Traffic load monitoring method, system and medium for advertisement information service

Abstract

The application provides a kind of advertisement information service-oriented traffic load monitoring method, system and medium, by obtaining the service node load state record sequence corresponding to the time stamp of advertisement request stream data sequence in the preset monitoring period;Advertisement request stream data sequence is decoupled and rhythm extraction is carried out to request source, and the request impact conduction time sequence description is generated by cascading propagation in combination with advertisement position request propagation topological structure, and the time sequence distribution feature representation is extracted, while the load change feature representation is extracted to load state record sequence;Call traffic load collaborative analysis network to carry out collaborative interaction mapping to two kinds of feature representations, generate associated response characteristic representation;Associated response characteristic representation is carried out load bearing situation deduction, and load bearing capacity boundary description is obtained;According to the boundary description, generate traffic scheduling guide instruction and push to advertisement request distribution component to trigger request shunt adjustment.The application can improve the accuracy of load bearing capacity boundary determination.

Description

Technical Field

[0001] This invention relates to the field of data processing technology, and in particular to a method, system and medium for monitoring traffic load for advertising information services. Background Technology

[0002] As the scale of internet advertising services continues to expand, the uneven temporal distribution of advertising request traffic and the dynamic fluctuations in service node load are becoming increasingly significant. Traffic load monitoring technology refers to the technical means of continuously observing the arrival patterns of advertising requests and the resource occupancy status of service nodes to support scheduling decisions. Currently, the common approach is to collect request frequency statistics and resource utilization records separately, and determine whether the load has exceeded the limit based on static thresholds and trigger traffic splitting operations. However, the instantaneous surge in request traffic requires a response delay before it can be reflected in the load status records, and the load recovery process cannot be fed back to the distribution side in a timely manner. This separation of request-side and load-side monitoring means that scheduling operations often cannot be initiated until the load has significantly deteriorated, making it difficult to make adaptive interventions in advance during the impact accumulation phase. Summary of the Invention

[0003] In view of this, the present invention provides a method, system, and medium for monitoring traffic load in advertising information services. The technical solution of the embodiments of the present invention is implemented as follows: On one hand, the present invention provides a traffic load monitoring method for advertising information services, the method comprising: Acquire the sequence of advertising request stream data generated within a preset monitoring period and the sequence of service node load status records that correspond to the timestamps of the advertising request stream data sequence. The request traffic situation is extracted from the advertising request stream data sequence to obtain the temporal distribution feature representation of the advertising request stream data sequence. The load change situation is extracted from the service node load status record sequence to obtain the load change feature representation of the service node load status record sequence. The temporal distribution feature representation and the load change feature representation are input into the traffic load joint situation analysis stage. The pre-configured traffic load collaborative analysis network is invoked to perform collaborative interactive mapping between the temporal distribution feature representation and the load change feature representation, generating a correlation response characteristic representation that characterizes the dynamic relationship between advertising request traffic and load status. By performing load-bearing situation simulation on the associated response characteristics, the load-bearing capacity boundary description of the service node under the continuous impact of the advertising request stream data sequence is obtained; Based on the load capacity boundary description, a traffic scheduling guidance instruction for the advertising information service is generated, and the traffic scheduling guidance instruction is pushed to the advertising request distribution component to trigger request diversion adjustment operation.

[0004] On the other hand, the present invention provides a computer system including a memory and a processor, wherein the memory stores a computer program that can run on the processor, and the processor executes the program to implement the steps in the above method.

[0005] Thirdly, the present invention provides a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the steps in the above method.

[0006] This invention acquires advertising request stream data sequences and service node load status record sequences with timestamp correspondence within a preset monitoring period, enabling simultaneous consideration of request-side time distribution and load-side state changes based on dual information. Temporal distribution feature representations are extracted from the advertising request stream data sequences, and load change feature representations are extracted from the load status record sequences. Both are input into the traffic-load joint situational analysis stage, effectively avoiding response fragmentation and situational lag caused by isolated analysis of requests and loads. A pre-configured traffic-load collaborative analysis network is invoked to perform collaborative interactive mapping of the two types of feature representations, generating a correlation response characteristic representation that characterizes the dynamic relationship between request traffic and load status. This allows for bidirectional interactive expression of the unidirectional impact of traffic on load and the reverse constraint of load on traffic within the same representation space, improving the accuracy of the correlation characterization. Load bearing situation extrapolation processing is performed on this correlation response characteristic representation to obtain a boundary description of the load bearing capacity of service nodes under continuous traffic impact, providing a forward-looking boundary constraint basis for scheduling decisions. Based on this boundary description, a traffic scheduling guidance instruction is generated and pushed to the advertising request distribution component to trigger a traffic diversion adjustment operation. This allows the scheduling decision to be directly linked to the load bearing situation prediction results, ensuring that the service node load remains in a stable operating range while guaranteeing the timeliness of request response. Attached Figure Description

[0007] The accompanying drawings, which are incorporated in and constitute a part of this specification, illustrate embodiments consistent with the present invention and, together with the specification, serve to explain the technical solutions of the present invention.

[0008] Figure 1 This is a schematic diagram illustrating the implementation process of a traffic load monitoring method for advertising information services provided in an embodiment of the present invention.

[0009] Figure 2 This is a schematic diagram of the composition structure of a load monitoring device provided in an embodiment of the present invention.

[0010] Figure 3 This is a schematic diagram of the hardware entity of a computer system provided in an embodiment of the present invention. Detailed Implementation

[0011] To make the objectives, technical solutions, and advantages of the present invention clearer, the technical solutions of the present invention will be further described in detail below with reference to the accompanying drawings and embodiments. The described embodiments should not be regarded as limitations on the present invention. All other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0012] This invention provides a method for monitoring traffic load in advertising information services, which can be executed by a processor of a computer system. The computer system can refer to devices with data processing capabilities, such as servers, laptops, tablets, and desktop computers.

[0013] Figure 1 This is a schematic diagram illustrating the implementation process of a traffic load monitoring method for advertising information services provided in an embodiment of the present invention, as shown below. Figure 1 As shown, the method includes: Step S100: Obtain the advertising request stream data sequence generated within the preset monitoring period and the service node load status record sequence that has a timestamp correspondence with the advertising request stream data sequence.

[0014] The preset monitoring period is a time window with a fixed length pre-configured by the monitoring strategy. Within this preset monitoring period, the advertising request distribution component of the advertising information service system writes each arriving advertising request record to the request log storage area in real time. Each advertising request record contains at least a timestamp of the request arrival time and an advertising slot identifier. Arranging all advertising request records within the preset monitoring period in ascending order of their arrival timestamps constitutes the advertising request stream data sequence. Each sequence element in the advertising request stream data sequence corresponds to a complete advertising request record, and the order of the sequence elements strictly follows the chronological relationship of the timestamps carried in each advertising request record.

[0015] The advertising information service system refers to a distributed computer business system deployed in a network environment to receive advertisers' placement requests and display or return advertising content to end users. This system consists of multiple subsystems working together, including an advertising request access gateway, an advertising slot management module, an advertising material storage and placement module, an advertising request distribution component, a service node cluster, and an advertising performance monitoring data collection module. The advertising information service system receives advertising requests from terminal applications or web browsers. These requests carry request context information such as advertising slot identifiers, user identifiers, terminal device type identifiers, and network environment parameters. Based on the advertising slot identifier and user identifier, the system retrieves matching advertising placement strategies, selects advertising material content that meets the placement conditions from the advertising material storage area, and encapsulates the display link or rendering data of the advertising material content into an advertising response message and returns it to the request initiator. In the complete advertising request processing chain, the advertising request distribution component is responsible for distributing arriving advertising requests to specific service nodes in the downstream service node cluster according to preset routing rules. Each service node in the service node cluster undertakes a series of computational tasks, including parsing advertising requests, retrieving placement strategies, matching materials, and encapsulating responses. The advertising information service system also maintains records of business traffic associations between ad placements, describing the traffic transmission relationship when a user clicks on one ad placement and is redirected to another. The operational status data of the entire advertising information service system, including ad request flow data sequences and service node load status record sequences, is collected and aggregated in real time through data collection agents distributed across various subsystems.

[0016] The service node load status record sequence is extracted from load status data collected from service nodes that handle ad request processing. A load status collection agent process runs on each service node. This agent process collects runtime resource usage indicators of the service node at preset sampling intervals. These indicators include one or more of the following: CPU utilization, memory usage, network throughput, and the number of concurrent requests. The load status collection agent process inputs the multi-dimensional resource usage indicator values ​​collected each time into a preset load level determination logic and outputs a load level marker corresponding to that sampling time. The load level marker uses a discrete level identification system, dividing the load into four discrete levels: light load, medium load, heavy load, and overload. Each level corresponds to a preset resource usage indicator threshold range. Arranging the load level markers for all sampling times within the preset monitoring period in ascending order of the sampling timestamp constitutes the service node load status record sequence. There is a timestamp correspondence between the ad request stream data sequence and the service node load status record sequence.

[0017] Step S200: Extract the request traffic situation from the ad request stream data sequence to obtain the temporal distribution feature representation of the ad request stream data sequence, and extract the load change situation from the service node load status record sequence to obtain the load change feature representation of the service node load status record sequence. Input the temporal distribution feature representation and the load change feature representation into the traffic load joint situation analysis stage.

[0018] As one embodiment, step S200 may specifically include the following steps S210 to S260: Step S210: Perform request source decoupling on the ad request stream data sequence. Based on the ad slot identifier carried in each ad request record, split the ad request stream data sequence into multiple single-source request subsequences. Each single-source request subsequence corresponds to an independent ad slot identifier and contains all request arrival time records of that ad slot within the preset monitoring period.

[0019] An ad placement identifier is a unique identifier assigned to each ad display position in the advertising information service system. This identifier is generated during the system initialization configuration phase and remains unchanged throughout the ad placement's lifecycle. During request source decoupling, a mapping container structure is created with the ad placement identifier as the key and a list of empty request arrival times as the values. Next, the ad request stream data sequence is traversed sequentially. During this traversal, the ad placement identifier field value carried in the current ad request record is read, and the timestamp of the request arrival time in that record is extracted and appended to the list of request arrival times corresponding to the ad placement identifier in the mapping container structure. Finally, after completing the full traversal of the ad request stream data sequence, the list of request arrival times corresponding to each ad placement identifier in the mapping container structure is output as an independent single-source request subsequence. Each single-source request subsequence contains only all request arrival time records associated with a specific ad placement identifier, and the original arrival time order of each request arrival time record is maintained within the single-source request subsequence.

[0020] Step S220: Perform request arrival rhythm extraction for each single-source request subsequence, analyze the time interval distribution pattern between adjacent request arrival times in the single-source request subsequence, and transform the sparse-dense alternation pattern presented in the time interval distribution pattern into the request arrival rhythm descriptor corresponding to the ad slot. The request arrival rhythm descriptor is used to characterize the periodic aggregation tendency and intermittent sparsity tendency of the request traffic of the ad slot.

[0021] Step S220 is executed independently for each single-source request subsequence obtained in step S210. First, time interval calculation is performed on the single-source request subsequence. Starting from the arrival time of the second request in the subsequence, the absolute time difference between the current arrival time and the previous arrival time is calculated sequentially, generating a time interval sequence corresponding to the single-source request subsequence. Second, distribution pattern analysis is performed on the time interval sequence. This analysis uses a density estimation method based on a sliding time window. Specifically, the preset monitoring period is divided into multiple time slices of equal length, the length of which is preset according to the business rhythm characteristics of the advertising information service system. Within each time slice, the number of request arrival times falling within that slice's range is counted, and this number is used as the request density indicator value for that slice. The request density indicator values ​​of all time slices are arranged in chronological order to form a request density time series sequence. Third, rhythmic pattern recognition is performed on the request density time series sequence. This rhythmic pattern recognition is achieved by detecting continuous time slice intervals with density values ​​higher than a preset clustering threshold and continuous time slice intervals with density values ​​lower than a preset sparsity threshold. Continuous time fragment intervals with density values ​​consistently above the clustering threshold are marked as clustered periods, and continuous time fragment intervals with density values ​​consistently below the sparsity threshold are marked as sparse periods. Finally, the alternation pattern of the identified clustered and sparse periods on the time axis is encoded as a request arrival rhythm descriptor. The encoded structure of the request arrival rhythm descriptor includes the ad slot identifier, the start time fragment index and duration fragment length of the clustered period, the start time fragment index and duration fragment length of the sparse period, and the number of alternation cycles between the clustered and sparse periods.

