Intelligent regulation system for matching supply and demand of public services based on social perception data

Abstract

The application discloses a public service supply and demand matching intelligent regulation and control system based on social perception data, relates to the technical field of public service supply and demand matching, and comprises a data construction module, a service mapping module, a disturbance identification module, an abnormality judgment module, a strategy execution module and a regulation and control feedback module; the disturbance identification module performs time series clustering analysis on a data record set, identifies a data section affected by external non-service events, and labels the data section as a disturbed social perception data section; the abnormality judgment module acquires associated behavior characteristic information of each disturbed social perception data section, evaluates the data abnormality degree of each disturbed social perception data section by constructing an abnormality degree generation model after acquisition, and classifies according to the evaluation result. The application solves the problem of resource misadjustment caused by external service event interference in social perception data, and realizes accurate identification and dynamic hierarchical adjustment of public service regulation and control response.

Description

Technical Field

[0001] This invention relates to the field of public service supply and demand matching technology, and specifically to an intelligent regulation system for public service supply and demand matching based on social perception data. Background Technology

[0002] Intelligent regulation of public service supply and demand based on social perception data refers to the dynamic perception of public demand changes (such as transportation, healthcare, and education) across different times, locations, and contexts by collecting and analyzing multi-source social perception data from social platforms, mobile devices, location services, and user feedback. This is combined with existing service supply capacity and artificial intelligence algorithms to achieve intelligent matching and dynamic regulation of public service resources, thereby improving the accuracy and responsiveness of public services. This regulation method breaks through the traditional allocation mechanism relying on static statistics and human experience. It can identify demand hotspots and supply-demand imbalances in real time under complex environments such as emergencies, peak periods, and spatial heterogeneity, and quickly generate executable allocation strategies. The reason for using social perception data for intelligent regulation is that in modern society, information spreads rapidly and population mobility is high. Relying solely on historical data or fixed rules is insufficient to accurately reflect the public's immediate needs. Social perception data, with its broad scope, real-time nature, and contextual sensitivity, can provide more detailed profiles of public needs and dynamic trend analysis, thus providing solid data support for the scientific allocation of public services and achieving optimal resource allocation and maximization of public interests.

[0003] Existing intelligent regulation and control technologies for public service supply and demand matching based on social perception data typically achieve dynamic optimization of public service resource allocation by constructing a closed-loop system that includes multiple links such as data collection, perception analysis, demand forecasting, supply and demand matching, and intelligent scheduling. First, in the data collection phase, the system collects social perception data related to public behavior, location, emotions, and demands in real time from multiple sources, including social media, mobile applications, location systems, smart sensors, and user interaction platforms. Next, in the perception analysis phase, technologies such as natural language processing, sentiment analysis, and clustering modeling are used to identify users' implicit needs and contextual characteristics, and to construct user or regional profiles. Then, in the demand forecasting phase, the system uses historical and real-time data, along with time series analysis and deep learning methods, to predict the intensity and trend of service demand for specific regions or groups over a future period. Next, in the supply and demand matching phase, the predicted demand is intelligently compared with current service resource data, taking into account factors such as resource distribution, service capacity, accessibility, and user priority, and a preliminary allocation plan is generated through a matching algorithm. Finally, in the intelligent scheduling phase, mechanisms such as reinforcement learning, multi-objective optimization, or rule engines are used to dynamically adjust the plan, ensuring optimal matching and rapid response under conditions of limited resources and complex scenarios, thereby achieving intelligent, precise, and efficient services. This entire process works in tandem, enabling the system to have a closed-loop capability from "perception-understanding-prediction-decision-control," greatly enhancing the adaptability and responsiveness of the public service system to changes in the complex social environment.

[0004] The existing technology has the following shortcomings:

[0005] When a social media event, misinformation spreads rapidly online, or malicious hype spreads quickly, users' attention and behavior tend to concentrate on a particular type of public service (such as a sudden surge in searches for hospital appointments, traffic control, or government services), resulting in a significant peak in the social perception data. Because the system assumes user behavior represents genuine service demand, this data is misjudged as an extreme imbalance between regional service supply and demand, triggering the control system to immediately implement control strategies, including resource allocation and service redirection. However, since this behavioral peak is not actually caused by changes in the load of service locations but is driven by non-real events outside the service, the system, without a "behavior-service causal confidence" identification mechanism, cannot distinguish the degree of deviation between such abnormal behavior and genuine service demand, thus mistakenly inputting false perception signals into the control model. Existing intelligent regulation technologies for matching the supply and demand of public services based on social perception data cannot dynamically adjust the intensity of the supply and demand matching response based on the degree of data anomalies caused by external non-service events. This results in large-scale misallocation of system resources to service areas with no actual pressure, while service areas with real demand are unable to maintain normal operation due to resource diversion. This leads to a series of negative consequences, such as regulation failure, imbalance of urban service systems, and decreased public trust in intelligent regulation mechanisms.

[0006] The information disclosed in the background section is only intended to enhance the understanding of the background of this disclosure, and therefore may include information that does not constitute prior art known to those skilled in the art. Summary of the Invention

[0007] The purpose of this invention is to provide an intelligent regulation system for matching the supply and demand of public services based on social perception data, so as to solve the problems mentioned in the background art.

[0008] To achieve the above objectives, the present invention provides the following technical solution: a public service supply and demand matching intelligent regulation system based on social perception data, including a data construction module, a service mapping module, a disturbance identification module, an anomaly judgment module, a strategy execution module, and a regulation feedback module;

[0009] The data construction module collects social perception data from multiple sources, and after collection, it performs data cleaning, format unification and structuring to form a structured data record sequence;

[0010] The service mapping module establishes a mapping relationship between the structured data record sequence and public service types. It classifies each data record through the association rules between behavior tags and service intentions, and generates a set of data records containing service category identifiers.

[0011] The disturbance identification module performs time-series clustering analysis on the data record set to identify data segments affected by external non-service events and marks these data segments as disturbed socially perceived data segments.

[0012] The anomaly detection module acquires the associated behavioral feature information of each disturbed social perception data segment, and after acquisition, evaluates the degree of data anomaly of each disturbed social perception data segment by constructing an anomaly degree generation model, and classifies it according to the evaluation results.

[0013] The strategy execution module executes dynamic adjustment strategies for supply and demand matching response intensity corresponding to each category, based on the classification results.

[0014] The regulation and feedback module acquires data on the operation of public services after implementing dynamic regulation strategies for the intensity of each supply and demand matching response. Based on the data on the operation of public services, it updates the classification results of the disturbed social perception data segments to enable subsequent classification judgments and adjustments to regulation strategies.

