A Text-Enhanced Time Series Anomaly Detection Method Based on Semantic Interference Suppression
By combining trend decoupling and multi-scale coding with dynamic memory fusion, the problem of semantic interference in multimodal time series anomaly detection is solved, achieving efficient detection of short-term mutations and local perturbations, and improving the interpretability and robustness of the detection results.
Patent Information
- Application Number
- CN202611511493.4
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2026-09-28
- Publication Date
- 2026-11-17
AI Technical Summary
In multimodal time series anomaly detection, existing technologies struggle to distinguish between semantically normal and anomalous fluctuations, are susceptible to textual interference, may disrupt normal time series patterns through direct fusion, mask local mutations with low-frequency trends, cause fine-grained information leakage through cross-scale interactions, and are unable to fully reflect the impact of complex events from a single source.
By employing trend decoupling, multi-scale downsampling, and scale-independent encoding, dynamic memory integrates exogenous events, endogenous statistics, and domain prior text. Cross-modal interactive parallel one-dimensional convolution is combined with partially visible cross-scale mask decoding and low-frequency trend backinjection, adaptive convolution smoothing, and anomaly score calculation.
It effectively separates low-frequency trends from local residual information, enhances the detection capability for short-term mutations and local disturbances, improves the interpretability and robustness of detection results, reduces the impact of normal background noise, and improves the accuracy and stability of anomaly detection.
Smart Images

Figure CN123116099A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the fields of artificial intelligence, time series data processing, and multimodal information fusion, specifically to a text-enhanced time series anomaly detection method based on semantic interference suppression. Background Technology
[0002] Time series anomaly detection aims to identify observations or patterns that deviate from normal evolution patterns in time series data. Existing methods are mainly classified into prediction-based methods, reconstruction-based methods, contrastive learning-based methods, and large model-based methods based on their modeling paradigms. Among them, reconstruction-based methods recover the original sequence by learning a compressed representation of normal data and score anomalies based on the reconstructed residuals.
[0003] With the development of multimodal data acquisition methods, time series anomaly detection has begun to incorporate external event text. Numerical modalities reflect the system's state trajectory, while external texts such as event reports, weather warnings, or operation logs provide contextual information about system fluctuations. Although numerical changes caused by external events may significantly deviate from general statistical patterns, if they can be consistently explained by the semantics of events within the same or adjacent time range, the changes are considered interpretable fluctuations. If numerical values deviate significantly but lack corresponding external semantic support, or if existing text cannot explain the deviation, then they are considered anomalous fluctuations under conditions of semantic deficiency.
[0004] Existing text augmentation anomaly detection methods typically aim to improve the consistency and reconstruction accuracy between time series and text representations, using text as a general auxiliary feature for concatenation, attention interaction, or gating fusion. However, due to differences in sampling frequency, information density, and semantic granularity between continuous time series signals and sparse event text, and the potential for redundant, mismatched, or weakly correlated content in the actual input text, direct fusion can easily disrupt normal time series pattern modeling and lead to cross-modal negative transfer. Existing hard gating methods may also remove explanatory context under complex semantic overlap conditions, making it difficult to balance noise suppression and effective semantic preservation.
[0005] Furthermore, current text from a single source cannot fully reflect the impact of complex events on time series; when directly encoding multiple scales on the original sequence, low-frequency trends may mask local abrupt changes, and unconstrained cross-scale interactions may cause fine-grained information leakage; point-to-point errors are easily affected by natural background noise and small phase shifts. Therefore, a time series anomaly detection method is needed that can utilize historical event patterns and domain priors, quantify cross-modal conflicts, and adaptively control the intensity of text fusion. Summary of the Invention
[0006] In view of the above problems, the present invention provides a text-enhanced time series anomaly detection method based on semantic interference suppression, which solves the technical problems of difficulty in distinguishing between normal semantics and abnormal fluctuations and easy text interference in the prior art.
[0007] This invention provides a text-enhanced time-series anomaly detection method based on semantic interference suppression, comprising the following steps: Step S1: Obtain the multivariate time series and the corresponding external event text, and obtain the domain prior text; the multivariate time series includes meteorological information at multiple time points, and the external event text includes meteorological text at multiple time points; the domain prior text records the influence of the types of external events on the multivariate time series; Trend decoupling is performed on the input multivariate time series to obtain low-frequency trend terms and residual sequences; multi-scale downsampling and scale-independent encoding are performed on the residual sequences to obtain multi-scale time series features; Step S2: Generate endogenous statistical text based on the statistics of the time segments of the multivariate time series; Encode and semantically fuse exogenous event text, endogenous statistical text, and domain prior text respectively to obtain enhanced text features; Step S3: Perform parallel one-dimensional convolution on the multi-scale temporal features and enhanced text features respectively to obtain consensus benchmark features and dynamic complementary features. Perform cross-modal interaction and gated residual interpolation on the consensus benchmark features and dynamic complementary features respectively to obtain fused features. Step S4: The fused features are split by scale and upsampled, decoded under partial visible cross-scale mask constraints, and then reconstructed by back-injection of the low-frequency trend term and adaptive convolution smoothing. Step S5: Calculate the time-point anomaly score based on the multi-scale reconstruction error and memory pattern deviation of the reconstructed sequence and the multivariate time series, and threshold the time-point anomaly score to obtain the anomaly detection result.
[0008] Preferably, in step S1, the meteorological information includes evaporation, average wind speed, precipitation, average temperature, maximum temperature, minimum temperature, maximum water temperature, minimum water temperature, average surface temperature, minimum surface temperature, snow depth, snowfall, 2-minute wind direction angle, 5-second wind direction angle, maximum wind speed within 2 minutes, and maximum wind speed within 5 seconds; the meteorological text includes text records of storms, hurricanes, and flood-related events.
[0009] Preferably, in step S1, the expressions for the low-frequency trend term and the residual sequence are:
[0010] in, Indicates low-frequency trend terms; Represents multivariate time series; This represents one-dimensional average pooling; Indicates the pooling kernel size; Indicates the pooling step size;
[0011] in, Represents the residual sequence; In step S1, the expression for the multi-scale temporal features is:
[0012]
[0013]
[0014] in, Represents the temporal features at the i-th scale; Indicates Transformer encoder; Indicates a scale-independent mask; Represents the embedding feature at scale i; This indicates a block embedding operation; This represents the i-th scale multi-scale time series sequence obtained by downsampling; This represents the position code of the i-th scale; N represents the total number of scales. This represents the scale-independent mask value between the a-th feature label and the b-th feature label; These represent the a-th feature label and the b-th feature label, respectively.
