancement strategy data is generated. Based on the representation data and enhancement strategy data, an enhanced proof is generated and knowledge distillation deployment is completed. Online monitoring is conducted to detect drift and compliance threshold triggers for degradation, rollback, and retraining.

[0016] In this embodiment, the process of collecting original time series, service topology, and covariates, performing event segmentation through peak-valley detection and cumulative sum-abrupt detection, and constructing spectral data by combining differentiable causal discovery, time-delay Granger calculus, and short-time Fourier transform specifically includes: The original time series is a multivariate time series obtained by recording the values ​​and timestamps of each sensor and service indicator at a fixed or variable sampling interval. A unified time axis is established with the smallest time granularity. The original time series with different sampling rates are interpolated to fill the gaps according to the linear ratio of adjacent sampling points, and missing markers are written at the filled positions. The business topology is a data structure of objects and connections corresponding to the original time series. It is used to present the static associations and directional constraints between devices, sites, regions or process links. It includes entity identifiers, entity types, connection edges between entities and edge attributes. Based on the entity identifiers, the channels and sensor sources in the original time series are mapped and verified to form an entity-channel correspondence table. The covariates are exogenous variables that affect the original time series and originate from external systems or public data sources. The covariates are synchronized with the original time series on the time axis. Gaps are filled by linear interpolation of adjacent sampling points and missing markers are added. The original time series is subjected to joint event segmentation by peak and valley detection and cumulative and abrupt change detection. Peak and valley detection finds local maxima and minima within a sliding window and calculates historical noise levels by segmentation according to covariates to set trigger thresholds. Cumulative and abrupt change detection progressively accumulates the difference from the baseline and compares it with the trigger threshold. If the difference exceeds the trigger threshold, it is marked as an abrupt change point. Continuous abrupt change points are paired with adjacent peaks and valleys to determine the start and end of the event and generate the event segmentation results. Based on the event segmentation results, differentiable causal discovery is performed to obtain the initial results of directed dependencies between variables. The initial results are verified by the time-delay Granger test and the main lag order is added. The time-delay Granger test is used to compare whether the prediction error decreases significantly before and after the introduction of candidate lagged variables to determine whether there is a causal effect. At the same time, covariates are added as exogenous nodes. Exogenous nodes only output influence to variables within the system and are not affected by reverse effects. The output is a causal graph containing directed edges and lag information. Construct a topology graph containing nodes, edges, and edge attributes based on the entity-channel correspondence table; The original time series is segmented based on the event segmentation results, and after windowing, a short-time Fourier transform is performed on each segment, which are then aggregated to form a spectrogram. The causal graph, topological graph, and spectrogram, along with the event segmentation results, are combined into spectral data.

[0017] This implementation method achieves unified structural modeling of the original time series, business topology, and covariates by constructing event segmentation results and multidimensional graph data, effectively enhancing the correlation and causal interpretability among multi-source data, and providing a structured reference for subsequent enhancement operations.

[0018] In this embodiment, the compliance verification of pre-defined enhancement operators based on spectral data, including causal graph Markov equivalence preservation, physical invariant constraint satisfaction, and invariance of dominant frequency and harmonic ratio, and outputting compliance space data, specifically includes: Based on the spectral data, preset enhancement operators are determined. The preset enhancement operators are a set of operations that perform controlled perturbations on the time series, including at least one of time warp, amplitude affine, controlled cross-channel mixing, phase perturbation, bandpass resampling and missing replay. Candidate parameter ranges and step granularity are set for each type of enhancement operator. The preset enhancement operators are checked for causal graph Markov equivalence preservation. Based on the directed edge relations and conditional independence relations of the causal graph, the enhancement is limited to the channel set that is adjacent or in the same Markov blanket of the causal graph. Adding or deleting paths that would change the conditional independence relations is prohibited. Cross-channel mixing is only allowed between channel pairs that are directly connected in the causal graph. Time warp and amplitude affine are subject to consistent rules between parent and child channels to avoid changes in directional relations. Those that do not meet the requirements are judged as non-compliant. The physical invariant constraints of the preset enhancement operators are checked. Based on the entity connections and static constraints in the topology diagram, it is checked whether the entity relationships of each channel are maintained after enhancement, and whether the dimensional and conservation relationships given by the business invariant list are destroyed. Parameters that touch the constraint boundary are truncated according to the boundary value and the reason for truncation is recorded. Candidates that violate the constraints are eliminated. The preset enhancement operator is checked to ensure that the main frequency and harmonic ratio remain unchanged. Based on the main frequency position, main peak bandwidth and harmonic ratio of each order in the spectrum at the event segment level, the frequency domain energy adjustment is limited to the allowed frequency band. It is prohibited to introduce new main peaks or move recorded main peaks beyond the tolerance range. The allowed frequency bands for bandpass resampling are given by the set of frequency bands marked on the spectrum. Candidates that exceed the allowed frequency bands are eliminated. Finally, the list of enhanced operators obtained after screening, the corresponding upper and lower limits of parameters, and the list of allowed parameters are combined and output as compliant spatial data.

