Semantic micro-cluster based time series situation awareness system and method

Abstract

The application relates to a time sequence situation awareness system and method based on semantic micro clusters, which collects work order texts in a current time window, obtains work order embedding vectors and time efficiency quality, constructs a disjoint set forest, calculates a rotation tensor according to work order embedding vectors and time efficiency quality of members of the disjoint set forest, obtains a stabilized centroid according to the rotation tensor, updates a merging relationship of the disjoint set forest according to the stabilized centroid, constructs an interval tree, performs interval overlap query on the interval tree to obtain a concurrent semantic micro cluster set in an active time interval, calculates a situation level according to a stabilized centroid of the concurrent semantic micro cluster set and a time sequence change amount of a member number, generates situation early warning data when the situation level meets a warning condition, and eliminates hot event virtual high caused by event trajectory flickering, fracture and false merging in a monitoring large screen.

Description

Technical Field

[0001] This invention relates to the field of temporal situation awareness, specifically to a temporal situation awareness system and method based on semantic microclusters. Background Technology

[0002] Customer service hotlines collect a large number of user call requests and generate work order texts. These work order texts are clustered according to semantic content and statistically analyzed by time to identify requests that occur in clusters over a period of time and to perceive their trends. Current event situational awareness methods process incoming work order texts window by window, with each window containing newly added work order texts. Within each window, several clusters are generated, with the arithmetic mean of the member vectors within each cluster serving as the cluster center. Clusters from adjacent time windows are then linked based on the proximity of their cluster centers, forming a sequence showing the evolution of the same event over time. This sequence is then displayed on a monitoring dashboard, showing the occurrence and development of concentrated requests, allowing operations personnel to understand the distribution and changes in requests.

[0003] However, in the actual operation of hotspot detection at the minute-level, operators observed that the evolution trajectory of the same continuous event on the monitoring screen frequently flickered and broke. A request event lasting for hours or even days was split into multiple events, counted and presented separately, or reappeared as a new event after being interrupted during the tracking process. At the same time, two originally unrelated requests were merged into the same event, causing the event content to become mixed. Due to the accumulation of the above situations, the number of hotspot events on the monitoring screen was artificially inflated, making it difficult for operators to reliably track the entire occurrence and development process of the same request event as a single object continuously. The above phenomena are more common during off-peak hours and in situations where there is a sudden increase in small-scale segmented requests, which has become an urgent problem to be addressed in the actual operation of event situational awareness.

[0004] To address the aforementioned shortcomings, a technical solution is provided. Summary of the Invention

[0005] To address the technical problems raised in the background, this invention is proposed. This invention provides a temporal situational awareness system and method based on semantic micro-clusters.

[0006] This invention is achieved through the following technical solution: a temporal situational awareness method based on semantic micro-clusters, comprising: Collect work order texts within the current time window, perform multi-dimensional semantic tagging on the work order texts to obtain work order embedding vectors and timeliness / quality, classify the work order embedding vectors into semantic micro-clusters according to semantic proximity, and construct a disjoint set forest; For each semantic micro-cluster in the disjoint set forest, calculate the rotation tensor based on the work order embedding vector and timeliness quality of its members, obtain the stable centroid based on the rotation tensor, update the merging relation of the disjoint set forest based on the stable centroid, and construct an interval tree with the active time interval of the semantic micro-cluster as the index and the stable centroid as the node value. Perform interval overlap query on the interval tree to obtain the set of concurrent semantic micro-clusters in the active time interval. Calculate the situation level based on the stable centroid and the time-series change in the number of members of the concurrent semantic micro-cluster set. When the situation level meets the warning conditions, generate situation warning data.

[0007] Furthermore, the calculation steps for timeliness quality are as follows: Collect the work order text within the current time window, perform word segmentation and noise removal on the work order text to obtain a regularized work order text; Multidimensional semantic tagging is performed on the regularized work order text to extract multidimensional semantic tags containing subject, object, time, location, request and emotion. The multidimensional semantic tags are encoded into work order embedding vectors, and the timeliness quality is calculated based on the access time of the work order text within the current time window.

[0008] Furthermore, the steps for constructing a disjoint set forest are as follows: The work order embedding vectors are assigned to the corresponding semantic micro-clusters according to their semantic proximity to the current centroid of the existing semantic micro-clusters. For work order embedding vectors whose semantic proximity is lower than the preset assignment threshold, new semantic micro-clusters are created. The work order embedding vectors assigned to the same semantic micro-cluster are merged into the same set of the disjoint set forest to obtain the disjoint set forest.

[0009] Furthermore, the analytical steps for stabilizing the centroid are as follows: For each semantic micro-cluster in the disjoint set forest, collect the work order embedding vectors and timeliness quality of the members of the semantic micro-cluster, obtain the current centroid of the semantic micro-cluster, and the anchor centroid determined and fixed by the mean of the first preset number of work order embedding vectors to which the semantic micro-cluster is assigned. The stable centroid of the semantic microcluster is calculated based on the current centroid, anchor centroid, work order embedding vector of the member, and timeliness quality.

[0010] Furthermore, the steps for constructing the interval tree are as follows: Calculate the deviation angle between the stabilized centroid and the anchor centroid. When the deviation angle exceeds the preset splitting threshold, split the semantic micro-cluster into a new set in the disjoint set forest and register the belonging relationship between the parent set and the child set. When the deviation angle between the stabilized centroids of two coexisting semantic micro-clusters is lower than the preset merging threshold, merge the sets corresponding to the two semantic micro-clusters in the disjoint set forest. For each semantic microcluster, the active time interval of the semantic microcluster in the time window sequence is used as the key and the stable centroid is used as the value, and then inserted into the interval tree.

[0011] Furthermore, the steps for calculating the stable centroid of the semantic microcluster are as follows: The work order embedding vectors of the members of the semantic micro-cluster are used as particles with corresponding timeliness quality. The gyration tensor of the particles is calculated with the current centroid of the semantic micro-cluster as the reference point. The gyration tensor is decomposed into features to obtain the principal gyration radius square sequence.

[0012] Furthermore, the steps for calculating the stable centroid of semantic microclusters also include: The normalized variance of the principal gyration radius square sequence is used as the shape anisotropy degree; according to the parallel axis shift theorem, the product of the sum of time-dependent mass and the square of the Euclidean distance from the current centroid to the anchor centroid is used as the drift inertia increment; a first interval and a second interval are set, and the boundary between the first interval and the second interval is determined by the situation level output in the previous time window. When the shape anisotropy degree is in the first interval, the conformation identifier is set to the first value, and when the shape anisotropy degree is in the second interval, the conformation identifier is set to the second value.

[0013] Furthermore, the steps for calculating the stable centroid of semantic microclusters also include: The reciprocal of the sum of the trace of the rotation tensor and the preset inertia reference is used as the inertial impedance coefficient. Calculate the mean of the work order embedding vectors belonging to the semantic micro-cluster in the current time window, weighted by timeliness and quality. Move the current centroid along the direction from the current centroid to the mean, allowing a step size to obtain the moved centroid. The allowable step size is the product of the difference between the mean and the current centroid and the inertial impedance coefficient. When the conformation identifier is the first value, the moving centroid is used as the stable centroid. When the conformation identifier is the second value, the moving centroid and the anchor centroid are convexly combined according to the convex combination coefficient to obtain the stable centroid. The convex combination coefficient is the ratio of the sum of the shape anisotropy degree and the drift inertia increment to the sum plus the preset regression constant.

[0014] Furthermore, the steps for calculating the situation level are as follows: Perform an interval overlap query on the interval tree to obtain a set of concurrent semantic micro-clusters in the active time interval; For each semantic microcluster in the concurrent semantic microcluster set, the centroid drift rate is calculated based on the displacement of the stable centroid of the semantic microcluster between adjacent time windows, and the size growth rate is calculated based on the increment of the number of members of the semantic microcluster between adjacent time windows. The centroid drift rate and the size growth rate are used as the temporal change of the semantic microcluster. The situation level is calculated based on the temporal changes of each semantic micro-cluster in the concurrent semantic micro-cluster set.

[0015] Furthermore, the steps for outputting situational early warning data are as follows: When the situation level exceeds the level threshold corresponding to the warning conditions, situation warning data is generated, which includes the stable centroid of the corresponding semantic micro-cluster, the active time interval, and the situation level.

[0016] A temporal situational awareness system based on semantic micro-clusters includes: The semantic tagging module is used to collect work order texts within the current time window, perform multi-dimensional semantic tagging on the work order texts, obtain work order embedding vectors and timeliness quality, classify the work order embedding vectors into semantic micro-clusters according to semantic proximity, and construct a disjoint set forest. The centroid stabilization module is used to calculate the rotation tensor for each semantic micro-cluster in the disjoint set forest based on the work order embedding vector and timeliness quality of its members, obtain the stabilized centroid based on the rotation tensor, update the merging relationship of the disjoint set forest based on the stabilized centroid, and construct an interval tree with the active time interval of the semantic micro-cluster as the index and the stabilized centroid as the node value. The situation warning module is used to perform interval overlap queries on the interval tree to obtain the set of concurrent semantic micro-clusters in the active time interval. It calculates the situation level based on the stable centroid and the time-series change in the number of members of the concurrent semantic micro-clusters. When the situation level meets the warning conditions, it generates situation warning data.

[0017] Compared with the prior art, the beneficial effects of the present invention are: By converting work order text within the current time window into work order embedding vectors and timeliness quality, and then aggregating them using a disjoint set forest, a structured organization of request content is achieved. A rotation tensor reflecting the spatial distribution characteristics of the point cloud is calculated using the work order embedding vectors and timeliness quality of each member within the semantic micro-cluster. The stable centroid is then calculated from the rotation tensor. Adjusting the merging state of each set in the disjoint set forest using the stable centroid helps to smooth out abrupt changes in centroid position caused by fluctuations in small-sample work orders and suppress positional shifts caused by the evolution of long-cycle events. Combined with an interval tree indexed by the active time interval for interval overlap queries, concurrent semantic micro-cluster sets existing in the same time period are directly identified. The situation level is calculated based on the difference between the displacement of the stable centroid and the number of members, ensuring that continuously occurring request events maintain a stable temporal correlation between adjacent time windows. This eliminates the artificially inflated hotspot events in the monitoring dashboard caused by flickering, broken, and mis-merging of event trajectories.

[0018] By setting anchor centroids in each semantic microcluster based on a predetermined number of vectors initially entering, and combining shape anisotropy (reflecting the degree of isotropy of the point cloud) and drift inertia increment (reflecting the centroid's deviation from the starting point), a quantitative characterization of the internal conformational state of the semantic microcluster is achieved. Stable condensed conformations and small-sample degenerate conformations are distinguished by conformational identifiers. In small-sample degenerate conformations, the moving centroid and anchor centroid are convexly combined using convex combination coefficients, and the inertial impedance coefficient constrains the centroid's movement step towards the weighted average of the current time window, enabling targeted convergence of unreliable current position estimates. Simultaneously, the centroid drift rate is obtained from the displacement of stable centroids between adjacent time windows, and the size growth rate is obtained from the difference in the number of members. After normalization, these are weighted and synthesized into a situational level, assessing the severity of events from two dimensions: semantic shift and request surge, preventing the underestimation of rapidly evolving events caused by single-dimensional indicators. Attached Figure Description

[0019] To more clearly illustrate the technical solutions of the embodiments of the present invention, the accompanying drawings used in the description of the embodiments will be briefly introduced below. The following drawings are not drawn to scale according to the actual size, but are intended to show the main idea of ​​the present invention.