[0015] Preferably, in the service mapping module, keyword information, behavior location path information, and time tag information are extracted from the structured data record sequence. Combined with a preset service semantic dictionary and public service scenario tag library, a mapping relationship between structured behavior information and public service types is established. Based on the established mapping relationship, the behavior tags in each data record are matched with the service intent according to rules. After matching, the data record is assigned a service category identifier to generate a set of data records containing service category identifiers.

[0016] Preferably, in the disturbance identification module, the data record set is divided into several data sequences arranged in chronological order according to the service category identifier. Clustering processing based on threshold division is performed on the change in the number of behaviors in each data sequence to identify data segments where the number of behaviors continuously increases and exceeds a preset growth threshold. Among them, data segments where the number of behaviors increases beyond the preset growth threshold but the path trajectory field does not contain the coordinate points in the corresponding service area coordinate set, and the keyword field in the data segment appears at a synchronously increasing frequency in external public information sources, are identified as data segments affected by external non-service events, and are marked as disturbed social perception data segments.

[0017] Preferably, in the anomaly determination module, the associated behavioral feature information of each disturbed social perception data segment is obtained, and normalized after acquisition. After normalization, a dimensionless feature vector is constructed, and an anomaly degree generation model is constructed based on the dimensionless feature vector to generate an anomaly degree evaluation value for each disturbed social perception data segment. The anomaly degree evaluation value of each disturbed social perception data segment is compared with a pre-set threshold range. The data anomaly degree of each disturbed social perception data segment is evaluated based on the comparison result, and each disturbed social perception data segment is divided into a pseudo-disturbance segment, a light disturbance segment, and a heavy disturbance segment based on the evaluation result.

[0018] Preferably, in the anomaly detection module, the associated behavioral characteristic information of each disturbed social perception data segment is obtained. The associated behavioral characteristic information includes four types of characteristic information, specifically:

[0019] The percentage of user behavior records located within the public service facility area per unit time is specifically: the ratio between the number of user behavior records whose geographical coordinates fall within the public service facility area and the total number of records within that time unit in each preset time unit.

[0020] The percentage of data records with a complete behavioral conversion path is specifically defined as the ratio between the number of data records that simultaneously include the four behavioral fields of query, click, navigation, and terminal access in the behavioral sequence and the total number of records within that time unit.

[0021] The average growth rate of the frequency of high-frequency keywords in external public information sources is specifically: the average increase in the time-series change value of the frequency of keywords extracted from the disturbed social perception data segment in external information sources within the corresponding time period.

[0022] The maximum difference between the rate of change of the number of behaviors in each behavior cluster unit after clustering is specifically: the maximum difference between the rate of change of the number of behaviors in several spatial behavior units obtained after clustering by inputting the geographic location field in the behavior record into the clustering process, within adjacent time windows.

[0023] Preferably, the associated behavioral feature information of each disturbed social perception data segment is normalized, and a dimensionless feature vector is constructed after normalization, specifically as follows:

[0024] The values ​​of each feature in the associated behavioral feature information are scaled relative to the maximum and minimum values ​​of each feature in the preset historical sample set, and the values ​​of each feature are mapped to a standardized numerical range between zero and one.

[0025] After normalizing the various feature information, the proportion of user behavior records within the scope of public service facilities, the proportion of data records with complete behavior conversion paths, the average growth rate of the frequency of high-frequency keywords in external public information sources, and the maximum difference between the behavior quantity change rate of each behavior cluster unit after clustering are combined into a dimensionless feature vector in a preset order.

[0026] Preferably, in the anomaly determination module, an anomaly degree generation model is constructed based on dimensionless feature vectors to generate anomaly degree assessment values ​​for each disturbed social perception data segment, specifically:

[0027] The dimensionless feature vector of each disturbed social perception data segment is mapped to the labeled feature vector in the historical sample set. Each sample in the historical sample set includes a set of dimensionless feature vectors and their corresponding public service response status labels.

[0028] An anomaly level generation model is constructed using supervised learning. The model consists of an input layer, a cross-feature construction layer, and a weighted fusion layer. The input layer receives a dimensionless feature vector. The cross-feature construction layer performs feature combination operations on features of each dimension to generate several interaction terms. The weighted fusion layer weights each interaction term based on the feature combination weights learned during the training phase and outputs a single evaluation value representing the degree of data anomaly.

[0029] During the training phase, a loss function is constructed by minimizing the error between the model output value and the sample label, and the model parameters are iteratively optimized using backpropagation.

[0030] After the model training is completed, the dimensionless feature vector of any newly received disturbed social perception data segment is input into the anomaly degree generation model to obtain the corresponding anomaly degree evaluation value.

[0031] Preferably, an anomaly level generation model is constructed using supervised learning. This model consists of an input layer, a cross-feature construction layer, and a weighted fusion layer, wherein:

[0032] The input layer is used to receive the dimensionless feature vector after normalization. The dimensionless feature vector includes the proportion of user behavior records within the scope of public service facilities per unit time, the proportion of data records with complete behavior conversion paths, the average growth rate of the frequency of high-frequency keywords in external public information sources, and the maximum difference between the behavior quantity change rates of each behavior cluster unit after clustering.

[0033] The cross-feature construction layer performs pairwise combinations on the feature information of each dimension in the dimensionless feature vector to form feature interaction terms, which are used to capture the mutual influence relationship between various behavioral features.

[0034] The weighted fusion layer performs a weighted summation on all feature interaction terms based on the feature combination weight parameters determined during the training phase, and outputs a single evaluation value representing the degree of data anomaly of each disturbed socially perceived data segment.

[0035] Preferably, the anomaly assessment value of each disturbed social sensing data segment is compared with a pre-set threshold range. Based on the comparison results, the anomaly degree of each disturbed social sensing data segment is assessed, and based on the assessment results, each disturbed social sensing data segment is divided into a false disturbance segment, a minor disturbance segment, and a major disturbance segment, specifically:

[0036] When the anomaly assessment value is less than the minimum value of the preset threshold range, the data anomaly level of the disturbed social perception data segment is determined to be low, and it is then classified as a false disturbance segment.

[0037] When the abnormality assessment value is greater than or equal to the minimum value of the preset threshold interval and less than or equal to the maximum value of the preset threshold interval, the abnormality of the disturbed social perception data segment is determined to be moderate, and it is classified as a slightly disturbed segment.

[0038] When the anomaly assessment value is greater than the maximum value of the preset threshold range, the data anomaly of the disturbed social perception data segment is determined to be severe, and it is then classified as a heavily disturbed segment.

[0039] Preferably, in the strategy execution module, based on the classification results, a dynamic adjustment strategy for the supply and demand matching response intensity corresponding to each category is executed, specifically as follows:

[0040] For the disturbed social perception data segments that are classified as pseudo-interference segments, the original resource allocation plan remains unchanged based on the service type corresponding to the data segment, and the behavioral characteristic patterns of the data segment are recorded in the interference behavior feature database for subsequent interference pattern comparison and identification.