[0019] This implementation method performs triple checks on the enhancement operator, including causal graphs, physical invariants, and spectral consistency, to ensure the acceptability and compliance of the enhancement operation in the logical, structural, and frequency domains, significantly reducing the abnormal disturbances and risk propagation caused by the enhancement.

[0020] In this embodiment, the step of calling a Mamba-based three-domain cooperative transformer within the operator and intensity range defined by the compliant spatial data, jointly generating enhancement parameters in the time domain, frequency domain, and event domain under graph mask gating, and performing compliant projection and rejection sampling to produce enhanced view data specifically includes: A three-domain cooperative transformer based on Mamba is established, wherein the three-domain cooperative transformer is a parameterized data transformation module implemented using the selective state-space network Mamba; Within the upper and lower limits and step granularity of the parameters given in the compliant spatial data, candidate parameter sequences are generated for time warp, amplitude affine, phase perturbation, bandpass resampling, controlled cross-channel mixing and missing replay, and graph mask gating is generated. The graph mask gating is used to mark channel pairs as allowed or prohibited based on the causal graph adjacency relationship to limit cross-channel effects. Enhancements are implemented in a fixed order: within the allowed event segments, time warp is applied first, followed by amplitude affine projection; compliant projection is performed on out-of-bounds values, whereby values ​​exceeding the upper or lower limit are replaced with the corresponding boundary values ​​and the replacement position is recorded; phase perturbation and bandpass resampling are applied sequentially within the allowed frequency bands; candidates not in the allowed frequency bands are directly discarded; and operations involving cross-channel effects are masked and prohibited using a map mask; controlled cross-channel mixing is applied within the allowed cross-channel mapping and event segment range; and non-adjacent channel pairs in the causal graph are not processed. Rejection sampling is implemented, which involves verifying the consistency of the results after enhancement. Samples that have event boundaries exceeding the limits, cross-channel relationships and map masks conflicting, or main peaks moving out of the allowed frequency band are marked as rejected and removed. Enhancement results are synthesized in the order of "time domain → frequency domain → event domain", and enhancement view data is generated and written with operator identifier, final parameter value, active channel, event segment and frequency band number.

[0021] This implementation uses a Mamba-based three-domain cooperative transformer to jointly generate enhancement parameters in the time, frequency, and event domains. Combined with graph mask gating and compliant projection mechanisms, it ensures both structural consistency and diversity of the enhancement samples, thereby improving enhancement quality and controllability.