[0020] Figure 1 This is a system block diagram of a temporal situational awareness system based on semantic micro-clusters. Figure 2 This is a flowchart of a temporal situational awareness method based on semantic micro-clusters; Figure 3 Schematic diagram for determining the stable centroid; Figure 4 A schematic diagram illustrating the conformational differentiation of the rotation tensor; Figure 5 This is a schematic diagram of a forest of disjoint sets. Figure 6 This is a diagram illustrating interval overlap queries. Detailed Implementation

[0021] The technical solutions in the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are also within the scope of protection of the present invention.

[0022] Example 1 The underlying cause of the aforementioned problems is as follows: The existing method uses the arithmetic mean of member vectors within a cluster as the cluster center and associates clusters in adjacent time windows based on the proximity of their cluster centers using a fixed similarity threshold. When the number of related work orders within a single time window is small, the number of member vectors participating in the averaging is insufficient. The cluster center position shifts significantly with minor changes in individual members, disrupting the proximity-based association, which manifests as flickering and broken evolution trajectories on the monitoring screen. When the event duration is long, the user's description of the same event changes as the event progresses. The cluster center is continuously shifted as it is updated window-by-window, gradually deviating from the initial position of the member vectors. Work orders in subsequent time windows whose similarity to the shifted cluster center falls below the fixed threshold are classified as new clusters, causing the same event to be split into multiple events. Furthermore, the shifted cluster center's similarity to adjacent event work orders increases, incorporating those adjacent event work orders and resulting in the incorrect merging of different events.

[0023] To resolve the two types of instability mentioned above, such as Figure 2 As shown, the present invention proposes a temporal situational awareness method based on semantic micro-clusters, including steps S10, S20 and S30.

[0024] Step S10: Collect the work order text within the current time window, perform multi-dimensional semantic tagging on the work order text to obtain the work order embedding vector and timeliness quality, classify the work order embedding vector into semantic micro-clusters according to semantic proximity, and construct a disjoint set forest that represents the member affiliation and merging relationship of the semantic micro-clusters; Step S10 includes steps S11, S12 and S13.

[0025] Step S11: Collect the work order text within the current time window, perform word segmentation and noise removal on the work order text to obtain a regularized work order text.

[0026] The work order text refers to the text record output by the customer service hotline acceptance system at the work order acceptance interface, recording a single user call request. The data type is a variable-length character sequence. Each work order text includes an access time given by the acceptance system's clock. The access time is measured in units of time, and its numerical range falls between the start and end times of the current time window. Specifically, the word segmentation and noise reduction methods are as follows: the work order text is segmented into word sequences according to the target domain word segmentation dictionary; punctuation marks and stop words are removed from the word sequences; synonymous colloquial expressions are grouped into standardized words according to the target domain thesaurus; and the words are arranged in their original order to obtain a regularized work order text, which is the standardized word sequence after noise removal. The target domain word segmentation dictionary is obtained as follows: Based on a general Chinese word segmentation dictionary, the frequency of occurrence of each string is counted from historical work order texts in the target scenario. Strings with a frequency higher than a preset frequency threshold and not included in the general Chinese word segmentation dictionary are selected as candidate domain words. Department names, business type names, and administrative division names in the service item directory are selected as supplementary domain words. The candidate domain words and supplementary domain words are then incorporated into the general Chinese word segmentation dictionary to obtain the target domain word segmentation dictionary. The preset frequency threshold is obtained as follows: A set of historical work order texts in the target scenario is taken, and candidate domain words are extracted for different frequency values. The validity of each candidate domain word as a domain word is manually verified. The minimum frequency value that ensures the proportion of verified valid candidate domain words is not less than a preset proportion is taken as the preset frequency threshold. For example, the preset proportion is 90%. The target domain thesaurus is obtained as follows: different colloquial expressions pointing to the same service item are grouped together from historical work order texts in the target scenario. The standardized item name in each group is designated as the standardized word of that group. The correspondence between the colloquial expressions and their standardized words constitutes the target domain thesaurus. Step S11 uses word segmentation and denoising instead of directly encoding the work order text because the work order text contains a large number of colloquial particles and punctuation noise. Direct encoding would result in two requests with similar semantic content having significantly different work order embedding vectors. Word segmentation and denoising ensure that requests with the same semantic content fall into similar standardized word sequences, and after subsequent encoding, their positions are adjacent, providing a basis for the aggregation of work orders for the same event.

[0027] Step S12: Perform multi-dimensional semantic tagging on the regularized work order text, extract multi-dimensional semantic tags containing subject, object, time, location, request and emotion, encode the multi-dimensional semantic tags into work order embedding vectors, and calculate the timeliness quality based on the access time of the work order text within the current time window.

[0028] Specifically, the multidimensional semantic labeling method is as follows: the standardized work order text is input into the semantic labeling model, which outputs multidimensional semantic labels. Each multidimensional semantic label is a set of value fields, recording dimensions such as the request initiator, the target object, the time involved, the location of the occurrence, the request category, and the sentiment tendency. The value for each dimension is an item from a predefined set of values ​​for that dimension. The semantic labeling model is a sequence labeling model, consisting of a word embedding layer, a bidirectional long short-term memory network layer, and a conditional random field layer. The word embedding layer maps each standardized word in the standardized work order text to a word vector, and the bidirectional long short-term memory network layer performs forward and backward encoding on the word vector sequence. The context representation of each position is obtained. The Conditional Random Field layer outputs the label of each position in terms of subject, object, time, location, request, and emotion according to the context representation of each position. The semantic labeling model uses the work order text and manually annotated multidimensional semantic labels as training data. During training, the regularized work order text is input into the semantic labeling model to obtain the output labels of each dimension. The cross-entropy between the manually annotated labels and the output labels is used as the loss function. The gradient of the loss with respect to the parameters of each parameter of the word embedding layer, the bidirectional long short-term memory network layer, and the conditional random field layer is calculated by the error backpropagation algorithm. The parameters of each layer are updated by the gradient descent method and iterated until the loss no longer decreases.The work order embedding vector is equal to a d-dimensional real vector obtained by concatenating the values ​​of each dimension of the multi-dimensional semantic label in dimensional order and inputting it into the semantic coding model for mapping, where d is the output dimension set by the semantic coding model. The semantic coding model includes a label embedding layer, several fully connected layers, and an L2 norm normalization layer. The label embedding layer maps the values ​​of each dimension of the multi-dimensional semantic label into embedding vectors and concatenates them in dimensional order. Several fully connected layers perform nonlinear mapping on the concatenation result to obtain a d-dimensional real vector. The L2 norm normalization layer divides the d-dimensional real vector by its Euclidean norm to obtain a d-dimensional real unit vector normalized by the L2 norm. The semantic coding model uses work order pairs with the same event and work order pairs with different events as training data, with the constraint that the vector dot product of work order pairs with the same event is greater than the vector dot product of work order pairs with different events. During training, for each triple consisting of a work order pair with the same event and a work order pair with different events, the interval loss is used as the loss function. The interval loss is equal to the larger of zero and the following result, where the result is the preset interval minus the same event and work order pairs with different events. The vector dot product of one event work order pair, plus the vector dot products of different event work order pairs, is used to calculate the gradient of the loss with respect to the parameters of the label embedding layer and each fully connected layer using the backpropagation algorithm. The parameters are then updated using gradient descent, iterating until the loss no longer decreases. The preset interval is the minimum difference required in the interval loss where the vector dot product of the same event work order pair exceeds the vector dot products of different event work order pairs. The preset interval is obtained by training the semantic coding model with different values ​​between zero and one, and statistically analyzing the discrimination accuracy of the semantic coding model for the same event work order pair and different event work order pairs on the validation set. The value that maximizes the discrimination accuracy is taken as the preset interval. For example, the preset interval is 0.2. The input of the semantic coding model is the concatenation result of multi-dimensional semantic labels, and the output is a d-dimensional real unit vector normalized by the L2 norm. This ensures that the work order embedding vectors of work orders with similar request content are adjacent in the embedding space, and that the distance between vectors in the embedding space and each element of the subsequent rotation tensor are dimensionless and fall within a bounded range.

[0029] Timeliness quality refers to the relative position of a single work order within the current time window. Timeliness quality is equal to the difference between the work order text's access time and the start time of the current time window, divided by the time window length. The value of timeliness quality ranges from zero to one. A value of zero corresponds to a work order accessed at the start time of the current time window, and a value of one corresponds to a work order accessed at the end time of the current time window. Timeliness quality is calculated based on the time window in which the work order text's access time falls. The calculated value is retained along with the work order's embedding vector and is not recalculated when the time window changes. When the sum of the timeliness quality of all work order embedding vectors participating in the same calculation is zero, the timeliness quality of that group of work order embedding vectors is uniformly set to one. The time window length is obtained as follows: it is determined based on the longest response delay required for hotspot discovery and the average arrival rate of work orders in the target scenario. The time window length does not exceed the longest response delay, and the average number of work orders within each time window obtained by dividing the time window by this time window length is not less than the minimum number of work orders required for clustering. The minimum number of work orders is the preset number required to determine the anchor centroid in step S21. For example, the time window length is five minutes. The time window length is calibrated in the offline stage in the above manner, and the calibrated time window length remains unchanged as a fixed value set by the system during online operation.

[0030] The work order embedding vector is also assigned a timeliness quality because work orders that are accessed closer to the end of the time window can better reflect the semantic content of the current stage of the event. Using timeliness quality as a weight in the subsequent calculation of the rotation tensor and centroid can make the calculation results tilt towards the semantics of newer requests within the time window, so that the centroid characterization results fit the semantics of the current stage of the event.

[0031] like Figure 5 The diagram shown is a schematic of a disjoint set forest.

[0032] Reference Figure 5 This diagram illustrates a disjoint forest of rooted trees. The arrowed lines connecting nodes represent pointers to their parent nodes; the node a parent pointer points to is the root node, which serves as the representative element of that semantic micro-cluster. The "derivation source" field next to the root node represents the representative element of the parent semantic micro-cluster from which this semantic micro-cluster is derived at the time of splitting. It shows two independent rooted trees—the parent set and the child set—and the tracing record of the parent set's representative element through the "derivation source" field.

[0033] Step S13: Assign the work order embedding vector to the corresponding semantic micro-cluster according to its semantic proximity to the current centroid of the existing semantic micro-cluster. Create a new semantic micro-cluster for work order embedding vectors whose semantic proximity is lower than the preset assignment threshold. Merge the work order embedding vectors assigned to the same semantic micro-cluster into the same set of the disjoint set forest to obtain the disjoint set forest that represents the membership and merging relationship of the semantic micro-cluster.