[0041] For the disturbed social perception data segments that are classified as lightly disturbed segments, resource scheduling adjustment operations with limited ranges are performed on the service types corresponding to the data segments. Specifically, upper and lower limits for resource allocation are set, and within these limits, the quantity of resources, scheduling frequency, and service node coverage are proportionally adjusted. At the same time, a behavior change trend monitoring program is launched to dynamically evaluate the results of resource allocation adjustment.

[0042] For disturbed social perception data segments that are classified as heavily disturbed segments, the automatic resource scheduling process for the service type corresponding to the data segment is suspended, and a manually preset static resource allocation scheme is switched to perform fixed configuration distribution for service nodes, service time periods and service content. At the same time, the behavioral data input channel and feedback channel for this service type are suspended to prevent the current abnormal data from continuing to participate in the control decision.

[0043] The technical effects and advantages provided by the present invention in the above technical solution are as follows:

[0044] 1. This invention establishes a precise mapping relationship between structured user behavior and public service types by introducing a service mapping module and a disturbance identification module. Furthermore, based on a joint analysis mechanism of behavior quantity growth thresholds, behavior path trajectories, and keyword frequencies from external information sources, it accurately identifies disturbance behavior segments that do not originate from genuine service demands. Compared to traditional control methods that rely solely on behavior frequency for scheduling responses, this solution effectively identifies perceived noise caused by external service events, significantly reducing the probability of "false peaks" being misjudged as service pressure peaks by the system. This fundamentally avoids large-scale misallocation of control resources, ensuring the authenticity and rationality of scheduling behavior.

[0045] 2. This invention designs a four-factor behavioral characteristic index system, including dimensions such as service field behavioral density, behavioral path integrity, external keyword diffusion trend, and regional behavioral clustering degree. A dimensionless data input model is generated through normalization and feature vector construction. Based on the constructed supervised learning-based anomaly degree generation model, the system can automatically output the anomaly degree assessment value for each disturbed segment and classify it into three categories—false disturbance, minor disturbance, and severe disturbance—based on a preset threshold range. On this basis, the system executes differentiated strategies such as maintaining the original plan, dynamic scheduling, or emergency static solutions, achieving precise level-based responses to data segments with different intensities of disturbance, greatly improving the adaptability, flexibility, and stability of the control strategy.

[0046] 3. This invention specifically incorporates a control feedback module, which uses service operation data after the execution of the control strategy as feedback input to dynamically update the classification results of disturbance segments, thereby achieving system self-correction and strategy adjustment. This closed-loop feedback mechanism enables the system to not only make judgments during initial identification but also to verify and correct them based on actual effects after control execution, gradually improving the accuracy of disturbance behavior judgment and the effectiveness of strategy selection. Furthermore, the behavioral characteristics of pseudo-disturbance segments are archived into an interference behavior feature library, supporting subsequent comparison and intervention optimization. This endows the control system with long-term evolution capabilities and knowledge accumulation capabilities, improving the overall intelligence level and decision reliability. Attached Figure Description

[0047] To more clearly illustrate the technical solutions in the embodiments of this application or the prior art, the drawings used in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments recorded in this invention. For those skilled in the art, other drawings can be obtained based on these drawings.

[0048] Figure 1This is a schematic diagram of the modules of the intelligent regulation system for matching the supply and demand of public services based on social perception data according to the present invention. Detailed Implementation

[0049] Exemplary embodiments will now be described more fully with reference to the accompanying drawings. However, these exemplary embodiments can be implemented in many forms and should not be construed as limited to the examples set forth herein; rather, they are provided so that the description of this disclosure will be more complete and fully convey the concept of the exemplary embodiments to those skilled in the art.

[0050] This invention provides, for example Figure 1 The intelligent regulation and control system for matching the supply and demand of public services based on social perception data shown includes a data construction module, a service mapping module, a disturbance identification module, an anomaly judgment module, a strategy execution module, and a regulation feedback module.

[0051] The data construction module collects social perception data from multiple sources. After collection, the data is cleaned, formatted, and structured to form a structured data record sequence. The social perception data is used to characterize user behavior and service intent.

[0052] "Collecting social perception data from multiple sources" can be achieved by deploying data access strategies across various digital platforms, sensor networks, and third-party information interfaces. Specific methods include: collecting location trajectory and travel route data by calling the behavior log interfaces of public transportation platforms (such as map navigation and travel services); extracting behavioral correlation information such as keyword posting frequency, forwarding behavior, and comment structure by accessing the public information flow APIs of social media platforms; and collecting service query records, access frequency, and user operation sequences by parsing access logs from systems such as government services, medical appointments, and community platforms. Simultaneously, user-authorized data sources, such as user interaction records from government apps or city service applications, can also be introduced. All data sources are synchronized and frequency-adjusted through data collection scheduling logic configured in the software to ensure timeliness and coverage.

[0053] Raw data collected often suffers from issues such as messy formatting, redundant fields, missing values, and time errors. Therefore, software workflows can automatically perform data cleaning and structuring. Specifically, this includes: first, using null value detection and field standardization rules to uniformly map and complete fields from different data sources; second, using regular expressions to clean redundant text and illegal characters; third, using timestamp correction algorithms to align data from different sources in time sequence; and finally, introducing preset field templates (such as "user identifier-behavior type-time point-location coordinates-service intent") to parse unstructured data into standard format records, ultimately generating a unified formatted sequence of structured data records. This sequence is saved in JSON, CSV, or vector array format and used as input for subsequent analysis modules.

[0054] Collecting and structuring multi-source social perception data is a fundamental prerequisite for intelligent regulation. User behavior patterns, service intentions, and areas of interest are fragmented and heterogeneous. Without transforming these into a unified data set, effective behavioral understanding, supply and demand assessment, and regulatory response are impossible. Furthermore, the diverse sources of social perception data, without standardization, can lead to the same event appearing in multiple formats, affecting the consistency and accuracy of analysis results. Cleaning and structuring eliminate irrelevant and interfering data, improving data quality and ensuring that subsequent modules have a high-quality, quantifiable, and comparable data foundation for service mapping, disturbance identification, and anomaly classification. Therefore, this process is not only a necessary step in the technical logic but also a core prerequisite for ensuring the correct execution of regulatory strategies.

[0055] The service mapping module establishes a mapping relationship between the structured data record sequence and public service types. It classifies each data record through the association rules between behavior tags and service intentions, and generates a set of data records containing service category identifiers.

[0056] In this embodiment, in the service mapping module, keyword information, behavior location path information, and time tag information are extracted from the structured data record sequence. Combined with a preset service semantic dictionary and public service scenario tag library, a mapping relationship between structured behavior information and public service types is established. Based on the established mapping relationship, the behavior tags in each data record are matched with the service intent according to rules. After matching, the data record is assigned a service category identifier to generate a set of data records containing service category identifiers.