[0015] Preferably, in step S2, the step of generating endogenous statistical text based on the statistics of the time segments of the multivariate time series specifically includes: calculating the mean, standard deviation, maximum value, and minimum value of the time segments of the multivariate time series; filling the mean, standard deviation, maximum value, and minimum value into a structured prompt template according to preset fields to generate endogenous statistical text; The steps of encoding and semantically fusing exogenous event text, endogenous statistical text, and domain prior text to obtain enhanced text features specifically include: The exogenous event text and the endogenous statistical text are encoded to obtain joint text features; Encode the domain prior text to obtain domain prior features; Historical semantic features are obtained by retrieving historical semantic prototypes from the dynamic memory using the joint text features. Enhanced text features are obtained by fusing the domain prior features and historical semantic features.
[0016] Preferably, the expression for the joint text feature is:
[0017] in, Indicates joint text features; Indicates a text encoder; Represents the text of an external event; This represents endogenous statistical text; The expression for the domain prior features is:
[0018] in, Represents prior features of the domain; Represents prior text of the domain; The expression for the historical semantic feature is:
[0019]
[0020] in, Indicates historical semantic features; Indicates from the set of relevance weights The top-k high-weight indexes selected from the list; This represents the relevance weight of the i'-th memory unit; This represents the i'th learnable historical semantic prototype; This represents the relevance weight of the i'-th memory unit; Represents an exponential function; Represents the cosine similarity function; d represents the learnable pattern index of the i'-th memory unit; d represents the feature dimension; K represents the total number of memory units in the dynamic memory bank; Represents the learnable pattern index of the j-th memory unit; The expression for the enhanced text feature is:
[0021] in, This indicates enhanced text features; Indicates the layer normalization function; Indicates joint text features; This represents a multi-head cross-attention operation; Q, K, and V represent the query, key, and value of the cross-attention operation, respectively; C represents a comprehensive contextual feature pool formed by concatenating domain prior features and historical semantic features.
[0022] Preferably, in step S3, the expressions for the consensus benchmark feature and the dynamic complementary feature are:
[0023]
[0024] in, This indicates the temporal dynamic complementary characteristics; Indicates the characteristics of the time-series consensus benchmark; This represents a one-dimensional convolution used to extract dynamic complementary features; This represents a one-dimensional convolution used to extract consensus benchmark features; Represents multi-scale temporal features. This indicates the dynamic complementary features of the text; Represents the benchmark features of textual consensus; This represents a one-dimensional convolution used to extract dynamic complementary features; This represents a one-dimensional convolution used to extract consensus benchmark features; In step S3, the specific steps of performing cross-modal interaction and gated residual interpolation on the consensus benchmark features and dynamic complementary features to obtain the fused features include: The consensus benchmark features are used to perform cross-modal interaction to obtain the baseline representation; the dynamic complementary features are used to perform cross-modal interaction to obtain the joint representation; the mean square offset of the joint representation relative to the baseline representation is used as the coupling conflict divergence, and the increment of the joint representation relative to the baseline representation is gated residual interpolation with a security fusion weight that monotonically decreases as the coupling conflict divergence increases to obtain the fusion feature.
[0025] Preferably, the expression for the baseline representation is:
[0026] in, Indicates baseline characterization; This represents cross-modal attention operations; Q, K, and V represent query, key, and value, respectively. The expression for the joint representation is:
[0027] in, Indicates joint representation; This indicates a cross-modal attention operation that sequentially receives queries, keys, and values; The expression for the coupling conflict divergence is:
[0028] Where D represents the coupling conflict divergence; This represents the eigenvalues of the joint representation in the i''th dimension; The baseline represents the eigenvalues in the i''th dimension. The expression for the security fusion weight is:
[0029] in, Indicates the security fusion weight; This refers to the Softplus function; Represents the learnable scalar scaling parameter; The expression for the fusion feature is:
[0030] in, Indicates fusion characteristics; This indicates element-wise multiplication.
[0031] Preferably, in step S4, the expression for splitting and upsampling the fused features by scale is:
[0032] in, Represents the upsampled features at the i-th scale; This represents the linear interpolation operation corresponding to the i-th scale; This represents the i-th scale fusion feature obtained by splitting from the fusion features; The expression for decoding under partially visible multi-scale mask constraints is:
[0033]
[0034] in, This represents the partially visible cross-scale mask value of the query position at scale i to the context position at scale j during decoding; i and j represent the query scale index and the accessed scale index, respectively. This represents the scale number corresponding to the i-th position; This represents the scale number corresponding to the j-th position; This represents the decoding result at the i-th scale; This represents a Transformer decoder constrained by a partially visible cross-scale mask. Represents the upsampled features at the i-th scale; In step S4, the expression for the reconstructed sequence obtained after back-injection and adaptive convolution smoothing of the low-frequency trend term is:
[0035]
[0036] in, This represents the reconstruction result at the i-th scale; This represents the low-frequency trend term downsampled to the i-th scale; Indicates the reconstructed sequence; This indicates an adaptive convolution smoothing operation; This indicates the result to be smoothed after merging the reconstruction results from various scales.
[0037] Preferably, in step S5, the expression for the multi-scale reconstruction anomaly score is:
[0038]
[0039]
[0040]
[0041] in, This represents the local mean at the k-th scale; This represents a one-dimensional average pooling operation; Represents the true multivariate time series at the k-th scale; This indicates that the pooling kernel size is 5; This indicates that the pooling step size is 1; This represents the point-to-point squared error at the k-th scale; Represents the dynamic tolerance boundary at the k-th scale; The scaling factor represents the normal local fluctuations; Represents a positive stabilizing term that improves numerical stability; This represents the reconstructed anomaly score after deducting normal fluctuations at the k-th scale; Represents a linear rectified function; Indicates the multi-scale reconstruction anomaly score; N represents the total number of scales; Represents the interpolation function; Represents the reconstructed anomaly score at the k-th scale; This indicates that the output length is set to the original multivariate time series length L; This indicates that a linear interpolation mode is used; The expression for the memory pattern deviation is:
[0042]
[0043] in, q represents the deviation from the memory pattern; q represents the current input query vector. This represents the inner product of the query vector q and the memory feedback prototype m; and Let represent the L2 norms of q and m respectively; m represents the memory feedback prototype. This represents the set of memory cell indexes with the highest relevance weights; This represents the relevance weight of the i'-th memory unit; This represents the i'th historical semantic prototype; The expression for the time-point anomaly score is:
[0044] in, Indicates the outlier score at each time point; Indicates the multi-scale reconstruction anomaly score; Indicates the degree of deviation from the memory pattern; This represents the weighting coefficient of memory pattern deviation in the outlier score at each time point.