[0022] Step S230: Input the request arrival rhythm descriptors corresponding to all ad slots into the preset ad slot request propagation topology. The ad slot request propagation topology is constructed with ad slot identifiers as nodes and business traffic associations between ad slots as directed edges. In the ad slot request propagation topology, the request arrival rhythm descriptors of each node are cascaded and propagated according to the propagation direction of the directed edges to generate a request impact propagation timing description for each ad slot node within the preset monitoring period.

[0023] As one embodiment, step S230 may specifically include the following steps S231 to S236: Step S231: Read the business traffic association records between ad slots from the preset ad slot association relationship storage area. Based on the traffic direction from the source ad slot identifier to the target ad slot identifier in the business traffic association records, construct an ad slot request propagation topology with ad slot identifier as the node and traffic direction as the directed edge. Each directed edge in the ad slot request propagation topology is marked with a traffic intensity attribute.

[0024] The ad placement association storage area maintains business traffic association records between various ad placements within the advertising information service system. Each business traffic association record includes a source ad placement identifier field, a target ad placement identifier field, and a traffic intensity attribute flag field. The source ad placement identifier field indicates the ad placement that initiated the user's redirection behavior, the target ad placement identifier field indicates the ad placement the user is guided to, and the traffic intensity attribute flag field is used to quantitatively describe the proportion of user request traffic from the source ad placement to the target ad placement relative to the total user request traffic from the source ad placement. The value of the traffic intensity attribute flag is obtained through statistical analysis of user redirection behavior logs within a historical preset duration window. During the statistical analysis, the number of user requests that clicked from the source ad placement and redirected to the target ad placement is counted, divided by the total number of user requests that clicked on the source ad placement within the same historical preset duration window, and the resulting quotient is used as the value of the traffic intensity attribute flag. Step S231, in constructing the ad placement request propagation topology, firstly, creates a graph data structure with ad placement identifiers as node identifiers. Each node in the graph data structure contains a node identifier field and a node attribute field. The node attribute field stores the request arrival rhythm descriptor corresponding to that ad placement. Secondly, it reads the business traffic association records in the ad placement association storage area one by one. For each record, it creates a directed edge in the graph data structure from the node corresponding to the source ad placement identifier to the node corresponding to the target ad placement identifier, and appends the traffic intensity attribute to the edge attribute of this directed edge. After completing the reading of all business traffic association records and the creation of directed edges, the resulting graph data structure is the ad placement request propagation topology.

[0025] Step S232: Load the request arrival rhythm descriptor corresponding to each ad slot into the node in the ad slot request propagation topology that matches the ad slot identifier, as the initial rhythm state description of the node. The initial rhythm state description includes the aggregation period tendency description and sparse period tendency description of the ad slot request traffic.

[0026] For each node in the ad slot request propagation topology, a search and matching process is performed based on the ad slot identifier corresponding to that node in the request arrival rhythm descriptor set output in step S220. The retrieved request arrival rhythm descriptor is assigned to the node attribute field of that node as its initial rhythm state description. The initial rhythm state description includes the clustering period tendency description and the sparse period tendency description of the ad slot corresponding to that node. The clustering period tendency description consists of the start time fragment index and the duration fragment length of the clustering period of the ad slot, while the sparse period tendency description consists of the start time fragment index and the duration fragment length of the sparse period of the ad slot. After loading the initial rhythm state descriptions of all nodes, each node in the ad slot request propagation topology has a state expression describing its own independent request rhythm characteristics.

[0027] Step S233: For each node in the ad slot request propagation topology, backtrack along all incoming edge directions of the node to obtain the request arrival rhythm descriptor of the upstream neighboring node, and perform propagation fusion of the request arrival rhythm descriptor of the upstream neighboring node with the traffic intensity attribute mark on the corresponding incoming edge to generate the incremental description of the request impact propagated from upstream to the current node.

[0028] For any node in the ad placement request propagation topology, firstly, obtain the set of incoming edges consisting of all incoming edges pointing to that node. For each incoming edge in the set, determine the starting node of that incoming edge, which is the upstream neighbor of the current node. Read the current rhythm state description from the node attribute field of the upstream neighbor. This rhythm state description is the initial rhythm state description of the upstream neighbor in the first round of propagation, and in subsequent rounds of propagation, it is the rhythm state description of the upstream neighbor updated in the previous round of propagation. Secondly, read the value of the traffic intensity attribute marker attached to the incoming edge. Thirdly, perform a propagation fusion operation: multiply the clustering time tendency description and the sparse time tendency description contained in the rhythm state description of the upstream neighbor by the value of the traffic intensity attribute marker, respectively, to obtain the clustering tendency propagation component and the sparse tendency propagation component after traffic intensity scaling. The aggregation tendency propagation components, scaled by the flow intensity, corresponding to all incoming edges of the current node are summed time-by-time fragment to obtain the aggregated aggregation tendency propagation total description. Similarly, the sparse tendency propagation components, scaled by the flow intensity, corresponding to all incoming edges of the current node are summed time-by-time fragment to obtain the aggregated sparse tendency propagation total description. Both the aggregated aggregation tendency propagation total description and the aggregated sparse tendency propagation total description are used together as the incremental description of the request impact propagated from upstream to the current node.

[0029] Step S234: Overlay the incremental description of the impact of the upstream request to the current node with the node’s own initial rhythm state description to obtain the updated rhythm state description of the node in the current transmission round. The rhythm state overlay process makes the node’s rhythm state description reflect both its original request characteristics and the impact transmission characteristics of the upstream request.

[0030] Specifically, for the clustering period tendency description included in the initial rhythm state description of the current node, its clustering tendency value at each time slice is added to the total amount of clustering tendency propagation after convergence in the corresponding time slice of the request influence increment description generated in step S233, resulting in a superimposed clustering tendency description. For the sparse period tendency description included in the initial rhythm state description of the current node, its sparse tendency value at each time slice is added to the total amount of sparse tendency propagation after convergence in the corresponding time slice of the request influence increment description generated in step S233, resulting in a superimposed sparse tendency description. The superimposed clustering tendency description and the superimposed sparse tendency description are combined to form the updated rhythm state description of the current node in the current propagation round.

[0031] Step S235: Repeatedly execute the propagation round iteration along the directed edge direction in the ad slot request propagation topology. After each round of propagation round iteration is completed, the updated rhythm state description of each node is used as the input state for the next round of propagation round iteration, until the preset propagation convergence round limit is reached, and the final rhythm state description of each node after the propagation convergence is obtained.

[0032] The stopping condition for the propagation round iteration is that the number of propagation rounds executed so far reaches a preset propagation convergence round limit. The value of the preset propagation convergence round limit is determined based on the depth of the directed edge hierarchy of the ad placement request propagation topology, for example, it is set to be positively correlated with the number of nodes on the longest directed path in the ad placement request propagation topology. In each round of propagation round iteration, steps S233 and S234 are executed sequentially for each node according to the direction order of the directed edges in the ad placement request propagation topology. After completing one round of propagation round iteration, the rhythm state description of each node in the ad placement request propagation topology is updated, and this updated rhythm state description will be used as the input state for the next round of propagation round iteration. When the preset propagation convergence round limit is reached, the iteration process stops, and the updated rhythm state description held by each node in the ad placement request propagation topology at this time is taken as the final rhythm state description of each node.

[0033] Step S236: Expand the final rhythm state description of each ad slot node after the completion of the propagation convergence into a request impact intensity indicator on each time segment according to the time axis, and arrange the request impact intensity indicators in chronological order as a request impact propagation time sequence description of each ad slot node within the preset monitoring period.

[0034] The final rhythmic state description of any ad placement node includes a superimposed clustering tendency description and a superimposed sparsity tendency description. The clustering tendency value of the superimposed clustering tendency description at each time segment and the sparsity tendency value of the superimposed sparsity tendency description at the corresponding time segment are merged. During merging, the clustering tendency value is taken as the positive impact intensity component, and the sparsity tendency value is inverted and taken as the negative impact intensity component. The positive and negative impact intensity components are algebraically summed to obtain the request impact intensity indication at that time segment. The request impact intensity indications at all time segments are arranged in chronological order to form the request impact propagation timing description of the ad placement node within a preset monitoring period. The request impact propagation timing description reflects the distribution of composite request impact intensity experienced by the ad placement node in each time segment after comprehensively considering its own independent request rhythmic characteristics and the request impact propagation characteristics from upstream traffic.

[0035] Step S240: Extract the set of request aggregation nodes for service nodes from the request impact propagation timing description. The set of request aggregation nodes contains all ad slot nodes in the ad slot request propagation topology that point to service nodes. Synchronously overlay the request impact propagation timing descriptions of each node in the set of request aggregation nodes to generate a timing profile description of the aggregation request traffic for service nodes.

[0036] In the ad slot request propagation topology, there exists a special type of node whose outgoing edges point to service nodes. This means that user requests for ad slots corresponding to these nodes will ultimately be processed directly by the service nodes. Step S240 first traverses the outgoing edge direction of the ad slot request propagation topology, filtering out all ad slot nodes whose outgoing edges point to service nodes, and aggregating the ad slot identifiers of these ad slot nodes into a request aggregation node set. Next, for each ad slot node in the request aggregation node set, the request impact propagation timing description generated in step S236 is obtained. Then, a synchronous overlay operation is performed on the request impact propagation timing descriptions of all request aggregation nodes, that is, the request impact intensity indicators at the same time fragment position of the request impact propagation timing descriptions of each request aggregation node are summed time-fragment by time fragment to obtain the total aggregation request impact intensity value at each time fragment. The total aggregation request impact intensity values ​​of all time fragments are arranged in chronological order of the time fragments to form the aggregation request traffic timing profile description facing the service node.

[0037] Step S250: Perform traffic impact peak and valley morphology analysis on the converged request traffic time sequence profile description, identify the impact accumulation segment with continuously rising traffic amplitude, the impact release segment with continuously decreasing traffic amplitude, and the impact plateau segment with stable traffic amplitude in the converged request traffic time sequence profile description, and use the accumulation steepness description of the impact accumulation segment and the release steepness description of the impact release segment as the time sequence distribution feature representation of the advertising request stream data sequence.

[0038] As one embodiment, step S250 may specifically include the following steps S251 to S256: Step S251: Convert the time-series profile description of the aggregation request traffic into a traffic amplitude sequence description that is continuously arranged along the time axis. Each sequence position in the traffic amplitude sequence description corresponds to the aggregation request traffic amplitude in a time segment.

[0039] The convergence request traffic time-series profile description generated in step S240 is a discrete data sequence indexed by time slices. Each data element in this discrete data sequence is the total convergence request impact intensity value at the corresponding time slice. Step S251 directly uses this discrete data sequence as a traffic amplitude sequence description. Each sequence position in the traffic amplitude sequence description corresponds to a time slice, and the value at that sequence position is the convergence request traffic amplitude of that time slice.

[0040] Step S252: Perform flow change direction determination on the flow amplitude sequence description, calculate the change direction mark of the flow amplitude between adjacent sequence positions in the flow amplitude sequence description one by one. The change direction mark includes the amplitude increasing mark, the amplitude decreasing mark and the amplitude stationary mark, and arrange the change direction mark in chronological order to form the flow change direction mark sequence.