[0057] In the service mapping module, three core dimensions of information can be extracted from the structured data record sequence using software: first, keyword information, including text fields such as user search terms, page click terms, and tag terms; second, behavioral location path information, referring to spatial data such as the location trajectory, station coordinates, and route changes corresponding to user behavior; and third, time tag information, including the specific time point of the behavior and the duration of the behavior. The software system can combine a preset service semantic dictionary and a public service scenario tag library. The former contains words related to common public services (such as medical care, transportation, government affairs, and education) and their synonym extensions, while the latter defines the keyword patterns, behavioral path characteristics, and time density characteristics that should be matched for various service scenarios in the form of rules. In the specific implementation process, the system first performs field parsing and standardization processing on each structured data record, and then compares its keywords with the entries in the service semantic dictionary through rule matching, similarity calculation, or preset keyword matching methods. It also performs spatial cross-validation by combining the location covered by the user path with the service area in the scenario library, and finally determines whether the concentration of behavior conforms to a certain type of service characteristics in the time dimension. Records that satisfy multiple rules are assigned corresponding service category identifiers, forming a mapping relationship library to support subsequent classification and control decisions.

[0058] The reason for establishing a mapping relationship between behavioral keywords, path trajectories, and time tags with the service semantic dictionary and scenario tag library is that user behavior itself in social perception data does not directly point to specific service types, but rather manifests as various scattered and ambiguous behavioral characteristics. Without a clear correspondence mechanism between semantics, behavior, and service, the system cannot accurately determine whether a data record reflects the true intention for a particular type of public service. For example, behaviors such as "searching for nearby hospitals," "continuously visiting multiple government service pages," and "staying in a certain administrative region for an extended period" can only be transformed into a clear representation of service demand after being mapped to specific semantic tags and service types. Furthermore, service scenarios often exhibit regional concentration and temporal clustering; relying solely on a single dimension (such as keywords) may lead to numerous misjudgments. Therefore, combining the aforementioned three-dimensional information with the tag library can maximize the accuracy of service identification and provide a well-structured and semantically clear foundational data support for subsequent disturbance identification and control strategy execution. This mechanism is a crucial prerequisite for ensuring the intelligent response of the entire system.

[0059] After establishing the mapping relationship between structured behavioral data and service types, the system can perform rule matching between behavioral tags and service intents for each data record based on this mapping relationship. Specifically, the system first performs semantic parsing on the behavioral tags in the structured data records, extracting keywords, phrase structures, and context-related fields. It then matches these tags with intent categories in the service semantic dictionary using methods such as semantic classification, semantic vector similarity calculation, and keyword weight comparison. Simultaneously, it calls the corresponding spatial features and temporal patterns from the mapping table, combining them with the behavioral paths and time tags carried in the records for multi-dimensional rule cross-validation. If a record significantly corresponds to a certain type of service intent in terms of semantics, space, or time rules, the system labels the record with the corresponding service category identifier using logical matching rules. The entire process can be implemented in software by a set of rule parsing engines (based on template matching, similarity thresholds, and priority sorting). After matching, the system automatically aggregates all data records assigned service category identifiers into a standardized data record set for subsequent modules to perform time-series analysis, clustering judgment, and other operations.

[0060] The fundamental purpose of this approach is to accurately transform complex and diverse user behavior data from various sources into structured inputs that correspond one-to-one with specific public service needs, providing data support with practical business semantics for subsequent judgment and control strategies. In socially perceived data, the same behavioral label may appear in multiple contexts. Without rule matching and service semantic mapping, the system cannot identify the true intent behind these behaviors. For example, a user searching for "ID card" may intend to visit a government service hall, or it may simply be for information retrieval; only by combining their location behavior path and time tags can we infer whether they have a genuine motivation to seek government services. Therefore, performing rule matching based on mapping relationships not only improves the accuracy of service classification but also avoids wasting resources processing misclassified data in subsequent stages. More importantly, this mechanism enables the intelligent control system to execute precise responses based on "service logic," rather than blindly following the raw fluctuations of perceived data. This effectively mitigates the risk of misleading control caused by the ambiguity of user behavior and is a crucial step in ensuring the effectiveness of supply and demand judgments.

[0061] The disturbance identification module performs time-series clustering analysis on the data record set to identify data segments affected by external non-service events and marks these data segments as disturbed socially perceived data segments.

[0062] In this embodiment, in the disturbance identification module, the data record set is divided into several data sequences arranged in chronological order according to the service category identifier. Clustering processing based on threshold division is performed on the change in the number of behaviors in each data sequence to identify data segments whose number of behaviors increases continuously and exceeds a preset growth threshold. Among them, data segments whose number of behaviors increases beyond the preset growth threshold but whose path trajectory field does not contain coordinate points in the corresponding service area coordinate set, and whose keyword field appears at a synchronously increasing frequency in external public information sources, are identified as data segments affected by external non-service events, and are marked as disturbed social perception data segments.

[0063] To divide the data record set into multiple time-series data sequences, the system first groups all data records assigned service category identifiers by identifier value, with each group representing a specific public service type (such as healthcare, transportation, and government services). Within each service category group, the system reads the timestamp field from the data records and sorts them in ascending order of timestamp to form a continuous time series. Each data record retains its original fields such as timestamp, keywords, and path information. After sorting, the system assembles the records for each service category into a structured time series array for subsequent behavioral fluctuation analysis. This process is efficiently completed through the software's data indexing mechanism and sorting algorithms (such as quicksort or bucket sort based on the time field), ensuring that the data for each service category forms a logically continuous input stream on the timeline.

[0064] For each sorted service category data sequence, the system constructs a sliding window with fixed time intervals (e.g., every minute, every five minutes, etc.) to count the number of behavior records in each time segment, forming a behavior quantity change curve. Subsequently, the system sets a growth rate threshold, which is the upper limit of normal variation calculated based on the historical average behavior quantity fluctuation range and fluctuation rate of the service category. If the month-on-month growth rate of behavior quantity in several consecutive time segments consistently exceeds this preset threshold, the system merges these consecutive segments into a single data segment. The essence of clustering is threshold-based aggregation; that is, when the behavior quantity changes between adjacent time segments meet the growth rules, they are grouped into the same high-variability segment. This process can be implemented using a behavior growth rate calculation function within the software combined with a sliding window mechanism, offering flexible time granularity and a category-adaptive threshold strategy. The finally labeled continuously growing data segments serve as candidate perturbation segments for subsequent determination of whether they are affected by external factors.

