Time series anomaly detection method based on multi-window cluster symbolization

Abstract

The application discloses a time series anomaly detection method based on multi-window clustering symbolization, and relates to the technical field of anomaly detection, and comprises the following steps: a cross-window feature drift capture baseline is established, an input time series is segmented according to a sliding time window, feature distribution of each time window is extracted, and a continuous mapping model is constructed; a transition gradient of a clustering centroid is extracted according to feature distribution change between adjacent time windows; and a dynamic drift monitoring matrix used for describing dynamic change of the centroid is generated. The application realizes continuous update of symbol mapping in time and space by cross-window feature drift capture and synchronous remapping of the centroid, eliminates faults and failures caused by sudden change of the centroid, and improves accuracy and stability of anomaly detection; through abnormal mapping freezing control and symbol consistency closed loop, a self-healing regulation mechanism is constructed, so that the system has self-recovery and self-stabilization capability, and continuity, reliability and robustness of the detection process are ensured.

Description

Technical Field

[0001] This invention relates to the field of anomaly detection technology, and specifically to a time series anomaly detection method based on multi-window clustering symbolization. Background Technology

[0002] Time series anomaly detection based on multi-window clustering and symbolic representation refers to an intelligent anomaly identification method that integrates multi-scale segmentation of time series data, feature clustering, and symbolic representation. Its core idea is as follows: First, the time series is subjected to dimensionality reduction (e.g., PAA segmentation) within different time windows, extracting multi-dimensional features such as mean, trend, and fluctuation within each window. Then, a clustering algorithm (e.g., K-means) is used to cluster the multi-window features, mapping similar features to the same symbols, thus achieving a symbolic representation of the time series. Next, a word set is constructed based on the obtained symbolic sequence, and multiple indicators such as frequency, coverage, and distance are used to measure the degree of anomaly in the symbolic segments, forming an interpretable anomaly score. This method achieves a unified characterization of local trends and global changes in the sequence through multi-window clustering, and reduces the complexity of high-dimensional data through symbolic representation, improving the model's anomaly identification capability and generalization accuracy in complex scenarios such as medical signals, industrial monitoring, and IoT sensing.

[0003] However, existing technologies still have the following shortcomings: During the real-time clustering update phase, the abrupt changes in feature distribution triggered by the rolling time window can cause discontinuous transitions in cluster centroids within a short period, leading to the failure of the spatiotemporal reference of the symbol mapping table. At this time, old symbols are still referenced in the buffer, while the new centroids have not yet been synchronously replaced, creating a symbol index gap. This gap prevents features from being correctly mapped to the valid symbol space for some time segments, resulting in gaps or repeated coding areas in the symbol sequence, which in turn triggers a cascading failure of the anomaly detection chain. In severe cases, the anomaly distribution model may drift, causing the overall anomaly identification results to skew towards the normal range, leading to widespread misjudgments or system freezes, thus compromising the continuity and traceability of time series analysis. Summary of the Invention

[0004] To address the above problems, the purpose of this invention is to provide a time series anomaly detection method based on multi-window clustering symbolization, the technical solution of which is as follows: A time series anomaly detection method based on multi-window clustering symbolization includes the following steps: Step 1: Establish a cross-window feature drift capture baseline. Divide the input time series into segments according to the sliding time window, extract the feature distribution of each time window, and construct a continuous mapping model. Extract the transition gradient of the cluster centroid based on the feature distribution changes between adjacent time windows, and generate a dynamic drift monitoring matrix to characterize the dynamic changes of the centroid. Step 2: Based on the dynamic drift monitoring matrix, a centroid synchronous remapping network is constructed. The reverse phase registration algorithm is used to continuously and smoothly adjust the discontinuous transitions of the cluster centroids between adjacent time windows. The gradient information of the dynamic drift monitoring matrix is ​​converted into symbol mapping correction weights to establish a traceable symbol transfer path between the old and new cluster centroids. Step 3: Based on the symbol transfer path, generate a time-delay adaptive write-back chain, monitor the synchronization offset of the cluster centroid in real time, and use a hysteresis feedback structure to dynamically write back and update the spatiotemporal index in the symbol mapping table, so that the symbol mapping has a gradual decay characteristic during the write-back process, thereby realizing the continuous and smooth transition of the symbol mapping table. Step 4: Based on the stable output of the time-delay adaptive write-back chain, execute abnormal mapping freeze control, lock the symbol mapping table of the high-risk area of ​​centroid transition, and set the freeze constraint threshold according to the decay weight of the write-back chain to prevent the symbol space from collapsing locally when the cluster centroid changes abruptly. Step 5: Based on the output of the anomaly mapping freeze control, construct a symbol consistency verification closed loop. Use a dual-channel anomaly measurement mechanism to cross-compare symbol drift and anomaly score deviation, and feed the consistency correction signal back to the cross-window feature drift capture baseline, thereby forming a self-healing synchronization loop. This achieves stable closed-loop control of dynamic updating of cluster centroids, thereby improving the continuity and reliability of time series anomaly detection.

[0005] Optionally, the steps for establishing a cross-window feature drift capture baseline include: The input time series is segmented into continuous time window sequences according to the time order, and an overlap ratio is set between adjacent time windows to maintain the continuity of time series features. For each time window, extract multi-dimensional feature indicators such as statistical mean, fluctuation range, trend direction, short-term stability and local change rate from the data to construct a feature vector with time series representativeness; A continuous mapping model is constructed based on the feature vectors of adjacent time windows. By comparing the similarity and change magnitude of the feature distributions of adjacent windows, the continuous evolution relationship of features over time is recorded. The transition gradient of the cluster centroid is extracted based on the continuous mapping model, and a dynamic drift monitoring matrix is ​​generated with the gradient as a reference. The dynamic drift monitoring matrix is ​​arranged with the time window as the row and the feature dimension as the column, and is used to record the drift direction and magnitude of each feature dimension.

[0006] Optionally, the dynamic drift monitoring matrix introduces a time-related weighting factor during the generation process, so that the influence weight of the cluster centroid transition gradient in the most recent time window is higher than that in the distant window in the matrix update, thereby enhancing the response capability to real-time feature drift and reflecting the direction and magnitude of centroid changes in real time when abrupt changes occur in the time series, providing a stable time reference for subsequent centroid synchronous remapping.

[0007] Optionally, the steps of constructing a centroid synchronization remapping network based on the dynamic drift monitoring matrix include: The time window feature drift information in the dynamic drift monitoring matrix is ​​analyzed to determine the direction and intensity of change of the cluster centroid within each time window, and the non-continuous transition region of the centroid is identified based on the gradient magnitude. Based on the continuous feature distribution information recorded in the dynamic drift monitoring matrix, a synchronous reference relationship between the old centroid and the new centroid is constructed, and the spatial positions of the two are reversed to form a continuous transition path in the feature space. The gradient information in the dynamic drift monitoring matrix is ​​converted into sign mapping correction weights, and the sign mapping relationship is dynamically corrected based on the magnitude and direction of the centroid. Based on the symbol mapping correction weights, a traceable symbol transfer path is established between the old and new cluster centroids to achieve continuous transition of the symbol mapping table in the time and space dimensions.

[0008] Optionally, the symbol mapping correction weight is adaptively adjusted according to the gradient magnitude of each time window in the dynamic drift monitoring matrix. When the change magnitude of the centroid is small, the symbol mapping relationship remains unchanged. When the change magnitude of the centroid exceeds the preset threshold, the correction weight is gradually increased so that the symbol mapping can achieve smooth decay and update within the continuous time window, thereby avoiding the local collapse of the symbol space during the centroid transition process.