[0034] Both the disjoint set forest and the interval tree in step S24 are data structures that are continuously maintained across time windows. They are updated one time window at a time starting from the system's initial time window and are not rebuilt when the time window switches. The existing semantic micro-clusters refer to semantic micro-clusters that have been created in previous time windows and are still registered in the disjoint set forest. The current centroid of an existing semantic micro-cluster is the stable centroid of the semantic micro-cluster output in step S22 in the previous time window and retained with the semantic micro-cluster in the current time window. When the current time window is the system's initial time window, there are no existing semantic micro-clusters. All work order embedding vectors accessed in the current time window trigger the creation of new semantic micro-clusters.

[0035] Specifically, the semantic proximity is calculated as follows: take the inner product of the work order embedding vector and the current centroid of an existing semantic micro-cluster, divide the inner product by the product of the Euclidean norm of the work order embedding vector and the Euclidean norm of the current centroid of the existing semantic micro-cluster, and obtain the semantic proximity between the work order embedding vector and the existing semantic micro-cluster. The semantic proximity ranges from negative one to one, and the larger the value, the closer the semantic content of the two is.

[0036] The method for determining inclusion is as follows: For a work order embedding vector, calculate its semantic proximity with the current centroid of all existing semantic microclusters. Select the existing semantic microcluster with the largest semantic proximity as a candidate semantic microcluster. When the semantic proximity of the candidate semantic microcluster is greater than or equal to the preset inclusion threshold, the work order embedding vector is included in the candidate semantic microcluster. When the semantic proximity of the work order embedding vector with all existing semantic microclusters is less than the preset inclusion threshold, a new semantic microcluster is created based on the work order embedding vector. The preset inclusion threshold is obtained as follows: Collect multiple sets of work order pairs belonging to the same event and multiple sets of work order pairs belonging to different events. For each set of work order pairs, calculate the semantic proximity between the work order embedding vectors of the two work orders. Statistically calculate the lower bound of the semantic proximity distribution of work order pairs for the same event and the upper bound of the semantic proximity distribution of work order pairs for different events. Take the median between the lower bound of the semantic proximity distribution of work order pairs for the same event and the upper bound of the semantic proximity distribution of work order pairs for different events as the preset inclusion threshold. The lower bound of the distribution refers to the value at a lower quantile after arranging the semantic proximity of the work order pairs from smallest to largest. The upper bound of the distribution refers to the value at a higher quantile after arranging the semantic proximity of the work order pairs from smallest to largest. Quantiles are used instead of minimum and maximum values ​​to avoid the influence of individual outliers. For example, the lower bound is taken as the value at the 5th percentile, and the upper bound is taken as the value at the 95th percentile. The determination method of the lower and upper bounds of the distribution in subsequent parts of this specification is consistent with this. For example, the preset inclusion threshold is 0.7. After all the work order embedding vectors accessed in the current time window have been included in an existing semantic micro-cluster or a newly created semantic micro-cluster, the construction of the disjoint set forest for the current time window is completed.

[0037] The disjoint forest is a set data structure consisting of several rooted trees. Each work order embedding vector is a node, and each rooted tree corresponds to a semantic micro-cluster. The root node is the node whose parent pointer points to itself, and it serves as the representative element of that semantic micro-cluster. Each node stores a pointer to its parent node, and each root node also stores a derivation source field. This derivation source field records the representative element of the parent semantic micro-cluster from which the semantic micro-cluster is derived at the time of splitting. The derivation source field of semantic micro-clusters not obtained by splitting is set to null. Work order embedding vectors belonging to the same semantic micro-cluster are connected to the same rooted tree through a merging operation. The merging operation involves merging the parent pointer of the root node of the rooted tree with fewer nodes. The root node of a rooted tree with a large number of nodes is pointed to. When querying the semantic micro-cluster to which a work order embedding vector belongs, the parent nodes are visited sequentially along the parent pointer from the node corresponding to the work order embedding vector until the root node is reached. The root node visited is the representative element. All nodes traversed from the starting node to the root node constitute the path from the node to the root. After the query is completed, the parent pointers of each node on the path are changed to point to the root node. The root node of each rooted tree also stores the number of nodes in the rooted tree. The number of nodes is updated synchronously during inclusion, merging, and splitting. The number of members in the disjoint set forest is equal to the total number of work order embedding vectors that have been included, and the number of rooted trees is equal to the number of existing semantic micro-clusters. The single cost of merging operations and querying representative elements in the disjoint set forest is close to constant, which can support frequent merging of a large number of work orders within a minute-level time window. Its derivation source field provides a structural carrier for tracing the derivation source when the same event is split into parent and child events.

[0038] The work order embedding vector physically corresponds to the location point of the user request in the embedding space after semantic tagging and encoding. Multiple requests triggered by the same real-world event cluster into nearby point groups due to similar semantic content, while requests triggered by different real-world events are located in different point groups. Semantic proximity is used to measure the closeness between the work order embedding vector and the current centroid of the semantic micro-cluster. This is based on the fact that the training constraints of the semantic encoding model make the vector dot product of work order pairs with the same event greater than the vector dot product of work order pairs with different events. Therefore, the angle between the same event request and the current centroid of the semantic micro-cluster is small, indicating high semantic proximity, while the angle between different event requests is large, indicating low semantic proximity. Disjoint set forests are used to handle member affiliation because path compression and merging by node number control the height of the rooted tree, and the single cost of merging and querying representative elements is close to constant. If independent member lists are used, each merge requires traversing and moving members, resulting in a large cumulative traversal within a single window, and the splitting derivation relationship cannot be preserved when events are split.

[0039] The disjoint set forest organizes the embedded vectors of work orders accessed in the current time window into several rooted trees according to semantic content. This allows scattered requests with the same topic to be aggregated into objects that can be processed as a whole in subsequent steps. The granularity of the processed objects is increased from a single work order to a semantic micro-cluster. Subsequent centroid stabilization, merging, and splitting of semantic micro-clusters can be accomplished through operations on the representative elements. When performing derivation source tracing, starting from the representative element of the semantic micro-cluster to be traced, the derivation source field of its root node is read to obtain the representative element of its parent semantic micro-cluster at the split time. Then, the derivation source field of the parent semantic micro-cluster is read, and so on, tracing upwards level by level until the derivation source field is empty. The representative elements traversed are connected in sequence to reconstruct the derivation chain of the event from the initial semantic micro-cluster. The evolution process of the same event is maintained as a traceable derivation chain. The derivation source field is thus read and used by the derivation source tracing operation, rather than just recorded. Thus, step S13 lays the organizational foundation for subsequent centroid stabilization and temporal tracing at the semantic micro-cluster level.

[0040] Step S20: Receive the disjoint forest set; for each semantic micro-cluster in the disjoint forest set, calculate the rotation tensor based on the work order embedding vector and timeliness quality of its members; obtain the stable centroid based on the rotation tensor; update the merge relation of the disjoint forest set based on the stable centroid; and construct an interval tree with the active time interval of the semantic micro-cluster as the index and the stable centroid as the node value. Step S20 includes steps S21, S22, S23, and S24.

[0041] Step S21: For each semantic micro-cluster in the disjoint set forest, collect the work order embedding vectors and timeliness quality of the members of the semantic micro-cluster, obtain the current centroid of the semantic micro-cluster, and the anchor centroid determined and fixed by the mean of the preset number of work order embedding vectors first assigned to the semantic micro-cluster.

[0042] Specifically, the method for collecting members is as follows: a semantic micro-cluster is identified by the representative element of each rooted tree in the disjoint set forest. All nodes of the rooted tree are traversed along the parent pointer, and the work order embedding vector and timeliness quality corresponding to each node are extracted. The current centroid refers to the centroid position of the semantic micro-cluster before entering step S22. For semantic micro-clusters that have existed in previous time windows and have been continuously registered in the disjoint set forest, the current centroid is the stable centroid of the semantic micro-cluster output in step S22 in the previous time window and retained with the semantic micro-cluster to the current time window. For semantic micro-clusters newly created in the current time window, the current centroid is the vector obtained by multiplying the work order embedding vectors of all current members of the semantic micro-cluster by their respective timeliness quality, summing them, and then dividing by the sum of the timeliness quality. Anchor centroid refers to a fixed reference vector that serves as the semantic starting point of the semantic micro-cluster. When the number of members in the semantic micro-cluster is less than a preset number, the anchor centroid is the arithmetic mean of the work order embedding vectors of all current members of the semantic micro-cluster. When the number of members in the semantic micro-cluster reaches the preset number, the anchor centroid is determined by the arithmetic mean of the preset number of work order embedding vectors that were first assigned to the semantic micro-cluster and remains fixed. The preset number refers to the minimum number of members required to determine the anchor centroid. The preset number is obtained by collecting member assignment sequences of multiple semantic micro-clusters that have been confirmed as single events in the target scenario. For each member assignment sequence, members are added one by one starting from the first member, and the arithmetic mean of the current work order embedding vectors of all members is calculated. The number of members corresponding to the first time the Euclidean norm of the displacement before and after adding a new member is less than one-tenth of the mean of the Euclidean distance between all pairs of work order embedding vectors of the current members of the semantic micro-cluster is recorded. The maximum value of the number of members corresponding to all member assignment sequences is taken as the preset number. The aforementioned preset number is calibrated in the manner described above during the offline phase, and the calibrated preset number remains unchanged as a fixed value set by the system during online operation.

[0043] The current centroid and the anchor centroid are used to carry the semantic micro-cluster separately because the current centroid reflects the latest semantic position of the semantic micro-cluster as the time window moves and is used to follow the event evolution, while the anchor centroid freezes the initial semantic position of the semantic micro-cluster and is used to constrain the positional offset in long-term evolution, so that the semantic micro-cluster can both follow the event evolution and not deviate from the semantic starting point of the event.

[0044] like Figure 3 The diagram shown illustrates the principle for determining the stable centroid.

[0045] Reference Figure 3This diagram illustrates the process of calculating a stable centroid from the current centroid, anchor centroid, member work order embedding vectors, and timeliness / quality. The dots labeled as anchor centroids are determined and fixed by the mean of the first preset number of work order embedding vectors to which the semantic micro-cluster is assigned; the dots labeled as the current centroid reflect the current semantic position of the micro-cluster. The diagram also demonstrates the process of moving the current centroid along the direction pointing to the weighted mean of the current time window to obtain a moving centroid. The arrows in this direction represent the allowed step size, which is the product of the difference between the mean and the current centroid and the inertial impedance coefficient. Furthermore, the diagram shows the processing when the conformation identifier is the second value, i.e., a small-sample degenerate conformation. The line connecting the moving centroid and the anchor centroid represents a convex combination of the moving centroid and the anchor centroid according to the convex combination coefficient. The position point obtained by the convex combination is the stable centroid, indicating that the stable centroid is constrained to the vicinity of the semantic starting point.

[0046] Step S22: Calculate the stable centroid of the semantic micro-cluster based on the current centroid, anchor centroid, work order embedding vector of the member, and timeliness quality. Step S22 includes steps S221, S222, and S223.

[0047] Step S221: The work order embedding vectors of the members of the semantic micro-cluster are used as mass points with corresponding timeliness quality. The current centroid of the semantic micro-cluster is used as the reference point to calculate the gyration tensor of the mass point. The gyration tensor is a second-order moment matrix of the deviation vector of each member's work order embedding vector relative to the current centroid, weighted by timeliness quality. The gyration tensor is subjected to eigenvalue decomposition to obtain the principal gyration radius square sequence.