[0045] Preferably, the method further includes the step of training a sequence anomaly detection model, comprising: constructing a loss function using reconstruction loss, morphological alignment loss, cross-modal contrast loss, and decoupling loss, and training the model based on the loss function; The expression for the reconstruction loss is:
[0046] in, Represents the reconstruction loss; N' represents the total number of error elements averaged by the current reconstruction loss; i''' represents the index of the error element; This represents the i'''th reconstructed residual; This represents the Huber loss corresponding to the i'''th reconstruction residual; 'Indicates the focus modulation index; The expression for the morphological alignment loss is:
[0047] in, This indicates the shape alignment loss; This represents the DILATE loss function used to constrain sequence shape and time alignment. Indicates the reconstructed sequence; This represents the input multivariate time series; The expression for the cross-modal contrast loss is:
[0048] in, Indicates cross-modal contrast loss; This represents the contrastive loss from temporal global embedding to textual global embedding; This represents the contrastive loss from text global embedding to temporal global embedding; The expression for the decoupling loss is:
[0049] in, N''' represents the decoupling loss; N''' represents the number of features involved in the average calculation of the current decoupling loss. Represents the cosine similarity function The total loss of the model is:
[0050] in, This represents the total loss of the model; , and These represent the weight coefficients for morphological alignment loss, cross-modal contrast loss, and decoupling loss, respectively.
[0051] Compared with the prior art, the present invention has at least the following beneficial effects: (1) By performing trend decoupling, multi-scale downsampling and scale-independent encoding on multivariate time series, this invention can effectively separate low-frequency trend terms and local residual information, reduce the interference of trend changes and cross-scale information aliasing on the modeling of abnormal details, and improve the ability to extract abnormal features such as short-term mutations and local disturbances.
[0052] (2) This invention achieves dynamic judgment of cross-modal information conflict and text reliability by dynamically fusing exogenous event text, endogenous statistical text, domain prior text and historical semantic prototypes through memory fusion, and comparing the offset continuity between low-risk benchmark representation and joint representation, so as to enable credible text to participate in temporal reconstruction and anomaly discrimination, thereby improving the interpretability and robustness of detection results.
[0053] (3) This invention performs multi-scale decoding under partial visible cross-scale mask constraints, and calculates the anomaly score by combining low-frequency trend back-injection, adaptive convolutional smoothing and multi-scale reconstruction error and memory pattern deviation after dynamic tolerance. This can reduce the impact of normal background noise and interpretable fluctuations on the detection results and improve the accuracy and stability of anomaly detection. Attached Figure Description
[0054] The accompanying drawings are for illustrative purposes only and are not intended to limit the scope of the invention.
[0055] Figure 1 The flowchart shows the text-enhanced time series anomaly detection method based on semantic interference suppression provided by this invention.
[0056] Figure 2 The overall framework diagram of the text-enhanced time series anomaly detection model provided by this invention is shown. Detailed Implementation
[0057] To better understand the above-described objectives, features, and advantages of the present invention, the invention will be further described in detail below with reference to the accompanying drawings and specific embodiments. It should be noted that, unless otherwise specified, the embodiments of the present invention and the features thereof can be combined with each other. Furthermore, the present invention can be implemented in other ways different from those described herein; therefore, the scope of protection of the present invention is not limited to the specific embodiments disclosed below.
[0058] The text-enhanced time series anomaly detection method based on semantic interference suppression provided by this invention takes a multivariate time series, the corresponding exogenous event text, and the domain prior text as input, and sequentially performs trend decoupling, multi-scale temporal coding, endogenous statistical text generation, dynamic memory semantic fusion, cross-modal interference suppression, partially visible cross-scale decoding, and anomaly scoring. For example... Figure 2 As shown, the model quantifies cross-modal conflicts by comparing the offsets of baseline representations and joint representations, and continuously adjusts the intensity of dynamic complementary semantics entering the fused features based on the cross-modal conflicts.
[0059] To illustrate the effectiveness of the method proposed in this invention, the following detailed description of the above technical solution is provided through a specific embodiment. Given a multivariate time series containing multiple continuous observations, each observation having multiple variables, the multivariate time series is denoted as... The invention obtains external event text that is synchronized with or adjacent to the corresponding time window. The text-enhanced time series anomaly detection model adopted in this invention is trained. During the training phase, the model is trained using normal multivariate time series and their corresponding texts; during the inference phase, the test multivariate time series is input. and corresponding external event text The system outputs anomaly scores at each time point and thresholds them, where 1 indicates an anomaly and 0 indicates a normal anomaly.
[0060] First, the inference stage of the text-enhanced time series anomaly detection model of this invention is described in detail. For example... Figure 1 As shown, a text-enhanced time series anomaly detection method based on semantic interference suppression is disclosed. The specific implementation steps are as follows: Step S1: Obtain the multivariate time series and the corresponding external event text, and obtain the domain prior text; the multivariate time series includes meteorological information at multiple time points, and the external event text includes meteorological text at multiple time points; the domain prior text records the influence of the types of external events on the multivariate time series; Trend decoupling is performed on the input multivariate time series to obtain low-frequency trend terms and residual sequences; multi-scale downsampling and scale-independent encoding are performed on the residual sequences to obtain multi-scale time series features; In some embodiments, five subsets of meteorological events from the CrisisTS dataset—California, Carolina, Florida, New York, and Texas—are used. Each dataset contains daily sampled meteorological information and corresponding meteorological texts from January 2016 to December 2019. The meteorological information includes 16 indicators: evaporation, average wind speed, precipitation, average temperature, maximum temperature, minimum temperature, maximum water temperature, minimum water temperature, average surface temperature, minimum surface temperature, snow depth, snowfall, 2-minute wind direction angle, 5-second wind direction angle, maximum wind speed within 2 minutes, and maximum wind speed within 5 seconds. The meteorological texts include textual records of events such as storms, hurricanes, and floods.
[0061] In some embodiments, the multivariate time series and the corresponding external event text can be preprocessed, including statistically analyzing the missing proportion of each variable and deleting variables with a missing proportion exceeding 90%; filling the missing values of the retained variables with linear interpolation and smoothing them with one-dimensional Kalman filtering.
[0062] During training, the first 60% of the dataset for each region can be used as the training set, and the last 40% as the test set. During the testing phase, events with an adjacent interval of no more than 7 time steps are grouped into the same event group. Under the condition of a fixed random seed of 42, 50% of the event groups are randomly selected to mask the external event text, while the text records of the remaining event groups are retained to simulate input scenarios with semantic missingness in certain time ranges.