[0009] Optionally, the step of generating a time-delay adaptive write-back chain based on the symbol transfer path includes: Based on the symbol transfer path, the synchronization offset of the cluster centroid in the time dimension is monitored. By comparing the current time window centroid position with the reference centroid position, the offset reflecting the synchronization accuracy is obtained. Based on the cluster centroid synchronization offset, a feedback structure with hysteresis characteristics is constructed on the time axis to enable the symbol mapping table write-back process to have time delay response capability, so as to prevent over-correction caused by short-term mutations. The dynamic write-back update process is initiated based on the hysteresis feedback structure. The spatiotemporal index of the symbol mapping table is written back sequentially according to the offset direction and magnitude, and a progressive decay weight is introduced to achieve a smooth transition. The output of the time-delay adaptive write-back chain is subjected to stability verification and synchronization calibration. The update is refined through recursive feedback to ensure that the synchronization deviation between the symbol mapping table and the cluster centroid converges to a predetermined threshold.

[0010] Optionally, the progressive decay weight is dynamically adjusted according to the time distance of the cluster centroid synchronization offset. As the time window spacing increases, the update magnitude is gradually reduced, so that the symbol mapping table forms a smooth decay trajectory from near to far during the write-back process, thereby maintaining the temporal continuity and spatial stability of the symbol mapping when the cluster centroid changes continuously.

[0011] Optionally, the steps for performing anomaly mapping freeze control based on the stable output of the time-delay adaptive write-back chain include: After the time-delay adaptive write-back chain stabilizes, the spatiotemporal evolution of the symbol map is detected. Based on the rate, direction, and magnitude of symbol map changes, high-risk areas of cluster centroid transitions are identified, and a frozen monitoring list is established. The freeze constraint threshold is determined based on the decay weight of the adaptive write-back chain with time delay. For high-risk areas that exceed the threshold, symbol mapping table locking is performed to make the freeze range correspond to the risk level. After the symbol mapping table is locked, stability maintenance is performed on the frozen region by adjusting the decay rate of surrounding symbols to maintain boundary smoothness and monitoring the drift convergence status. Once the drift amplitude of the frozen region returns to stability, the frozen region is gradually unlocked based on the write-back chain output signal to achieve smooth recovery and continuous connection of the symbol mapping table.

[0012] Optionally, the unlocking and recovery of the frozen region is achieved by dynamically evaluating the decay weight of the time-delay adaptive write-back chain. When the decay weight recovers to a stable threshold range, the update frequency is gradually increased in a progressive decay manner, so that the symbol mapping of the frozen region and the mapping change of the adjacent non-frozen region can be smoothly transitioned in the time dimension, and the synchronous correspondence with the cluster centroid is re-established.

[0013] Optionally, the steps for constructing a symbolic consistency verification closed loop based on the output of the anomaly mapping freeze control include: After the abnormal mapping freeze control is completed, the symbol mapping table state output by the freeze control phase is analyzed. By comparing the historical state with the current state, key features reflecting symbol consistency are extracted to obtain the baseline parameters of the symbol space. Based on the baseline parameters of the symbol space, a dual-channel anomaly measurement mechanism is constructed to detect the continuity deviation of symbol drift from the time dimension and evaluate the deviation of anomaly scores from the spatial dimension, so as to achieve joint evaluation of the symbol mapping consistency state. The output results of the dual-channel anomaly measurement mechanism are cross-compared to generate a consistency correction signal, and the delay adaptive write-back chain and centroid synchronization remapping process are adjusted according to the deviation between the time channel and the spatial channel, respectively. The consistency correction signal is fed back to the cross-window feature drift capture baseline to form a self-healing synchronization loop, thereby achieving stable closed-loop control of dynamic updating of cluster centroids.

[0014] The beneficial effects of the technical solution provided by this invention include at least the following: This invention introduces a cross-window feature drift capture baseline and centroid synchronous remapping mechanism during time series clustering and symbolization, enabling continuous tracking and stable control of the dynamic changes in cluster centroids. Through the synergistic effect of the dynamic drift monitoring matrix and the time-delay adaptive write-back chain, the symbol mapping table maintains smooth temporal continuity and consistent spatial correspondence during centroid updates, fundamentally eliminating symbol fragmentation and index failure problems caused by centroid mutations. This technology gives the symbolized representation of time series spatiotemporal continuity, allowing the symbol sequence to stably reflect changes in the original feature distribution, thereby significantly improving the accuracy and stability of anomaly detection.

[0015] This invention achieves self-healing regulation of the dynamic evolution of the symbol space by constructing a closed loop of anomaly mapping freeze control and symbol consistency verification. The system automatically executes freeze protection upon detecting abrupt changes in cluster centroids, gradually unlocks the symbol space during the recovery phase based on the write-back chain decay weight, and performs consistency correction feedback through a dual-channel anomaly measurement mechanism, forming an adaptive closed-loop adjustment structure throughout the detection process. This mechanism enables the symbol space to possess self-recovery and self-stabilization capabilities in highly dynamic environments, ensuring that the anomaly detection model maintains output consistency and reliability under continuous input conditions, thereby improving the robustness and interpretability of time series analysis. Attached Figure Description

[0016] To more clearly illustrate the technical solutions in the embodiments of this application or the prior art, the drawings used in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments recorded in this invention. For those skilled in the art, other drawings can be obtained based on these drawings.

[0017] Figure 1 This is a flowchart of the time series anomaly detection method based on multi-window clustering symbolization according to the present invention. Detailed Implementation

[0018] Exemplary embodiments will now be described more fully with reference to the accompanying drawings. However, these exemplary embodiments can be implemented in many forms and should not be construed as limited to the examples set forth herein; rather, they are provided so that the description of this disclosure will be more complete and fully convey the concept of the exemplary embodiments to those skilled in the art.

[0019] This invention provides, for example Figure 1 The time series anomaly detection method based on multi-window clustering symbolization shown includes the following steps: Step 1: Establish a cross-window feature drift capture baseline. Divide the input time series into segments according to the sliding time window, extract the feature distribution of each time window, and construct a continuous mapping model. Extract the transition gradient of the cluster centroid based on the feature distribution changes between adjacent time windows, and generate a dynamic drift monitoring matrix to characterize the dynamic changes of the centroid. The specific implementation process of this step is as follows: The input time series is segmented sequentially to form a continuous sequence of time windows. Each time window contains a fixed number of sampling points or spans of time, reflecting the local dynamic characteristics within that time period. During segmentation, the overlap ratio of time windows is finely controlled to ensure sufficient information overlap between adjacent windows, thus guaranteeing a smooth transition in temporal continuity. After dividing the time windows, basic features are extracted from the data within each window, including statistical mean, fluctuation amplitude, upward or downward trend, short-term stability, and local rate of change, among other multi-dimensional indicators. This allows each time window to be expressed as a temporally representative feature vector. This sliding segmentation and feature extraction method maintains a continuous observation channel for features over time, providing a solid data foundation for subsequent dynamic modeling of feature distributions.

[0020] After obtaining the feature vectors for each time window, a continuous mapping model is constructed based on the correlation between these feature vectors to describe the continuous evolution of feature distributions between adjacent time windows. Specifically, by comparing the feature distributions of adjacent time windows, the similarity and magnitude of change in their statistical distribution patterns are extracted to form a quantitative reference for the continuity of feature distribution. This continuous mapping model not only records the directional information of feature evolution over time but also reflects the degree of correlation between different feature dimensions over time, enabling local changes in the time series to be integrated into a continuous spatiotemporal trajectory. In this process, to avoid feature jumps caused by abrupt changes between time windows, the continuous mapping model adopts a step-by-step transition approach between adjacent windows in establishing the mapping relationship. That is, when the feature distribution of each window is mapped to the next window, some continuous distribution information from the previous window is retained to ensure the smoothness and traceability of the overall feature evolution.