[0022] In this embodiment, the multi-granularity InfoNCE comparison learning and overlay of the enhanced view data are described. Figure 1 The joint loss of consistent prediction, prospective prediction, and cyclic consistency yields characterization data, specifically including: The enhanced view data and the corresponding original time series are loaded, and a training batch is formed according to the sample identifier and event segment number. Each sequence is input into a shared time encoder to obtain an intermediate representation, which is then mapped into a normalized vector by a projection head. The projection head is a mapping module composed of a fully connected layer and a ReLU activation function. Using the original time series and augmented view data of the same sample and the same time span as positive sample pairs, and different samples or different event segments within the same batch as negative sample pairs, multi-granularity InfoNCE contrastive learning is implemented for the augmented view data. The multi-granularity includes sample level pairing of whole sequences, event level pairing by event segment number, and segment level pairing by sliding window. Cosine similarity is calculated for each granularity and temperature scaling is applied. The contrastive losses of sample level, event level, and segment level are accumulated according to preset weights. Superimposed on comparative learning Figure 1 To achieve joint loss, based on the causal graph, the induced relationship results are constructed from the channel correlation statistics of the encoder output, and compared with the adjacent relationships in the causal graph item by item. The deviation is included in the penalty term. The look-ahead prediction consistent joint loss is superimposed, and a look-ahead head is set. The look-ahead head is a lightweight prediction module that outputs the statistics and interval values ​​of the next time period for the original time series and the enhanced view data respectively, and compares the differences under the same statistical caliber, and includes the differences in the penalty term. The system employs a superimposed cyclic consistent joint loss and sets up an inverse mapping head, which is a lightweight restoration module. The original morphological sequence is input and output as the augmented view data. The difference between the original morphological sequence and the original temporal sequence at the corresponding time point is included in the penalty term. In one forward calculation, the contrastive loss and the joint loss are accumulated to form the total loss. In the backward update, only the parameters of the temporal encoder, the projection head, the look-ahead head, and the inverse mapping head are updated until convergence. After training is completed, the representation data is output.

[0023] This implementation method uses multi-granularity InfoNCE comparison learning and... Figure 1 The design of joint loss for consistency, prediction consistency, and cycle consistency enables the model to maintain robust feature representation capabilities under enhanced views, effectively improving representation accuracy and robustness to multiple types of temporal disturbances.

[0024] In this embodiment, the process of using enhanced view data as the evaluation object, calling the improved Chronos model for multi-step quantile prediction and uncertainty estimation, and combining the upper confidence bound strategy to determine the enhancement type and intensity, and generating enhancement strategy data, specifically includes: The improved Chronos model includes a discretization encoder, a language modeling decoder, and an inverse quantization restorer. The discretization encoder is used to discretize the original time series and enhanced view data into a sequence of symbols and align them with covariates. The language modeling decoder is used to predict future symbols under covariate conditions through autoregression and is trained with cross-entropy loss. The inverse quantization restorer is used to restore the predicted sequence of symbols into real-valued trajectories and generate quantile statistics. Using compliance space data as a candidate pool, evaluation combinations are generated according to enhancement type, intensity, allowed event segments and frequency bands. The original time series segments and enhanced view segments are segmented from the historical window corresponding to each evaluation combination and the covariates are aligned. The data is then input into a discretization encoder for event adaptive discretization. The distribution of each channel is statistically analyzed in segments according to the event segmentation results. Adaptive bucketing is established within the segments to map continuous values ​​to symbols. Conditional cue vectors are generated for covariates and injected with the symbols aligned with the symbol time axis. Values ​​that touch the compliance boundary are written into boundary symbols. Missing symbols are written into missing test locations and the length of continuous missing values ​​is recorded. The original time series symbols, enhanced view symbols and covariate cue vectors are output. The original time-series symbols, enhanced view symbols, and covariate cues are concatenated in time and fed into the language modeling decoder. The autoregressive method predicts the future symbol sequence. During training, the cross-entropy loss is used to supervise the next symbol. Unenhanced control samples and enhanced samples are simultaneously included in the same batch. After calculating the loss separately, the parameters are updated with equal weight. During the decoding stage, temperature sampling and kernel sampling are performed within the upper and lower limits of the parameters in the compliant space data to generate multiple future symbol trajectories. At the same time, non-compliant future symbol trajectories are pruned according to the allowed event segments and frequency bands, and the symbol trajectory data is output. The symbol trajectory data is input into the dequantization restorer, which gradually restores the real-value trajectory according to the reverse rule of the discretization mapping. At each time step, multiple real-value trajectories are sorted to obtain the quantile values ​​of the preset quantile set, and the quantile bandwidth and trajectory variance are calculated as uncertainty indicators. The quantile prediction is aligned with the actual observation of the corresponding time window, and the quantile loss is gradually accumulated according to the quantile point as the prediction cost. The amount of reduction of quantile loss relative to the unenhanced control sample is used as the prediction improvement amount, and the uncertainty indicators are recorded together as the evaluation result. The evaluation results employ an upper confidence bound strategy for decision-making, using the predicted improvement amount as the benefit item and superimposing quantile bandwidth and trajectory variance as risk penalties to obtain a single score. Within the scope of compliance space data, the enhancement type and intensity with the highest score are selected and organized into enhancement strategy data. The enhancement strategy data includes the selected type, intensity, historical window identifier, predicted improvement amount, uncertainty index, and decision score.