[0063] The domain prior text is used to record the impact of possible exogenous event types on multivariate time series, without indicating whether the corresponding event has occurred at a specific time point. Specifically, the domain to which the dataset belongs and the distribution of potential typical events are input into a large language model using fixed prompt words to obtain domain prior text that has not been manually filtered. The domain prior text includes disaster types, variable responses, and response delays, thereby distinguishing it from exogenous event texts that describe meteorological events at specific time points or in the vicinity of a time range.
[0064] This invention performs one-dimensional average pooling on the input multivariate time series to obtain a low-frequency trend term; the residual sequence is obtained from the difference between the multivariate time series and the low-frequency trend term, expressed as:
[0065] in, Indicates low-frequency trend terms; Represents multivariate time series; This represents one-dimensional average pooling; Indicates the pooling kernel size; This indicates the pooling step size.
[0066]
[0067] in, This represents the residual sequence.
[0068] Subsequently, the residual sequence is subjected to non-overlapping average pooling and block embedding at different scales to obtain embedding features at each scale; the embedding features at each scale are encoded and concatenated according to the scale-independent mask to obtain multi-scale temporal features.
[0069] The expressions for the embedding features at each scale are as follows:
[0070] in, Represents the embedding feature at scale i; This indicates a block embedding operation; This represents the i-th scale multi-scale time series sequence obtained by downsampling; This represents the position code of the i-th scale; N represents the total number of scales.
[0071] The expression for the scale-independent mask is:
[0072] in, This represents the scale-independent mask value between the a-th feature label and the b-th feature label; These represent the a-th feature label and the b-th feature label, respectively. It represents an infinite quantity.
[0073] The scale-independent mask enables each feature tag to interact with self-attention only with feature tags of the same scale.
[0074] The process of encoding the embedding features at each scale includes: inputting the embedding features at each scale and the scale-independent mask into the Transformer encoder to obtain the temporal features at each scale, expressed as follows:
[0075] in, Represents the temporal features at the i-th scale; Indicates Transformer encoder; This indicates a scale-independent mask.
[0076] Different scales By concatenating the data in scale order, multi-scale temporal features are obtained. .
[0077] This step yields low-frequency trend terms and multi-scale time-series features. The low-frequency trend terms are used for trend back-injection in subsequent steps, and the obtained multi-scale time-series features are used for cross-modal interaction in subsequent steps.
[0078] Step S2: Generate endogenous statistical text based on the statistics of the time segments of the multivariate time series; Encode and semantically fuse exogenous event text, endogenous statistical text, and domain prior text respectively to obtain enhanced text features; In this step, the present invention first fills a structured prompt template with the statistics of time segments of a multivariate time series to generate endogenous statistical text.
[0079] The multivariate time series time segment represents a multivariate time series within a preset time window, and the statistics specifically include: mean, standard deviation, maximum value, and minimum value.
[0080] The calculated mean, standard deviation, maximum, and minimum values are filled into a structured prompt template according to preset fields to generate endogenous statistical text. The endogenous statistical text converts the numerical statistical state of the current time segment into textual semantics, which, together with the exogenous event text describing the current or nearby meteorological events, represents the current window.
[0081] The steps of encoding and semantically fusing exogenous event text, endogenous statistical text, and domain prior text to obtain enhanced text features specifically include: The exogenous event text and the endogenous statistical text are encoded to obtain joint text features; Encode the domain prior text to obtain domain prior features; Historical semantic features are obtained by retrieving historical semantic prototypes from the dynamic memory using the joint text features. Enhanced text features are obtained by fusing the domain prior features and historical semantic features.
[0082] Specifically, by jointly encoding exogenous event text and endogenous statistical text, the expression for the joint text features is as follows:
[0083] in, Indicates joint text features; Indicates a text encoder; Represents the text of an external event; This indicates endogenous statistical text.
[0084] In some embodiments, the text encoder may employ a GPT-2 with full parameter tuning based on task data.
[0085] Encoding the domain prior text yields the following expression for the domain prior features:
[0086] in, Represents prior features of the domain; Represents the prior text of the domain.
[0087] Historical semantic prototypes are retrieved from the dynamic memory using the joint text features to obtain historical semantic features, as described in the following details.
[0088] The dynamic memory of this invention includes K learnable pattern indexes. and K learnable historical semantic prototypes Using joint text features as the query, the relevance between the joint text features and each pattern index is calculated based on cosine similarity. After feature dimension square root scaling and Softmax normalization, the relevance weight of the i-th memory unit is obtained, expressed as:
[0089] in, This represents the relevance weight of the i'-th memory unit; Represents an exponential function; Represents the cosine similarity function; d represents the learnable pattern index of the i'-th memory unit; d represents the feature dimension; K represents the total number of memory units in the dynamic memory bank; This represents the learnable pattern index of the j-th memory unit.
[0090] Select the top-k memory units with the highest relevance weights, and perform weighted aggregation on the corresponding historical semantic prototypes to obtain the historical semantic features, expressed as:
[0091] in, Indicates historical semantic features; Indicates from the set of relevance weights The top-k high-weight indexes selected from the list; This represents the relevance weight of the i'-th memory unit; This represents the i'th learnable historical semantic prototype.
[0092] Multi-head cross-attention is performed using joint text features as the query and a comprehensive context feature pool as the key and value. Enhanced text features are then obtained through residual connections and layer normalization, expressed as follows:
[0093] in, This indicates enhanced text features; Indicates the layer normalization function; Indicates joint text features; This represents a multi-head cross-attention operation; Q, K, and V represent the query, key, and value of the cross-attention operation, respectively; C represents a comprehensive contextual feature pool formed by concatenating domain prior features and historical semantic features.
[0094] Step S3: Perform parallel one-dimensional convolution on the multi-scale temporal features and enhanced text features respectively to obtain consensus benchmark features and dynamic complementary features. Perform cross-modal interaction and gated residual interpolation on the consensus benchmark features and dynamic complementary features respectively to obtain fused features. In this step, the expressions for the consensus benchmark feature and the dynamic complementary feature are:
[0095]
[0096] in, This indicates the temporal dynamic complementary characteristics; Indicates the characteristics of the time-series consensus benchmark; This represents a one-dimensional convolution used to extract dynamic complementary features; This represents a one-dimensional convolution used to extract consensus benchmark features; Represents multi-scale temporal features. This indicates the dynamic complementary features of the text; Represents the benchmark features of textual consensus; This represents a one-dimensional convolution used to extract dynamic complementary features; This represents a one-dimensional convolution used to extract consensus benchmark features.