[0048] Specifically, the work order embedding vector of each member of this semantic micro-cluster is considered as a mass point in the embedding space. The position of the mass point is the work order embedding vector, and the mass of the mass point is the timeliness quality corresponding to that work order embedding vector. The rotation tensor is calculated according to the following formula: , in the formula, For the rotation tensor, there is a A real symmetric matrix of order 1; The number of members of the semantic micro-cluster is derived from the total number of nodes of the semantic micro-cluster collected in step S21. For this semantic micro-cluster The work order embedding vector of each member is, A 3D real unit vector, the value of which comes from the output of step S12; For the first The timeliness quality of each member, with a value ranging from zero to one, is derived from the output of step S12. This is the current centroid of the semantic micro-cluster, and its value comes from step S21; For the first The deviation vector of each member's work order embedding vector relative to the current centroid. The outer product of the deviation vector and itself results in a... The formula is constructed as follows: The rotation tensor originates from the rotation tensor used in polymer physics to characterize the spatial conformation of chain-like molecules. For each member, the product of its deviation vector relative to the current centroid and itself is taken. After weighting by the time-dependent mass of that member, the product is summed over all members and then normalized by dividing by the sum of time-dependent masses. The diagonal elements of the resulting matrix represent the mass-weighted mean square distribution of the member's deviation from the current centroid in each coordinate direction, while the off-diagonal elements represent the correlation of the offsets in different coordinate directions. The entire rotation tensor characterizes the spatial distribution of the point cloud of the semantic microcluster members relative to the current centroid. Normalization by dividing by the sum of time-dependent masses ensures that the rotation tensor is not affected by the absolute value of the number of members but only reflects the distribution shape of the deviation.

[0049] Eigenvalues ​​are decomposed into d non-negative eigenvalues. The d non-negative eigenvalues ​​are arranged in descending order to form the principal gyration radius square sequence. The j-th element of the principal gyration radius square sequence is the quality-weighted mean square distribution of the member point cloud along the j-th principal direction. Considering that the number of members in this semantic micro-cluster is usually much smaller than the dimension d of the embedding vector, the number of non-zero eigenvalues ​​of the rotation tensor does not exceed the number of members. Specifically, when the current centroid is taken as the time-weighted average of the members' time-dependent quality, the number of non-zero eigenvalues ​​does not exceed the number of members minus one. When the current centroid is taken as the stable centroid of the previous time window, the number of non-zero eigenvalues ​​does not exceed the number of members. Eigenvalue decomposition can be performed on the Gram matrix formed by the member deviation vectors to obtain non-zero eigenvalues, and the remaining eigenvalues ​​are set to zero to reduce the amount of computation. The element in the i-th row and j-th column of the Gram matrix is ​​equal to the inner product of the square root of the time-dependent quality of the i-th member and its deviation vector, and the product of the square root of the time-dependent quality of the j-th member and its deviation vector, divided by the sum of the time-dependent qualities of all members. The non-zero eigenvalues ​​of the Gram matrix formed in this way are equal to the non-zero eigenvalues ​​of the rotation tensor. When the semantic micro-cluster contains only one member or the work order embedding vectors of all members overlap, all deviation vectors are zero vectors, the gyration tensor is a zero matrix, and all elements of the principal gyration radius square sequence are zero.

[0050] The embedded vector point cloud of an employee's order physically corresponds to the set of location points formed in the embedding space after semantic encoding of multiple requests for the same real-world event. An event that has been formed and is semantically focused corresponds to a point cloud with similar distribution in all directions, while an event consisting of only a few requests or in the semantic transition stage corresponds to a point cloud that is only distributed in a few directions. The rotation tensor is used to characterize the distribution of point clouds. The basis for this is that the rotation tensor of the mass system is equal to the mass-weighted second moment of the members relative to the reference point, and its eigenstructure gives the distribution scale of the point cloud along each principal direction. When the current centroid is taken as the time-weighted mean of the members, the rotation tensor is the time-weighted covariance matrix, and the eigenvalues ​​are the time-weighted variances along each principal direction. When the time-weights of the members are equal, it is equivalent to the covariance matrix of the point cloud position coordinates. When the current centroid is taken as the stable centroid of the previous time window, the rotation tensor contains a component contributed by the offset of the current centroid relative to the weighted mean, in addition to the time-weighted second central moment. This component increases with the increase of the offset, so that the rotation tensor can characterize a higher distribution and anisotropy, which is consistent with the judgment that the reliability of the current centroid estimation of the semantic microcluster decreases. If we only use the scalar of the average distance from the member to the current centroid, then point clouds that are similarly distributed in all directions and degenerate point clouds that are distributed only in a single direction may have the same average distance and become indistinguishable.

[0051] The rotation tensor and the principal gyration radius square sequence quantitatively represent the spatial distribution of the member point cloud using a matrix and its eigenvalues, making the current conformational state of the semantic micro-cluster a quantity that can be calculated and compared in subsequent steps: a point cloud composed of a small number of members has non-zero distribution only in a few principal directions, and the principal gyration radius square sequence shows a few large values ​​and most zero values; the point cloud of formed events has similar distribution in each principal direction, and the principal gyration radius square sequence shows a similar distribution. Since the rotation tensor is time-weighted by the time mass and normalized by dividing by the sum of the time mass, the principal gyration radius square sequence only reflects the distribution shape, and the conformations of semantic micro-clusters of different sizes can be compared under the same standard. The trace of the rotation tensor is equal to the sum of the elements of the principal gyration radius square sequence, that is, the total mass-weighted mean square distribution of the member point cloud relative to the current centroid, which is used in step S223 to calculate the inertial impedance coefficient to constrain the centroid movement amplitude.

[0052] like Figure 4 The diagram shown is a schematic of the configuration of the rotation tensor.

[0053] Reference Figure 4The diagram illustrates the process of distinguishing between stable condensed conformations and small-sample degenerate conformations using the rotation tensor of a point cloud composed of member work order embedding vectors. The left side of the diagram shows a compact point cloud with similar distribution in all principal directions, corresponding to an isotropic compact spherical conformation. The squared principal gyration radii sequence obtained from this point cloud have similar values ​​for each element, and the shape anisotropy tends to zero, indicating that this semantic micro-cluster is in a stable condensed conformation. The conformation identifier is set to the first value. The right side of the diagram shows a sparse point cloud with distribution only in a few principal directions, corresponding to a one-dimensional extended conformation. The squared principal gyration radii sequence obtained from this point cloud exhibits a few large values ​​and most zero values, and the shape anisotropy tends to one, indicating that this semantic micro-cluster is in a small-sample degenerate conformation. The conformation identifier is set to the second value. The distribution patterns of the point clouds on both sides of the diagram represent the mass-weighted mean square distribution of the point clouds along the principal directions obtained by the eigenvalue decomposition of the rotation tensor. This shows that the intrinsic structure of the rotation tensor distinguishes the degenerate point cloud containing only a few requests from the compact point cloud containing formed events.

[0054] Step S222: The normalized variance of the squared sequence of the principal gyration radius is taken as the shape anisotropy; according to the parallel axis shift theorem, the product of the sum of the time-dependent mass and the square of the Euclidean distance from the current centroid to the anchor centroid is taken as the drift inertia increment; a first interval and a second interval are set, and the boundary between the first interval and the second interval is determined by the situation level output in step S30 of the previous time window. When the shape anisotropy is in the first interval, the configuration identifier is set to the first value, and when the shape anisotropy is in the second interval, the configuration identifier is set to the second value.

[0055] Specifically, the anisotropy of shape is calculated using the following formula: , in the formula, For shape anisotropy; The dimension of the embedding vector; The first of the main gyration radius square sequence Each element has a value derived from step S221; The arithmetic mean of the sequence of squared principal gyration radii is equal to the trace of the gyration tensor divided by . The principle behind this formula is as follows: the shape anisotropy originates from the relative shape anisotropy in polymer physics, which characterizes the degree to which chain conformations deviate from isotropy. The derivation starts with the variance of the principal gyration radius square sequence. The variance is the sum of the squares of the differences between each element and the arithmetic mean divided by the number of elements. Dividing the variance by the square of the arithmetic mean yields the dimensionless relative dispersion. Dividing by the dimension minus one normalizes the relative dispersion to zero to one. When all elements of the principal radius of gyration square sequence are equal, the member point cloud unfolds identically in all principal directions, corresponding to an isotropic compact spherical conformation with zero variance and zero shape anisotropy. When only one element of the principal radius of gyration square sequence is non-zero and the rest are zero, the member point cloud unfolds only in a single principal direction, corresponding to a one-dimensional extended conformation. Let the unique non-zero element be a positive value. The arithmetic mean is equal to this positive value divided by d. The variance is equal to the square of this positive value multiplied by the dimension minus one, then divided by the square of d. The variance divided by the square of the arithmetic mean gives the dimension minus one, and dividing by the dimension minus one again gives one. The shape anisotropy ranges from zero to one. When all elements of the principal radius of gyration square sequence are zero, i.e., the arithmetic mean is zero, the shape anisotropy is taken as one, representing the value for the degenerate conformation. The drift inertia increment is equal to the product of the sum of the time-dependent masses of the members of the semantic microcluster and the square of the Euclidean distance from the current centroid to the anchor centroid. This is based on the parallel axis shift theorem of classical mechanics, which states that when the reference point of the rotational inertia is shifted from the centroid to another reference point, the increment of the rotational inertia is equal to the product of the sum of the masses and the square of the distance between the two reference points. When the current centroid is taken as the weighted average of the time-dependent masses of the members, i.e., the centroid of the particle system, the drift inertia increment is equal to the increment of the rotational inertia of the semantic microcluster when the reference point is shifted from this centroid to the anchor centroid. The rotational inertia is equal to the product of the trace of the rotation tensor and the sum of the time-dependent masses of the members of the semantic microcluster. When the current centroid is taken as the stabilized centroid of the previous time window, the drift inertia increment, in the form of the parallel axis shift theorem, quantifies the positional offset of the current centroid relative to the anchor centroid as a quantity with the same shape as the rotational inertia increment.

[0056] The conformation identifier is determined as follows: the boundary boundary is equal to the product of the baseline boundary boundary and the following result, which is the sum of the value one and the situation level output by the semantic microcluster in step S30 in the previous time window. The situation level ranges from zero to one. When the semantic microcluster does not have a situation level in the previous time window, the situation level is zero. When the boundary boundary is greater than one, the situation level is one. When the shape anisotropy is less than or equal to the boundary boundary, the semantic microcluster is determined to be in a stable condensed conformation and the conformation identifier is set to the first value. When the shape anisotropy is greater than the boundary boundary, the semantic microcluster is determined to be in a small sample degenerate conformation and the conformation identifier is set to the second value. The benchmark boundary is obtained as follows: multiple sets of semantic micro-cluster samples that have been confirmed as formed single events and multiple sets of semantic micro-cluster samples that have been confirmed as small sample degeneration are collected respectively. The shape anisotropy degree is calculated for each set of samples. The upper bound of the shape anisotropy degree distribution of the formed single event samples and the lower bound of the shape anisotropy degree distribution of the small sample degeneration samples are statistically analyzed. The median between the upper bound and the lower bound of the distribution is taken as the benchmark boundary. For example, the benchmark boundary is taken as 0.5.