[0065] Grouping data records by service category and performing threshold-based clustering within each group aims to independently monitor behavioral trends across different service dimensions, thereby achieving accurate identification and minimizing false positives. Different categories of public services exhibit significantly different usage rhythms and behavioral fluctuation characteristics. For example, medical services may see high-frequency concentrations at specific times, while government services show stronger regularity. Uncategorized processing could dilute abnormal behavior in individual services, affecting identification sensitivity. Furthermore, short-term increases in the number of behaviors do not necessarily represent a rise in actual demand. Only when the increase exceeds the threshold set by the historical patterns of the service category and exhibits continuity can it potentially indicate external event disturbances. Therefore, identifying continuous high-growth segments through clustering helps the system capture potential data anomalies early, providing a data foundation for subsequent determination of the existence of external non-service events, and also providing a precise starting point for intelligent control strategies.

[0066] To identify data segments affected by external non-service events, the system first performs spatial matching calculations between the path trajectory field and the service area coordinate set on data segments with continuously increasing behavior. Specifically, the latitude and longitude coordinates in the path trajectory field of each data record are compared with a pre-configured service area coordinate set. If most records within a data segment do not fall within the service area coordinate set, the data segment is deemed to lack actual service access behavior. Simultaneously, the system also synchronously captures event keyword frequency data from external public information sources (such as social media platforms and news APIs), extracts high-frequency rising terms of external words within the specified time period, and cross-compares them with high-frequency words in the keyword field of the data segment. If the keyword overlap exceeds a set threshold and the frequency change trend remains consistent over time, the behavior segment is determined to be associated with an externally disseminated event. Finally, the system identifies behavior segments that simultaneously satisfy both spatial deviation and keyword fluctuation matching as data segments affected by external non-service events and marks them as disturbed social perception data segments for subsequent intervention assessment and control strategy decision-making.

[0067] This identification logic essentially combines the two characteristics of "spatial behavior deviation" and "external semantic synchronization" as judgment criteria. The path trajectory field reflects whether user behavior truly approaches or intervenes in public service facilities, while the service area coordinate set is a set of high-confidence coordinates extracted from service operation and maintenance data. If user behavior has no spatial intersection with service facilities, it indicates that the behavior has not actually reached the service site. At the same time, external public information sources reflect the dissemination path of events or information. When certain public opinion information, misinformation, or hot topic hype emerge, although users may not actually go to the service point, their online behavior (search, click, forward) will quickly concentrate. This data appearance can easily be misjudged as an increase in real demand. By combining the two dimensions of spatial deviation and semantic synchronization, these "perceptual fluctuation segments" that do not have a real intention to access can be separated, thereby accurately identifying which data is amplified by external non-service events, providing an effective criterion for eliminating interference factors in the control strategy.

[0068] The fundamental reason for this design is that existing supply and demand control models typically judge changes in service pressure based on fluctuations in user behavior. If the system mistakenly interprets "perceived behavior" as "real service demand," it triggers a series of measures such as resource reallocation and service priority adjustments. However, if these behaviors originate from non-service events, such as misleading public opinion or collective forwarding on social media platforms, it not only leads to ineffective responses from control resources but may also create reverse pressure on areas with genuine demand, ultimately causing system-level control disorder. Therefore, it is essential to construct a mechanism that can perceive the "causal source behind the behavior." By determining whether spatial behavior actually reaches the service area and observing whether the behavior is highly synchronized with online events, the disrupted behavior segment can be identified without relying on subjective user feedback. This mechanism is a crucial step in ensuring that control strategies are based solely on "real load" rather than "false perceptions," and it is the core logic for improving the robustness, stability, and credibility of the intelligent control system.

[0069] The anomaly detection module acquires the associated behavioral feature information of each disturbed social perception data segment, and after acquisition, evaluates the degree of data anomaly of each disturbed social perception data segment by constructing an anomaly degree generation model, and classifies it according to the evaluation results.

[0070] In this embodiment, in the anomaly determination module, the associated behavioral feature information of each disturbed social perception data segment is obtained and normalized after acquisition. After normalization, a dimensionless feature vector is constructed, and an anomaly degree generation model is constructed based on the dimensionless feature vector to generate an anomaly degree evaluation value for each disturbed social perception data segment. The anomaly degree evaluation value of each disturbed social perception data segment is compared with a pre-set threshold range. The data anomaly degree of each disturbed social perception data segment is evaluated based on the comparison result, and each disturbed social perception data segment is divided into a pseudo-disturbance segment, a light disturbance segment, and a heavy disturbance segment based on the evaluation result.

[0071] In this embodiment, the anomaly determination module acquires the associated behavioral feature information of each disturbed social perception data segment. The associated behavioral feature information includes four types of feature information, specifically:

[0072] The percentage of user behavior records located within the public service facility area per unit time is specifically: the ratio between the number of user behavior records whose geographical coordinates fall within the public service facility area and the total number of records within that time unit in each preset time unit.

[0073] The percentage of data records with a complete behavioral conversion path is specifically defined as the ratio between the number of data records that simultaneously include the four behavioral fields of query, click, navigation, and terminal access in the behavioral sequence and the total number of records within that time unit.

[0074] The average growth rate of the frequency of high-frequency keywords in external public information sources is specifically: the average increase in the time-series change value of the frequency of keywords extracted from the disturbed social perception data segment in external information sources within the corresponding time period.

[0075] The maximum difference between the rate of change of the number of behaviors in each behavior cluster unit after clustering is specifically: the maximum difference between the rate of change of the number of behaviors in several spatial behavior units obtained after clustering by inputting the geographic location field in the behavior record into the clustering process, within adjacent time windows.

[0076] To acquire the associated behavioral characteristics of each affected social perception data segment, an automated extraction mechanism based on behavioral log data and external information sources can be constructed. First, the structured social perception data is divided into several sub-segments according to time windows, with each sub-segment's data records serving as a processing unit. For the "percentage of user behavior records located within the public service facility's area within a unit of time," a geospatial matching algorithm can be used to spatially match the latitude and longitude fields of each record with a preset set of public service facility coordinates, determining whether it falls within the service facility's buffer zone. The number of such records is then counted and divided by the total number of records in the current time window to generate the percentage value. For the "percentage of data records with complete behavioral conversion paths," a complete sequence containing the four steps of "query → click → navigation → terminal access" can be identified from the behavioral fields, and the proportion of such data records to the total number of records in the current window is calculated. For the "frequency of high-frequency keywords appearing in external public information sources,"... The "average growth rate" can be calculated by extracting a set of high-frequency keywords from the current disturbed data segment, constructing a list of corresponding query terms, and then obtaining the frequency of these keywords over several consecutive time periods by accessing external information sources (such as government platform public opinion interfaces, news APIs, etc.). The growth rate can then be calculated using differential or moving average methods, and the mean can be obtained. For the "maximum difference between the rate of change of behavior quantity in each behavior cluster unit after clustering," spatial clustering (such as K-means or DBSCAN) can be constructed based on the geographic location field. The rate of change of behavior records in each spatial cluster unit between the current time window and the previous window can be statistically analyzed, and the maximum difference among these rates of change can be calculated. All these steps can be automated through software programs, preserving temporal and spatial characteristics while extracting multi-dimensional behavioral indicators, providing structured input data for subsequent normalization and model calculations.