[0097] The specific steps for obtaining the fused features by performing cross-modal interaction and gated residual interpolation on the consensus benchmark features and dynamic complementary features respectively include: The consensus benchmark features are used to perform cross-modal interaction to obtain the baseline representation; the dynamic complementary features are used to perform cross-modal interaction to obtain the joint representation; the mean square offset of the joint representation relative to the baseline representation is used as the coupling conflict divergence, and the increment of the joint representation relative to the baseline representation is gated residual interpolation with a security fusion weight that monotonically decreases as the coupling conflict divergence increases to obtain the fusion feature.
[0098] Specifically, firstly, cross-modal interaction is performed on the time-series consensus baseline features and the text-based consensus baseline features to obtain the baseline representation, expressed as:
[0099] in, Indicates baseline characterization; This represents cross-modal attention operations; Q, K, and V represent query, key, and value, respectively.
[0100] Based on the interaction of consensus benchmark features, cross-modal interaction is performed on dynamic complementary features to obtain a joint representation, expressed as:
[0101] in, Indicates joint representation; This indicates a cross-modal attention operation that sequentially receives queries, keys, and values.
[0102] Subsequently, secure fusion weights are calculated based on the mean square offset of the joint representation relative to the baseline representation, and fusion features are obtained by gated residual interpolation, specifically including: Calculate the mean square offset of the joint representation relative to the baseline representation in the d-dimensional feature space, and use the obtained mean square offset as the coupling conflict divergence:
[0103] Where D represents the coupling conflict divergence; This represents the eigenvalues of the joint representation in the i''th dimension; The baseline represents the eigenvalues in the i''th dimension.
[0104] Calculate the secure fusion weights based on the coupling conflict divergence:
[0105] in, Indicates the security fusion weight; This refers to the Softplus function; This represents the learnable scalar scaling parameter.
[0106] The increment of the joint representation relative to the baseline representation is gated residual interpolated using secure fusion weights, and then the fused features are obtained after layer normalization:
[0107] in, Indicates fusion characteristics; This indicates element-wise multiplication.
[0108] When the enhanced text features are consistent with the multi-scale temporal features, the mean squared offset between the joint representation and the baseline representation is small, and the secure fusion weights allow dynamic complementary semantics to participate in the fusion. When the external event text contains redundant, mismatched, or weakly related content, the mean squared offset between the two branches increases, and the secure fusion weights continuously decrease, thereby smoothly suppressing the increment of the joint representation relative to the baseline representation. This approach does not use hard gating to delete text context, but rather retains consensus benchmark information and temporally dynamic complementary information, continuously controlling text increments that may cause cross-modal negative transfer.
[0109] Through this step, the consensus benchmark interacts across modalities to form a stable baseline representation, while the dynamic complementary cross-modal interaction forms a joint representation containing additional semantics. The offset between the two directly determines the fusion strength of the additional semantics. The resulting fused features preserve reliable textual interpretations while suppressing unreliable textual interference, and serve as input for multi-scale decoding in subsequent steps.
[0110] Step S4: The fused features are split by scale and upsampled, decoded under partial visible cross-scale mask constraints, and then reconstructed by back-injection of the low-frequency trend term and adaptive convolution smoothing. In this step, the expression for splitting and upsampling the fused features by scale is:
[0111] in, Represents the upsampled features at the i-th scale; This represents the linear interpolation operation corresponding to the i-th scale; This represents the i-th scale fusion feature obtained by splitting from the fusion features.
[0112] A partially visible cross-scale mask is constructed based on the dimensional relationships between scales, expressed as follows:
[0113] in, This represents the partially visible cross-scale mask value of the query position at scale i to the context position at scale j during decoding; i and j represent the query scale index and the accessed scale index, respectively. This represents the scale number corresponding to the i-th position; This represents the scale number corresponding to the j-th position.
[0114] Decoding is performed under partially visible multi-scale mask constraints, expressed as follows:
[0115] in, This represents the decoding result at the i-th scale; This represents a Transformer decoder constrained by a partially visible cross-scale mask. This represents the upsampling feature at the i-th scale.
[0116] The mask allows the coarse-scale global structure to guide the recovery of the fine-scale step by step, while preventing fine-grained information from leaking directly to the coarse scale.
[0117] The low-frequency trend term is back-injected into the decoding results at each scale according to the scale to obtain the reconstruction results at each scale. Adaptive convolution smoothing is then applied to the reconstruction results at each scale to obtain the reconstruction sequence, expressed as:
[0118]
[0119] in, This represents the reconstruction result at the i-th scale; This represents the low-frequency trend term downsampled to the i-th scale; Indicates the reconstructed sequence; This indicates an adaptive convolution smoothing operation; This indicates the result to be smoothed after merging the reconstruction results from various scales.
[0120] In this step, the multi-scale residual features recover local dynamics under directed cross-scale context constraints, while the low-frequency trend term is used to recover the overall evolutionary components of the multivariate time series. The resulting reconstructed sequence corresponds to the input multivariate time series at each time point and is used for loss calculation and multi-scale reconstruction error calculation in subsequent steps.
[0121] Step S5: Calculate the time-point anomaly score based on the multi-scale reconstruction error and memory pattern deviation of the reconstructed sequence and the multivariate time series, and threshold the time-point anomaly score to obtain the anomaly detection result.
[0122] In this step, the multi-scale reconstruction anomaly score after dynamic tolerance is first calculated based on the squared errors of the reconstructed sequence and the multivariate time series at each scale. Specifically, this includes: For the k-th scale, calculate the point-to-point squared error between the true scale sequence and the corresponding reconstructed sequence:
[0123] in, This represents the point-to-point squared error at the k-th scale; This represents the k-th scale true multivariate time series used for anomaly scoring; This represents the reconstructed sequence at the k-th scale.
[0124] The local mean at the k-th scale is calculated using one-dimensional average pooling with a kernel size of 5 and a step size of 1:
[0125] in, This represents the local mean at the k-th scale; This represents a one-dimensional average pooling operation; Represents the true multivariate time series at the k-th scale; This indicates that the pooling kernel size is 5; This indicates that the pooling step size is 1.
[0126] Construct a dynamic tolerance boundary based on the fluctuation of the true-scale sequence relative to the local mean:
[0127] in, Represents the dynamic tolerance boundary at the k-th scale; The scaling factor represents the normal local fluctuations; This represents a positive stabilizing term that improves numerical stability.
[0128] Subtract the dynamic tolerance boundary from the point-to-point squared error, and use ReLU to retain the portion exceeding the dynamic tolerance boundary:
[0129] in, This represents the reconstructed anomaly score after deducting normal fluctuations at the k-th scale; This represents the linear rectification function.