[0041] Specifically, starting from the second sequence position in the flow amplitude sequence description, the difference between the flow amplitude at the current sequence position and the flow amplitude at the previous sequence position is calculated sequentially. This difference is compared to a preset amplitude stability tolerance value: if the difference is greater than a positive value of the amplitude stability tolerance value, the change direction marker corresponding to the current sequence position is set to an increasing amplitude marker; if the difference is less than a negative value of the amplitude stability tolerance value, the change direction marker corresponding to the current sequence position is set to a decreasing amplitude marker; if the absolute value of the difference is less than or equal to the amplitude stability tolerance value, the change direction marker corresponding to the current sequence position is set to a stable amplitude marker. For the first sequence position in the flow amplitude sequence description, its change direction marker can be set to an empty marker or, by default, the change direction marker of its subsequent adjacent positions can be used. Arranging the change direction markers corresponding to all sequence positions in chronological order constitutes the flow change direction marker sequence.

[0042] Step S253: Perform continuous same-direction marker merging on the flow change direction marker sequence, merge multiple markers with the same change direction and consecutively adjacent positions in the flow change direction marker sequence into an independent flow pattern segment, and divide the segment into an impact accumulation segment, an impact release segment, or an impact platform segment according to the change direction marker type contained in the segment.

[0043] The finite state automaton contains three states: accumulation state, release state, and plateau state. Transitions between states are triggered by changes in the marker type within the flow change direction marker sequence. The merging process is as follows: Initialize an empty list of flow pattern segments. Use the first valid change direction marker in the flow change direction marker sequence as the current segment marker type and record the starting sequence position index of this segment. Sequentially traverse each subsequent change direction marker in the flow change direction marker sequence. If the currently traversed change direction marker has the same marker type as the current segment, continue traversing; if the currently traversed change direction marker has a different marker type, terminate the extension of the current segment, encapsulate the marker type, starting sequence position index, and ending sequence position index of the current segment into a flow pattern segment record, append it to the flow pattern segment list, and use the currently traversed change direction marker as the new current segment marker type and the current traversal position as the new starting sequence position index, continuing the traversal process. After completing a full traversal of the flow change direction marking sequence, each segment record in the flow pattern segment list is classified according to its marking type: segments marked with increasing amplitude are classified as impact accumulation segments, segments marked with decreasing amplitude are classified as impact release segments, and segments marked with stable amplitude are classified as impact plateau segments.

[0044] Step S254: For each flow pattern segment divided into impact accumulation segments, extract the flow amplitude corresponding to the start time segment and the flow amplitude corresponding to the end time segment of the segment, and determine the accumulation steepness description of the impact accumulation segment based on the number of time segments spanned by the segment and the amplitude difference between the start and end flow amplitudes.

[0045] For any impact accumulation segment, the flow amplitude corresponding to the starting sequence position of the segment is read from the flow amplitude sequence description and recorded as the starting amplitude; the flow amplitude corresponding to the ending sequence position of the segment is read and recorded as the ending amplitude. The amplitude difference between the ending amplitude and the starting amplitude is calculated; the amplitude difference is the positive value obtained by subtracting the starting amplitude from the ending amplitude. The number of time segments spanned by the segment is obtained, i.e., the difference between the ending sequence position index and the starting sequence position index of the segment. The cumulative steepness description is determined by dividing the amplitude difference by the number of time segments spanned by the segment, and the quotient is used as the cumulative steepness description of the impact accumulation segment. The magnitude of the cumulative steepness description is positively correlated with the growth rate of the flow amplitude per unit time segment.

[0046] Step S255: For each flow pattern segment divided into impact release segments, extract the flow amplitude corresponding to the start time segment and the flow amplitude corresponding to the end time segment of the segment, and determine the release steepness description of the impact release segment based on the number of time segments spanned by the segment and the amplitude difference between the start and end flow amplitudes.

[0047] For any impact release segment, the flow amplitude corresponding to the starting sequence position of the segment is read from the flow amplitude sequence description and recorded as the starting amplitude; the flow amplitude corresponding to the ending sequence position of the segment is read and recorded as the ending amplitude. The amplitude difference between the starting and ending amplitudes is calculated; the amplitude difference is a positive value obtained by subtracting the ending amplitude from the starting amplitude. The number of time segments spanned by the segment is obtained, i.e., the difference between the ending sequence position index and the starting sequence position index of the segment. The release steepness description is determined by dividing the amplitude difference by the number of time segments spanned by the segment, and the quotient is used as the release steepness description of the impact release segment. The magnitude of the release steepness description is positively correlated with the rate of decrease of the flow amplitude within a unit time segment.

[0048] Step S256: Arrange the cumulative steepness description of the impact accumulation segment and the release steepness description of the impact release segment according to the order of their respective segments on the time axis to form a morphological description sequence consisting of alternating cumulative steepness description and release steepness description, and use the morphological description sequence as the temporal distribution feature representation of the advertising request stream data sequence.

[0049] Based on the order of appearance of each traffic pattern segment obtained in step S253 on the time axis, the cumulative steepness description of each impact accumulation segment and the release steepness description of each impact release segment are extracted sequentially, and the steepness descriptions are arranged into a sequence structure according to their order of appearance. For impact platform segments appearing in the segmentation results, an empty description marker is inserted at the corresponding position in the pattern description sequence or the segment is skipped. The final sequence structure, which is composed of alternating cumulative steepness descriptions and release steepness descriptions, is the temporal distribution feature representation of the advertising request flow data sequence. The temporal distribution feature representation, in a compact serialized form, depicts the dynamic alternation pattern of impact accumulation rate and impact release rate of advertising request traffic at the service node entry point within the preset monitoring period.

[0050] Step S260: Perform load state transition chain reconstruction on the service node load state record sequence, connect adjacent load level marker transition events in the service node load state record sequence into a directed load state transition link according to the order of the transition, and extract the transition sub-links whose occurrence frequency meets the preset frequency condition from the directed load state transition link as the load change feature representation of the service node load state record sequence. The time-series distribution feature representation and the load change feature representation are jointly transmitted to the input entry of the traffic load joint situation analysis stage.

[0051] Step S260 first identifies load level marker transition events between adjacent sampling times in the service node load state record sequence. Each sequence element in the service node load state record sequence represents a sampling time and its corresponding load level marker. Starting from the second sequence element, the load level marker of the current sequence element is compared with that of the previous sequence element. If they are different, a load state transition event is generated, which includes the source load level marker, the target load level marker, and the timestamp of the transition. All identified load state transition events are concatenated according to the order of their timestamps. A directed connection is formed between every two adjacent load state transition events, with the source load level marker pointing to the target load level marker, thus forming a directed load state transition link.

[0052] Secondly, frequent hop sub-link extraction is performed on the directed links for load state transitions. Frequent hop sub-link extraction employs a sliding window-based sub-link frequency statistics method. Specifically, a sub-link length parameter is set, representing the number of consecutive load level marker transition events contained in the sub-link. The sub-link length parameter is typically 2 or 3. Using the sub-link length parameter as the window width, the window slides from the starting position to the ending position on the directed link for load state transitions, sequentially extracting consecutive load level marker transition event sequences within each window's coverage area. Each sequence is considered a hop sub-link. The frequency of occurrence of all extracted hop sub-links is counted, and hop sub-links with a frequency greater than or equal to a preset frequency condition are selected. The preset frequency condition can be an absolute frequency threshold or a relative frequency ratio threshold.

[0053] Finally, the selected transition sub-links that meet the preset frequency criteria are combined into a transition sub-link set, which is used as the load change feature representation of the service node load state record sequence. The load change feature representation reflects the typical load state transition patterns that repeatedly occur in the service node within the preset monitoring period. The time-series distribution feature representation generated in step S256 and the load change feature representation generated in step S260 are jointly transmitted to the input entry of the traffic load joint situational analysis stage.

[0054] Step S300: Invoke the pre-configured traffic load collaborative analysis network to perform collaborative interactive mapping of the time-series distribution feature representation and the load change feature representation, and generate a correlation response characteristic representation that represents the dynamic correlation between advertising request traffic and load status.

[0055] As one embodiment, step S300 may specifically include the following steps S310 to S360: Step S310: Load the time-series distribution feature representation and the load change feature representation simultaneously into the input access layer of the traffic load collaborative analysis network. The input access layer performs structure normalization and dimension alignment operations on the time-series distribution feature representation and the load change feature representation to generate a standardized input feature pair representation with a consistent representation form.

[0056] The traffic load collaborative analysis network is a neural network model with dual input and dual output channels. This neural network model consists of a stacked input access layer, a bidirectional cross-sensing processing sublayer, an interactive integration processing layer, and a time-series analysis output layer. The input access layer contains two parallel feature embedding transformation modules: a traffic feature embedding transformation module and a load feature embedding transformation module. The traffic feature embedding transformation module consists of two cascaded fully connected layers. The input dimension of the first fully connected layer matches the dimension of the time-series distribution feature representation, and the output dimension of the second fully connected layer is a preset unified feature embedding dimension. A non-linear activation function is set between the two fully connected layers. The load feature embedding transformation module has the same hierarchical structure as the traffic feature embedding transformation module, and its second fully connected layer also outputs a preset unified feature embedding dimension.

[0057] The process of structure normalization and dimension alignment is as follows: The temporal distribution feature representation is fed into the traffic feature embedding and transformation module. After linear transformation and nonlinear mapping by a nonlinear activation function in the first fully connected layer, it is then mapped to a unified feature embedding dimension through the second fully connected layer, outputting a normalized traffic-side feature sequence. Simultaneously, the load variation feature representation is fed into the load feature embedding and transformation module. After the same linear transformation, nonlinear mapping, and dimension mapping process, a normalized load-side feature sequence is output. The normalized traffic-side feature sequence and the normalized load-side feature sequence are paired to form a standardized input feature pair representation.

[0058] Step S320: The bidirectional cross-sensing processing sublayer inside the traffic load collaborative analysis network performs forward information permeation on the standardized input feature representation, calculates the attention response distribution of each time position point in the time-series distribution feature representation to each load state point in the load change feature representation, and generates a traffic-to-load attention mapping representation describing the direction of traffic action on the load based on the attention response distribution.

[0059] As one embodiment, step S320 may specifically include the following steps S321 to S325: Step S321: Decompose the traffic query vector sequence description corresponding to the time-series distribution feature representation from the standardized input feature pair representation, and decompose the load query vector sequence description corresponding to the load change feature representation from the standardized input feature pair representation. Each vector element in the traffic query vector sequence description corresponds to a time position in the request traffic fluctuation pattern sequence, and each vector element in the load query vector sequence description corresponds to a load state point in the load state transition event sequence description.

[0060] The bidirectional cross-sensing processing sublayer contains two structurally symmetrical but parameter-independent cross-attention calculation units: a forward cross-attention calculation unit and a backward cross-attention calculation unit. Step S321 is executed in the forward cross-attention calculation unit. The forward cross-attention calculation unit contains three linear transformation matrices: a query linear transformation matrix, a key linear transformation matrix, and a value linear transformation matrix. The flow-side feature sequence in the standardized input feature pair representation is sequentially multiplied with the query linear transformation matrix to generate a flow query vector sequence description. Each vector element in the flow query vector sequence description corresponds to a time position in the flow-side feature sequence. Simultaneously, the load-side feature sequence in the standardized input feature pair representation is sequentially multiplied with both the key linear transformation matrix and the value linear transformation matrix to generate the key vector sequence and value vector sequence in the load query vector sequence description. Each vector element in the load query vector sequence description corresponds to a load state point in the load-side feature sequence.