[0077] In this embodiment, the associated behavioral feature information of each disturbed social perception data segment is normalized, and a dimensionless feature vector is constructed after normalization, specifically as follows:

[0078] The values ​​of each feature in the associated behavioral feature information are scaled relative to the maximum and minimum values ​​of each feature in the preset historical sample set, and the values ​​of each feature are mapped to a standardized numerical range between zero and one.

[0079] Mapping the numerical values ​​of each feature in the associated behavioral information to a standardized range between zero and one can typically be achieved through interval scaling normalization. This involves extracting the maximum and minimum values ​​of each feature from a pre-defined historical sample set, and then using these as a benchmark to linearly map the current feature value. Specifically, the software system can perform batch statistics on historical data in the initial stage to form maximum and minimum value pairs (i.e., feature upper and lower limits) for each feature dimension. When data enters the anomaly detection process, the system performs the following operation on the feature values ​​of each disturbed social perception data segment: subtracting the minimum value of the feature from the historical samples, and then dividing by the range of the feature (maximum minus minimum), thus obtaining the normalized standardized value. In this way, all feature values ​​are mapped to the range [0,1], ensuring they participate in subsequent processing while maintaining dimensional consistency. This processing is necessary because the original values ​​of various features have inconsistent orders of magnitude and unit dimensions (e.g., one is a percentage, one is a frequency growth rate, and another might be a difference). If directly used as model input, it can easily cause dimensionality bias, calculation distortion, or allow a single feature to dominate the result during fusion. Therefore, the dimensionless input structure after standardizing the scale not only improves data compatibility but also provides fair and comparable input conditions for subsequent model evaluation, thereby enhancing the stability and accuracy of overall regulatory judgment.

[0080] After normalizing the various feature information, the proportion of user behavior records within the scope of public service facilities, the proportion of data records with complete behavior conversion paths, the average growth rate of the frequency of high-frequency keywords in external public information sources, and the maximum difference between the behavior quantity change rate of each behavior cluster unit after clustering are combined into a dimensionless feature vector in a preset order.

[0081] After normalizing the various feature information, the software system can automatically execute the feature vector construction logic. This involves arranging and combining the normalized feature values ​​in a pre-defined fixed order to form a multi-dimensional numerical vector, resulting in the final dimensionless feature vector. This operation can be achieved by defining a standard feature arrangement template. For example, the first dimension could be set as "the proportion of user behavior records within the scope of public service facilities," the second as "the proportion of complete behavior conversion paths," the third as "the average growth rate of external keywords," and the fourth as "the maximum difference in behavior change rates between cluster units." When processing any disturbed data segment, the system only needs to extract the corresponding four normalized values ​​according to this template order and concatenate them to form a four-dimensional vector. For instance, if the four normalized feature values ​​in a data segment are 0.62, 0.48, 0.74, and 0.91, then the corresponding feature vector is [0.62, 0.48, 0.74, 0.91]. This method ensures a unified input structure for all data segments, facilitating subsequent batch processing of the model, feature pattern comparison, and evaluation calculations. The reason for constructing this vector is that the determination of anomaly severity relies on the joint expression of multi-dimensional features; a single feature alone is insufficient to accurately reflect the intensity of the anomaly. Structured combination of multiple dimensionsless features into a vector not only enhances expressive power but also enables subsequent models to capture the interactions between multiple dimensions, thereby improving the accuracy of identifying complex perturbations.

[0082] In this embodiment, in the anomaly determination module, an anomaly degree generation model is constructed based on dimensionless feature vectors to generate anomaly degree evaluation values ​​for each disturbed social perception data segment, specifically:

[0083] The dimensionless feature vector of each disturbed social perception data segment is mapped to the labeled feature vector in the historical sample set. Each sample in the historical sample set includes a set of dimensionless feature vectors and their corresponding public service response status labels.

[0084] To achieve the goal of "matching the dimensionless feature vector of each disrupted socially perceived data segment with the labeled feature vectors in the historical sample set," a structured sample database can be constructed. In this database, each sample data record contains a set of dimensionless feature vectors extracted from historical disrupted data segments, as well as a service response status label (e.g., no response, minor allocation, full allocation) exhibited by that data segment in the actual supply and demand regulation response. The system first needs to archive similar data disturbances that occurred in the historical period and then label them by comparing them with the actual response of the public service system, either manually or through a rule engine. For example, if a disturbance segment does not trigger significant service allocation behavior after it occurs, it can be labeled as a "pseudo-disturbance segment"; if it triggers small-scale allocation but does not exceed the service tolerance range, it can be labeled as a "minor disturbance segment"; if it causes resource reallocation or affects the normal operation of services in other areas, it can be labeled as a "major disturbance segment." After sample preparation, during the model training phase, the dimensionless feature vectors extracted from the current disturbed data segment are input and compared with vectors in the sample library using vector space matching or distance metrics (such as Euclidean distance or cosine similarity) to find similar samples or as input for supervised learning training sets. This establishes training sample pairs of "input features → known response labels," enabling the model to learn the actual service response relationships caused by various combinations of behavioral features, thereby improving the model's ability to identify different disturbance patterns. For example, if a set of features in the sample is "low proportion of service facility behavior, very few complete conversion paths, extremely high external growth rate of keywords, and high spatial change rate," and the label is "pseudo-disturbance segment," the model will gradually learn that although this feature combination is "active," it is not equivalent to a real supply-demand imbalance. This process can be fully implemented through the database structure, training framework, and data interface of the software system. The fundamental purpose of this is to establish an "empirical correlation" between data behavior and service response, giving the model an empirical basis and ensuring that its output anomaly assessment value is closer to the real urban operating state.

[0085] An anomaly level generation model is constructed using supervised learning. The model consists of an input layer, a cross-feature construction layer, and a weighted fusion layer. The input layer receives a dimensionless feature vector. The cross-feature construction layer performs feature combination operations on features of each dimension to generate several interaction terms. The weighted fusion layer weights each interaction term based on the feature combination weights learned during the training phase and outputs a single evaluation value representing the degree of data anomaly.

[0086] During the training phase, a loss function is constructed by minimizing the error between the model output value and the sample label, and the model parameters are iteratively optimized using backpropagation.