[0130] The reconstructed anomaly scores at each scale are linearly interpolated to the original sequence length and then summed.
[0131] in, Indicates the multi-scale reconstruction anomaly score; Represents the interpolation function; Represents the reconstructed anomaly score at the k-th scale; This indicates that the output length is set to the original multivariate time series length L; This indicates that a linear interpolation mode is used.
[0132] Subsequently, based on the cosine distance between the current query vector and the memory feedback prototype, the memory pattern deviation is obtained. The multi-scale reconstruction anomaly score and the memory pattern deviation are fused to obtain the time-point anomaly score, which is then thresholded to obtain the anomaly detection result, as detailed below.
[0133] Let q be the query vector corresponding to the current input, and m be the retrieved memory feedback prototype. The memory feedback prototype is formed by weighting the relevance weights obtained from dynamic memory retrieval with the corresponding historical semantic prototypes:
[0134] Where m represents the memory feedback prototype; This represents the set of memory cell indexes with the highest relevance weights; This represents the relevance weight of the i'-th memory unit; This represents the i'th historical semantic prototype.
[0135] Calculate the cosine distance between the current query vector and the memory feedback prototype to obtain the memory pattern deviation:
[0136] in, q represents the deviation from the memory pattern; q represents the current input query vector. This represents the inner product of the query vector q and the memory feedback prototype m; and Let represent the L2 norms of q and m, respectively.
[0137] The anomaly score at each time point is obtained by adding the multi-scale reconstruction anomaly score to the memory pattern deviation score adjusted by weighting coefficients:
[0138] in, Indicates the outlier score at each time point; Indicates the multi-scale reconstruction anomaly score; Indicates the degree of deviation from the memory pattern; This represents the weighting coefficient of memory pattern deviation in the outlier score at each time point.
[0139] The anomaly score at each time point is compared with an anomaly detection threshold. When the anomaly score at a corresponding time point is higher than the threshold, 1 is output and that time point is classified as an anomaly; when the anomaly score at a corresponding time point is not higher than the threshold, 0 is output and that time point is classified as normal. This yields the anomaly detection results corresponding to each time point of the input multivariate time series.
[0140] In this step, the model utilizes four types of losses to complete normal temporal reconstruction, morphological alignment, cross-modal alignment, and feature decoupling. During the inference phase, it simultaneously measures numerical reconstruction deviation and historical semantic pattern deviation. A dynamic tolerance boundary is used to subtract normal local fluctuations, and the thresholded time-point anomaly scores form the final anomaly detection result.
[0141] The present invention also includes the step of training the sequence anomaly detection model, comprising: constructing a loss function using reconstruction loss, morphological alignment loss, cross-modal contrast loss and decoupling loss, and training the model based on the loss function.
[0142] The details of the losses are as follows: Define the reconstruction residual as The residuals are reconstructed using the standard Huber loss metric:
[0143] in, This represents the Huber loss corresponding to the reconstructed residual e; e represents the difference between the input multivariate time series and the reconstructed sequence. Represents the absolute value of e; This represents the threshold used in Huber loss to distinguish between the quadratic error interval and the linear error interval. When When using quadratic terms, when When using a linear growth term.
[0144] Applying focus weights to the Huber loss yields the reconstruction loss:
[0145] in, Represents the reconstruction loss; N' represents the total number of error elements averaged by the current reconstruction loss; i''' represents the index of the error element; This represents the i'''th reconstructed residual; This represents the Huber loss corresponding to the i'''th reconstruction residual; 'Indicates the focus modulation index.
[0146] DILATE loss is used as the morphological alignment loss:
[0147] in, This indicates the shape alignment loss; This represents the DILATE loss function used to constrain sequence shape and time alignment. Indicates the reconstructed sequence; This represents the input multivariate time series.
[0148] The fused features and enhanced text features are respectively subjected to mean pooling and then projected through the corresponding multilayer perceptron to obtain the global embedding for cross-modal comparison:
[0149] in, This represents the temporal global embedding obtained after mean pooling and multilayer perceptron projection of the fused features; This represents the global text embedding obtained after applying mean pooling and multilayer perceptron projection to the enhanced text features. and These represent the temporal-side multilayer perceptron and the text-side multilayer perceptron, respectively. This indicates mean pooling; Indicates fusion characteristics; This indicates enhanced text features.
[0150] For the m-th time-series global embedding within the batch and the nth text global embedding Calculate the cosine similarity:
[0151] in, This represents the cosine similarity between the m-th temporal global embedding and the n-th text global embedding within a batch. This represents the i-th temporal global embedding; This indicates the global embedding of the j-th text. Represents global embedding of timing The second norm; Indicates global text embedding The 2-norm.
[0152] The InfoNCE loss from temporal global embedding to text global embedding is:
[0153] in, The contrastive loss represents the difference between temporal global embedding and text global embedding; N'' represents the number of samples in the batch. Represents the natural logarithm function; This represents the cosine similarity between the k-th time-series sample and its paired text sample; This represents the cosine similarity between the k-th time series sample and the l-th text sample; This indicates the temperature parameter for comparison and learning.
[0154] The InfoNCE loss from text global embedding to temporal global embedding is:
[0155] in, This represents the contrastive loss from text global embedding to temporal global embedding; Let represent the cosine similarity between the l-th time series sample and the k-th text sample.
[0156] The cross-modal contrast loss is obtained by averaging the InfoNCE losses from the two directions:
[0157] in, This represents the cross-modal contrast loss.
[0158] Decoupling loss is used to reduce the cosine similarity between the consensus baseline features and dynamic complementary features of both temporal and textual modalities:
[0159] in, N''' represents the decoupling loss; N''' represents the number of features involved in the average calculation of the current decoupling loss. This represents the cosine similarity function.
[0160] The total loss of the model is:
[0161] in, This represents the total loss of the model; Indicates the losses incurred during reconstruction; This indicates the shape alignment loss; Indicates cross-modal contrast loss; Indicates the decoupling loss; , and These represent the weight coefficients for morphological alignment loss, cross-modal contrast loss, and decoupling loss, respectively.
[0162] In this embodiment, the sliding window length is set to 128, the batch size to 32, the data block length to 8, the hidden layer dimension to 96, the number of attention heads to 4, the number of encoder and decoder layers to 2 each, and the convolutional kernel sizes for the multi-scale feature extraction process to 12, 8, 4, and 2; the total model loss is... , and All are set to 0.01. The learning rate for fine-tuning the large language model is set to... The dynamic memory capacity and number of events were both set to 32, the number of queries was set to 5, the top-k number of retrieved related items was set to 10, and the state decay rate was set to 0.95. Pure temporal warm-up and multimodal joint training were each executed for 30 rounds, and the initial learning rate of the backbone network was set to... The early stopping patience value is set to 3. The model is implemented using PyTorch on a single A10 GPU.