[0061] Step S322: Perform a scaling dot product-based attention measurement on each traffic query vector in the traffic query vector sequence description and each load query vector in the load query vector sequence description, and generate a two-dimensional attention description table. The value of the entry at the intersection of a certain row and a certain column in the two-dimensional attention description table represents the attention response intensity of the corresponding traffic time position to the corresponding load state point.

[0062] For example, suppose the traffic query vector sequence description contains M traffic query vectors, and the key vector sequence in the load query vector sequence description contains N key vectors. For the i-th traffic query vector and the j-th key vector, calculate their vector dot product, which is the sum of the products of the corresponding dimensional components of the two vectors. Divide the result of the vector dot product by the square root of the uniform feature embedding dimension, and use the quotient as the unnormalized raw score of attention. Repeat the above calculation for all combinations of M traffic query vectors and N key vectors to obtain a two-dimensional raw score table of attention with M rows and N columns. The value of the entry at the intersection of the i-th row and j-th column in the two-dimensional raw score table is the unnormalized representation of the attention response intensity of the i-th traffic time position to the j-th load state point.

[0063] Step S323: Perform row-wise distribution normalization on the two-dimensional attention level description table so that the sum of the attention response intensity of the same flow time position represented by each row of the two-dimensional attention level description table to all load state points satisfies the preset normalization constraint condition, thus obtaining the normalized attention level description table.

[0064] Row-based normalization employs a flexible maximum normalization function, independently applied to each row of the two-dimensional attention score table. For the i-th row, all N original attention scores are used as input. The flexible maximum normalization function calculates the exponent value of each original attention score with the natural constant as the base, then calculates the sum of all N exponent values ​​for that row. Finally, each exponent value is divided by this sum to obtain the normalized attention description value corresponding to each column position in that row. After row-based normalization, the sum of all normalized attention description values ​​for each row equals 1, satisfying the preset normalization constraint. The results of normalization for all rows are combined to form a normalized attention description table.

[0065] Step S324: Based on the distribution of attention response intensity in each row of the normalized attention level description table, extract the load status point identifier corresponding to the column with the highest attention response intensity in each row, and construct a one-way attention pointing relationship description from the traffic time position corresponding to that row to the load status point identifier.

[0066] For the i-th row of the normalized attention description table, iterate through the N normalized attention description values ​​in that row and determine the column index of the normalized attention description value with the largest value. Locate the corresponding load state point in the load lookup vector sequence description based on this column index and obtain the identifier of that load state point. Construct a one-way attention pointing relationship description from the i-th traffic time position to the identifier of this load state point. The one-way attention pointing relationship description includes the time-series index of the traffic time position, the identifier of the load state point, and the corresponding normalized attention description value. Perform the above operation on each row of the normalized attention description table to obtain the one-way attention pointing relationship descriptions corresponding to all traffic time positions.

[0067] Step S325: Arrange the descriptions of the one-way attention-direction relationships corresponding to all traffic time positions in chronological order of traffic time positions, and depict the arrangement results on a two-dimensional plane as the trajectory of the focus of attention migrating over time. Represent this trajectory as a traffic-to-load attention mapping diagram.

[0068] The descriptions of unidirectional attention relationships corresponding to M traffic time positions are arranged in ascending order according to the temporal index of the traffic time position. The load state point identifier in each unidirectional attention relationship description constitutes a temporal sequence of attention points. The horizontal axis is set as the temporal index of the traffic time position, and the vertical axis is set as the discrete value of the load state point identifier. Connecting the coordinate points corresponding to each attention point in the progressive order of the traffic time positions on a two-dimensional plane forms a polygonal line trajectory. This polygonal line trajectory is the traffic-to-load attention mapping diagram, which visually displays the migration trajectory of the attention focus of advertising request traffic in the load state space of the service node within a preset monitoring period.

[0069] Step S330: The bidirectional cross-sensing processing sub-layer performs reverse information permeation on the standardized input feature representation, calculates the attention response distribution of each load state point in the load change feature representation to each time position point in the time-series distribution feature representation, and generates a load-to-flow attention mapping representation describing the direction of the load's reaction to the flow based on the attention response distribution.

[0070] Step S330 is executed in the reverse cross-attention calculation unit, and its execution process is symmetrical to that of step S320. The reverse cross-attention calculation unit also includes a query linear transformation matrix, a key linear transformation matrix, and a value linear transformation matrix. The load-side feature sequence in the standardized input feature pair description is multiplied with the query linear transformation matrix to generate a load query vector sequence description. The flow-side feature sequence in the standardized input feature pair description is then multiplied sequentially with the key linear transformation matrix and the value linear transformation matrix to generate the key vector sequence and value vector sequence in the flow query vector sequence description. For each load query vector in the load query vector sequence description and each key vector in the flow query vector sequence description, a scaling dot product-based attention measurement and row-wise distribution normalization are performed to obtain a reverse normalized attention description table. Based on the distribution of attention response intensity in each row of the reverse normalized attention description table, the flow time location identifier corresponding to the column with the highest attention response intensity in each row is extracted, constructing a one-way attention pointing relationship description from the load state point corresponding to that row to the flow time location identifier. The descriptions of the unidirectional attention-direction relationships corresponding to all load state points are arranged in the order of their appearance and depicted on a two-dimensional plane as the trajectory of the focus of attention as the load state evolves. This trajectory is represented as a load-direction traffic attention mapping diagram.

[0071] Step S340: Input the traffic-to-load attention mapping representation and the load-to-traffic attention mapping representation into the interactive integration processing layer of the traffic-load collaborative analysis network, perform a bidirectional information fusion operation based on the selection gating mechanism, and generate a set of collaborative interactive feature representations that simultaneously contain forward action descriptions and reverse action descriptions.

[0072] The interactive integration processing layer consists of gated linear units and feature concatenation units. Each gated linear unit contains two independent gated branches, processing flow-to-load attention map representations and load-to-flow attention map representations, respectively. Each gated branch comprises a fully connected layer and a non-linear activation function cascaded together. The output dimension of the fully connected layer is the same as the input feature dimension. The non-linear activation function compresses the output value of the fully connected layer to an open interval between 0 and 1, serving as the gate weight coefficients. The flow-to-load attention map representation is multiplied element-wise with the gate weight coefficients output from the first gated branch to obtain the gated forward action description. The load-to-flow attention map representation is multiplied element-wise with the gate weight coefficients output from the second gated branch to obtain the gated reverse action description. The feature concatenation unit concatenates and merges the gated forward action descriptions and the gated reverse action descriptions along the feature dimension to generate a set of collaborative interactive feature representations. Each feature vector element in the collaborative interaction feature representation set contains both the dimensional components describing the forward action and the dimensional components describing the reverse action at the corresponding time or state position.

[0073] Step S350: Perform time-series following tightness analysis on the collaborative interaction feature representation set, extract the response following time offset between the time position on the flow side and the time position on the load side, convert the response following time offset into a dimensionless response following hysteresis ratio description, and extract the decay trend of the interaction response intensity with relative time offset and convert it into a dimensionless response following intensity decay ratio description.

[0074] As one embodiment, step S350 may specifically include the following steps S351 to S355: Step S351: Locate a pair of collaborative interaction feature representations in the collaborative interaction feature representation set where the interaction response intensity reaches a local maximum. The pair of collaborative interaction feature representations consists of a flow-side time location feature representation element and a load-side time location feature representation element, and the interaction response intensity between the two satisfies the local peak determination condition.

[0075] The time-series analysis output layer includes an interaction response strength evaluation module. This module calculates an interaction response strength for each pair of traffic-side and load-side time-location feature representation elements in the collaborative interaction feature representation set. The interaction response strength is calculated by multiplying the corresponding dimension components of the traffic-side and load-side time-location feature representation vectors, summing the results, and then dividing by the product of the magnitudes of the two vectors. The criteria for determining local peak values ​​are as follows: For a pair of elements with position indices i and j in the set of collaborative interaction feature representations, the interaction response intensity is greater than that of the combination with position indices i-1 and j-1, and is greater than that of the combination with position indices i-1 and j+1, and is greater than that of the combination with position indices i-1 and j+1, and is greater than that of the combination with position indices i and j-1, and is greater than that of the combination with position indices i+1 and j-1, and is greater than that of the combination with position indices i+1 and j+1.

[0076] Step S352: Extract the timestamp indication information corresponding to the time location feature representation element on the flow side, and extract the timestamp indication information corresponding to the time location feature representation element on the load side. Calculate the time difference between the timestamp indication information of the time location feature representation element on the flow side and the timestamp indication information of the time location feature representation element on the load side. Divide the absolute time length of the time difference by the absolute time length of the preset baseline response hysteresis reference duration, and use the resulting quotient as the dimensionless response following hysteresis ratio description.

[0077] The timestamp information corresponding to the traffic-side time location feature element is the absolute time coordinate value of that traffic time location within a preset monitoring period. The timestamp information corresponding to the load-side time location feature element is the absolute time coordinate value of the sampling time corresponding to that load status point within a preset monitoring period. The time difference is the result of subtracting the timestamp information of the load-side time location from the timestamp information of the traffic-side time location, and taking the absolute value of this result to obtain the absolute time length value. The preset baseline response hysteresis reference duration is a fixed time length reference value determined based on the typical request processing latency of the advertising information service system. Dividing the absolute time length value by the absolute time length value of the baseline response hysteresis reference duration yields a unitless pure value, which is the response following hysteresis ratio description.

[0078] Step S353: Using a pair of collaborative interaction feature representations with local maximum interaction response intensity as the central reference position, define a time analysis window range of preset length along the forward and backward extension directions of the time axis, and extract the decrease in interaction response intensity between adjacent collaborative interaction feature representation pairs within the time analysis window range.

[0079] The preset length of the time analysis window is measured in units of time fragments. In the forward extension direction of the time axis, a preset number of collaborative interaction feature pairs are continuously selected from the central reference position along the direction of increasing time; in the backward extension direction of the time axis, a preset number of collaborative interaction feature pairs are continuously selected from the central reference position along the direction of decreasing time. The method for extracting the decrease in interaction response intensity is as follows: for any two temporally adjacent collaborative interaction feature pairs within the time analysis window, the difference between the interaction response intensity of the earlier collaborative interaction feature pair and the interaction response intensity of the later collaborative interaction feature pair is calculated, and this difference is used as the decrease in interaction response intensity between the two.

[0080] Step S354: Construct a decay trend depiction of the interaction response intensity as a function of the relative time offset based on the extracted decrease in interaction response intensity, and determine the relative time offset corresponding to the decrease in interaction response intensity to the preset decay reference boundary from the decay trend depiction.

[0081] Using the central reference position as the zero point of the relative time offset, a scatter sequence is constructed in a two-dimensional coordinate system, with the relative time offset corresponding to each collaborative interaction feature pair within the time analysis window as the abscissa and the interaction response intensity of that collaborative interaction feature pair as the ordinate. A piecewise linear interpolation method is used to connect the scatter sequence into a continuous polygonal line, which represents the attenuation trend. The preset attenuation reference boundary is the result of multiplying the interaction response intensity at the central reference position by a preset attenuation ratio coefficient. In the attenuation trend depiction, starting from the central reference position, along both the increasing and decreasing directions of the relative time offset, the point where the polygonal line first intersects the preset attenuation reference boundary is found, and the absolute value of the relative time offset corresponding to that point is extracted.

[0082] Step S355: Divide the absolute time length of the relative time offset by the absolute time length of the preset baseline intensity decay reference duration, and use the resulting quotient as the dimensionless response following intensity decay ratio description.

[0083] The absolute value of the relative time offset extracted in step S354 is used as the absolute time length value. The preset baseline intensity decay reference duration is a fixed time length reference value determined based on the typical load response decay characteristics of the advertising information service system. Dividing the absolute time length value by the absolute time length value of the baseline intensity decay reference duration yields a unitless pure value, which is the description of the response following intensity decay ratio.