[0057] Shape anisotropy and drift inertia increment correspond to the spread shape of the member point cloud and the degree of offset of the current centroid relative to the semantic starting point, respectively. The former reflects whether the semantic cluster is a formed event or a transitional state containing only a few requests, while the latter reflects the migration of event semantics over time. The basis for distinguishing conformations by shape anisotropy is the relative dispersion of the rotation tensor eigenvalues, which characterizes the degree to which the point cloud deviates from isotropy. Small sample point clouds spread only in a few principal directions have a large relative dispersion and high shape anisotropy, while formed point clouds with similar spread in all directions have a small relative dispersion and low shape anisotropy. The basis for measuring long-term offset by drift inertia increment is the parallel axis shift theorem. The offset of the current centroid relative to the anchor centroid is characterized by a rotational inertia increment equal to the product of the sum of masses and the square of the offset distance, making long-term evolutionary offset a calculable quantity. If only the threshold of the number of members is used to judge small samples, it will be impossible to identify transitional point clouds that have a large number of members but are still degenerating and spreading in a single direction, and the transitional state will be mistakenly treated as a morphology.

[0058] Shape anisotropy and conformational identifiers determine whether a semantic cluster is currently in a jittery degenerate state or a smoothly followable condensed state using a single identifier value. This allows subsequent centroid calculations to select different processing branches based on conformation. Since the boundary is determined by the situational level of the semantic cluster in the previous time window, a higher situational level corresponds to a larger boundary, making it easier for semantic clusters in active events to be classified as condensed states when shape anisotropy is equal. Consequently, during periods of high demand for active events, these clusters tend to follow the current semantic state rather than reverting to the anchor centroid. Shape anisotropy is a dimensionless quantity ranging from zero to one; drift inertia increment is also a dimensionless quantity, and its value increases as the sum of the time-dependent mass of the members of the semantic micro-cluster increases. Since the drift inertia increment only participates in convex combination in branches where the conformational identifier is the second value, i.e., small sample degenerate conformations, this branch corresponds to semantic micro-clusters with fewer members and a smaller sum of time-dependent mass. At this time, the drift inertia increment and shape anisotropy are on comparable numerical orders. Subsequent step S223 then uses a preset regression constant to absorb the typical order of magnitude of the sum of shape anisotropy and drift inertia increment, so that the convex combination coefficient falls within a certain range. Thus, step S222 transforms the two types of causes of dual instability into a definite conformational identifier and a calculable drift inertia increment, respectively.

[0059] Step S223: The reciprocal of the sum of the trace of the rotation tensor and the preset inertia reference is used as the inertial impedance coefficient; the mean of the work order embedding vectors belonging to the semantic micro-cluster in the current time window is calculated by weighting by timeliness and quality; the current centroid is moved by an allowable step along the direction pointing from the current centroid to the mean to obtain the moving centroid, the allowable step being the product of the difference between the mean and the current centroid and the inertial impedance coefficient; when the conformation identifier is a first value, the moving centroid is used as the stable centroid; when the conformation identifier is a second value, the moving centroid and the anchor centroid are convexly combined according to the convex combination coefficient to obtain the stable centroid, the convex combination coefficient being the ratio of the sum of the shape anisotropy and the drift inertia increment to the sum plus the preset regression constant.

[0060] Specifically, the inertial impedance coefficient is equal to the reciprocal of the sum of the trace of the rotation tensor and the preset inertial reference. The trace of the rotation tensor is derived from step S221 and is the mass-weighted mean square distribution of the member point cloud relative to the current centroid. The preset inertial reference refers to the reference inertia that ensures the inertial impedance coefficient falls within the range of greater than zero and not exceeding one. The preset inertial reference is obtained by taking the sum of the value one and the median of the traces of the rotation tensors of multiple semantic microclusters with sufficient members and confirmed as a single event in the target scene as the preset inertial reference. The value one ensures that the preset inertial reference is not less than one, thereby ensuring that the inertial impedance coefficient is within the range of greater than zero and not exceeding one.

[0061] The current time window weighted mean refers to the average of the work order embedding vectors belonging to the semantic micro-cluster in the current time window, weighted by timeliness quality. It is equal to the vector obtained by multiplying each work order embedding vector belonging to the semantic micro-cluster in the current time window by its respective timeliness quality, summing the results, and then dividing by the sum of these timeliness quality values. The work order embedding vectors belonging to the semantic micro-cluster in the current time window are determined by the access time of their corresponding work order text falling within the current time window. When no work order embedding vectors belong to the semantic micro-cluster in the current time window, the current time window weighted mean is taken as the current centroid of the semantic micro-cluster. The allowable step size refers to the displacement vector that the current centroid moves towards the current time window weighted mean. The allowable step size is equal to the product of the difference vector between the current time window weighted mean and the current centroid and the inertial impedance coefficient. The moving centroid is equal to the sum of the current centroid and the allowable step size. Since the inertial impedance coefficient ranges from greater than zero to no more than one, the moving centroid falls on the line segment between the current centroid and the current time window weighted mean without crossing the current time window weighted mean. For a semantic micro-cluster newly created in the current time window, its current centroid is equal to the average of all current members weighted by timeliness and quality, which is equal to the weighted average of the current time window. At this time, the stride is allowed to be a zero vector, and the moving centroid is equal to the current centroid.

[0062] The convex combination coefficient is calculated as follows: the sum of shape anisotropy and drift inertia increment is taken as the conformational drift sum. The convex combination coefficient is equal to the ratio of the conformational drift sum to the result obtained by adding a preset regression constant to the conformational drift sum. The value of the convex combination coefficient is not less than zero and less than one. The preset regression constant is a constant that makes the convex combination coefficient of a typical semantic micro-cluster fall near half. The preset regression constant is obtained by taking the median of the conformational drift sums calculated for multiple semantic micro-clusters that have been confirmed as single events in the target scene as the preset regression constant. The calculation branch of the stabilization centroid is as follows: when the conformational identifier is the first value, i.e., a stable condensed conformation, the moving centroid is taken as the stabilization centroid; when the conformational identifier is the second value, i.e., a small sample degenerate conformation, the stabilization centroid is equal to the product of the moving centroid and the result obtained by subtracting the convex combination coefficient from one, plus the product of the anchor centroid and the convex combination coefficient, i.e., a convex combination is performed between the moving centroid and the anchor centroid according to the convex combination coefficient. When the trace of the rotation tensor is zero, the inertial impedance coefficient is taken as the reciprocal of the preset inertia reference, and the moving centroid is calculated in the above manner, without affecting subsequent branches. In step S223, the stable centroid is pulled towards the anchor centroid under the degenerate conformation. This is because the degenerate conformation indicates that the member point cloud consists of only a small number of requests or unidirectional expansion, and the current centroid position estimate is unreliable. Pulling back towards the anchor centroid can suppress position jumps caused by unreliable estimates. When the drift inertia increment is large, the convex combination coefficient increases accordingly, and the stable centroid is more biased towards the anchor centroid. If the offset of the current centroid relative to the anchor centroid does indeed originate from the real evolution of the event semantics, then in the subsequent step S23, when the deviation angle exceeds the preset splitting threshold, the semantic microcluster is split and the current stable centroid is used as the anchor centroid of the new semantic microcluster. The real evolution will not be suppressed for a long time due to continuous pullback.

[0063] The trace, shape anisotropy, and drift inertia increment of the rotation tensor correspond to the rotational inertia per unit time mass of the member point cloud, the deviation of the spread shape, and the inertia offset of the current centroid relative to the semantic starting point, respectively. The centroid movement amplitude is constrained by the reciprocal of the sum of the trace of the rotation tensor and the preset inertia benchmark. This is based on the principle that in classical mechanics, the larger the rotational inertia of a system of particles, the smaller the change in its motion state under the same action. Therefore, semantic microclusters with larger traces of the rotation tensor (i.e., larger rotational inertia per unit time mass) have smaller centroid movement amplitudes within the window. Convex combinations pull the centroid of the degenerate conformation towards the anchor centroid, based on the principle that the variance of the current centroid estimate of the degenerate point cloud is large. Convex combinations towards a fixed anchor centroid are equivalent to shrinking an unreliable current estimate with a reliable semantic starting point. Both the inertial impedance coefficient and the convex combination coefficient are dimensionless quantities. The inertial impedance coefficient ranges from greater than zero to no more than one, while the convex combination coefficient ranges from not less than zero to less than one; their magnitudes are comparable. If the same fixed movement amplitude is applied to all semantic micro-clusters without distinguishing their conformations, large point clouds will drift and overshoot between windows due to excessive movement amplitude, and degenerate point clouds will randomly jump due to lack of pullback. The dual instability cannot be suppressed at the same time.

[0064] Compared to the current centroid, which is simply the weighted average of the current time window, the stabilized centroid, in semantic microclusters with a large member point cloud distribution, is constrained in terms of inter-window displacement due to the larger trace of the rotation tensor and the smaller inertial impedance coefficient. The stable centroid position changes smoothly between adjacent time windows, and the temporal association established based on positional proximity is no longer interrupted by large-scale centroid jumps. In semantic microclusters with small sample degradation, the stabilized centroid is constrained to the vicinity of the semantic starting point due to its convex combination towards the anchor centroid, and minor changes in individual members no longer cause large-scale centroid displacement. Because the stabilized centroid is constrained by both the current semantics and the anchor centroid, the migration of event semantics during long-term evolution is reflected in the stabilized centroid with a controlled movement amplitude, preventing the same event from being classified as a new cluster or included in adjacent events due to continuous centroid drift. The stabilized centroid, as a node value, is received by subsequent steps to update the merging relationship and construct the interval tree, ensuring that the same event remains a single object with continuous position between windows, and maintaining the continuous evolution trajectory of the same event on the monitoring screen without flickering or breaks.

[0065] Step S23: Calculate the deviation angle between the stabilized centroid and the anchor centroid. When the deviation angle exceeds a preset splitting threshold, split the semantic micro-cluster into a new set in the disjoint set forest and register the parent set and child set affiliation. When the deviation angle between the stabilized centroids of two coexisting semantic micro-clusters is lower than a preset merging threshold, merge the sets corresponding to the two semantic micro-clusters in the disjoint set forest.

[0066] Specifically, the deviation angle is calculated as follows: Take the inner product of the stabilized centroid and the anchor centroid, divide this inner product by the product of the Euclidean norm of the stabilized centroid and the Euclidean norm of the anchor centroid, and obtain the cosine of the deviation angle. The deviation angle is equal to the inverse cosine of the cosine of the deviation angle, and the value of the deviation angle ranges from zero to pi. The stabilized centroid originates from step S223, and the anchor centroid originates from step S21.