[0163] This invention uses VUS-ROC and VUS-PR as evaluation metrics, denoted as ROC and PR, respectively. Each method was tested three times independently on five datasets, and the average values were taken. The results are shown in Table 1. Each cell in the table gives the ROC / PR.
[0164] Table 1
[0165] As shown in Table 1, SISAD achieves the highest VUS-ROC and VUS-PR values on the California, Carolina, New York, and Texas datasets; it also achieves the highest VUS-ROC value of 0.8276 and a VUS-PR of 0.2204 on the Florida dataset. These results demonstrate that our method outperforms the comparison methods on most datasets and key metrics.
[0166] This invention conducted component ablation experiments. Four variants were obtained by removing dynamic memory and attention retrieval, removing joint decoupling fusion, removing dynamic memory, endogenous text, and domain priors, and removing joint representation branches. Results on the California, Carolina, and New York datasets are shown in Table 2, with each cell in the table providing the ROC / PR.
[0167] Table 2
[0168] As shown in Table 2, the complete model outperforms all ablation variants on both metrics across the three datasets, indicating that dynamic memory, textual information, joint decoupling fusion, and joint representation branches all effectively contribute to the anomaly detection results. The text perturbation experiments of this invention also show that conflict divergence and anomaly scores increase when irrelevant text is input; however, when a mixture of event text and irrelevant text is input, the anomaly score generally approaches that of the original event text scenario, demonstrating that the semantic interference suppression module can preserve effective event semantics and weaken the influence of irrelevant information.
[0169] While the specific embodiments of the present invention depict actions or steps in a particular order, this should be understood as requiring such actions or steps to be performed in the specific order shown or in sequential order, or requiring all illustrated actions or steps to be performed to achieve the desired result. In certain environments, multitasking and parallel processing may be advantageous. Similarly, although several specific implementation details are included in the above discussion, these should not be construed as limiting the scope of this disclosure. Certain features described in the context of individual embodiments may also be implemented in combination in a single implementation. Conversely, various features described in the context of a single implementation may also be implemented individually or in any suitable sub-combination in multiple implementations.
[0170] 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 changes or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in the present invention should be included within the scope of protection of the present invention.
Claims
1. A text enhancement time series anomaly detection method based on semantic interference suppression, characterized in that, Includes the following steps: Step S1: Obtain the multivariate time series and the corresponding external event text, and obtain the domain prior text; the multivariate time series includes meteorological information at multiple time points, and the external event text includes meteorological text at multiple time points; the domain prior text records the influence of the types of external events on the multivariate time series; Trend decoupling is performed on the input multivariate time series to obtain low-frequency trend terms and residual sequences; multi-scale downsampling and scale-independent encoding are performed on the residual sequences to obtain multi-scale time series features; Step S2: Generate endogenous statistical text based on the statistics of the time segments of the multivariate time series; Encode and semantically fuse exogenous event text, endogenous statistical text, and domain prior text respectively to obtain enhanced text features; Step S3: Perform parallel one-dimensional convolution on the multi-scale temporal features and enhanced text features respectively to obtain consensus benchmark features and dynamic complementary features. Perform cross-modal interaction and gated residual interpolation on the consensus benchmark features and dynamic complementary features respectively to obtain fused features. Step S4: The fused features are split by scale and upsampled, decoded under partial visible cross-scale mask constraints, and then reconstructed by back-injection of the low-frequency trend term and adaptive convolution smoothing. Step S5: Calculate the time-point anomaly score based on the multi-scale reconstruction error and memory pattern deviation of the reconstructed sequence and the multivariate time series, and threshold the time-point anomaly score to obtain the anomaly detection result.
2. The method of claim 1, wherein, In step S1, the meteorological information includes evaporation, average wind speed, precipitation, average temperature, maximum temperature, minimum temperature, maximum water temperature, minimum water temperature, average surface temperature, minimum surface temperature, snow depth, snowfall, 2-minute wind direction angle, 5-second wind direction angle, maximum wind speed within 2 minutes, and maximum wind speed within 5 seconds; the meteorological text includes text records of storms, hurricanes, and flood-related events.
3. The method of claim 2, wherein the method further comprises: In step S1, the expressions for the low-frequency trend term and the residual sequence are: wherein, represents a low frequency trend term; represents a multivariate time series; represents one-dimensional average pooling; represents a pooling kernel size; represents a pooling stride; wherein denotes the residual sequence; In step S1, the expression for the multi-scale temporal features is: wherein, denotes the i-th scale temporal feature; denotes the Transformer encoder; denotes the scale-independent mask; denotes the i-th scale embedding feature; denotes the patch embedding operation; denotes the i-th scale multi-scale temporal sequence obtained by down-sampling; denotes the position encoding of the i-th scale; N denotes the total number of scales; denotes the scale-independent mask value between the a-th feature mark and the b-th feature mark; denote the a-th feature mark and the b-th feature mark, respectively.
4. The method of claim 3, wherein, In step S2, the step of generating endogenous statistical text based on the statistics of the time segments of the multivariate time series specifically includes: calculating the mean, standard deviation, maximum value, and minimum value of the time segments of the multivariate time series; filling the mean, standard deviation, maximum value, and minimum value into a structured prompt template according to preset fields to generate endogenous statistical text; The steps of encoding and semantically fusing exogenous event text, endogenous statistical text, and domain prior text to obtain enhanced text features specifically include: The exogenous event text and the endogenous statistical text are encoded to obtain joint text features; Encode the domain prior text to obtain domain prior features; Historical semantic features are obtained by retrieving historical semantic prototypes from the dynamic memory using the joint text features. Enhanced text features are obtained by fusing the domain prior features and historical semantic features.
5. The method of claim 4, wherein the method further comprises: The expression for the joint text feature is: wherein, represents joint text features; represents text encoder; represents exogenous event text; represents endogenous statistical text; The expression for the domain prior features is: wherein, represents a domain prior feature; represents a domain prior text; The expression for the historical semantic feature is: in, Indicates historical semantic features; Indicates from the set of relevance weights The top-k high-weight indexes selected from the list; This represents the relevance weight of the i'-th memory unit; This represents the i'th learnable historical semantic prototype; This represents the relevance weight of the i'-th memory unit; Represents an exponential function; Represents the cosine similarity function; d represents the learnable pattern index of the i'-th memory unit; d represents the feature dimension; K represents the total number of memory units in the dynamic memory bank; Represents the learnable pattern index of the j-th memory unit; The expression for the enhanced text feature is: in, This indicates enhanced text features; Indicates the layer normalization function; Indicates joint text features; This represents a multi-head cross-attention operation; Q, K, and V represent the query, key, and value of the cross-attention operation, respectively; C represents a comprehensive contextual feature pool formed by concatenating domain prior features and historical semantic features.