[0021] After obtaining the continuous mapping model, the transition gradients of the cluster centroids are extracted based on the changing relationships reflecting the differences in features between adjacent time windows in the model, to characterize the feature drift trend of the time series during the sliding update process. This process identifies the trajectory of the centroids in the time dimension by comparing the relative positional offsets of adjacent time windows in the feature space, and quantifies this trajectory in the form of gradients. The transition gradients not only reveal the direction and speed of change of the overall features of the time series, but also reflect the instantaneous fluctuations caused by local anomalies, thus distinguishing between normal smooth transitions and discontinuous abrupt changes. In this step, the transition gradients of the cluster centroids, as the core parameters describing the dynamic evolution of the time series, form a one-to-one correspondence with the distribution changes in the continuous mapping model, providing the basic input for subsequent dynamic drift monitoring. In this way, the temporal structure of the time series is further abstracted into a measurable change trajectory, enabling continuous drift capture in the time dimension.

[0022] After obtaining the transition gradient of the cluster centroid, a dynamic drift monitoring matrix is ​​generated using this gradient information as the core reference to characterize the dynamic changes of the centroid. This matrix uses time windows as rows and feature dimensions as columns, with each element recording the drift magnitude and direction information of each feature dimension within the corresponding time window. By introducing time-related weighting factors into the matrix, the drift changes of the nearest window can have a higher impact on the overall monitoring results, thereby enhancing the response to real-time changes. The dynamic drift monitoring matrix not only quantitatively describes the intensity of changes in the cluster centroid over different time periods but also displays the coupling relationship between the time dimension and the feature dimension in a two-dimensional form, enabling a unified expression of change patterns at different time scales. After generation, the matrix continuously receives update information from the continuous mapping model and the cluster centroid gradient, forming an adaptive dynamic monitoring structure to continuously track the evolution of the feature distribution during real-time rolling of the time series. When the input time series experiences abrupt changes or drifts, the matrix can instantly reflect the suddenness and directionality of the centroid changes, providing a stable time reference for subsequent centroid synchronization remapping and sign mapping correction.

[0023] By executing the above steps sequentially, a step-by-step construction from raw time-series data to a dynamic drift monitoring matrix is ​​achieved. This enables the entire cross-window feature drift capture baseline to accurately characterize the dynamic changes of cluster centroids while maintaining temporal continuity. This process, by establishing a continuous observation and dynamic recording mechanism in the time dimension, ensures the spatiotemporal consistency of subsequent centroid synchronization and symbol mapping updates, fundamentally improving the stability and traceability of the real-time anomaly detection process.

[0024] Step 2: Based on the dynamic drift monitoring matrix, a centroid synchronous remapping network is constructed. The reverse phase registration algorithm is used to continuously and smoothly adjust the discontinuous transitions of the cluster centroids between adjacent time windows. The gradient information of the dynamic drift monitoring matrix is ​​converted into symbol mapping correction weights to establish a traceable symbol transfer path between the old and new cluster centroids. The specific implementation process of this step is as follows: After obtaining the dynamic drift monitoring matrix, the drift information of the time windows recorded within it is analyzed to determine the direction and intensity of change in the spatial distribution of the cluster centroids within each time window. The drift value of each cell in the dynamic drift monitoring matrix reflects the degree of difference in feature distribution between different time windows. Therefore, by continuously reading the gradient changes of the matrix along the time axis, the motion trajectory of the cluster centroids over the entire time series can be extracted. This motion trajectory is used to identify discontinuous transition segments that may exist in the centroids during the rolling of the time window. When the gradient amplitude within a certain time period exceeds a preset change threshold, the system marks that interval as a potential centroid discontinuity region for subsequent synchronous remapping operations. Through this centroid transition identification method based on gradient change intensity, centroid abrupt changes can be accurately located and detected in advance without disrupting temporal continuity, providing a temporal basis for subsequent smooth adjustments.

[0025] After identifying the discontinuous transition region of the centroid, a synchronous reference relationship is constructed between the old and new centroids based on the continuous feature distribution information recorded in the dynamic drift monitoring matrix. This reference relationship uses a time window as the primary index, matching the clustered centroid positions in the previous time window with the centroid positions in the subsequent time window, and combining this with the aforementioned drift gradient direction to form a temporal mapping alignment. By introducing the concept of reverse phase registration, the relative displacement between the old and new centroids is reversed, allowing the spatial position of the new centroid to gradually trace back towards the direction of change of the old centroid until a continuous transition path is formed in the feature space. Essentially, this process transforms the temporal gradient information of the distribution changes in the dynamic drift monitoring matrix into an adjustment vector for centroid synchronization, enabling the centroids to gradually transition rather than abruptly migrate in the time dimension, thereby eliminating the spatiotemporal discontinuity caused by discontinuous transitions.

[0026] After smoothly aligning the old and new centroids, the gradient information in the dynamic drift monitoring matrix is ​​further transformed into symbol mapping correction weights. Specifically, based on the centroid adjustment magnitude and direction corresponding to each time window, the gradient values ​​at each time point in the drift monitoring matrix are mapped as weight coefficients to guide the correction of the correspondence between symbols and features in the symbol space. At this point, the magnitude of the weight value directly reflects the degree of influence of centroid drift on the symbol mapping table: when the gradient change is small, the weight value tends to stabilize, and the symbol mapping relationship remains unchanged; while when the gradient change is large, the weight value increases, and the symbol mapping relationship will be partially corrected to adapt to the new feature distribution. Through this weighted mapping correction mechanism, the symbol space can exhibit adaptive characteristics when dealing with dynamic changes in the centroid, thereby ensuring the temporal consistency of symbols. The introduction of this weight also has a delayed smoothing effect, that is, gradually attenuating the drift influence over multiple consecutive windows, avoiding local collapse of the symbol space when abruptly changing in a single window.

[0027] After generating the symbol mapping correction weights, a traceable symbol transfer path is established between the old and new cluster centroids to achieve a continuous symbol mapping transition from the old centroid distribution to the new centroid distribution. This symbol transfer path, with the time window sequence as its main axis, comprehensively correlates the time series gradient in the dynamic drift monitoring matrix, the centroid smoothing adjustment results, and the symbol mapping correction weights, ensuring that the symbol mapping relationship at each time step can be dynamically updated based on the drift information of the previous time step. This continuous and traceable symbol transfer path not only maintains the stable evolution of the symbol mapping table in the time dimension but also achieves the correspondence and coherence between feature distribution and symbol representation in the spatial dimension. As the time series continues to be input, new symbols will gradually replace old symbols along the predetermined transfer path, thus forming a symbol evolution mechanism that both preserves historical references and adapts to real-time changes. This mechanism ensures time synchronization between cluster centroid updates and symbol mapping updates, preventing symbol gaps and abnormal drifts caused by update delays or inconsistencies.

[0028] Through the above steps, the entire process of centroid synchronization and remapping based on a dynamic drift monitoring matrix was realized. This process, based on the characteristic drift information of the time series, establishes a synchronization mechanism that maintains the stability and continuity of the symbol space under dynamic changes in the cluster centroid through continuous reference, smooth alignment, weight correction, and symbol transfer path construction. This implementation establishes a coordinated relationship between temporal continuity, spatial smoothness, and symbol consistency, significantly reducing the risk of symbol mapping failure caused by centroid mutations in real-time clustering updates, thus providing a stable and reliable foundation for subsequent symbol write-back and anomaly detection.