[0067] The split determination method is as follows: when the deviation angle between the stable centroid and the anchor centroid is greater than the preset split threshold, it is determined that the current semantics of the semantic micro-cluster has deviated from the semantic starting point to the point of constituting a new event. A split is then performed in the disjoint set forest. A subset is formed by the work order embedding vectors that have been assigned to the semantic micro-cluster in the current time window. This subset is treated as a new semantic micro-cluster, and its anchor centroid is the current stable centroid. The original semantic micro-cluster serves as the parent set, retaining its remaining members and the original anchor centroid. The split execution method is as follows: each node assigned to the semantic micro-cluster in the current time window is extracted from the original root tree. Any extracted node is taken as the root node of the subset, with its parent pointer pointing to itself. The remaining nodes are then... The parent pointers of the extracted nodes all point to the root node of the tree, forming the rooted tree corresponding to the subset; the parent set retains the remaining nodes, and the parent pointers of each retained node are all changed to point to the root node of the parent set, and the number of nodes saved in the two rooted trees are updated according to the number of extracted and retained nodes respectively; the representative element of the parent set at the split time is recorded in the derivation source field of the root node of the subset tree, thereby registering the split derivation relationship between the representative element of the parent set and the representative element of the subset; no split is performed when the deviation angle is less than or equal to the preset split threshold, and no split is performed when there is no work order embedding vector in the current time window that is assigned to the semantic micro-cluster, or when the work order embedding vector assigned to the current time window is all the members of the semantic micro-cluster. After the split is executed, the subset becomes a newly added semantic micro-cluster. Its active time interval starts and ends with the current time window number. The stable centroid of the subset in the current time window is taken as its anchor centroid, which is the stable centroid at the split time. The current centroid, rotation tensor, and other parameters of the subset are recalculated according to the subset's own members from the next time window. The parent set retains the remaining members after the split. The current centroid of the parent set in the next time window is taken as the current centroid before the split, not the stable centroid that triggered the split. The active time interval of the parent set retains the values ​​before the split. The other parameters of the parent set are recalculated according to the retained members from the next time window. If the number of members of the parent set still does not reach the preset number after the split, the anchor centroid of the parent set continues to be updated by rolling the arithmetic mean of the work order embedding vectors of all its current members until the number of members reaches the preset number and then remains unchanged. The preset splitting threshold refers to the deviation angle boundary for determining whether a semantic micro-cluster has evolved into a new event. The preset splitting threshold is obtained by collecting multiple sets of stable centroid and anchor centroid sample pairs belonging to the same event evolution process and multiple sets of stable centroid and anchor centroid sample pairs belonging to different events. For each sample pair, the deviation angle is calculated, and the upper bound of the deviation angle distribution of the same event evolution sample pairs and the lower bound of the deviation angle distribution of the different event sample pairs are statistically analyzed. The median between the upper bound and the lower bound of the distribution is taken as the preset splitting threshold. For example, the preset splitting threshold is taken as one-quarter of pi.

[0068] The merging determination method is as follows: For every two semantic microclusters currently coexisting in the disjoint set forest, calculate the deviation angle between the stable centroids of the two semantic microclusters. The calculation method for the deviation angle is the same as described above. When the deviation angle between the stable centroids of the two semantic microclusters is less than a preset merging threshold, it is determined that the two semantic microclusters correspond to the same event. In the disjoint set forest, the parent pointer of the root node of the rooted tree corresponding to the semantic microcluster with fewer members is pointed to the root node of the rooted tree corresponding to the semantic microcluster with more members, and they are merged into one set. The number of nodes stored in the rooted tree after merging is updated to the sum of the number of nodes in both sets. The anchor centroid of the merged semantic microcluster is taken from the set with more members. The anchor centroid of the micro-cluster; the stable centroid of the merged semantic micro-cluster in the current time window is taken from the stable centroid of the semantic micro-cluster with more members. The starting endpoint of the active time interval of the merged semantic micro-cluster is taken from the smaller of the starting endpoints of the active time intervals of the two semantic micro-clusters, and the ending endpoint is taken from the larger of the ending endpoints of the active time intervals of the two semantic micro-clusters. The current centroid and other parameters of the merged semantic micro-cluster are recalculated from the next time window according to all members after the merger. The semantic micro-cluster with fewer members no longer exists as an independent semantic micro-cluster after the merger, and its corresponding node in the interval tree is deleted by step S24. Merging is not performed when the deviation angle is greater than or equal to the preset merging threshold. The preset merge threshold refers to the deviation angle boundary for determining whether two semantic micro-clusters correspond to the same event. The preset merge threshold is obtained by collecting multiple sets of stable centroid sample pairs of two semantic micro-clusters belonging to the same event fragment and multiple sets of stable centroid sample pairs of two semantic micro-clusters belonging to different events. For each sample pair, the deviation angle is calculated, and the upper bound of the deviation angle distribution for sample pairs of the same event fragment and the lower bound of the deviation angle distribution for sample pairs of different events are statistically analyzed. The median between these two distributions is taken as the preset merge threshold. For example, the preset merge threshold is one-sixth of pi. After splitting and merging, the merge relationship of the disjoint set forest is updated based on the stable centroids.

[0069] The deviation angle physically corresponds to the angle between the stabilized centroid vector and the anchor centroid vector in the embedding space. As the same event evolves over time, its current semantic direction deviates less from its semantic starting point, while its direction deviates more when it has evolved into another event. Two fragments belonging to the same event have stable centroid directions that are close. The semantic deviation is measured by the deviation angle rather than Euclidean distance because the difference in the request semantic content after the work order embedding vector is normalized by the L2 norm is reflected in the directional difference, and the angle can eliminate the difference in vector magnitude and only reflect the directional difference. If the split is not explicitly determined and the shifted stabilized centroid continues to absorb work orders, the similarity between the evolved request and the original semantic starting point of the same event decreases, and it will be implicitly judged as a new cluster, causing event fragmentation and loss of derivation relationship.

[0070] The updated disjoint set forest implements the splitting and merging of semantic microclusters into a data structure using set operations. When the same event splits due to semantic evolution, the splitting result is recorded as a parent-child event rather than a new, unrelated event, maintaining the traceability of the event evolution process in the data structure. Two semantic microclusters with similar stable centroid directions are merged, and fragments of the same event that were originally divided into two semantic microclusters due to short-term fluctuations are re-merged into a single event. The subsets obtained from the splitting use the current stable centroid as the anchor centroid, and the new event has an independent and frozen semantic starting point from the moment of splitting, so subsequent tracking is no longer affected by the semantic starting point of the original event. Thus, step S23 transforms the evolution of events and fragment merging from implicit similarity fluctuations into explicit structural operations, eliminating the artificially high number of hot events caused by frequent splitting and incorrect merging of the same event.

[0071] Step S24: For each semantic microcluster, use the active time interval of the semantic microcluster in the time window sequence as the key and the stable centroid as the value, and insert it into the interval tree.

[0072] Specifically, a time window sequence refers to a sequence of time windows numbered chronologically, obtained by dividing continuous time into fixed-length segments. Each time window corresponds to an integer sequence number. An active time interval refers to the range of time window sequences in which the semantic micro-cluster is active. The active time interval is equal to a closed interval formed by the time window sequence number where the first member of the semantic micro-cluster is assigned as the starting endpoint and the time window sequence number where the most recent work order embedding vector is assigned as the ending endpoint. Both endpoints of the active time interval are integer sequences. When a work order embedding vector is assigned to the semantic micro-cluster in the current time window, the ending endpoint of the active time interval of the semantic micro-cluster is updated to the current time window sequence number. When no work order embedding vector is assigned to the semantic micro-cluster in the current time window, the ending endpoint retains its original value.

[0073] The interval tree is a balanced binary search tree data structure using intervals as keys. Each node stores the active time interval of a semantic micro-cluster as the key and the stable centroid of that semantic micro-cluster as the value. Nodes are organized into a balanced binary search tree by sorting the start endpoints of the active time intervals. Each node also stores the maximum value of the end endpoints of all active time intervals in its subtree rooted at that node, used to support interval overlap queries. The number of nodes in the interval tree equals the number of existing semantic micro-clusters. The endpoint values ​​of the active time intervals range from the system's start time window number to the current time window number. By using active time intervals as keys and storing the maximum values ​​of subtree endpoints, the interval tree ensures that the cost of querying all semantic micro-clusters overlapping with a certain time window interval increases logarithmically with the number of semantic micro-clusters, enabling rapid location of concurrent relationships for a large number of semantic micro-clusters within a minute-level time window.

[0074] The insertion operation is as follows: For each semantic micro-cluster, identify the semantic micro-cluster using the representative element of the rooted tree corresponding to the semantic micro-cluster in the disjoint set forest. If the semantic micro-cluster does not yet have a corresponding node in the interval tree, insert the node into the interval tree using the active time interval of the semantic micro-cluster as the key and the stable centroid of the semantic micro-cluster as the value, according to the insertion rules of the balanced binary search tree, and update the maximum value of the subtree endpoints stored in each node along the insertion path. If the semantic micro-cluster already has a corresponding node in the interval tree and the starting endpoint of its active time interval has not changed, update the key of the node to the current active time interval and update the value of the node to the current active time interval. The stable centroid of the current node is determined, and the maximum value of the subtree endpoints of each node is updated along the path from the node to the root node. If the starting endpoint of the active time interval of the semantic microcluster changes, the node is first deleted according to the deletion rules of the balanced binary search tree, and then re-inserted with the current active time interval as the key and the current stable centroid as the value. The maximum value of the subtree endpoints of each node is updated along the deletion path and the insertion path. For semantic microclusters that are merged into other semantic microclusters and no longer exist independently in the merging process in step S23, the corresponding nodes in the interval tree are deleted, and the maximum value of the subtree endpoints of each node is updated along the deletion path.

[0075] The active time interval physically corresponds to the time window span of an event on the timeline from its occurrence to the most recent received request. A continuous event may span multiple consecutive time windows, and multiple events may coexist within the same time window. An interval tree is constructed using the active time interval as the key and the stable centroid as the value. This is based on the principle that when a balanced binary search tree is ordered by its starting endpoint and each node maintains the maximum value of its subtree endpoints, non-overlapping subtrees can be pruned within logarithmic cost. The time span of an event is modeled as a closed interval, based on the fact that the active periods of the same event are continuous on the timeline, and can be completely defined by the start and end time window indices. If all semantic microclusters are compared one by one, the number of comparisons for locating concurrent events within a single window increases linearly with the number of semantic microclusters, making it difficult to complete within a minute-level window.

[0076] The interval tree organizes all semantic micro-clusters into a fast-searchable temporal index based on their active time intervals, eliminating the need for sequential comparisons of time periods when locating concurrent semantic micro-clusters within the same time period. Each node in the interval tree also stores a stable centroid, allowing the stable centroid to be retrieved simultaneously with the location of concurrent semantic micro-clusters. Continuous events spanning multiple windows are maintained as a single node in the interval tree, rather than being repeatedly registered per window. Therefore, step S24 provides a temporal retrieval foundation for subsequently calculating temporal changes and assessing the situation among concurrent events, supporting minute-level concurrent situation identification.

[0077] Step S30: Receive the interval tree, perform interval overlap query on the interval tree to obtain semantic micro-clusters with overlapping active time intervals, calculate the situation level based on the stable centroids and temporal changes in the number of members of the overlapping semantic micro-clusters, and generate and output situation warning data when the situation level meets the warning conditions; Step S30 includes steps S31, S32 and S33.