[0084] Step S360: Combine the response following hysteresis ratio description and the response following intensity decay ratio description as a characterization of the associated response properties.

[0085] The response following hysteresis ratio description generated in step S352 and the response following strength decay ratio description generated in step S355 are packaged into a two-dimensional data set structure, which is the correlation response characteristic representation. The response following hysteresis ratio description in the correlation response characteristic representation reflects the degree to which changes in ad request traffic affect the load state of the service node relative to the system baseline response hysteresis duration. The response following strength decay ratio description reflects the rate at which the interaction response strength between ad request traffic and the load state decays as the time misalignment between the two increases.

[0086] Step S400: Perform load bearing situation deduction on the associated response characteristic characterization to obtain the load bearing capacity boundary description of the service node under the continuous impact of the advertising request stream data sequence.

[0087] As one embodiment, step S400 may specifically include the following steps S410 to S460: Step S410: Extract the response following hysteresis ratio description and the response following strength decay ratio description from the associated response characteristic characterization, and simultaneously transmit the response following hysteresis ratio description and the response following strength decay ratio description to the preset load bearing situation inference processor's initial state configuration interface.

[0088] The load endurance situation simulation processor is a numerical simulation engine built on differential dynamics systems. Its core is a simulation calculation module that defines the state simulation space and state evolution rules. The initial state configuration interface is an input port provided by the load endurance situation simulation processor for receiving external input parameters. Step S410 parses the associated response characteristic representation generated in step S360, separating the values ​​described by the response following hysteresis ratio and the response following intensity decay ratio, and passes these two values ​​as configuration parameters to the initial state configuration interface.

[0089] Step S420: The load bearing situation simulation processor generates an initial state vector representation for load bearing situation analysis based on the response following hysteresis ratio description and the response following intensity decay ratio description. The initial state vector representation corresponds to an initial state coordinate point in the state simulation space of the load bearing situation simulation processor.

[0090] The state projection space of the load capacity situation simulation processor is a multi-dimensional continuous space. Each dimension of this multi-dimensional continuous space corresponds to different characteristics of the service node's load state, including the load saturation dimension, the load change rate dimension, and the load recovery resilience dimension. The initial state vector representation is a vector with the same number of dimensions as this multi-dimensional continuous space. Each component of this vector is generated by the response-following hysteresis ratio description and the response-following strength decay ratio description through a preset initial state mapping function. The initial state mapping function maps the response-following hysteresis ratio description and the response-following strength decay ratio description to the initial values ​​of the load saturation dimension and the load change rate dimension, respectively, and sets the initial value of the load recovery resilience dimension to a preset constant value. The components of each dimension of the initial state vector representation are located on the corresponding coordinate axes of the state projection space, i.e., the initial state coordinate points are determined.

[0091] Step S430: Within the state deduction space of the load bearing situation deduction processor, take the initial state vector representation as the deduction starting point, use the response following hysteresis ratio as the basis for adjusting the deduction step size, and use the response following intensity decay ratio as the basis for deduction direction deflection to perform multi-step continuous state deduction, generating a deduction path sequence composed of multiple intermediate state vector representations.

[0092] As one embodiment, step S430 may specifically include the following steps S431 to S435: Step S431: Establish initial state coordinate points corresponding to the initial state vector representation in the state deduction space of the load bearing situation deduction processor. The coordinate values ​​of each dimension of the initial state coordinate points are determined by the projection of each dimension component of the initial state vector representation onto the coordinate system of the state deduction space.

[0093] The coordinate system of the state deduction space is an orthogonal coordinate system with the load saturation dimension, the load change rate dimension, and the load recovery elasticity dimension as coordinate axes. The component values ​​of the initial state vector in the load saturation dimension are used as the first-dimensional coordinate values ​​of the initial state coordinate points, the component values ​​of the initial state vector in the load change rate dimension are used as the second-dimensional coordinate values ​​of the initial state coordinate points, and the component values ​​of the initial state vector in the load recovery elasticity dimension are used as the third-dimensional coordinate values ​​of the initial state coordinate points.

[0094] Step S432: Determine the corresponding step size adjustment coefficient based on the response following hysteresis ratio description. There is a pre-set monotonically decreasing mapping relationship between the step size adjustment coefficient and the hysteresis degree represented by the response following hysteresis ratio description. Perform a multiplication scaling operation on the preset basic inference step size and the step size adjustment coefficient to obtain the actual inference step size applicable to the current inference process.

[0095] A pre-defined monotonically decreasing mapping relationship is stored in the parameter storage area of ​​the load bearing situation simulation processor. This monotonically decreasing mapping relationship is specifically a piecewise linear mapping function, whose domain is the interval consisting of all possible values ​​described by the response-following hysteresis ratio, and its range is a closed interval between 0 and 1. The larger the value described by the response-following hysteresis ratio, the more significant the hysteresis, and the closer the corresponding step size adjustment coefficient is to zero; the smaller the value described by the response-following hysteresis ratio, the milder the hysteresis, and the closer the corresponding step size adjustment coefficient is to 1. The preset basic simulation step size is a fixed step size value determined based on the characteristic scale of the state simulation space. Multiplying the value of the basic simulation step size by the value of the step size adjustment coefficient yields the actual simulation step size.

[0096] Step S433: Using the response following intensity decay ratio description as the direction deflection adjustment reference, the preset basic inference direction vector is mapped to the direction deflection adjustment reference by a deflection angle, and the actual inference direction vector applicable to the current inference process is calculated. The deflection angle of the actual inference direction vector relative to the basic inference direction vector is determined by the response following intensity decay ratio description.

[0097] As one embodiment, step S433 may specifically include the following steps S4331 to S4335: Step S4331: The response following strength decay ratio is directly used as a dimensionless indicator of the degree of decay. The degree of decay indicator is positively correlated with the rate of decay of the response following strength reflected by the response following strength decay ratio.

[0098] The larger the value described by the response following strength decay ratio, the slower the rate at which the interaction response strength decays with relative time offset, and the greater the corresponding decay degree indication; the smaller the value described by the response following strength decay ratio, the faster the rate at which the interaction response strength decays with relative time offset, and the smaller the corresponding decay degree indication.

[0099] Step S4332: Obtain the pre-set basic deduction direction vector in the state deduction space. The direction of the basic deduction direction vector is the dominant direction of the load bearing situation in the state deduction space toward the stable convergence region.

[0100] The basic inference direction vector is a unit direction vector determined based on the global geometry of the state inference space during the initialization phase of the load bearing situation inference processor. The direction of this unit direction vector points to a preset stable convergence region in the state inference space. The stable convergence region is a coordinate region in the state inference space characterized by lower values ​​of the load saturation dimension, lower values ​​of the load change rate dimension, and higher values ​​of the load recovery elasticity dimension.

[0101] Step S4333: Query the pre-configured deflection angle mapping relationship reference table according to the decay degree indicator, and obtain the direction deflection angle corresponding to the decay degree indicator from the deflection angle mapping relationship reference table. The direction deflection angle is used to represent the deviation of the actual inferred direction vector from the basic inferred direction vector.

[0102] The deflection angle mapping reference table is stored in the parameter storage area of ​​the load bearing situation simulation processor. This reference table uses the discretized value range of the decay degree indicator as an index to store the corresponding directional deflection angle. The unit of the directional deflection angle is angle measurement. The larger the decay degree indicator, the smaller the corresponding directional deflection angle, indicating that the actual simulation direction vector is closer to the basic simulation direction vector; the smaller the decay degree indicator, the larger the corresponding directional deflection angle, indicating that the actual simulation direction vector deviates more significantly from the basic simulation direction vector.

[0103] Step S4334: In the state deduction space, with the basic deduction direction vector as the reference direction axis and the direction deflection angle as the rotation amplitude, perform an angle rotation operation on the basic deduction direction vector in a rotation plane perpendicular to the reference direction axis to obtain the actual deduction direction vector after deflection adjustment.

[0104] The normal vector of the rotation plane is determined by the cross product of the basic derivation direction vector and a preset auxiliary reference vector. Within the rotation plane, with the reference direction axis as the starting axis, a rotation transformation is performed according to the rotation amplitude specified by the direction deflection angle. The direction of the rotation transformation is either a preset clockwise or counterclockwise direction. The direction of the vector after the rotation transformation is the direction of the actual derivation direction vector.

[0105] Step S4335: Perform vector length normalization on the actual deduction direction vector, adjust the magnitude of the actual deduction direction vector to the preset unit vector length standard, so that the actual deduction direction vector only provides deduction direction guidance and does not interfere with the magnitude of the actual deduction step size.

[0106] Calculate the Euclidean magnitude of the actual derivation direction vector in the current coordinate system. Divide each coordinate component of the actual derivation direction vector by this magnitude to obtain a normalized vector with a magnitude of 1. This normalized vector is the final actual derivation direction vector.

[0107] Step S434: Starting from the initial state coordinate point, move the actual deduction step by the actual deduction direction vector to reach the first intermediate state coordinate point, and vectorize the first intermediate state coordinate point as the first intermediate state vector representation.

[0108] In the state deduction space, each coordinate component of the actual deduction direction vector is multiplied by the actual deduction step size to obtain the coordinate components of the displacement vector. The coordinate values ​​of the initial state coordinate points are then added to the corresponding coordinate components of the displacement vector to obtain the coordinate values ​​of the first intermediate state coordinate point. The coordinate values ​​of the first intermediate state coordinate point constitute the dimensional components of the first intermediate state vector.

[0109] Step S435: Take the first intermediate state vector representation as the new starting point of the deduction, and repeat the deduction step of moving the vector along the actual deduction direction by the actual deduction step size until the number of deduction steps reaches the preset upper limit of the number of deduction steps. Vectorize the intermediate state coordinate points obtained in each deduction, and arrange all vectorized representations in the order of generation into a deduction path sequence.

[0110] The preset upper limit for the number of deduction steps is an integer value predetermined based on the scale and granularity of the state deduction space. In each deduction step, starting from the current deduction starting point, the system moves along the actual deduction direction vector determined in step S433 by the actual deduction step length determined in step S432 to obtain the next intermediate state coordinate point. The next intermediate state coordinate point is then vectorized and used as the new deduction starting point to enter the next deduction step. The deduction process terminates when the number of deduction rounds executed equals the preset upper limit for the number of deduction steps. The initial state vector representation and the intermediate state vector representation generated in each deduction round are arranged in the order of generation to form a deduction path sequence.

[0111] Step S440: Perform state convergence determination on the deduced path sequence, monitor the state difference between two adjacent intermediate state vector representations in the deduced path sequence, and when the state difference decreases to below the preset convergence stability reference boundary, mark the currently reached intermediate state vector representation as a stable convergence state representation.

[0112] When calculating the state difference, for two adjacent intermediate state vector representations in the deduction path sequence, the absolute value of the component differences between them in each coordinate dimension is calculated, and the sum of the absolute values ​​of the component differences in each dimension is taken as the state difference. The convergence stability reference boundary is a preset numerical threshold. Starting from the second vector representation in the deduction path sequence, the state difference between it and the previous vector representation is calculated sequentially, and the state difference is compared with the convergence stability reference boundary. When the state difference is less than the convergence stability reference boundary for the first time, the latter of the two adjacent vector representations is marked as a stable convergent state representation, and the state convergence determination process is terminated. If no state difference is less than the convergence stability reference boundary after traversing the entire deduction path sequence, the last vector representation in the deduction path sequence is marked as a stable convergent state representation.

[0113] Step S450: Obtain the spatial coordinate points occupied by the stable convergent state representation in the state deduction space, calculate the shortest spatial geometric interval between the spatial coordinate points and the preset load-bearing boundary hypersurface, and at the same time obtain the preset load-bearing boundary reference scale distance. Divide the shortest spatial geometric interval distance by the load-bearing boundary reference scale distance, and use the resulting quotient as the dimensionless load-bearing remaining ratio description characterizing the remaining bearing capacity.