[0025] This implementation constructs an improved Chronos model, inputting the original time series and augmented views into a language modeling decoder after symbolic encoding. It utilizes cross-entropy loss to predict multi-step future symbol sequences and combines this with a dequantization recovery module for quantile value restoration, effectively improving the predictability and quantitative evaluation capability of the augmented data. Furthermore, it introduces uncertainty estimation indices, including quantile bandwidth and trajectory variance, and designs an upper confidence bound strategy to fuse prediction improvement and risk indices for augmentation combination scoring, thereby achieving refined selection and control of augmentation type and intensity. This strategy significantly differs from traditional augmentation methods that rely on empirical settings or random perturbations. It possesses the ability to optimize the augmentation path based on prediction performance feedback, ensuring that augmented samples improve model performance while maintaining controllable risk. It solves problems such as strong blindness and unstable generalization effects in augmentation strategies, exhibiting stronger adaptability and stability in practical applications.

[0026] In this embodiment, the process of generating an enhanced proof based on representation data and enhancement strategy data, completing knowledge distillation deployment, and online monitoring of drift and compliance threshold triggering degradation, rollback, and retraining specifically includes: Based on representation data and enhancement strategy data, enhancement proof data is constructed. The enhancement proof data includes the enhanced representation vector, enhancement type, enhancement strength, predicted improvement amount, uncertainty index and corresponding historical window identifier, which are used to record the basis and expected effect of enhancement behavior. The enhanced proof data is used for archiving of enhancement schemes during the deployment process. During the deployment system operation phase, real-time data is continuously collected and compared with the representation content in the enhanced proof. If the representation result of the real-time data is detected to deviate from the corresponding representation content, and the deviation exceeds the set drift threshold, the current enhancement strategy is considered to be invalid. When the drift threshold is triggered, a compliance response mechanism is initiated based on the status of the enhancement strategy. The compliance response mechanism includes: performing a downgrade process, which involves pausing the current enhancement strategy and switching to a standard enhancement template that has been verified to be stable during multiple rounds of training; performing a rollback process, which involves switching the current data processing flow to a path without enhancement or the last effective enhancement scheme; and performing a reconstruction process, which uses the current drift data as input to regenerate new map data, compliance space data, and enhancement strategy data.

[0027] This implementation method enhances the generation and archiving of proof documents, and combines drift detection and response mechanisms to achieve continuous monitoring and dynamic adjustment of the enhancement strategy in actual deployment, possessing long-term stable operation and self-recovery capabilities under abnormal scenarios.

[0028] Example 1: To verify the feasibility of this invention in practice, it was applied to the production monitoring and predictive optimization scenario of a compressor-chilled water combined power supply system in a chemical industrial park. The park consists of multiple screw compressors, cooling towers, plate heat exchangers, and circulating pumps, forming a clear business topology. Forty-eight key sensors were deployed on-site, and five covariates—temperature, humidity, wind speed, sunshine duration, and electricity price—were simultaneously input. The sampling period was one minute, and data was collected continuously for sixty days.

[0029] The raw data exhibits strong non-stationary characteristics: planned start-ups and shutdowns lead to slow transitions, load switching driven by peak-valley electricity prices causes rapid changes, and cold waves cause abrupt changes in the distribution of covariates. During implementation, multi-source data were first aligned on a unified time axis and missing positions were marked. Peak-valley detection, combined with cumulative and abrupt change detection, was used to delineate event boundaries. Then, differentiable causal discovery yielded directed dependencies, and time-delay Granger calculus was used to supplement the main lag orders. Short-time Fourier transforms were performed on event segments to extract the dominant frequency position, peak bandwidth, and harmonic ratio. These three types of information, along with event segmentation, were fused into a "causal-time-delay-spectrum" spectral data. Based on this spectral data, operators such as time warp, amplitude affine transformation, phase perturbation, bandpass resampling, controlled cross-channel mixing, and missing replay were subjected to three types of compliance checks: Markov equivalence preservation, physical invariance satisfaction, and dominant frequency / harmonic ratio invariance. This yielded a compliance space for usable operators and intensity intervals. Subsequently, a three-domain collaborative transformer based on Mamba is used to jointly generate time-domain, frequency-domain, and event-domain parameters under graph mask gating, and an enhanced view is produced by combining compliant projection and rejection sampling. Multi-granularity InfoNCE is used to conduct sample-level, event-level, and fragment-level comparative learning, and these are then overlaid. Figure 1 The joint loss of consistent prediction, forward prediction, and cyclic consistency is used to obtain a robust characterization. Then, an improved Chronos is used for multi-step quantile prediction and uncertainty estimation. The above confidence bound strategy adaptively selects the type and strength of reinforcement within the compliance space to form an "reinforcement strategy" and generates an reinforcement proof sheet containing operators, strength, and compliance evidence. The teacher model is distilled to obtain a lightweight student model, which is then deployed online to monitor distribution drift and structural drift. When a threshold is triggered, the model is automatically downgraded or rolled back.