[0029] Step 3: Based on the symbol transfer path, generate a time-delay adaptive write-back chain, monitor the synchronization offset of the cluster centroid in real time, and use a hysteresis feedback structure to dynamically write back and update the spatiotemporal index in the symbol mapping table, so that the symbol mapping has a gradual decay characteristic during the write-back process, thereby realizing the continuous and smooth transition of the symbol mapping table. The specific implementation process of this step is as follows: Based on the established symbol transfer path, the synchronization offset of the cluster centroids over time is continuously monitored. Since a traceable symbol transfer path has been established in the previous stage, the real-time offset of the centroids within each time window can be continuously observed based on the recorded dynamic position changes of the centroids along this path. By comparing the centroid position of the current time window with the reference centroid position for the corresponding time period in the symbol transfer path, the offset reflecting the synchronization accuracy can be obtained. The direction and magnitude of this offset are used to measure the difference between the current state of the symbol map and the ideal synchronization state, thus providing a quantitative reference for subsequent dynamic write-back. This process ensures that the update behavior of the symbol map is always based on the centroid synchronization state, keeping the rhythm of symbol updates strictly consistent with the dynamic changes of the cluster centroids, fundamentally preventing mapping discontinuities caused by centroid drift lag.

[0030] Based on the obtained cluster centroid synchronization offsets, a feedback structure with hysteresis characteristics is constructed on the time axis, enabling the symbol map's write-back process to have a time-delayed response capability. This hysteresis feedback structure introduces a temporal buffer between the symbol update operation and the centroid offset detection, preventing the map from immediately reacting drastically to short-term abrupt changes. Instead, it gradually adjusts the write-back based on the persistence and stability of the offsets. This avoids over-correction due to instantaneous fluctuations and ensures a smooth transition of the symbol map during continuous drift phases. This time-delay mechanism essentially establishes a time buffer during the symbol update process, allowing the symbol map to exhibit elastic-like recovery characteristics when facing non-stationary feature distributions, thereby maintaining the stability of the overall symbol space in dynamic environments.

[0031] After constructing the hysteresis feedback structure, a dynamic write-back update process is initiated. This process uses the cluster centroid synchronization offset as input reference and sequentially writes back the mapping relationships of each time window in the symbol transfer path according to the offset direction and magnitude. Specifically, the time-delay adaptive write-back chain progressively updates the spatiotemporal index of the symbol map table in the time dimension according to the order of previous to subsequent windows, so that the reference position of each symbol can migrate accordingly with the dynamic adjustment of the centroid. To ensure the smoothness of the update, a progressive decay weight is introduced into the write-back chain during the update process. That is, as the write-back distance time window increases, the update impact gradually weakens, thus forming a smooth decay effect from near to far. In this way, the symbol map table will not experience abrupt changes during the write-back process, but will exhibit continuous gradual characteristics, enabling a natural transition between old and new symbols in time and space. The core of this process lies in the dual effect of time delay and decay mechanism, which allows the adjustment of the symbol map to respond promptly to the dynamic changes of the centroid while maintaining sufficient stability.

[0032] After completing the dynamic write-back update, the stability of the entire time-delay adaptive write-back chain's output is verified and synchronized. Specifically, the latest state of the written-back symbol map is compared with the reference trajectory of the symbol transfer path to evaluate its consistency in the temporal dimension and its deviation convergence in the spatial dimension. When a small residual offset is found in some symbol maps, the write-back chain will perform a local refinement update again through recursive feedback until the synchronization deviation between the symbol map and the cluster centroid is lower than a predetermined threshold. Through this closed-loop adjustment process, the time-delay adaptive write-back chain can continuously optimize its response delay parameters and decay rate, achieving stable evolution of the symbol map over the entire time series. Ultimately, the symbol map exhibits a smooth write-back trajectory with gradual decay in time and maintains structural consistency with the centroid in space, thus forming an adaptive dynamic update mechanism.

[0033] Through the execution of the above steps, the complete construction and dynamic operation of the symbol transfer path and the time-delay adaptive write-back chain are achieved. This implementation introduces a hysteresis feedback structure and a gradual decay mechanism, enabling the symbol mapping table to maintain a controllable response delay in the time dimension and continuous symbol migration characteristics in the spatial dimension when dealing with continuous updates to cluster centroids. This ensures a smooth transition and traceability throughout the symbolization process. This method exhibits high adaptability and robustness in dynamic time-series environments, significantly reducing the risk of symbol mapping chaos caused by unstable centroid updates, and providing a continuous and reliable symbol data input foundation for subsequent anomaly detection.

[0034] Step 4: Based on the stable output of the time-delay adaptive write-back chain, execute abnormal mapping freeze control, lock the symbol mapping table of the high-risk area of ​​centroid transition, and set the freeze constraint threshold according to the decay weight of the write-back chain to prevent the symbol space from collapsing locally when the cluster centroid changes abruptly. The specific implementation process of this step is as follows: After the time-delay adaptive write-back chain achieves stable output, the spatiotemporal evolution of the symbol map table is monitored in real time to identify high-risk regions for centroid transitions. Since the time-delay adaptive write-back chain has already implemented gradual decay updates to the symbol map table in the previous stage, the rate, direction, and magnitude of symbol map changes in its output can reflect the synchronous stability of the centroids. When an abnormal increase in the symbol map update rate, nonlinear fluctuations in the decay curve, or excessively rapid accumulation of local weights are detected within a certain time window, it indicates a risk of abrupt centroid mutation in the corresponding region. At this point, based on the dynamic drift monitoring matrix and historical reference data in the symbol transfer path, the map table for this risk region is calibrated, establishing a freeze monitoring list containing high-risk symbol indices, time intervals, and characteristic directions. This sub-step enables early spatial localization of potential centroid discontinuous changes, providing constraints for subsequent freeze control.

[0035] After high-risk areas are identified and recorded, a corresponding freeze constraint threshold is determined based on the decay weight of the time-delay adaptive write-back chain. This threshold is used as a criterion to selectively lock the symbol map. This decay weight reflects the response strength and impact range of symbol updates in the time dimension, and therefore can serve as a dynamic parameter for freeze control. When the decay weight exceeds the preset risk threshold, the system determines that the symbol mapping in that area is in a non-steady-state phase. At this point, a freeze mechanism is immediately initiated, temporarily locking the corresponding symbol index and its adjacent time window, thus stopping its updates by the time-delay adaptive write-back chain for a short period. To prevent the locking range from becoming too large and causing symbol space imbalance, the freeze operation is preferentially performed on local segments with the largest weight gradient and concentrated drift direction, while the write-back chain continues to update normally in stable areas. Through this adaptive freeze strategy based on decay weight, the freeze range strictly corresponds to the risk level, ensuring both the accuracy of freeze control and avoiding the interruption of the overall symbol map update.

[0036] After the symbol map is locked, stability maintenance operations are performed on the symbol reference relationships within the frozen region to ensure that the symbol map still possesses interpretability and temporal consistency during the frozen state. Specifically, during the freezing period, the symbol map of high-risk regions is maintained in its pre-freezing state, and the decay rate of symbols surrounding this region is adjusted through the hysteresis feedback structure of the time-delay adaptive write-back chain, so that the mapping changes between the frozen and unfrozen regions exhibit a continuous transition boundary effect. Simultaneously, the symbol index of the frozen region undergoes small-amplitude synchronization checks every time delay period to confirm whether the drift level of the region has returned to stability. When the check results indicate that the drift magnitude has gradually converged to a normal range, the symbol map of the frozen region enters a waiting-to-unlock state so that it can re-participate in the write-back update during the subsequent recovery phase. This process not only maintains symbol consistency during the freezing period but also prevents semantic gaps between the frozen and unfrozen regions, thereby ensuring the overall continuity of the symbol space.