[0087] During the training phase, a loss function is constructed by minimizing the error between the model output value and the sample labels, and the model parameters are iteratively optimized using backpropagation. This can be implemented as a standard supervised learning process in software. Specifically, the dimensionless feature vector of each sample in the historical sample set is used as the model input, and the public service response state label (e.g., 0 for pseudo-perturbation, 1 for light perturbation, and 2 for heavy perturbation) is used as the model's target output. The current output value is calculated through forward propagation. Then, the error between the predicted output and the actual label is calculated, typically quantified using the mean squared error (MSE) or cross-entropy loss function. This error function measures the degree of deviation in the model's current predictive ability. The error is then propagated back to each layer of the model using the backpropagation algorithm, gradually adjusting the weight parameters and optimizing its learning direction. The backpropagation process relies on the chain rule, updating the weights layer by layer to minimize the overall error. In each iteration, samples are re-inputted, and parameters are updated until the loss function converges or the preset training rounds are reached. Taking anomaly level generation models as an example, if the model predicts a data segment as a "slightly perturbed segment" (output value 1.2) but labels it as a "pseudo-perturbed segment" (value 0), the loss function will output a large error. After multiple rounds of optimization, the model will gradually learn to match these dimensionless feature combinations with the "pseudo-perturbed segment". This process can be flexibly built and executed at the software level using common deep learning frameworks (such as TensorFlow or PyTorch), completely independent of hardware characteristics or external devices. The core purpose of this method is to enable the model to automatically learn the non-linear relationship between features and responses through training on a large number of samples, thereby possessing a high-precision anomaly level judgment capability when facing new data.

[0088] After the model training is completed, the dimensionless feature vector of any newly received disturbed social perception data segment is input into the anomaly degree generation model to obtain the corresponding anomaly degree evaluation value, which is used for subsequent classification judgment and response strategy execution.

[0089] After model training is complete, the dimensionless feature vector of any newly received disturbed social perception data segment is input into the anomaly level generation model to obtain the corresponding anomaly level evaluation value. This can be achieved on the software platform through the deployed model call interface or integrated online inference service. Specifically, the following steps are taken: First, the four related behavioral features of the newly received data segment are extracted according to a predetermined process and normalized to construct a dimensionless feature vector. Then, this feature vector is passed as input to the trained anomaly level generation model. This model has learned the mapping relationship between features and anomaly level during the training phase and can quickly output a real value as the anomaly level evaluation value. This value usually falls within a preset range (e.g., between 0 and 1), indicating the level of anomaly in the data segment. For example, if the model outputs an anomaly level evaluation value of 0.82 for a certain data segment, and the preset threshold range for a "re-disturbed segment" is 0.8 to 1.0, then the data segment can be classified as a "re-disturbed segment," triggering a high-intensity response strategy. This process is typically performed in software using batch inference or streaming processing, and can be integrated into the real-time control process of a public service supply and demand control platform to achieve automatic judgment and immediate response. The significance of this is that it avoids potential misjudgments caused by manually set static rules, accurately capturing abnormal behavior patterns induced by external non-service events through a data-driven approach, thus enabling the intelligent control system to have stronger identification capabilities and adaptability when facing complex social dynamics.

[0090] The above method generates anomaly assessment values ​​for each affected socially perceived data segment, aiming to achieve accurate quantitative identification of non-service-related disruptive behaviors through a data-driven approach. This method introduces a supervised learning model, utilizing the correspondence between known behavioral characteristics and public service response states in historical samples to uncover deep-seated correlations between different combinations of behavioral characteristics and anomaly levels, avoiding subjective misjudgments caused by manually set rules. The model enhances feature representation through cross-feature construction, and a weighted fusion layer further extracts key influencing factors, making the output assessment value more discriminative. Its physical meaning lies in the fact that this assessment value reflects the "anomaly risk level" of a particular affected data segment in the dimension of service authenticity. A higher value indicates that the behavioral data is more likely to be driven by non-service-related inducements, and the more severely it deviates from the true supply and demand state; a lower value means that the behavioral data is more credible and can be directly used for regulatory strategy formulation. Therefore, this assessment value, as an input variable for regulatory decisions, can not only improve the accuracy of supply and demand matching responses but also prevent resource misallocation and erroneous response triggers, enhancing the robustness and intelligence of the public service regulation system in complex social and public opinion environments.

[0091] The magnitude of the "anomaly assessment value of each disturbed socially perceived data segment" physically reflects the degree to which the user behavior characteristics corresponding to that data segment deviate from the actual service demand state, that is, the intensity of the anomaly in the data's deviation from the actual supply and demand relationship of public services. A smaller assessment value indicates a higher degree of matching between the user behavior characteristics (such as geographical distribution, completeness of the behavior chain, and keyword popularity diffusion) and the actual service location or process, making the behavior more "credible," and the system can consider that the data segment has not been disturbed by external events. A moderate assessment value indicates that some behavioral characteristics may be distorted by external content disturbances but have not completely deviated from the actual service chain, requiring careful handling. A larger assessment value indicates that the behavior in that data segment is highly concentrated but lacks spatial focus, the behavior path is incomplete, and keywords have an abnormal surge in external media, indicating that it is highly likely to be driven by non-service-related events, forming a "false supply and demand signal." Directly guiding resource response in this case will bring the risk of misallocation of resources or service mismatch. Therefore, this assessment value is essentially used to quantify the "risk degree of behavioral data deviating from the actual service," providing a basis for judgment in subsequent supply and demand matching and control strategies.

[0092] In this embodiment, a supervised learning approach is used to construct an anomaly level generation model. The anomaly level generation model consists of an input layer, a cross-feature construction layer, and a weighted fusion layer, wherein:

[0093] The input layer is used to receive the dimensionless feature vector after normalization. The dimensionless feature vector includes the proportion of user behavior records within the scope of public service facilities per unit time, the proportion of data records with complete behavior conversion paths, the average growth rate of the frequency of high-frequency keywords in external public information sources, and the maximum difference between the behavior quantity change rates of each behavior cluster unit after clustering.

[0094] To achieve the goal of "the input layer receiving the normalized dimensionless feature vector," the system first relies on the pre-built feature extraction and normalization process. In this process, the system extracts four core indicator data from structured behavioral data: the proportion of user behavior records within the public service facility area per unit time, the proportion of data records with complete behavior conversion paths, the frequency growth rate of keywords in external information sources, and the maximum difference in the rate of change of behavior quantity within cluster units. Subsequently, through normalization techniques such as interval scaling, each feature value is mapped to a standardized interval (e.g., 0 to 1), eliminating differences in the original data units and dimensions. At this point, the system assembles the four normalized feature values ​​into a one-dimensional vector in a preset order, forming a dimensionless feature vector. During model execution, this feature vector serves as model input, passed to the model's input layer through programming interfaces (such as tensor stream input channels, vector cache queues, etc.). To illustrate with a concrete example, if the normalized values ​​of the four features of a data segment are 0.76, 0.52, 0.33, and 0.81, the system will construct a vector of the form [0.76, 0.52, 0.33, 0.81] and input it into the model's input layer for further processing. This mechanism ensures a unified model input structure and clear data dimensions, contributing to the stability and accuracy of subsequent feature interaction modeling. Furthermore, since the input vector has had its units and dimensions removed, its numerical differences fully reflect the relative performance of each behavioral feature within the current data segment, effectively enhancing the anomaly assessment model's sensitivity and discriminative ability to detect differences in features within disturbed data segments.