[0030] Table 1. Overview of Scenarios and Data

[0031] Table 2. Summary Results of Comparative Experiments

[0032] Table 3 Comparison of Ablation Data for Key Components

[0033] A clear analysis can be made based on the data in Tables 1, 2, and 3. Regarding quantile loss, this invention reduces the P50 and P90 by approximately 14.5 and 19.3 percentage points respectively compared to the unenhanced baseline, indicating a significant convergence of extreme errors. Compared to empirical rules, it further reduces the P50 and P90 by 7.1 and 9.5 percentage points respectively, demonstrating that the three-domain synergistic enhancement is more stable when event boundaries and periodic components are simultaneously controlled. The MAPE is reduced to 7.1, reflecting that the average relative error of multi-step prediction can still maintain a good level under complex load switching and weather disturbances. The anomaly detection AUROC is improved to 0.919, indicating greater sensitivity to low-occurrence, weak signal anomalies, which is beneficial for early warning and energy efficiency optimization. In terms of engineering auditability, due to the addition of three types of compliance verification and enhanced proof documents, the audit failure rate is zero, and the average audit time is only 0.09 seconds, enabling rapid tracing of the complete chain of "operator—intensity—compliance evidence—benefit indicators" during production sampling inspections. Ablation analysis revealed that the absence of any key constraint or strategy led to varying degrees of performance degradation. Removing compliance checks resulted in a significant deterioration of P90, along with frequency migration and dimensional conflicts, demonstrating that the "compliance space" is a prerequisite for robustness. Canceling the UCB strategy weakened the adaptive optimization capability for new operating conditions, manifesting as an overall performance regression. In actual operation, a sudden change in covariate distribution occurred during a cold wave window. After triggering the drift threshold, the system automatically rolled back to the most recent stable enhancement strategy based on the evidence. The error recovered to within the threshold within two hours, and the entire process was auditable and traceable.

[0034] As can be seen from the embodiments, without destroying the causal structure, physical invariants, and the dominant frequency-harmonic relationship, this invention improves the generalization of representation, prediction accuracy, and anomaly detection capability through "compliance space constraints + three-domain collaborative enhancement + multiple consistency learning + quantile prediction and UCB closed loop". It also constructs a traceable, rollback, and self-healing engineering closed loop with enhanced proof and online drift monitoring, meeting the dual requirements of effectiveness and compliance for high-reliability scenarios such as the Industrial Internet of Things.

[0035] The above description is only a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any equivalent substitutions or modifications made by those skilled in the art within the scope of the technology disclosed in the present invention, based on the technical solution and inventive concept of the present invention, should be covered within the scope of protection of the present invention.