[0037] After the freeze control phase ends, the symbol mapping table of the frozen region is unlocked and restored to re-establish its association with the synchronous update of the cluster centroids. The unlocking operation is not immediate but rather a gradual transition based on the output signal of the adaptive write-back chain with time delay. Specifically, before the frozen region is unlocked, the system first re-evaluates the decay weight of the write-back chain. After confirming that it has recovered to a stable range, the update frequency is restored in a step-by-step decay manner, ensuring a smooth temporal transition between the symbol mapping of the frozen region and the mapping changes of the surrounding non-frozen regions. During this restoration process, the symbol index of the frozen region will again receive synchronization signal input from the symbol transfer path to ensure that the unlocked symbol space completely corresponds to the latest cluster centroid distribution. Through this gradual unlocking and write-back mechanism, the entire symbol mapping table achieves continuous restoration after the freeze control is completed, avoiding structural conflicts that occur at the moment of unlocking. Ultimately, the entire process of abnormal mapping freeze control enables the symbol space to automatically identify risk areas, proactively implement freeze protection, dynamically maintain boundary smoothness, and smoothly restore when facing centroid mutations, thereby effectively preventing local collapse of the symbol space and maintaining the stability and traceability of the time-series symbolized representation.

[0038] Through the implementation of the above steps, the entire process of anomaly mapping freeze control based on a time-delay adaptive write-back chain is realized. This implementation method, through the collaborative design of high-risk region identification, attenuation weight threshold setting, frozen state maintenance, and gradual unlocking and recovery, enables the symbol mapping table to possess adaptive protection capabilities when dealing with sudden changes in cluster centroids. The entire process forms a delayed response mechanism in the time dimension and establishes a continuous transition structure in the spatial dimension, thereby significantly improving the robustness and stability of the time series anomaly detection system in non-stationary environments, laying a solid foundation for subsequent symbol consistency verification and the construction of a self-healing synchronization loop.

[0039] Step 5: Based on the output of the anomaly mapping freeze control, construct a symbol consistency verification closed loop. Use a dual-channel anomaly measurement mechanism to cross-compare symbol drift and anomaly score deviation, and feed back the consistency correction signal to the cross-window feature drift capture baseline, thereby forming a self-healing synchronization loop to achieve stable closed-loop control of dynamic updating of cluster centroids, so as to improve the continuity and reliability of time series anomaly detection. The specific implementation process of this step is as follows: After the anomaly mapping freeze control is completed, a comprehensive analysis of the symbol mapping table output during the freeze control phase is performed to identify potential drift regions and bias accumulation points in the symbol space. Because the freeze control locks some symbol regions to prevent local collapse caused by centroid mutations, although overall stability is maintained, slight semantic shifts may still occur during the unlocking and recovery phase. Therefore, it is necessary to compare the historical and current states of the entire symbol space and extract key features reflecting symbol consistency by tracking the evolution of symbol indexes in the time-delay adaptive write-back chain and freeze control process. These features include the temporal connection relationships between symbols, the correspondence of symbols in spatial distribution, and the continuity trend of symbol mapping over time. Through systematic analysis of these features, a set of benchmark parameters describing the overall stability of the symbols can be obtained, which will serve as the input basis for subsequent consistency verification loops.

[0040] After obtaining the baseline parameters of the symbol space, a dual-channel anomaly measurement mechanism is constructed to simultaneously evaluate the consistency of the symbol mapping from both temporal and spatial dimensions. Specifically, the temporal channel is used to detect continuity deviations in symbol drift. By comparing the temporal extension of symbols at different stages—before freezing, during freezing, and after thawing—the degree of continuity maintenance of the symbol mapping during temporal evolution is determined. The spatial channel is used to measure deviations in anomaly scores. By comparing the differences between the symbol distribution and the cluster centroid distribution in the same time slice, the matching degree between the symbol space and the feature space is evaluated. The dual-channel structure allows the verification process to simultaneously capture the changes in both temporal drift and spatial deviation, thereby achieving comprehensive consistency measurement. When the temporal channel detects anomalies between symbols while the spatial channel remains matched, it indicates a lag in symbol updates; conversely, when the spatial channel shows significant deviation while the temporal channel remains stable, it indicates that the local adjustment of the cluster centroids is too rapid. Through cross-comparison of the two channels, the source of symbol consistency problems can be accurately identified, providing a targeted basis for subsequent correction feedback.

[0041] After the dual-channel anomaly measurement mechanism is implemented, the deviation results from the two channels are cross-compared to generate a consistency correction signal. This correction signal integrates the detection results from the temporal and spatial channels, reflecting the current overall stability of the symbol mapping through weighted fusion. When the correction signal shows that the temporal channel deviation is higher than the threshold, it indicates that the symbol drift is too fast. In this case, the temporal weight of the correction signal will drive a decrease in the update rate of the delayed adaptive write-back chain. When the correction signal shows that the spatial channel deviation is higher than the threshold, it indicates that the cluster centroid synchronization offset is too large. In this case, the spatial weight of the correction signal will prompt the centroid synchronization remapping process to readjust smoothly. At the global level of the symbol mapping table, the correction signal is gradually transmitted through dynamic feedback, enabling the symbol mapping to regain rhythmic balance in the temporal dimension and re-establish corresponding consistency in the spatial dimension. This correction process has adaptive characteristics, which can not only quickly fine-tune for local drifts but also automatically maintain the overall stability of the symbol mapping in long-term series.

[0042] The consistency correction signal is fed back to the cross-window feature drift capture baseline, forming a self-healing synchronization loop that enables stable closed-loop control of dynamic updates to the cluster centroid. This self-healing synchronization loop, driven by the consistency correction signal, injects real-time deviation information from the symbol mapping into the feature drift capture baseline, allowing the feature drift monitoring process to dynamically adjust its capture strategy based on feedback from the symbol level. When symbol drift signals repeatedly appear in the same interval, the cross-window feature drift capture baseline automatically adjusts the overlap ratio of the time windows and the frequency of feature distribution evaluation to improve the capture accuracy of centroid changes in that region. Simultaneously, after symbol consistency is restored, the correction signal gradually decays, allowing the system to return to normal monitoring mode. Through this closed-loop feedback structure, a continuous interactive and mutually corrective relationship is formed between the symbol space and the feature space, giving the entire anomaly detection process self-healing capabilities. When external disturbances cause temporary centroid drift, the symbol consistency verification closed loop can automatically detect, actively adjust, and ultimately restore the system to a stable state, thus achieving true dynamic self-healing synchronization control.

[0043] Through the implementation of the above steps, the entire process of constructing a closed-loop system for symbolic consistency verification based on anomaly mapping freeze control output is realized. This implementation combines a dual-channel anomaly measurement mechanism with adaptive feedback control, enabling dynamic temporal and spatial coupling of the symbolic mapping table, cluster centroid distribution, and feature drift capture baseline, forming a self-healing synchronization structure. This closed-loop design not only effectively improves the continuity and accuracy of anomaly detection but also ensures long-term stable operation of the system when facing complex non-stationary time series, significantly enhancing the robustness and interpretability of time series anomaly detection methods.