6. The text-enhanced time series anomaly detection method based on semantic interference suppression according to claim 5, characterized in that, In step S3, the expressions for the consensus benchmark feature and the dynamic complementary feature are: in, This indicates the temporal dynamic complementary characteristics; Indicates the characteristics of the time-series consensus benchmark; This represents a one-dimensional convolution used to extract dynamic complementary features; This represents a one-dimensional convolution used to extract consensus benchmark features; Represents multi-scale temporal features. This indicates the dynamic complementary features of the text; Represents the benchmark features of textual consensus; This represents a one-dimensional convolution used to extract dynamic complementary features; This represents a one-dimensional convolution used to extract consensus benchmark features; In step S3, the specific steps of performing cross-modal interaction and gated residual interpolation on the consensus benchmark features and dynamic complementary features to obtain the fused features include: The consensus benchmark features are used to perform cross-modal interaction to obtain the baseline representation; the dynamic complementary features are used to perform cross-modal interaction to obtain the joint representation; the mean square offset of the joint representation relative to the baseline representation is used as the coupling conflict divergence, and the increment of the joint representation relative to the baseline representation is gated residual interpolation with a security fusion weight that monotonically decreases as the coupling conflict divergence increases to obtain the fusion feature.
7. The text-enhanced time series anomaly detection method based on semantic interference suppression according to claim 6, characterized in that, The expression for the baseline representation is: in, Indicates baseline characterization; This represents cross-modal attention operations; Q, K, and V represent query, key, and value, respectively. The expression for the joint representation is: in, Indicates joint representation; This indicates a cross-modal attention operation that sequentially receives queries, keys, and values; The expression for the coupling conflict divergence is: Where D represents the coupling conflict divergence; The joint representation is in the first... eigenvalues of dimension; Indicates baseline characterization at the 1st eigenvalues of dimension The expression for the security fusion weight is: in, Indicates the security fusion weight; This refers to the Softplus function; Represents a learnable scalar scaling parameter; The expression for the fusion feature is: in, Indicates fusion characteristics; This indicates element-wise multiplication.
8. The text-enhanced time series anomaly detection method based on semantic interference suppression according to claim 7, characterized in that, In step S4, the expression for splitting and upsampling the fused features by scale is: in, Represents the upsampled features at the i-th scale; This represents the linear interpolation operation corresponding to the i-th scale; This represents the i-th scale fusion feature obtained by splitting from the fusion features; The expression for decoding under partially visible multi-scale mask constraints is: in, This represents the partially visible cross-scale mask value of the query position at scale i to the context position at scale j during decoding; i and j represent the query scale index and the accessed scale index, respectively. This represents the scale number corresponding to the i-th position; This represents the scale number corresponding to the j-th position; This represents the decoding result at the i-th scale; This represents a Transformer decoder constrained by a partially visible cross-scale mask. Represents the upsampled features at the i-th scale; In step S4, the expression for the reconstructed sequence obtained after back-injection and adaptive convolution smoothing of the low-frequency trend term is: in, This represents the reconstruction result at the i-th scale; This represents the low-frequency trend term downsampled to the i-th scale; Indicates the reconstructed sequence; This indicates an adaptive convolution smoothing operation; This indicates the result to be smoothed after merging the reconstruction results from various scales.
9. The text-enhanced time series anomaly detection method based on semantic interference suppression according to claim 8, characterized in that, In step S5, the expression for the multi-scale reconstruction anomaly score is: in, This represents the local mean at the k-th scale; This represents a one-dimensional average pooling operation; Represents the true multivariate time series at the k-th scale; This indicates that the pooling kernel size is 5; This indicates that the pooling step size is 1; This represents the point-to-point squared error at the k-th scale; Represents the dynamic tolerance boundary at the k-th scale; The scaling factor represents the normal local fluctuations; Represents a positive stabilizing term that improves numerical stability; This represents the reconstructed anomaly score after deducting normal fluctuations at the k-th scale; Represents a linear rectified function; Indicates the multi-scale reconstruction anomaly score; N represents the total number of scales; Represents the interpolation function; Represents the reconstructed anomaly score at the k-th scale; This indicates that the output length is set to the original multivariate time series length L; This indicates that a linear interpolation mode is used; The expression for the memory pattern deviation is: in, q represents the deviation from the memory pattern; q represents the current input query vector. This represents the inner product of the query vector q and the memory feedback prototype m; and Let represent the L2 norms of q and m respectively; m represents the memory feedback prototype. This represents the set of memory cell indexes with the highest relevance weights; This represents the relevance weight of the i'-th memory unit; This represents the i'th historical semantic prototype; The expression for the time-point anomaly score is: in, Indicates the outlier score at each time point; Indicates the multi-scale reconstruction anomaly score; Indicates the degree of deviation from the memory pattern; This represents the weighting coefficient of memory pattern deviation in the outlier score at each time point.
10. The text-enhanced time series anomaly detection method based on semantic interference suppression according to claim 9, characterized in that, It also includes the step of training a sequence anomaly detection model, including: constructing a loss function using reconstruction loss, morphological alignment loss, cross-modal contrast loss and decoupling loss, and training the model based on the loss function; The expression for the reconstruction loss is: in, N' represents the reconstruction loss; N' represents the total number of error elements averaged by the current reconstruction loss. Indicates the index of the error element; Indicates the first One reconstructed residual; Indicates the first Huber loss corresponding to each reconstructed residual; 'Indicates the focus modulation index; The expression for the morphological alignment loss is: in, This indicates the shape alignment loss; This represents the DILATE loss function used to constrain sequence shape and time alignment. Indicates the reconstructed sequence; This represents the input multivariate time series; The expression for the cross-modal contrast loss is: in, Indicates cross-modal contrast loss; This represents the contrastive loss from temporal global embedding to textual global embedding; This represents the contrastive loss from text-based global embedding to temporal-based global embedding; The expression for the decoupling loss is: in, Indicates the decoupling loss; This indicates the number of features involved in the calculation of the current average decoupling loss; Represents the cosine similarity function The total loss of the model is: in, This represents the total loss of the model; , and These represent the weight coefficients for morphological alignment loss, cross-modal contrast loss, and decoupling loss, respectively.