Claims

1. A multivariate time-series data augmentation method based on contrastive learning, characterized in that, Includes the following steps: The original time series, business topology and covariates are collected, and the events are segmented by peak and valley detection, cumulative sum and mutation detection. The graph data is constructed by combining differentiable causal discovery, time-delay Granger and short-time Fourier transform. Based on the graph data, compliance checks are performed on the preset enhancement operator to ensure that the causal graph Markov equivalence is preserved, the physical invariant constraints are satisfied, and the main frequency and harmonic ratio remain unchanged, and compliance space data is output. Within the range of operators and intensities defined by the compliant spatial data, a three-domain cooperative transformer based on Mamba is invoked to jointly generate enhancement parameters in the time domain, frequency domain, and event domain under graph mask gating, and to perform compliant projection and rejection sampling to produce enhanced view data. Multi-granularity InfoNCE contrastive learning is performed on the enhanced view data, and the joint loss of graph consistency, look-ahead prediction consistency and cycle consistency is overlaid to obtain the representation data; Using enhanced view data as the evaluation object, the improved Chronos model is called to perform multi-step quantile prediction and uncertainty estimation. Combined with the upper confidence bound strategy, the enhancement type and intensity are determined, and enhancement strategy data is generated. Based on the representation data and enhancement strategy data, an enhanced proof is generated and knowledge distillation deployment is completed. Online monitoring is conducted to detect drift and compliance threshold triggers for degradation, rollback, and retraining.

2. The multivariate time-series data augmentation method based on contrastive learning according to claim 1, characterized in that, The process involves collecting raw time series data, service topology, and covariates, performing event segmentation through peak-valley detection, cumulative sum, and abrupt change detection, and constructing spectral data using differentiable causal discovery, time-delay Granger calculus, and short-time Fourier transform. Specifically, this includes: The original time series is a multivariate time series. A unified time axis is established with the smallest time granularity. The gaps in the original time series with different sampling rates are filled by interpolation according to the linear ratio of adjacent sampling points, and missing markers are written at the filled positions. The business topology is a data structure of objects and connections corresponding to the original time series, including entity identifiers, entity types, connection edges between entities and edge attributes. Based on the entity identifiers, the channels and sensor sources in the original time series are mapped and verified to form an entity-channel correspondence table. The covariates are exogenous variables that affect the original time series and originate from external systems or public data sources. The covariates are synchronized with the original time series on the time axis. Gaps are filled by linear interpolation of adjacent sampling points and missing markers are added. The original time series is subjected to joint event segmentation by peak and valley detection and cumulative and abrupt change detection. Peak and valley detection finds local maxima and minima within a sliding window and calculates historical noise levels by segmentation according to covariates to set trigger thresholds. Cumulative and abrupt change detection progressively accumulates the difference from the baseline and compares it with the trigger threshold. If the difference exceeds the trigger threshold, it is marked as an abrupt change point. Continuous abrupt change points are paired with adjacent peaks and valleys to determine the start and end of the event and generate the event segmentation results. Based on the event segmentation results, differentiable causal discovery is performed to obtain the initial results of directed dependencies between variables. The initial results are verified by the time-delay Granger test and the main lag order is added. At the same time, covariates are added as exogenous nodes, and a causal graph containing directed edges and lag information is output. Construct a topology graph containing nodes, edges, and edge attributes based on the entity-channel correspondence table; The original time series is segmented based on the event segmentation results, and after windowing, a short-time Fourier transform is performed on each segment, which are then aggregated to form a spectrogram. The causal graph, topological graph, and spectrogram, along with the event segmentation results, are combined into spectral data.

3. The multivariate time-series data augmentation method based on contrastive learning according to claim 1, characterized in that, Based on the spectral data, compliance checks are performed on the preset enhancement operator to ensure causal graph Markov equivalence preservation, physical invariant constraint satisfaction, and invariance of dominant frequency and harmonic ratio, outputting compliance space data, specifically including: Based on the spectral data, preset enhancement operators are determined, and candidate parameter ranges and step granularity are set for each type of enhancement operator; The preset enhancement operators are checked for causal graph Markov equivalence preservation. Based on the directed edge relations and conditional independence relations of the causal graph, the enhancement is limited to the channel set that is adjacent or in the same Markov blanket of the causal graph. Adding or deleting paths that would change the conditional independence relations is prohibited. Cross-channel mixing is only allowed between channel pairs that are directly connected in the causal graph. Time warp and amplitude affine are subject to consistent rules between parent and child channels to avoid changes in directional relations. Those that do not meet the requirements are judged as non-compliant. The physical invariant constraints of the preset enhancement operators are checked. Based on the entity connections and static constraints in the topology diagram, it is checked whether the entity relationships of each channel are maintained after enhancement, and whether the dimensional and conservation relationships given by the business invariant list are destroyed. Parameters that touch the constraint boundary are truncated according to the boundary value and the reason for truncation is recorded. Candidates that violate the constraints are eliminated. The preset enhancement operator is checked to ensure that the main frequency and harmonic ratio remain unchanged. Based on the main frequency position, main peak bandwidth and harmonic ratio of each order in the spectrum at the event segment level, the frequency domain energy adjustment is limited to the allowed frequency band. It is prohibited to introduce new main peaks or move recorded main peaks beyond the tolerance range. The allowed frequency bands for bandpass resampling are given by the set of frequency bands marked on the spectrum. Candidates that exceed the allowed frequency bands are eliminated. Finally, the list of enhanced operators obtained after screening, the corresponding upper and lower limits of parameters, and the list of allowed parameters are combined and output as compliant spatial data.