[0044] This invention introduces a cross-window feature drift capture baseline and centroid synchronous remapping mechanism during time series clustering and symbolization, enabling continuous tracking and stable control of the dynamic changes in cluster centroids. Through the synergistic effect of the dynamic drift monitoring matrix and the time-delay adaptive write-back chain, the symbol mapping table maintains smooth temporal continuity and consistent spatial correspondence during centroid updates, fundamentally eliminating symbol fragmentation and index failure problems caused by centroid mutations. This technology gives the symbolized representation of time series spatiotemporal continuity, allowing the symbol sequence to stably reflect changes in the original feature distribution, thereby significantly improving the accuracy and stability of anomaly detection.

[0045] This invention achieves self-healing regulation of the dynamic evolution of the symbol space by constructing a closed loop of anomaly mapping freeze control and symbol consistency verification. The system automatically executes freeze protection upon detecting abrupt changes in cluster centroids, gradually unlocks the symbol space during the recovery phase based on the write-back chain decay weight, and performs consistency correction feedback through a dual-channel anomaly measurement mechanism, forming an adaptive closed-loop adjustment structure throughout the detection process. This mechanism enables the symbol space to possess self-recovery and self-stabilization capabilities in highly dynamic environments, ensuring that the anomaly detection model maintains output consistency and reliability under continuous input conditions, thereby improving the robustness and interpretability of time series analysis.

[0046] This invention can be applied to temperature monitoring scenarios in industrial equipment, enabling continuous anomaly detection and self-healing regulation of operating temperature changes in key components of a production line to prevent malfunctions caused by overheating or sensor drift. In this embodiment, by introducing cross-window feature drift capture, centroid synchronous remapping, time-delay adaptive write-back chain, anomaly mapping freeze control, and symbol consistency verification closed loop during the clustering and symbolization process of real-time temperature signals, the anomaly detection of temperature signals possesses continuity, stability, and self-recovery capabilities. The specific process includes the following sub-steps: First, the real-time temperature signals acquired by industrial equipment are segmented sequentially to form a continuous time window sequence. Each time window is set to 10 seconds in length, with a 50% overlap between adjacent windows to ensure the continuity and smoothness of temperature changes. Within each time window, four characteristic indicators—mean, fluctuation amplitude, instantaneous rate of rise, and stability—are extracted from the temperature signal to construct a feature vector representing the temperature change state within that time period. Subsequently, a continuous mapping model is established by comparing the feature vectors of adjacent windows to capture the smooth trend of temperature characteristics over time. When the continuous mapping model detects a sudden change in the feature distribution within a certain time period (e.g., the average temperature suddenly rises from 65℃ to 78℃), the system marks this segment as a high-drift zone and generates a dynamic drift monitoring matrix based on the direction of feature change. This matrix, with time windows as rows and feature types as columns, records the amplitude and direction of change for each feature, providing a time reference for subsequent centroid synchronization and sign correction.

[0047] After obtaining the dynamic drift monitoring matrix, the drift gradient information in the matrix is ​​analyzed to construct a centroid synchronization remapping relationship. This step matches the clustered centroids extracted from adjacent time windows and establishes a synchronization reference between the old and new centroids based on the aforementioned drift direction. Taking the temperature signal as an example, when the dynamic drift monitoring matrix shows that the mean characteristic gradient of the device temperature exceeds 10℃ / window, the system determines that there is a risk of centroid transition in this segment and immediately initiates a centroid smoothing adjustment process. By reverse-correcting the centroid change direction of the preceding and following time windows, the spatial position of the new centroid gradually approaches the trajectory of the old centroid change, thereby eliminating the spatiotemporal discontinuity caused by discontinuous transitions. At the same time, the drift gradient information is converted into symbol mapping correction weights to adjust the symbol correspondence of each temperature interval in the symbol table. For example, the symbol "A" originally corresponding to 65℃ extends to 68℃ after adjustment, while the symbol "B" originally corresponding to 70℃ extends upward to 74℃, ensuring the continuous expression of the temperature symbol sequence. Ultimately, a traceable symbolic transfer path is established between the old and new centroids, enabling a synchronous transition between the symbolic space and the feature space.

[0048] After establishing a traceable symbol transfer path, a time-delay adaptive write-back chain is generated based on this path to achieve continuous write-back and smooth updates of the symbol map table. At this point, the cluster centroid synchronization deviation is dynamically calculated based on the symbol offset of each time window. For example, during continuous monitoring, if the centroid change in equipment temperature exceeds 5°C within two adjacent windows but does not trigger an anomaly threshold, the time-delay adaptive write-back chain will adjust the write-back speed according to a decay coefficient of 0.6, ensuring that the symbol map update lags behind by half a window period. This avoids overcorrection caused by instantaneous temperature fluctuations and ensures stable adjustment of the symbol map over long-term trends. During the write-back process, the update impact decays exponentially with the extension of the time window; that is, the farther away from the current window, the smaller the symbol adjustment amplitude. In this way, the symbol map table remains spatially continuous and exhibits a gradually decaying smooth evolution in time, thus enabling the symbolic representation of the temperature signal to naturally adapt to the dynamic changes in equipment operation.

[0049] Based on the stable output of the time-delay adaptive write-back chain, a closed loop of anomaly mapping freeze control and symbol consistency verification is executed. When the system detects that the rate of change of the cluster centroid exceeds a set threshold (e.g., a temperature change exceeding 15℃ / 10 seconds), a freeze operation is performed on the corresponding symbol segment (e.g., segments "C" and "D") to prevent local collapse of the symbol space during abrupt changes. During the freeze, the system suspends symbol updates in that region and only adjusts the decay rate of neighboring symbols to maintain boundary smoothness. After the freeze is lifted, a dual-channel anomaly measurement mechanism is activated, monitoring symbol drift continuity in the time dimension and anomaly score deviation in the spatial dimension. For example, after the temperature drops from 78℃ to 70℃, the time channel detects that the symbol sequence "CDC" has returned to regularity, while the spatial channel confirms that the anomaly score in this temperature range has returned to the threshold. After cross-comparison of the two channels, a consistency correction signal is generated. This correction signal is fed back to the cross-window feature drift capture baseline to correct the feature capture strategy for future temperature changes, such as automatically adjusting the time window overlap ratio to 60% to enhance sensitivity to abrupt changes. Through this closed-loop feedback mechanism, the system can autonomously recover to a stable state after temperature fluctuations, ensuring the spatiotemporal consistency of the symbol mapping table and realizing the dynamic self-healing update of the cluster centroids.

[0050] Through the execution of the above steps, this embodiment achieves real-time dynamic clustering and symbolic self-healing detection of temperature signals from industrial equipment, enabling the system to maintain the continuity and reliability of detection results when facing non-stationary temperature changes. This method can also be extended to medical electrocardiogram (ECG) signal monitoring scenarios to identify abnormal variations in heart rhythm waveforms in real time, achieving highly efficient, robust, and adaptive symbolic anomaly detection.