[0114] The preset load-bearing boundary hypersurface is a pre-defined interface within the state simulation space, dividing it into a safe load-bearing area and an overload risk area. The preset load-bearing boundary hypersurface is constructed by fitting data after statistical analysis of service node overload events in historical operational data. In calculating the shortest spatial geometric interval, the spatial coordinates corresponding to the stable convergent state description are used as query points, and the shortest distance from the query point to the preset load-bearing boundary hypersurface is calculated using a dimensional projection distance accumulation method. The load-bearing boundary reference scale distance is a fixed reference value determined based on the characteristic scale of the state simulation space. Dividing the shortest spatial geometric interval distance by the load-bearing boundary reference scale distance yields a quotient that represents the remaining load-bearing ratio. A larger remaining load-bearing ratio indicates a more ample remaining load-bearing capacity for the service node.

[0115] Step S460: Combine the spatial coordinate points of the load-bearing capacity boundary description with the description of the stable convergence state.

[0116] The load-bearing capacity boundary description is a composite data structure containing two fields: the first field is a dimensionless numerical field used to store the description of the remaining load-bearing ratio; the second field is a coordinate array field used to store the component values ​​of the spatial coordinate points representing the stable convergence state on each coordinate axis of the state deduction space.

[0117] Step S500: Generate a traffic scheduling guidance instruction for the advertising information service based on the load capacity boundary description, and push the traffic scheduling guidance instruction to the advertising request distribution component to trigger the request diversion adjustment operation.

[0118] As one embodiment, step S500 may specifically include the following steps S510 to S560: Step S510: Perform boundary structure analysis on the load-bearing capacity boundary description, separate the spatial coordinate points of the load-bearing capacity remaining ratio description and the stable convergence state description from the load-bearing capacity boundary description, and use the load-bearing capacity remaining ratio description as a dimensionless remaining spatial ratio description that can be quantified and compared.

[0119] The boundary structure resolution operation is achieved by reading the various fields in the composite data structure describing the load-bearing capacity boundary. The value stored in the first field is read to obtain the remaining load-bearing ratio description. The coordinate array stored in the second field is read to obtain the component values ​​of the spatial coordinate points representing the stable convergence state on each coordinate axis. The value of the remaining load-bearing ratio description is then assigned to the dimensionless remaining spatial ratio description variable.

[0120] Step S520: Perform a policy correspondence search between the dimensionless remaining space ratio description and the preset traffic scheduling hierarchical policy set. In the traffic scheduling hierarchical policy set, find the scheduling policy template mark that matches the current value range of the dimensionless remaining space ratio description, and extract the diversion weight adjustment indication and diversion target route indication associated with the scheduling policy template mark.

[0121] As one embodiment, step S520 may specifically include the following steps S521 to S526: Step S521: Determine the numerical range of the dimensionless residual space ratio description. Compare the ratio value represented by the dimensionless residual space ratio description with multiple preset ratio range division thresholds one by one to determine the target ratio range identifier into which the ratio value falls.

[0122] The preset thresholds for dividing multiple ratio intervals include a first interval threshold, a second interval threshold, and a third interval threshold. The first interval threshold is less than the second interval threshold, and the second interval threshold is less than the third interval threshold. The ratio intervals are divided into four: the interval where the ratio value is less than the first interval threshold is designated as the first ratio interval; the interval where the ratio value is greater than or equal to the first interval threshold and less than the second interval threshold is designated as the second ratio interval; the interval where the ratio value is greater than or equal to the second interval threshold and less than the third interval threshold is designated as the third ratio interval; and the interval where the ratio value is greater than or equal to the third interval threshold is designated as the fourth ratio interval. The value described by the dimensionless residual space ratio is compared sequentially with the first interval threshold, the second interval threshold, and the third interval threshold to determine the ratio interval to which the value falls and obtain the corresponding ratio interval identifier.

[0123] Step S522: Use the target ratio range identifier as the retrieval key value to perform an index lookup operation in the traffic scheduling hierarchical strategy set. The traffic scheduling hierarchical strategy set stores the corresponding scheduling strategy template tag with the ratio range identifier as the index.

[0124] The traffic scheduling tiered strategy set is a key-value pair mapping storage structure. The key field in this storage structure stores the ratio range identifier, and the value field stores the scheduling strategy template tag. Using the target ratio range identifier determined in step S521 as the query key, a key-matching search is performed in the key-value pair mapping storage structure to retrieve the value field content corresponding to the query key.

[0125] Step S523: Extract the scheduling policy template tag that uniquely corresponds to the target ratio interval identifier from the set of traffic scheduling hierarchical policies, and output the scheduling policy template tag as the policy matching result.

[0126] If the key matching search in step S522 is successful, the obtained value field content is used as the scheduling policy template marker; if the key matching search fails, a preset default scheduling policy template marker is returned. The scheduling policy template marker is output as the policy matching result.

[0127] Step S524: Access the preset policy template content storage area according to the policy matching result, and obtain the traffic weight adjustment indication that is bound to the scheduling policy template mark from the policy template content storage area. The traffic weight adjustment indication is used to describe the direction and magnitude of the change in the traffic allocation ratio between different traffic paths.

[0128] The policy template content storage area is a structured data storage space that stores the detailed content definitions of multiple scheduling policy templates. Each scheduling policy template includes a template tag field, a traffic splitting weight adjustment indicator field, and a traffic splitting target route indicator field. The content of the traffic splitting weight adjustment indicator field is an adjustment operation description structure, which includes an operation type identifier and an operation magnitude identifier. The operation type identifier is either a weight increase or a weight decrease, and the operation magnitude identifier is either a small adjustment, a medium adjustment, or a large adjustment.

[0129] Step S525: Obtain the traffic splitting target route indication stored in the policy template content storage area that is bound to the scheduling policy template tag. The traffic splitting target route indication is used to describe the target service node cluster identifier that receives the traffic split.

[0130] The destination routing indication field contains a route description structure, which includes a list of cluster identifiers for the target service node clusters. This list of target service node cluster identifiers lists the unique identifiers of each service node cluster currently capable of handling the offloaded traffic.

[0131] Step S526: Return the traffic splitting weight adjustment instruction and the traffic splitting target route instruction as policy extraction results to the traffic scheduling guidance instruction generation process.

[0132] The traffic splitting weight adjustment instruction obtained in step S524 and the traffic splitting target route instruction obtained in step S525 are packaged into a policy extraction result data structure, and this policy extraction result data structure is passed to the subsequent instruction assembly stage.

[0133] Step S530: Based on the spatial coordinates of the stable convergence state, trace the reverse impact source in the pre-constructed ad placement and load impact correlation map, determine the set of ad request source markers that contribute most significantly to the current load capacity boundary, and combine the set of ad request source markers into the main impact source ad placement identifier set.

[0134] As one embodiment, step S530 may specifically include the following steps S531 to S536: Step S531: Map the spatial coordinate points of the stable convergence state representation to the influence distribution space defined by the correlation map of ad placement and load, and determine the precise mapping position of the spatial coordinate points of the stable convergence state representation in the influence distribution space.

[0135] The ad placement and load impact correlation graph is a pre-constructed high-dimensional vector space model. Each dimension of this model corresponds to an ad placement identifier, and each point in the space represents a load impact distribution state. When constructing the ad placement and load impact correlation graph, within a historical operating cycle, using a preset time window as the granularity, the proportion of ad request traffic generated by each ad placement within that time window to the total ad request traffic is statistically analyzed, and this proportion vector is used as a coordinate point in the impact distribution space. This statistical process is repeated, and the set of points formed by the proportion vectors corresponding to all time windows is used as the ad placement and load impact correlation graph. The spatial coordinate points representing the stable convergent state are mapped to the impact distribution space of the ad placement and load impact correlation graph using a preset spatial mapping transformation matrix. This spatial mapping transformation matrix is ​​determined during the initialization phase of the load bearing situation inference processor through typical point calibration and interpolation fitting methods. The mapped coordinate points are the precise mapping positions of the stable convergent state representation within the impact distribution space.

[0136] Step S532: Within the influence distribution space, define a reverse influence tracing range with a preset radius centered on the precise mapping location, and extract all ad placement influence vectors distributed within the reverse influence tracing range. Each ad placement influence vector has a specific vector direction and vector magnitude.

[0137] The preset radius is determined in advance based on the scale of the influence distribution space. Centered on the precise mapping location, and with the preset radius as the radius length, a hyperspherical region is defined within the influence distribution space as the reverse influence tracing range. The ad placement influence vector refers to the direction vector pointing from the reference coordinate point corresponding to each pre-marked ad placement identifier in the influence distribution space to the precise mapping location. The vector direction of each ad placement influence vector is from its corresponding reference coordinate point to the precise mapping location, and the vector magnitude is determined by the Euclidean distance between the reference coordinate point and the precise mapping location.

[0138] Step S533: Calculate the projection component value of the influence vector of each ad position in the direction pointing to the precise mapping position, and extract the positive part of the projection component value as the positive single influence contribution of the ad position to the current load capacity boundary.

[0139] For each ad slot identifier in the ad slot and load impact correlation graph, obtain its corresponding reference coordinate point and calculate the unit direction vector pointing from the reference coordinate point to the precise mapping position. Perform a vector dot product operation between the ad slot impact vector and the unit direction vector; the result of the vector dot product is the projection component value. If the projection component value is positive, it is taken as the positive single impact contribution of the ad slot to the current load capacity boundary; if the projection component value is negative or zero, the positive single impact contribution of the ad slot is zero.

[0140] Step S534: Sort the positive single influence contribution of all ad slots in descending order, and extract the ad request source tags of the ad slots that are ranked before the preset ranking threshold after sorting.

[0141] The preset ranking threshold is an integer value, representing the maximum number of ad slots to be extracted. All ad slots are sorted by their positive single-influence contribution in descending order of value. Starting from the first position in the sorting result, the preset ranking threshold number of ad slots are extracted sequentially, and their corresponding ad slot identifiers are extracted.

[0142] Step S535: Collect the extracted ad request source tags into an ad placement tag set, and perform deduplication and merging operations on the elements in the ad placement tag set to form the main impact source ad placement tag set.

[0143] All ad placement identifiers extracted in step S534 are placed into a set data structure, which automatically performs element deduplication. The deduplicated ad placement identifiers constitute the set of ad placement identifiers from the main impact sources.

[0144] Step S536: Output the set of main impact source ad slot identifiers to the assembly stage of the traffic scheduling guidance instruction, so as to form a complete scheduling instruction content together with the diversion weight adjustment instruction and the diversion target route instruction.

[0145] Step S540: Combine the traffic splitting weight adjustment instruction, the traffic splitting target routing instruction, and the set of main impact source ad slot identifiers into the core load part of the traffic scheduling guidance instruction, and add the instruction effective time limit flag and the instruction execution priority level flag to the core load part of the instruction to form a complete instruction description structure.

[0146] The instruction's effective timeout is marked as a timestamp value, indicating the expiration date of the traffic scheduling guidance instruction. The instruction execution priority level is marked as an enumerated value, taking the values ​​of high priority, medium priority, or low priority. The traffic splitting weight adjustment instruction, traffic splitting target route instruction, and the set of main impact source ad slot identifiers are used as three sub-fields of the instruction's core load portion. These, along with the instruction effective timeout mark field and the instruction execution priority level mark field, are assembled according to a preset instruction description structure format to generate a complete instruction description structure.

[0147] Step S550: Encode the complete instruction description structure into a scheduling message format that conforms to the communication interface agreement of the ad request distribution component, and generate a traffic scheduling guidance instruction message that can be directly parsed by the ad request distribution component.