4. The multivariate time-series data augmentation method based on contrastive learning according to claim 1, characterized in that, Within the operator and intensity range defined by the compliant spatial data, a Mamba-based three-domain cooperative transformer is invoked to jointly generate enhancement parameters in the time, frequency, and event domains under graph mask gating, and compliant projection and rejection sampling are performed to produce enhanced view data. Specifically, this includes: A three-domain cooperative transformer based on Mamba is established, wherein the three-domain cooperative transformer is a parameterized data transformation module implemented using the selective state-space network Mamba; Within the upper and lower limits and step granularity of the parameters given in the compliant spatial data, candidate parameter sequences are generated for time warp, amplitude affine, phase perturbation, bandpass resampling, controlled cross-channel mixing and missing replay, and graph mask gating is generated. The graph mask gating is to mark channel pairs as allowed or prohibited based on the causal graph adjacency relationship. Enhancements are implemented in a fixed order: within the allowed event segments, time warp is applied first, followed by amplitude affine projection; compliant projection is performed on out-of-bounds values, whereby values ​​exceeding the upper or lower limit are replaced with the corresponding boundary values ​​and the replacement position is recorded; phase perturbation and bandpass resampling are applied sequentially within the allowed frequency bands; candidates not in the allowed frequency bands are directly discarded; and operations involving cross-channel effects are masked and prohibited using a map mask; controlled cross-channel mixing is applied within the allowed cross-channel mapping and event segment range; and non-adjacent channel pairs in the causal graph are not processed. Rejection sampling is implemented, which involves verifying the consistency of the results after enhancement. Samples that have event boundaries exceeding the limits, cross-channel relationships and map masks conflicting, or main peaks moving out of the allowed frequency band are marked as rejected and removed. Enhancement results are synthesized in the order of "time domain → frequency domain → event domain", and enhancement view data is generated and written with operator identifier, final parameter value, active channel, event segment and frequency band number.

5. The multivariate time-series data augmentation method based on contrastive learning according to claim 1, characterized in that, The process involves performing multi-granularity InfoNCE comparative learning on the enhanced view data, overlaying joint losses for graph consistency, look-ahead prediction consistency, and cycle consistency to obtain representational data, specifically including: Load the enhanced view data and the corresponding original time series, form a training batch according to the sample identifier and event segment number, input each sequence into the shared time series encoder to obtain the intermediate representation, and then map it into a normalized vector through the projection head; Using the original time series and augmented view data of the same sample and the same time span as positive sample pairs, and different samples or different event segments within the same batch as negative sample pairs, multi-granularity InfoNCE contrastive learning is implemented for the augmented view data. The multi-granularity includes sample level pairing of whole sequences, event level pairing by event segment number, and segment level pairing by sliding window. Cosine similarity is calculated for each granularity and temperature scaling is applied. The contrastive losses of sample level, event level, and segment level are accumulated according to preset weights. While contrastive learning is performed, graph-consistent joint loss is superimposed. Based on the causal graph, induced relationship results are constructed from the channel correlation statistics of the encoder output. The results are compared with the adjacent relationships in the causal graph item by item, and the deviations are included in the penalty term. The look-ahead prediction consistent joint loss is superimposed, and a look-ahead head is set. The look-ahead head is a lightweight prediction module that outputs the statistics and interval values ​​of the next time period for the original time series and the enhanced view data respectively, and compares the differences under the same statistical caliber, and includes the differences in the penalty term. The system employs a superimposed cyclic consistent joint loss and sets up an inverse mapping head, which is a lightweight restoration module. The original morphological sequence is input and output as the augmented view data. The difference between the original morphological sequence and the original temporal sequence at the corresponding time point is included in the penalty term. In one forward calculation, the contrastive loss and the joint loss are accumulated to form the total loss. In the backward update, only the parameters of the temporal encoder, the projection head, the look-ahead head, and the inverse mapping head are updated until convergence. After training is completed, the representation data is output.