[0051] In another preferred embodiment, the present invention can be applied to medical electrocardiogram (ECG) monitoring scenarios to continuously detect, symbolically represent, and self-healingly synchronize changes in the dynamic changes of a patient's ECG waveform, thereby maintaining the stability and clinical interpretability of the detection results even in cases of arrhythmia, rhythm disturbances, or signal drift. This embodiment fully utilizes the steps proposed in this invention, such as cross-window feature drift baseline capture, centroid synchronous remapping, time-delay adaptive write-back chain, anomaly mapping freeze control, and symbol consistency verification closed loop, to achieve high-resolution dynamic modeling and anomaly identification of ECG signals. The specific process includes the following sub-steps: First, the acquired ECG signals are segmented sequentially to form a dynamic window sequence with continuous temporal overlap. To balance signal detail capture and real-time responsiveness, the window length is set to 2 seconds, with a 75% overlap between adjacent windows, ensuring the temporal integrity and continuity of key waveform features (such as the P wave, QRS complex, and T wave). Within each time window, multiple feature indicators are extracted, including the average length of the RR interval, peak amplitude, signal energy distribution, QRS rising slope, and heart rate variability, to construct the multidimensional feature vector for that window. Subsequently, based on the changing relationships between the feature vectors of adjacent windows, a continuous mapping model is constructed to describe the evolution trend of the ECG signal in the temporal dimension. For example, when the RR interval lengthens from 0.82 seconds to 1.05 seconds and the energy concentration decreases, the continuous mapping model will capture the dynamic trajectory of a slowed rhythm. In this way, the system forms a continuous feature channel across time windows, providing continuous temporal reference for subsequent centroid synchronization and abnormal symbol identification.

[0052] After establishing the dynamic drift monitoring matrix, it is analyzed to identify the dynamic change regions of cluster centroids in the ECG signal. The dynamic drift monitoring matrix is ​​structured with time windows as rows and feature dimensions as columns, where each cell reflects the magnitude of change of different features between adjacent windows. When the gradient magnitude of a certain region exceeds a set threshold (e.g., a change in heart rate variability exceeding 30%), the system determines that there may be a discontinuous transition of centroids in that segment. At this time, based on the distribution direction and amplitude information recorded in the dynamic drift monitoring matrix, a synchronous reference relationship between the old and new centroids is constructed, and the centroid position is smoothly corrected using the reverse phase registration method. For example, when the peak position of the QRS complex shifts significantly between two consecutive windows, the new centroid will gradually return along the direction of change of the old centroid to maintain the temporal consistency of the ECG waveform. At the same time, the gradient information of the drift matrix is ​​converted into symbol mapping correction weights to adjust the symbol mapping table, so that the symbolization results of the ECG waveform (e.g., using "A", "B", "C" to represent different rhythm patterns) maintain a smooth transition during centroid changes, preventing symbol breaks or duplicate encoding caused by short-term anomalies.

[0053] After obtaining a traceable symbol transfer path, a time-delayed adaptive write-back chain is generated based on this path to achieve dynamic write-back and self-balancing updates of the symbol map table. This write-back chain monitors the synchronization offset of the ECG signal centroid in real time, ensuring that the adjustment rhythm of the symbol map table aligns with the physiological changes in the ECG signal. For example, when a patient experiences a temporary drift in the ECG signal due to slight body movement or respiratory changes, the system automatically adjusts the write-back delay parameter based on the offset amplitude, causing the symbol map update to lag by half a time window, thus avoiding misjudging transient interference as rhythm abnormalities. During this process, a progressive decay weight is introduced, gradually weakening the impact of symbol updates farther from the current window, forming a smooth decay transition from near to far. In this way, the symbol map table exhibits a continuously decaying adaptive characteristic in time and maintains a consistent spatial correspondence with the ECG signal characteristics, ensuring that the symbolization results stably reflect the patient's physiological rhythm changes.

[0054] After the time-delay adaptive write-back chain stabilizes, anomaly mapping freeze control and symbol consistency verification closed-loop are executed to prevent symbol space collapse caused by centroid mutations or external interference. When the system detects that the rate of change of the ECG signal waveform exceeds a threshold (e.g., a sudden drop of 40% in the amplitude of the QRS main peak or a heart rate change exceeding 20 beats per minute), freeze control is immediately triggered to lock the symbol mapping for the corresponding time window, suspend updates, and dynamically adjust the decay rate of neighboring symbols to maintain a smooth transition. After the freeze is lifted, the system starts a dual-channel anomaly measurement mechanism. On the one hand, it detects the rhythmic continuity of the symbol sequence (e.g., the timing pattern of P-QRS-T) through the time dimension, and on the other hand, it detects the deviation of the symbol anomaly score (e.g., score drift in the arrhythmia segment) through the spatial dimension. When the dual-channel comparison results are consistent and the symbol mapping returns to stability, a consistency correction signal is generated and fed back to the cross-window feature drift capture baseline to correct future ECG signal feature capture strategies, such as automatically shortening the time window length to 1.5 seconds to improve the capture accuracy of high-frequency rhythms. This closed-loop feedback process enables the system to form a self-healing synchronization loop, meaning that when faced with signal abnormalities or external disturbances, the system can self-detect, self-correct, and restore stable operation, thereby continuously maintaining the accuracy and consistency of ECG signal abnormality detection.

[0055] Through the implementation of the above steps, this embodiment achieves a complete process from feature capture, centroid synchronization, sign smoothing to closed-loop self-healing in a medical electrocardiogram (ECG) signal monitoring scenario. This method can maintain the continuity of sign mapping and the stability of detection results even when ECG signals are subject to noise, drift, or sudden rhythm changes, avoiding misjudgments or missed detections due to short-term fluctuations. Simultaneously, the sign consistency verification closed loop constructed in this invention provides a real-time self-correction mechanism for medical scenarios, enabling the system to automatically adapt to the physiological characteristics and rhythm differences of different individuals, achieving high-precision identification of arrhythmias (such as premature beats, atrial fibrillation, bradycardia, etc.), and providing higher stability, interpretability, and security for medical monitoring equipment.

[0056] The foregoing has only described certain exemplary embodiments of the present invention by way of illustration. Undoubtedly, those skilled in the art can modify the described embodiments in various ways without departing from the spirit and scope of the present invention. Therefore, the foregoing drawings and descriptions are illustrative in nature and should not be construed as limiting the scope of protection of the claims of the present invention.

Claims

1. A time series anomaly detection method based on multi-window clustering symbolization, characterized in that, Includes the following steps: Step 1: Acquire real-time temperature signals during the operation of industrial equipment. Divide the real-time temperature signals into sliding time windows according to time sequence to form a continuous time window sequence. Extract the temperature mean, temperature fluctuation amplitude, temperature change rate, and temperature stability features from the temperature data within each time window to construct a temperature feature vector corresponding to each time window. Calculate the temperature feature drift based on the change in temperature feature vectors between adjacent time windows, and generate a dynamic drift monitoring matrix based on the temperature feature drift. The dynamic drift monitoring matrix is ​​used to record the amplitude and direction of change of different temperature feature dimensions within each time window. Step 2: Based on the dynamic drift monitoring matrix, cluster the temperature feature vectors in adjacent time windows to obtain the cluster centroids of the corresponding time windows, and calculate the change in the position of the cluster centroids between adjacent time windows. When the change in the position of the cluster centroid exceeds a preset threshold, the migration direction between the old and new cluster centroids is determined based on the change in the position of the cluster centroids. The position of the new cluster centroid is then gradually adjusted according to the migration direction, so that the new cluster centroid smoothly transitions from the position corresponding to the old cluster centroid to the new position. Symbol mapping correction weights are generated based on the temperature feature drift in the dynamic drift monitoring matrix, and the correspondence between the temperature feature interval and the symbol is adjusted using the symbol mapping correction weights to establish a symbol transfer path for the transition from the old symbol mapping relationship to the new symbol mapping relationship. Step 3: Calculate the synchronization offset between the cluster centroid of the current time window and the reference cluster centroid based on the symbol transfer path, and generate symbol mapping update weights based on the synchronization offset; perform attenuation processing on the symbol mapping update weights according to the distance between time windows, so that the adjustment amount of symbol mapping is smaller the farther away from the current time window; based on the attenuated symbol mapping update weights, gradually update the temperature feature interval and symbol correspondence in the symbol mapping table, so that the symbol mapping table is continuously adjusted as the temperature features change; Step four: During the symbol mapping update process, determine whether the temperature feature change is in a high-risk transition state based on the change in the cluster centroid position; when the change rate of the cluster centroid exceeds the preset freezing threshold, freeze the symbol mapping relationship of the corresponding temperature feature interval, keep the original symbol mapping relationship unchanged in the frozen area, and adjust the symbol mapping decay rate of the adjacent area according to the symbol change trend of the adjacent temperature feature interval; when the change rate of the cluster centroid recovers to a stable range, unfreeze the state and resume the symbol mapping update of the frozen area. Step 5: Perform consistency detection on the symbol mapping results before and after freezing. Detect continuous changes in the symbol sequence through the time dimension and deviations between the temperature feature distribution corresponding to the symbols and the anomaly detection scores through the spatial dimension. Generate a consistency correction signal based on the time and spatial dimension detection results, and adjust the time window partitioning parameters, temperature feature drift calculation parameters, or symbol mapping update weights according to the consistency correction signal to keep the cluster centroid update process of the industrial equipment temperature signal synchronized with the symbol mapping process.