[0078] like Figure 6 The diagram shown illustrates an interval overlap query.

[0079] Reference Figure 6 The diagram illustrates the physical process of executing overlapping queries between active time intervals and query intervals. The diagram shows the range of time window indices (i.e., the horizontal active time intervals) on the timeline for each semantic micro-cluster to be active. The degradation point interval formed by the current time window indices serves as the vertical query interval. The horizontal line segments traversed by the query intervals represent semantic micro-clusters where the active time intervals and query intervals overlap; these are collected as a set of concurrent semantic micro-clusters.

[0080] Step S31: Perform interval overlap query on the interval tree to obtain a set of concurrent semantic micro-clusters composed of semantic micro-clusters whose active time intervals overlap.

[0081] Specifically, the interval overlap query method is as follows: The query interval is defined by the degenerate point interval formed by using the current time window number as both the start and end endpoints. Starting from the root node of the interval tree, the tree is pruned downwards according to the maximum value of the subtree endpoints stored in the node. For each visited node, it is determined whether the node's active time interval overlaps with the query interval. The condition for overlap between two active time intervals is that the start endpoint of the first active time interval is not greater than the end endpoint of the second active time interval, and the start endpoint of the second active time interval is not greater than the end endpoint of the first active time interval. When the maximum value of the subtree endpoints of a node is less than the start endpoint of the query interval, the subtree of that node is pruned, because the end endpoints of all active time intervals within the pruned subtree are less than the start endpoint of the query interval and therefore cannot overlap. All semantic micro-clusters whose active time intervals overlap with the query interval are collected into a concurrent semantic micro-cluster set, which is the set of semantic micro-clusters that are active in the current time window.

[0082] The overlapping of active time intervals physically corresponds to the reality of multiple events coexisting within the same time period. Operations personnel focus on events that are currently unfolding; events whose active time intervals no longer cover the current time window are not included in the current situation assessment. The current active semantic micro-clusters are located using interval overlap queries, based on the principle that interval trees are pruned using the maximum value of their subtree endpoints, ensuring that subtrees that cannot overlap are pruned and not traversed. Two active time intervals overlap if and only if the starting endpoint of one does not exceed the ending endpoint of the other and they are simultaneously inverses, based on the necessary and sufficient condition that closed intervals have a common point on the number axis. If the situation is assessed for all historical semantic micro-clusters, it increases invalid computation and introduces events irrelevant to the current situation assessment.

[0083] The concurrent semantic micro-cluster set collects the semantic micro-clusters that are active in the current time window as the evaluation objects, so that subsequent situation assessments only apply to events that are currently unfolding rather than all historical events, and this screening is completed within a minute-level window with logarithmic cost; each semantic micro-cluster in the set coexists in the same time period, and its stable centroid is obtained along with the set, so the temporal change can be calculated without subsequent back lookups. Thus, step S31 narrows the scope of the situation assessment from all historical events to the current concurrent events.

[0084] Step S32: For each semantic microcluster in the concurrent semantic microcluster set, calculate the centroid drift rate based on the displacement of the stable centroid of the semantic microcluster between adjacent time windows, calculate the size growth rate based on the increment of the number of members of the semantic microcluster between adjacent time windows, and use the centroid drift rate and the size growth rate as the temporal change of the semantic microcluster.

[0085] Specifically, the centroid drift rate is equal to the Euclidean norm of the difference vector between the stable centroid of the current time window and the stable centroid of the previous time window of the semantic microcluster. The centroid drift rate reflects the migration amplitude of the event semantics between adjacent time windows. The stable centroid of the current time window and the stable centroid of the previous time window are both derived from the output of step S223 in the corresponding time window. When the semantic microcluster appears for the first time in the current time window and there is no stable centroid of the previous time window, the centroid drift rate is zero.

[0086] The growth rate is equal to the difference between the number of members in the current time window and the number of members in the previous time window for the semantic microcluster. The number of members is equal to the number of nodes in the rooted tree corresponding to the semantic microcluster in the disjoint forest. The growth rate reflects the increase or decrease in event requests between adjacent time windows. When the semantic microcluster appears for the first time in the current time window and does not have the number of members from the previous time window, the number of members in the previous time window is considered zero, and the growth rate is taken as the number of members in the current time window. The temporal change is a tuple consisting of the centroid drift rate and the growth rate. The centroid drift rate reflects semantic migration, and the growth rate reflects the increase or decrease in requests. Together, they characterize the temporal change of the semantic microcluster between adjacent time windows.

[0087] The centroid drift rate and scale growth rate physically correspond to the speed of semantic migration and the growth rate of request volume on the time axis, respectively. The intensity of the situation is reflected in both rapid semantic migration and rapid growth in request volume. Semantic migration is measured by the Euclidean norm of the displacement of the stable centroid in adjacent time windows, based on the fact that the stable centroid represents the current semantic position of the event, and the difference in position between adjacent windows is the amount of semantic migration within a unit window. Request growth is measured by the difference in the number of members in adjacent time windows, based on the fact that the number of members represents the cumulative number of work orders for the event, and the difference between adjacent windows is the amount of new requests within a unit window. If the situation is evaluated only based on the single dimension of the increase in the number of members, an evolutionary event with stable request volume but rapid semantic shift may be missed.

[0088] The temporal variation measures define the semantic migration and request increases / decreases of each concurrent semantic micro-cluster between adjacent time windows using two rate measures, making the severity of the event situation a calculable quantity. Since the centroid drift rate is based on a stable centroid suppressed by dual instability, it reflects the actual migration of event semantics rather than spurious displacements caused by centroid jitter; since the size growth rate is taken from the difference in the number of members after splitting and merging corrections, it reflects the actual increase / decrease in event requests rather than spurious fluctuations caused by erroneous merging or fragmentation. Therefore, step S32 provides an accurate measure of change for subsequently mapping the temporal variation measures to a situation level and generating early warnings.

[0089] Step S33: Calculate the situation level based on the temporal change of each semantic microcluster in the concurrent semantic microcluster set. When the situation level exceeds the level threshold corresponding to the warning condition, generate and output situation warning data containing the stable centroid, active time interval and situation level of the corresponding semantic microcluster.

[0090] Specifically, since the centroid drift rate and the scale growth rate have different physical dimensions and numerical magnitudes—the centroid drift rate being the dimensionless displacement in the embedded space and the scale growth rate being the number of requests per unit window—both are normalized before participating in the same situation level calculation. The normalized centroid drift rate is equal to the ratio of the centroid drift rate to the centroid drift rate normalization benchmark. The centroid drift rate is a Euclidean norm and is never less than zero. When this ratio is greater than one, the normalized centroid drift rate is set to one. The centroid drift rate normalization benchmark refers to the upper bound of the empirical value domain of the centroid drift rate, which is equal to the maximum value obtained after calculating the centroid drift rate for all processed semantic microclusters in the target scenario up to the end of the previous time window. When the current time window is the first time window and there are no processed semantic microclusters yet, the centroid drift rate normalization benchmark is set to zero. When the centroid drift rate normalization benchmark is zero, the normalized centroid drift rate is set to zero, and no warning is triggered in the current time window. The normalized scale growth rate is equal to the ratio of the scale growth rate to the normalized scale growth rate benchmark. When the scale growth rate is less than zero, the normalized scale growth rate is zero; when the ratio is greater than one, the normalized scale growth rate is one. The normalized scale growth rate benchmark refers to the upper bound of the empirical value domain of the scale growth rate, equal to the maximum value obtained after calculating the scale growth rate for all processed semantic microclusters in the target scenario up to the end of the previous time window. When the current time window is the first time window and there are no processed semantic microclusters yet, the normalized scale growth rate benchmark is zero; when the normalized scale growth rate benchmark is zero, the normalized scale growth rate is zero. Both the normalized centroid drift rate and the normalized scale growth rate range from zero to one.

[0091] The situation level is equal to the product of the normalized centroid drift rate and the centroid drift rate weight, plus the product of the normalized scale growth rate and the scale growth rate weight. The situation level ranges from zero to one. The centroid drift rate weight and scale growth rate weight are weight coefficients that characterize the proportion of the contribution of the two normalized rates to the situation. The sum of the two is equal to one. The method of obtaining them is as follows: collect multiple sets of situation samples that have been identified as requiring early warning and multiple sets of situation samples that have been identified as normal. Calculate the normalized centroid drift rate and normalized scale growth rate for each set of samples. Iterate through the centroid drift rate weight between zero and one with a step size of 0.1. Use the difference between the value of one and the centroid drift rate weight as the scale growth rate weight. Calculate the situation level of the two types of samples respectively. Take the set of values ​​that maximizes the ratio of the absolute value of the difference between the situation level mean of the samples requiring early warning and the normal samples to the sum of the standard deviations of the situation levels of the two types of samples as the centroid drift rate weight and scale growth rate weight. For example, the centroid drift rate weight is taken as half and the scale growth rate weight is taken as half.

[0092] The situation level is a weighted sum of the normalized centroid drift rate and the normalized scale growth rate, measuring the severity of the event situation. The level threshold is the boundary between situations requiring warning and routine situations. Both measures the same dimension and have values ​​ranging from zero to one, allowing for direct comparison: when both rates are zero, the situation level is zero, below the level threshold, and does not trigger a warning; when both rates reach the upper bound of their respective empirical value domains, the situation level is one, above the level threshold, and triggers a warning. The level threshold is obtained by collecting multiple sets of situation samples confirmed as requiring warning and multiple sets of situation samples confirmed as routine. The situation level is calculated for each set of samples, and the mean and standard deviation of the situation levels for the warning-requiring and routine samples are statistically analyzed. The level threshold is determined with the goal of equalizing the misjudgment rates for the warning-requiring and routine categories. The method for determining whether the situation level distributions of two types of samples approximately follow a normal distribution is as follows: Perform normality tests on the situation levels of the samples requiring early warning and the regular samples, respectively. At a given significance level, if both tests do not reject the normality hypothesis, the samples are considered to approximately follow a normal distribution; otherwise, they are considered not to approximately follow a normal distribution. When the situation level distributions of both types of samples approximately follow a normal distribution, making the misclassification rates of the two types equal is equivalent to making the standardized distances from the level threshold to the mean of the samples requiring early warning equal to the standardized distances to the mean of the regular samples. Therefore, the level threshold is equal to the sum of the following two products divided by the sum of the following two standard deviations: the first product is the product of the standard deviation of the situation level of the regular samples and the mean of the situation level of the samples requiring early warning; the second product is the product of the standard deviation of the situation level of the samples requiring early warning and the mean of the situation level of the regular samples. The two standard deviations are the standard deviations of the situation level of the samples requiring early warning and the standard deviations of the situation level of the regular samples. When the situation level distributions of the two types of samples are not approximately normal... When the situation appears to follow a normal distribution, the level threshold is taken as a value that equals the misclassification rates of the two types of situation levels: those of the warning samples and the regular samples. Specifically, candidate thresholds are selected one by one from zero to one, and for each candidate threshold, the proportion of samples with situation levels lower than the candidate threshold in the warning samples and the proportion of samples with situation levels higher than the candidate threshold in the regular samples are calculated. The candidate threshold that makes these two proportions equal is taken as the level threshold. If no candidate threshold makes the two proportions exactly equal, the candidate threshold that minimizes the absolute value of the difference between the two proportions is taken as the level threshold. For example, the level threshold is 0.6. When the situation level of a semantic micro-cluster is greater than the level threshold, the situation of the event corresponding to the semantic micro-cluster is determined to have reached the warning condition. Situation warning data containing the stable centroid, active time interval, and situation level of the semantic micro-cluster is generated and output. When the situation level is less than or equal to the level threshold, no situation warning data is generated.