[0148] The advertising request distribution component's communication interface predefines a structured message encoding format, which employs a nested key-value pair structure. During the encoding conversion process, each field in the instruction description structure is converted into a corresponding key-value pair entry according to the field mapping relationship agreed upon in the communication interface. For the traffic distribution weight adjustment indication field in the core load part of the instruction, its operation type identifier and operation magnitude identifier are encoded as independent sub-key-value pairs; for the traffic distribution target routing indication field, its target service node cluster identifier list is encoded as an array-type key-value pair; for the set of main impact source advertising slot identifiers, it is encoded as an array-type key-value pair. All key-value pairs are organized according to the root structure agreed upon in the communication interface to generate a traffic scheduling guidance instruction message conforming to the scheduling message format.

[0149] Step S560: Push the traffic scheduling guidance instruction message to the instruction receiving port of the ad request distribution component to initiate the traffic weight adjustment and route redirection operation of the ad request flow corresponding to the ad position identifier set of the main impact source within the ad request distribution component.

[0150] The ad request distribution component has an open command receiving port, which listens to a preset network communication address and port number to receive control command messages from external systems. Step S560 sends the traffic scheduling guidance command message generated in step S550 to the command receiving port of the ad request distribution component via a preset network communication protocol. Upon receiving the traffic scheduling guidance command message, the ad request distribution component parses the message body, extracting the traffic splitting weight adjustment instruction, the traffic splitting target route instruction, and the set of ad slot identifiers from the main impact sources. The request routing module within the ad request distribution component adjusts the traffic splitting weight allocation ratio of the ad request flows corresponding to each ad slot identifier in the set of ad slot identifiers from the main impact sources across different target service node clusters based on the operation type and operation magnitude identifiers in the traffic splitting weight adjustment instruction. Simultaneously, the ad request distribution component updates the destination service node address of the ad request flows corresponding to the set of ad slot identifiers from the main impact sources to the service node address in the target service node cluster specified by the traffic splitting target route instruction, thereby completing the request splitting adjustment operation.

[0151] Based on the foregoing embodiments, this invention provides a load monitoring device. The units and modules included in the device can be implemented by a processor in a computer device; of course, they can also be implemented by specific logic circuits. In the implementation process, the processor can be a central processing unit (CPU), a microprocessor unit (MPU), a digital signal processor (DSP), or a field programmable gate array (FPGA), etc.

[0152] Figure 2 This is a schematic diagram of the composition structure of a load monitoring device provided in an embodiment of the present invention, as shown below. Figure 2 As shown, the load monitoring device 200 includes: Data acquisition module 210 is used to acquire the advertising request stream data sequence generated within a preset monitoring period and the service node load status record sequence that has a timestamp correspondence with the advertising request stream data sequence; The feature extraction module 220 is used to extract the request traffic situation from the advertising request stream data sequence to obtain the temporal distribution feature representation of the advertising request stream data sequence, and to extract the load change situation from the service node load status record sequence to obtain the load change feature representation of the service node load status record sequence. The temporal distribution feature representation and the load change feature representation are input into the traffic load joint situation analysis stage. The feature mapping module 230 is used to call the pre-configured traffic load collaborative analysis network to perform collaborative interactive mapping of the time-series distribution feature representation and the load change feature representation, and generate a correlation response characteristic representation that represents the dynamic correlation between advertising request traffic and load status. The situation simulation module 240 is used to perform load bearing situation simulation on the characterization of associated response characteristics, and obtain the load bearing capacity boundary description of the service node under the continuous impact of the advertising request stream data sequence. The instruction distribution module 250 is used to generate traffic scheduling guidance instructions for advertising information services based on the load capacity boundary description, and push the traffic scheduling guidance instructions to the advertising request distribution component to trigger request diversion adjustment operations.

[0153] The descriptions of the apparatus embodiments above are similar to those of the method embodiments above, and have similar beneficial effects. In some embodiments, the functions or modules included in the apparatus provided by the present invention can be used to perform the methods described in the method embodiments above. For technical details not disclosed in the apparatus embodiments of the present invention, please refer to the descriptions of the method embodiments of the present invention for understanding.

[0154] It should be noted that, in the embodiments of the present invention, if the above-described traffic load monitoring method for advertising information services is implemented as a software functional module and sold or used as an independent product, it can also be stored in a computer-readable storage medium. Based on this understanding, the technical solution of the embodiments of the present invention, or the part that contributes to related technologies, can be embodied in the form of a software product. This software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the methods described in the various embodiments of the present invention. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), magnetic disks, or optical disks. Thus, the embodiments of the present invention are not limited to any specific hardware, software, or firmware, or any combination of hardware, software, and firmware.

[0155] Figure 3 A hardware entity diagram of a computer system provided as an embodiment of the present invention, such as... Figure 3 As shown, the hardware entity of the computer system 1000 includes a processor 1001 and a memory 1002, wherein the memory 1002 stores a computer program that can run on the processor 1001, and the processor 1001 executes the program to implement the steps in the method of any of the above embodiments.

[0156] The memory 1002 stores computer programs that can run on the processor. The memory 1002 is configured to store instructions and applications that can be executed by the processor 1001. It can also cache data to be processed or already processed (e.g., image data, audio data, voice communication data, and video communication data) of the processor 1001 and various modules in the computer system 1000. It can be implemented by flash memory or random access memory (RAM).

[0157] When processor 1001 executes a program, it implements the steps of the traffic load monitoring method for advertising information services described above. Processor 1001 typically controls the overall operation of computer system 1000.

[0158] This invention provides a computer storage medium storing one or more programs that can be executed by one or more processors to implement the steps of the traffic load monitoring method for advertising information services as described in any of the above embodiments.

[0159] The above description is merely an embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any changes or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in the present invention should be included within the scope of protection of the present invention.

Claims

1. A method for monitoring traffic load in advertising information services, characterized in that, The method includes: Acquire the sequence of advertising request stream data generated within a preset monitoring period and the sequence of service node load status records that correspond to the timestamps of the advertising request stream data sequence. The request traffic situation is extracted from the advertising request stream data sequence to obtain the temporal distribution feature representation of the advertising request stream data sequence. The load change situation is extracted from the service node load status record sequence to obtain the load change feature representation of the service node load status record sequence. The temporal distribution feature representation and the load change feature representation are input into the traffic load joint situation analysis stage. The pre-configured traffic load collaborative analysis network is invoked to perform collaborative interactive mapping between the time-series distribution feature representation and the load change feature representation, generating a correlation response characteristic representation that characterizes the dynamic relationship between advertising request traffic and load status. By performing load-bearing situation simulation on the aforementioned associated response characteristics, a description of the service node's load-bearing capacity boundary under the continuous impact of the advertising request stream data sequence is obtained. Based on the load capacity boundary description, a traffic scheduling guidance instruction for advertising information services is generated, and the traffic scheduling guidance instruction is pushed to the advertising request distribution component to trigger request diversion adjustment operation.

2. The method according to claim 1, characterized in that, The process of extracting request traffic patterns from the ad request stream data sequence to obtain a temporal distribution feature representation of the ad request stream data sequence, and extracting load change patterns from the service node load status record sequence to obtain a load change feature representation of the service node load status record sequence, and inputting the temporal distribution feature representation and the load change feature representation into the traffic load joint situational analysis stage, includes: The ad request stream data sequence is decoupled from the request source. Based on the ad slot identifier carried in each ad request record, the ad request stream data sequence is split into multiple single-source request subsequences. Each single-source request subsequence corresponds to an independent ad slot identifier and contains all request arrival time records of that ad slot within the preset monitoring period. For each single-source request subsequence, request arrival rhythm extraction is performed, the time interval distribution between adjacent request arrival times in the single-source request subsequence is analyzed, and the sparse-dense alternation pattern presented in the time interval distribution pattern is transformed into a request arrival rhythm descriptor corresponding to the ad slot. The request arrival rhythm descriptor is used to characterize the periodic aggregation tendency and intermittent sparsity tendency of the request traffic of the ad slot. The request arrival rhythm descriptors corresponding to all ad slots are input into a preset ad slot request propagation topology. The ad slot request propagation topology is constructed with ad slot identifiers as nodes and business traffic associations between ad slots as directed edges. In the ad slot request propagation topology, the request arrival rhythm descriptors of each node are cascaded and propagated according to the propagation direction of the directed edges to generate a request impact propagation timing description for each ad slot node within the preset monitoring period. Extract the request convergence node set for the service node from the request impact propagation timing description. The request convergence node set includes all ad slot nodes in the ad slot request propagation topology that point to the service node. Synchronously overlay the request impact propagation timing description of each node in the request convergence node set to generate a convergence request traffic timing profile description for the service node. Perform traffic impact peak and valley morphology analysis on the converged request traffic time series profile description to identify the impact accumulation segment with continuously rising traffic amplitude, the impact release segment with continuously decreasing traffic amplitude, and the impact plateau segment with stable traffic amplitude in the converged request traffic time series profile description. The accumulation steepness description of the impact accumulation segment and the release steepness description of the impact release segment are used as the time series distribution feature representation of the advertising request stream data sequence. The load state transition chain reconstruction is performed on the service node load state record sequence. The adjacent load level marker transformation events in the service node load state record sequence are connected in the order of transformation to form a directed load state transition link. Transition sub-links whose occurrence frequency meets the preset frequency condition are extracted from the directed load state transition link as the load change feature representation of the service node load state record sequence. The time-series distribution feature representation and the load change feature representation are jointly transmitted to the input of the traffic load joint situation analysis stage.

3. The method according to claim 2, characterized in that, The step involves inputting the request arrival rhythm descriptors corresponding to all ad slots into a preset ad slot request propagation topology. This topology is constructed using ad slot identifiers as nodes and directed edges representing business traffic associations between ad slots. Within this topology, the request arrival rhythm descriptors of each node are cascaded and propagated according to the direction of the directed edges, generating a request impact propagation timing description for each ad slot node within the preset monitoring period. This includes: Read the business traffic association records between ad slots from the preset ad slot association relationship storage area, and construct the ad slot request propagation topology structure with ad slot identifier as node and the traffic direction as directed edge pointing to the target ad slot identifier according to the traffic direction of the source ad slot identifier in the business traffic association record. Each directed edge in the ad slot request propagation topology structure is attached with a traffic intensity attribute tag. The request arrival rhythm descriptor corresponding to each ad slot is loaded onto the node in the ad slot request propagation topology that matches the ad slot identifier, as the initial rhythm state description of the node. The initial rhythm state description includes the aggregation period tendency description and the sparse period tendency description of the ad slot request traffic. For each node in the ad slot request propagation topology, backtrack along all incoming edge directions of the node to obtain the request arrival rhythm descriptor of the upstream neighboring node, and perform propagation fusion of the request arrival rhythm descriptor of the upstream neighboring node with the traffic intensity attribute mark on the corresponding incoming edge to generate an incremental description of the request impact propagated from upstream to the current node. The incremental description of the impact of the upstream request transmitted to the current node is superimposed with the initial rhythm state description of the node itself to obtain the updated rhythm state description of the node in the current transmission round. The rhythm state superposition process makes the rhythm state description of the node reflect both its original request characteristics and the impact transmission characteristics of the upstream request. In the advertising space request propagation topology, the propagation round iteration along the directed edge direction is repeatedly executed. After each round of propagation round iteration is completed, the updated rhythm state description of each node is used as the input state for the next round of propagation round iteration, until the preset propagation convergence round limit is reached, and the final rhythm state description of each node after the propagation convergence is obtained. The final rhythmic state description of each ad placement node after the completion of propagation convergence is expanded into a request impact intensity indication on each time segment according to the time axis, and the request impact intensity indication is arranged in chronological order to form the request impact propagation time sequence description of each ad placement node within the preset monitoring period.