2. The time series anomaly detection method based on multi-window clustering symbolization according to claim 1, characterized in that, The steps to establish a cross-window feature drift capture baseline include: The input time series is segmented into continuous time window sequences by sliding according to the time order, and an overlap ratio is set between adjacent time windows; Extract multidimensional feature indicators from the data within each time window and construct a feature vector that is representative of the time series. A continuous mapping model is constructed based on the feature vectors of adjacent time windows. By comparing the similarity and change magnitude of the feature distributions of adjacent windows, the continuous evolution relationship of features over time is recorded. The transition gradient of the cluster centroid is extracted based on the continuous mapping model, and a dynamic drift monitoring matrix is ​​generated with the gradient as a reference. The dynamic drift monitoring matrix records the drift direction and magnitude of each feature dimension with the time window as the row and the feature dimension as the column.

3. The time series anomaly detection method based on multi-window clustering symbolization according to claim 2, characterized in that, The dynamic drift monitoring matrix introduces a time-related weight factor during the generation process, making the influence weight of the cluster centroid transition gradient within the most recent time window in the matrix update higher than that of the distant window. This enhances the responsiveness to real-time feature drift and reflects the direction and magnitude of centroid changes in real time when abrupt changes occur in the time series.

4. The time series anomaly detection method based on multi-window clustering symbolization according to claim 2, characterized in that, The steps for constructing a centroid synchronous remapping network based on a dynamic drift monitoring matrix include: The time window feature drift information in the dynamic drift monitoring matrix is ​​analyzed to determine the direction and intensity of change of the cluster centroid within each time window, and the non-continuous transition region of the centroid is identified based on the gradient magnitude. Based on the continuous feature distribution information recorded in the dynamic drift monitoring matrix, a synchronous reference relationship between the old centroid and the new centroid is constructed, and the spatial positions of the two are reversed to form a continuous transition path in the feature space. The gradient information in the dynamic drift monitoring matrix is ​​converted into sign mapping correction weights, and the sign mapping relationship is dynamically corrected based on the magnitude and direction of the centroid. Based on the symbol mapping and weight correction, a traceable symbol transfer path is established between the old and new cluster centroids.

5. The time series anomaly detection method based on multi-window clustering symbolization according to claim 4, characterized in that, The symbol mapping correction weight is adaptively adjusted according to the gradient magnitude of each time window in the dynamic drift monitoring matrix. When the change magnitude of the centroid is small, the symbol mapping relationship remains unchanged. When the change magnitude of the centroid exceeds the preset threshold, the correction weight is gradually increased, so that the symbol mapping can achieve smooth decay and update within the continuous time window.

6. The time series anomaly detection method based on multi-window clustering symbolization according to claim 4, characterized in that, The steps for generating a time-delay adaptive write-back chain based on the symbol transfer path include: Based on the symbol transfer path, the synchronization offset of the cluster centroid in the time dimension is monitored. By comparing the current time window centroid position with the reference centroid position, the offset reflecting the synchronization accuracy is obtained. Based on the cluster centroid synchronization offset, a feedback structure with hysteresis characteristics is constructed on the time axis, enabling the symbol mapping table to have time delay response capability during the write-back process. The dynamic write-back update process is initiated based on the hysteresis feedback structure. The spatiotemporal index of the symbol mapping table is written back sequentially according to the offset direction and magnitude, and a progressive decay weight is introduced. The output of the time-delay adaptive write-back chain is subjected to stability verification and synchronization calibration. The update is refined through recursive feedback to ensure that the synchronization deviation between the symbol mapping table and the cluster centroid converges to a predetermined threshold.

7. The time series anomaly detection method based on multi-window clustering symbolization according to claim 6, characterized in that, The progressive decay weight is dynamically adjusted based on the time distance of the cluster centroid synchronization offset. As the time window spacing increases, the update magnitude is gradually reduced, so that the symbol mapping table forms a smooth decay trajectory from near to far during the write-back process.

8. The time series anomaly detection method based on multi-window clustering symbolization according to claim 1, characterized in that, The steps for implementing anomaly mapping freeze control based on the stable output of the time-delay adaptive write-back chain include: After the time-delay adaptive write-back chain stabilizes, the spatiotemporal evolution of the symbol map is detected. Based on the rate, direction, and magnitude of symbol map changes, high-risk areas of cluster centroid transitions are identified, and a frozen monitoring list is established. The freeze constraint threshold is determined based on the decay weight of the adaptive write-back chain with time delay. For high-risk areas that exceed the threshold, symbol mapping table locking is performed to make the freeze range correspond to the risk level. After the symbol mapping table is locked, stability maintenance is performed on the frozen region by adjusting the decay rate of surrounding symbols to maintain boundary smoothness and monitoring the drift convergence status. Once the drift amplitude of the frozen area returns to stability, the frozen area is gradually unlocked based on the output signal of the write-back chain.

9. The time series anomaly detection method based on multi-window clustering symbolization according to claim 8, characterized in that, The unlocking and recovery of the frozen region is achieved by dynamically evaluating the decay weight of the time-delay adaptive write-back chain. When the decay weight recovers to a stable threshold range, the update frequency is gradually increased in a progressive decay manner, so that the symbol mapping of the frozen region and the mapping change of the adjacent non-frozen region can be smoothly transitioned in the time dimension, and the synchronous correspondence with the cluster centroid is re-established.

10. The time series anomaly detection method based on multi-window clustering symbolization according to claim 8, characterized in that, The steps for constructing a closed-loop symbolic consistency verification based on the output of anomaly mapping freeze control include: After the abnormal mapping freeze control is completed, the symbol mapping table state output by the freeze control phase is analyzed. By comparing the historical state with the current state, key features reflecting symbol consistency are extracted to obtain the baseline parameters of the symbol space. Based on the baseline parameters of the symbol space, a dual-channel anomaly measurement mechanism is constructed to detect the continuity deviation of symbol drift from the time dimension and evaluate the deviation of anomaly scores from the spatial dimension. The output results of the dual-channel anomaly measurement mechanism are cross-compared to generate a consistency correction signal, and the delay adaptive write-back chain and centroid synchronization remapping process are adjusted according to the deviation between the time channel and the spatial channel, respectively. The consistency correction signal is fed back to the cross-window feature drift capture baseline to form a self-healing synchronization loop, thereby achieving closed-loop control of dynamic updating of cluster centroids.

Images