[0093] The situation level physically corresponds to the intensity of an event's combined semantic migration and request growth. Events with drastic situational increases require alerting to operations personnel. The situation level is obtained by normalizing and then weighting the indicators, based on the principle that normalizing and then weighting indicators with different dimensions ensures they fall within the same numerical range and are combined into a single indicator according to their contribution proportion. The alert is output by comparing the situation level with a level threshold, based on the principle that the level threshold is derived by back-calculating the two types of situation samples with the goal of equalizing the misclassification rates of both types. If normalization is not performed and the centroid drift rate and scale growth rate with different dimensions are directly summed, the larger numerical value dominates the summation result, while the smaller value contributes almost nothing to the situation level, leading to a distorted assessment.

[0094] The situational early warning data submits events that meet the warning conditions along with their stable centroids, active time intervals, and situational levels, enabling operators to obtain the semantic location, duration, and severity of events requiring current attention. Since the situational level is based on the centroid drift rate derived from the stable centroid suppressed by dual instability and the scale growth rate derived from the number of members corrected by splitting and merging, the situational assessment is no longer inflated by false displacements and scale fluctuations caused by centroid jitter, and the number of hot events is no longer artificially high. Because the situational level is synthesized by weight and compared with the level threshold derived by the constraint of equal false positive rates for the two types of situations, the warning trigger is valid for both semantic shift and request surge situations. Therefore, step S33 implements the suppression of centroid instability and the correction of member affiliation in the preceding steps into accurate situational early warnings, ensuring that the event situation presented on the monitoring screen is continuous, stable, and consistent with reality, eliminating the superficial problems described in the background technology, such as trajectory flickering and breakage, event fragmentation and erroneous association, and artificially high number of hot events.

[0095] Example 2 A temporal situational awareness system based on semantic micro-clusters includes: The semantic tagging module is used to collect work order texts within the current time window, perform multi-dimensional semantic tagging on the work order texts, obtain work order embedding vectors and timeliness quality, classify the work order embedding vectors into semantic micro-clusters according to semantic proximity, and construct a disjoint set forest. The centroid stabilization module is used to calculate the rotation tensor for each semantic micro-cluster in the disjoint set forest based on the work order embedding vector and timeliness quality of its members, obtain the stabilized centroid based on the rotation tensor, update the merging relationship of the disjoint set forest based on the stabilized centroid, and construct an interval tree with the active time interval of the semantic micro-cluster as the index and the stabilized centroid as the node value. The situation warning module is used to perform interval overlap queries on the interval tree to obtain the set of concurrent semantic micro-clusters in the active time interval. It calculates the situation level based on the stable centroid and the time-series change in the number of members of the concurrent semantic micro-clusters. When the situation level meets the warning conditions, it generates situation warning data.

[0096] The foregoing description is illustrative of the invention and should not be construed as limiting it. Although several exemplary embodiments of the invention have been described, those skilled in the art will readily understand that many modifications can be made to the exemplary embodiments without departing from the novel teachings and advantages of the invention. Therefore, all such modifications are intended to be included within the scope of the invention as defined in the claims. It should be understood that the foregoing description is illustrative of the invention and should not be construed as limiting it to the specific embodiments disclosed, and modifications to the disclosed embodiments and other embodiments are intended to be included within the scope of the appended claims. The invention is defined by the claims and their equivalents.

Claims

1. A temporal situational awareness method based on semantic micro-clusters, characterized in that, include: Collect work order texts within the current time window, perform multi-dimensional semantic tagging on the work order texts to obtain work order embedding vectors and timeliness / quality, classify the work order embedding vectors into semantic micro-clusters according to semantic proximity, and construct a disjoint set forest; For each semantic micro-cluster in the disjoint set forest, calculate the rotation tensor based on the work order embedding vector and timeliness quality of its members, obtain the stable centroid based on the rotation tensor, update the merging relation of the disjoint set forest based on the stable centroid, and construct an interval tree with the active time interval of the semantic micro-cluster as the index and the stable centroid as the node value. Perform interval overlap query on the interval tree to obtain the set of concurrent semantic micro-clusters in the active time interval. Calculate the situation level based on the stable centroid and the time-series change in the number of members of the concurrent semantic micro-cluster set. When the situation level meets the warning conditions, generate situation warning data.

2. The temporal situational awareness method based on semantic micro-clusters according to claim 1, characterized in that, The calculation steps for timeliness quality are as follows: Collect the work order text within the current time window, perform word segmentation and noise removal on the work order text to obtain a regularized work order text; Multidimensional semantic tagging is performed on the regularized work order text to extract multidimensional semantic tags containing subject, object, time, location, request and emotion. The multidimensional semantic tags are encoded into work order embedding vectors, and the timeliness quality is calculated based on the access time of the work order text within the current time window.

3. The temporal situational awareness method based on semantic micro-clusters according to claim 2, characterized in that, The steps to construct a disjoint set forest are as follows: The work order embedding vectors are assigned to the corresponding semantic micro-clusters according to their semantic proximity to the current centroid of the existing semantic micro-clusters. For work order embedding vectors whose semantic proximity is lower than the preset assignment threshold, new semantic micro-clusters are created. The work order embedding vectors assigned to the same semantic micro-cluster are merged into the same set of the disjoint set forest to obtain the disjoint set forest.

4. The temporal situational awareness method based on semantic micro-clusters according to claim 1, characterized in that, The analytical steps for stabilizing the centroid are as follows: For each semantic micro-cluster in the disjoint set forest, collect the work order embedding vectors and timeliness quality of the members of the semantic micro-cluster, obtain the current centroid of the semantic micro-cluster, and the anchor centroid determined and fixed by the mean of the first preset number of work order embedding vectors to which the semantic micro-cluster is assigned. The stable centroid of the semantic microcluster is calculated based on the current centroid, anchor centroid, work order embedding vector of the member, and timeliness quality.

5. The temporal situational awareness method based on semantic micro-clusters according to claim 4, characterized in that, The steps for constructing an interval tree are as follows: Calculate the deviation angle between the stabilized centroid and the anchor centroid. When the deviation angle exceeds the preset splitting threshold, split the semantic micro-cluster into a new set in the disjoint set forest and register the belonging relationship between the parent set and the child set. When the deviation angle between the stabilized centroids of two coexisting semantic micro-clusters is lower than the preset merging threshold, merge the sets corresponding to the two semantic micro-clusters in the disjoint set forest. For each semantic microcluster, the active time interval of the semantic microcluster in the time window sequence is used as the key and the stable centroid is used as the value, and then inserted into the interval tree.

6. The temporal situational awareness method based on semantic micro-clusters according to claim 4, characterized in that, The steps for calculating the stable centroid of semantic microclusters are as follows: The work order embedding vectors of the members of the semantic micro-cluster are used as particles with corresponding timeliness quality. The gyration tensor of the particles is calculated with the current centroid of the semantic micro-cluster as the reference point. The gyration tensor is decomposed into features to obtain the principal gyration radius square sequence.

7. The temporal situational awareness method based on semantic micro-clusters according to claim 6, characterized in that, The steps for calculating the stable centroid of semantic microclusters also include: The normalized variance of the principal gyration radius square sequence is used as the shape anisotropy degree; according to the parallel axis shift theorem, the product of the sum of time-dependent mass and the square of the Euclidean distance from the current centroid to the anchor centroid is used as the drift inertia increment; a first interval and a second interval are set, and the boundary between the first interval and the second interval is determined by the situation level output in the previous time window. When the shape anisotropy degree is in the first interval, the conformation identifier is set to the first value, and when the shape anisotropy degree is in the second interval, the conformation identifier is set to the second value.

8. The temporal situational awareness method based on semantic micro-clusters according to claim 7, characterized in that, The steps for calculating the stable centroid of semantic microclusters also include: The reciprocal of the sum of the trace of the rotation tensor and the preset inertia reference is used as the inertial impedance coefficient. Calculate the mean of the work order embedding vectors belonging to the semantic micro-cluster in the current time window, weighted by timeliness and quality. Move the current centroid along the direction from the current centroid to the mean, allowing a step size to obtain the moved centroid. The allowable step size is the product of the difference between the mean and the current centroid and the inertial impedance coefficient. When the conformation identifier is the first value, the moving centroid is used as the stable centroid. When the conformation identifier is the second value, the moving centroid and the anchor centroid are convexly combined according to the convex combination coefficient to obtain the stable centroid. The convex combination coefficient is the ratio of the sum of the shape anisotropy degree and the drift inertia increment to the sum plus the preset regression constant.

9. The temporal situational awareness method based on semantic micro-clusters according to claim 1, characterized in that, The steps for calculating the situation level are as follows: Perform an interval overlap query on the interval tree to obtain a set of concurrent semantic micro-clusters in the active time interval; For each semantic microcluster in the concurrent semantic microcluster set, the centroid drift rate is calculated based on the displacement of the stable centroid of the semantic microcluster between adjacent time windows, and the size growth rate is calculated based on the increment of the number of members of the semantic microcluster between adjacent time windows. The centroid drift rate and the size growth rate are used as the temporal change of the semantic microcluster. The situation level is calculated based on the temporal changes of each semantic micro-cluster in the concurrent semantic micro-cluster set.

10. The temporal situational awareness method based on semantic micro-clusters according to claim 9, characterized in that, The steps for outputting situational early warning data are as follows: When the situation level exceeds the level threshold corresponding to the warning conditions, situation warning data is generated, which includes the stable centroid, active time interval and situation level of the corresponding semantic micro-clusters.

11. A temporal situational awareness system based on semantic micro-clusters, characterized in that, The method for performing the temporal situational awareness method based on semantic micro-clusters as described in any one of claims 1-10 includes: The semantic tagging module is used to collect work order texts within the current time window, perform multi-dimensional semantic tagging on the work order texts, obtain work order embedding vectors and timeliness quality, classify the work order embedding vectors into semantic micro-clusters according to semantic proximity, and construct a disjoint set forest. The centroid stabilization module is used to calculate the rotation tensor for each semantic micro-cluster in the disjoint set forest based on the work order embedding vector and timeliness quality of its members, obtain the stabilized centroid based on the rotation tensor, update the merging relationship of the disjoint set forest based on the stabilized centroid, and construct an interval tree with the active time interval of the semantic micro-cluster as the index and the stabilized centroid as the node value. The situation warning module is used to perform interval overlap queries on the interval tree to obtain the set of concurrent semantic micro-clusters in the active time interval. It calculates the situation level based on the stable centroid and the time-series change in the number of members of the concurrent semantic micro-clusters. When the situation level meets the warning conditions, it generates situation warning data.

Images