4. The method according to claim 2, characterized in that, The process involves performing traffic surge peak-valley pattern analysis on the converged request traffic time-series profile description to identify surge accumulation segments with continuously rising traffic amplitude, surge release segments with continuously decreasing traffic amplitude, and surge plateau segments with stable traffic amplitude. The descriptions of the cumulative steepness of the surge accumulation segments and the release steepness of the surge release segments are used as temporal distribution features of the advertising request stream data sequence, including: The convergence request traffic time-series profile description is converted into a traffic amplitude sequence description arranged continuously along the time axis, where each sequence position in the traffic amplitude sequence description corresponds to the convergence request traffic amplitude in a time segment; The flow rate change direction is determined for the flow rate amplitude sequence description. The change direction markers between adjacent sequence positions in the flow rate amplitude sequence description are calculated one by one. The change direction markers include amplitude increasing markers, amplitude decreasing markers, and amplitude stabilizing markers. The change direction markers are arranged in chronological order to form a flow rate change direction marker sequence. The sequence of flow change direction markers is merged into a continuous same-direction marker merging process. Multiple markers with the same change direction and consecutively adjacent positions in the sequence of flow change direction markers are merged into an independent flow pattern segment. The segment is then divided into an impact accumulation segment, an impact release segment, or an impact platform segment based on the type of change direction markers contained in the segment. For each flow pattern segment divided into impact accumulation segments, the flow amplitude corresponding to the start time segment and the flow amplitude corresponding to the end time segment of the segment are extracted, and the accumulation steepness of the impact accumulation segment is determined based on the number of time segments spanned by the segment and the amplitude difference between the start flow amplitude and the end flow amplitude. For each flow pattern segment divided into impact release segments, the flow amplitude corresponding to the start time segment and the flow amplitude corresponding to the end time segment of the segment are extracted, and the release steepness of the impact release segment is determined based on the number of time segments spanned by the segment and the amplitude difference between the start flow amplitude and the end flow amplitude. The descriptions of the cumulative steepness of the impact accumulation segment and the descriptions of the release steepness of the impact release segment are arranged according to the order of their respective segments on the time axis, forming a morphological description sequence consisting of alternating descriptions of cumulative steepness and release steepness. This morphological description sequence is then used as a temporal distribution feature representation of the advertising request stream data sequence.

5. The method according to claim 1, characterized in that, The invocation of the pre-configured traffic load collaborative analysis network performs collaborative interactive mapping between the time-series distribution feature representation and the load change feature representation, generating a correlation response characteristic representation that characterizes the dynamic relationship between ad request traffic and load status, including: The time-series distribution feature representation and the load change feature representation are simultaneously loaded into the input access layer of the traffic load collaborative analysis network. The input access layer performs structure normalization and dimension alignment operations on the time-series distribution feature representation and the load change feature representation to generate a standardized input feature pair representation with a consistent representation form. The bidirectional cross-sensing processing sublayer within the traffic load collaborative analysis network performs forward information permeation on the standardized input feature representation, calculates the attention response distribution of each time point in the time-series distribution feature representation to each load state point in the load change feature representation, and generates a traffic-to-load attention mapping representation describing the direction of traffic action on the load based on the attention response distribution. The bidirectional cross-sensing processing sublayer performs reverse information permeation on the standardized input feature representation, calculates the attention response distribution of each load state point in the load change feature representation to each time position point in the time-series distribution feature representation, and generates a load-to-flow attention mapping representation describing the direction of load's reaction to flow based on the attention response distribution. The traffic-to-load attention mapping representation and the load-to-traffic attention mapping representation are input into the interactive integration processing layer of the traffic-load collaborative analysis network, and a bidirectional information fusion operation based on the selection gating mechanism is performed to generate a set of collaborative interactive feature representations that simultaneously contain forward action descriptions and reverse action descriptions. A time-series following tightness analysis is performed on the collaborative interaction feature representation set to extract the response following time offset between the traffic-side time position and the load-side time position, and the response following time offset is converted into a dimensionless response following hysteresis ratio description. The decay trend of the interaction response intensity with relative time offset is extracted and converted into a dimensionless response following intensity decay ratio description. The response following hysteresis ratio description and the response following intensity decay ratio description are used together as the characterization of the associated response characteristics.

6. The method according to claim 5, characterized in that, The process involves forward information permeation of the standardized input feature representation through a bidirectional cross-sensing processing sublayer within the traffic load collaborative analysis network, calculating the attention response distribution of each time point in the time-series distribution feature representation to each load state point in the load change feature representation, and generating a traffic-to-load attention mapping representation describing the direction of traffic's effect on the load based on the attention response distribution. This includes: From the standardized input feature pair description, a traffic query vector sequence description corresponding to the time-series distribution feature representation is extracted, and from the standardized input feature pair description, a load query vector sequence description corresponding to the load change feature representation is extracted. Each vector element in the traffic query vector sequence description corresponds to a time position in the request traffic fluctuation pattern sequence, and each vector element in the load query vector sequence description corresponds to a load state point in the load state transition event sequence description. For each traffic query vector in the traffic query vector sequence description and each load query vector in the load query vector sequence description, a scaled dot product-based attention measurement is performed to generate a two-dimensional attention description table. The value of the entry at the intersection of a certain row and a certain column in the two-dimensional attention description table represents the attention response intensity of the corresponding traffic time position to the corresponding load state point. The two-dimensional attention level description table is normalized by row distribution so that the sum of the attention response intensity of the same traffic time position represented by each row of the two-dimensional attention level description table to all load state points satisfies the preset normalization constraint condition, thus obtaining the normalized attention level description table. Based on the distribution of attention response intensity in each row of the normalized attention level description table, the load status point identifier corresponding to the column with the highest attention response intensity in each row is extracted, and a one-way attention pointing relationship description is constructed from the traffic time position corresponding to that row to the load status point identifier. The descriptions of the unidirectional attention-direction relationships corresponding to all traffic time locations are arranged in chronological order of traffic time locations, and the arrangement results are depicted on a two-dimensional plane as the trajectory of the focus of attention migrating over time. This trajectory is represented as the traffic-direction-load attention mapping diagram.

7. The method according to claim 5, characterized in that, The step of performing time-series following tightness analysis on the collaborative interaction feature representation set, extracting the response following time offset between the traffic-side time position and the load-side time position, and converting the response following time offset into a dimensionless response following hysteresis ratio description, and extracting the decay trend of the interaction response intensity with relative time offset and converting it into a dimensionless response following intensity decay ratio description, includes: In the set of collaborative interaction feature representations, a pair of collaborative interaction feature representations whose interaction response intensity reaches a local maximum is located. The pair of collaborative interaction feature representations consists of a flow-side time location feature representation element and a load-side time location feature representation element, and the interaction response intensity between the two satisfies the local peak determination condition. Extract the timestamp indication information corresponding to the time position feature representation element on the flow side, and extract the timestamp indication information corresponding to the time position feature representation element on the load side. Calculate the time difference between the timestamp indication information of the time position feature representation element on the flow side and the timestamp indication information of the time position feature representation element on the load side. Divide the absolute time length of the time difference by the absolute time length of the preset reference response hysteresis reference duration, and use the resulting quotient as the dimensionless response following hysteresis ratio description. Using the pair of cooperative interaction feature representations with local maximum interaction response intensity as the central reference position, a time analysis window of a preset length is defined along the forward and backward extension directions of the time axis, and the decrease in interaction response intensity between adjacent cooperative interaction feature representation pairs is extracted sequentially within the time analysis window. Based on the extracted decrease in interaction response intensity, a decay trend depiction of interaction response intensity as a function of relative time offset is constructed, and the relative time offset corresponding to when the interaction response intensity decreases to a preset decay reference boundary is determined from the decay trend depiction. Divide the absolute time length of the relative time offset by the absolute time length of the preset reference intensity decay duration, and use the resulting quotient as the dimensionless response following intensity decay ratio.

8. The method according to claim 1, characterized in that, The process of performing load-bearing situation deduction on the associated response characteristics to obtain a description of the service node's load-bearing capacity boundary under the continuous impact of the advertising request stream data sequence includes: The response following hysteresis ratio description and the response following intensity decay ratio description are parsed from the associated response characteristic characterization, and the response following hysteresis ratio description and the response following intensity decay ratio description are simultaneously transmitted to the initial state configuration interface of the preset load bearing situation inference processor. The load bearing situation simulation processor generates an initial state vector representation for load bearing situation analysis based on the response following hysteresis ratio description and the response following intensity decay ratio description. The initial state vector representation corresponds to an initial state coordinate point in the state simulation space of the load bearing situation simulation processor. Within the state deduction space of the load bearing situation deduction processor, the initial state vector is used as the deduction starting point, the response following hysteresis ratio is used as the deduction step size adjustment basis, and the response following intensity decay ratio is used as the deduction direction deflection basis to perform multi-step continuous state deduction, generating a deduction path sequence composed of multiple intermediate state vectors. The state convergence determination is performed on the inference path sequence, and the state difference between two adjacent intermediate state vector representations in the inference path sequence is monitored. When the state difference decreases to below the preset convergence stability reference boundary, the currently reached intermediate state vector representation is marked as a stable convergence state representation. Within the state deduction space, obtain the spatial coordinate points occupied by the stable convergence state representation, calculate the shortest spatial geometric interval between the spatial coordinate points and the preset load-bearing boundary hypersurface, and simultaneously obtain the preset load-bearing boundary reference scale distance. Divide the shortest spatial geometric interval distance by the load-bearing boundary reference scale distance, and use the resulting quotient as a dimensionless load-bearing remaining ratio characterizing the remaining bearing capacity. The spatial coordinate points of the load-bearing capacity boundary are combined with the description of the remaining load-bearing ratio and the description of the stable convergence state.

9. The method according to claim 8, characterized in that, Within the state deduction space of the load bearing situation deduction processor, multi-step continuous state deduction is performed, using the initial state vector representation as the deduction starting point, the response following hysteresis ratio as the basis for adjusting the deduction step size, and the response following intensity decay ratio as the basis for deduction direction deflection. This generates a deduction path sequence composed of multiple intermediate state vector representations, including: An initial state coordinate point corresponding to the initial state vector representation is established in the state deduction space of the load bearing situation deduction processor. The coordinate values ​​of each dimension of the initial state coordinate point are determined by the projection of each dimension component of the initial state vector representation into the coordinate system of the state deduction space. The corresponding step size adjustment coefficient is determined based on the response following hysteresis ratio description. There is a pre-set monotonically decreasing mapping relationship between the step size adjustment coefficient and the hysteresis degree represented by the response following hysteresis ratio description. The preset basic inference step size is multiplied and scaled with the step size adjustment coefficient to obtain the actual inference step size applicable to the current inference process. The response following intensity decay ratio is used as a reference value for directional deflection adjustment. The preset basic inference direction vector is mapped to the directional deflection adjustment reference value by a deflection angle. The actual inference direction vector applicable to the current inference process is calculated. The deflection angle of the actual inference direction vector relative to the basic inference direction vector is determined by the response following intensity decay ratio. Starting from the initial state coordinate point, move the actual deduction step by the distance along the actual deduction direction vector to reach the first intermediate state coordinate point, and vectorize the first intermediate state coordinate point as the first intermediate state vector representation. The first intermediate state vector representation is used as the new starting point for deduction. The deduction step is repeated by moving the actual deduction step by the actual deduction step size along the actual deduction direction vector until the number of deduction steps reaches the preset upper limit of the number of deduction steps. The intermediate state coordinate points obtained in each deduction are vectorized and all the vectorized representations are arranged in the order of generation to form the deduction path sequence.

10. A computer system comprising a memory and a processor, the memory storing a computer program executable on the processor, characterized in that, When the processor executes the program, it implements the steps of the method according to any one of claims 1 to 9.

11. A computer-readable storage medium having a computer program stored thereon, characterized in that, When executed by a processor, the computer program implements the steps of the method according to any one of claims 1 to 9.

Images