6. The multivariate time-series data augmentation method based on contrastive learning according to claim 1, characterized in that, The process involves using enhanced view data as the evaluation object, calling an improved Chronos model for multi-step quantile prediction and uncertainty estimation, combining an upper confidence bound strategy to determine the enhancement type and intensity, and generating enhancement strategy data. Specifically, this includes: The improved Chronos model includes a discretized encoder, a language modeling decoder, and an inverse quantization demultiplexer; Using compliance space data as a candidate pool, evaluation combinations are generated according to enhancement type, intensity, allowed event segments and frequency bands. The original time series segments and enhanced view segments are segmented from the historical window corresponding to each evaluation combination and the covariates are aligned. The data is then input into a discretization encoder for event adaptive discretization. The distribution of each channel is statistically analyzed in segments according to the event segmentation results. Adaptive bucketing is established within the segments to map continuous values ​​to symbols. Conditional cue vectors are generated for covariates and injected with the symbols aligned with the symbol time axis. Values ​​that touch the compliance boundary are written into boundary symbols. Missing symbols are written into missing test locations and the length of continuous missing values ​​is recorded. The original time series symbols, enhanced view symbols and covariate cue vectors are output. The original time-series symbols, enhanced view symbols, and covariate cues are concatenated in time and fed into the language modeling decoder. The autoregressive method predicts the future symbol sequence. During training, the cross-entropy loss is used to supervise the next symbol. Unenhanced control samples and enhanced samples are simultaneously included in the same batch. After calculating the loss separately, the parameters are updated with equal weight. During the decoding stage, temperature sampling and kernel sampling are performed within the upper and lower limits of the parameters in the compliant space data to generate multiple future symbol trajectories. At the same time, non-compliant future symbol trajectories are pruned according to the allowed event segments and frequency bands, and the symbol trajectory data is output. The symbol trajectory data is input into the dequantization restorer, which gradually restores the real-value trajectory according to the reverse rule of the discretization mapping. At each time step, multiple real-value trajectories are sorted to obtain the quantile values ​​of the preset quantile set, and the quantile bandwidth and trajectory variance are calculated as uncertainty indicators. The quantile prediction is aligned with the actual observation of the corresponding time window, and the quantile loss is gradually accumulated according to the quantile point as the prediction cost. The amount of reduction of quantile loss relative to the unenhanced control sample is used as the prediction improvement amount, and the uncertainty indicators are recorded together as the evaluation result. The evaluation results employ an upper confidence bound strategy for decision-making, using the predicted improvement amount as the benefit item and superimposing quantile bandwidth and trajectory variance as risk penalties to obtain a single score. Within the scope of compliance space data, the enhancement type and intensity with the highest score are selected and organized into enhancement strategy data.

7. The multivariate time-series data augmentation method based on contrastive learning according to claim 1, characterized in that, The process of generating enhanced proofs based on representation data and enhancement strategy data, deploying knowledge distillation, and online monitoring of drift and compliance threshold triggering degradation, rollback, and retraining specifically includes: Based on representational data and enhancement strategy data, construct enhanced proof single data; The enhanced proof data is used for archiving of enhancement schemes during the deployment process. During the deployment system operation phase, real-time data is continuously collected and compared with the representation content in the enhanced proof. If the representation result of the real-time data is detected to deviate from the corresponding representation content, and the deviation exceeds the set drift threshold, the current enhancement strategy is considered to be invalid. When the drift threshold is triggered, a compliance response mechanism is initiated based on the status of the enhancement strategy. The compliance response mechanism includes: performing a downgrade process, which involves pausing the current enhancement strategy and switching to a standard enhancement template that has been verified to be stable during multiple rounds of training; performing a rollback process, which involves switching the current data processing flow to a path without enhancement or the last effective enhancement scheme; and performing a reconstruction process, which uses the current drift data as input to regenerate new map data, compliance space data, and enhancement strategy data.

Images