[0095] The cross-feature construction layer performs pairwise combinations on the feature information of each dimension in the dimensionless feature vector to form feature interaction terms, which are used to capture the mutual influence relationship between various behavioral features.

[0096] To achieve the goal of "the cross-feature construction layer performing pairwise combinations of feature information from each dimension in the dimensionless feature vector to form feature interaction terms," ​​feature crossing technology can be used to construct a new set of interactive input features to enhance the model's ability to express implicit relationships between behavioral features. Specifically, the system first extracts the numerical elements of all dimensions from the input dimensionless feature vector, and then combines the feature values ​​of any two different dimensions. For example, using a product-crossing method, feature A is multiplied by feature B to generate a new cross feature A×B. If the vector is [0.76, 0.52, 0.33, 0.81], interaction terms such as 0.76×0.52 = 0.3952 and 0.76×0.33 = 0.2508 may be generated, forming a total of 6 cross terms. These cross terms constitute new input extension dimensions, serving as input to the next layer of the model to capture the nonlinear relationships between the original dimensions. The main purpose of this approach is to reveal the potential mutual reinforcement or weakening effects between multiple socially perceived behavioral features. For example, the intensity of user behavior clustering and the simultaneous increase in external keyword popularity may occur simultaneously, but evaluating them individually is insufficient to reflect their "synergistic anomaly." Constructing cross-features can significantly improve the model's ability to identify subtle perturbation patterns. Furthermore, this feature combination method has advantages such as strong versatility, clear implementation logic, and high scalability. It can be implemented in deep learning frameworks using conventional vector operation functions, greatly enhancing the model's ability to learn and adapt to complex behavioral feature co-changes.

[0097] The weighted fusion layer performs a weighted summation on all feature interaction terms based on the feature combination weight parameters determined during the training phase, and outputs a single evaluation value representing the degree of data anomaly of each disturbed socially perceived data segment.

[0098] The goal of "the weighted fusion layer, based on the feature combination weight parameters determined during the training phase, performs a weighted summation on all feature interaction terms and outputs a single evaluation value representing the degree of data anomaly for each disrupted socially sensed data segment" can be achieved by constructing a linear output layer in a feedforward neural network. During the training phase, the system automatically learns the importance of each feature interaction term based on the relationship between dimensionless feature interaction terms and actual service anomaly response labels in historical samples, using a loss function minimization strategy; that is, it determines the weight parameters corresponding to each cross-feature term. The weighted fusion layer takes all feature interaction terms generated by the previous layer as input, multiplies them by their corresponding weight parameters, performs a summation operation, and finally outputs a scalar value as the anomaly evaluation value for that disrupted socially sensed data segment. For example, if the previous layer generated 6 cross-feature terms... And has learned the corresponding weight parameters. The output value of this layer is ,Right now The core advantage of this mechanism lies in its ability to differentiate weighting among different feature interactions, thereby strengthening the modeling ability of key combination patterns that influence the degree of anomaly. Its implementation can be based on the fully connected layer (Linear Layer) modules of deep learning platforms such as TensorFlow and PyTorch, ensuring efficient computation and supporting subsequent gradient backpropagation optimization. In this way, not only is the trainability and generalization ability of the model guaranteed, but the output anomaly assessment value also possesses stability and interpretability, facilitating subsequent classification and judgment of spurious, minor, and major perturbations and the coordinated execution of response strategies.

[0099] In this embodiment, the anomaly assessment value of each disturbed social sensing data segment is compared with a pre-set threshold range. Based on the comparison results, the anomaly degree of each disturbed social sensing data segment is assessed, and based on the assessment results, each disturbed social sensing data segment is divided into a pseudo-disturbance segment, a lightly disturbed segment, and a heavily disturbed segment. Specifically:

[0100] When the anomaly assessment value is less than the minimum value of the preset threshold range, the data anomaly level of the disturbed social perception data segment is determined to be low, and it is then classified as a false disturbance segment.

[0101] This situation indicates that the clustering of user behavior observed in this data segment does not significantly deviate from historical behavior patterns or current service load. Such behavioral fluctuations are highly likely caused by short-term, minor information dissemination, occasional non-hotspot interference, or fluctuations in regular user behavior, and do not signify a conversion of actual service demand. Activating control strategies on this type of data segment would result in unnecessary resource waste and system disruption. Therefore, this type of data should be filtered out of the intelligent control process to prevent "over-response" and maintain the stability and discriminative capabilities of the control system.

[0102] When the abnormality assessment value is greater than or equal to the minimum value of the preset threshold interval and less than or equal to the maximum value of the preset threshold interval, the abnormality of the disturbed social perception data segment is determined to be moderate, and it is classified as a slightly disturbed segment.

[0103] This situation indicates that the data segment exhibits certain abnormal behavior characteristics. Although some features suggest concentrated behavior driven by external events, there is still a possibility of overlap with actual service demands. Such data segments implicitly suggest a situation where "some real demands are mixed within the abnormal clusters." Without intervention, this could lead to delayed service response in some areas. Therefore, a cautious control strategy should be adopted, such as appropriately allocating resources, delaying response triggers, and further verifying through secondary sensing channels, to maintain the system's responsiveness to potential real demands without causing misallocation of resources.

[0104] When the anomaly assessment value is greater than the maximum value of the preset threshold range, the data anomaly of the disturbed social perception data segment is determined to be severe, and it is then classified as a heavily disturbed segment.

[0105] This situation indicates that the data segment exhibits strong abnormal clustering characteristics and significantly deviates from the service carrying path. The frequency of keyword-driven public opinion dissemination has surged, and the spatial trajectory shows non-service-oriented synchronous behavior. This typically signifies data anomalies driven by non-service events such as public opinion shocks, misinformation, and hype. Failure to identify and isolate such situations promptly will severely mislead the control model, leading to large-scale resource misallocation and imbalances in the actual service area. The appropriate response strategy should be a complete blocking response suppression mechanism. Upon model identification, the input of this data segment to the intelligent control process should be immediately interrupted, and the results should be fed back to the control system for strategy correction.

[0106] The "pre-defined threshold interval" can be determined through statistical distribution analysis based on historical samples. Specifically, it first requires extracting labeled samples of disturbed social perception data from multiple historical periods. Each sample corresponds to an anomaly assessment value calculated by an anomaly generation model. These assessment values ​​are used as the overall sample set, and statistical analysis methods (such as percentile method, distribution fitting, and quantile breakpoint extraction) are applied to model their distribution characteristics. By observing the clustering and dispersion intervals of the assessment values, for example, the 30th percentile value can be set as the upper limit for the "pseudo-disturbance segment," and the 70th percentile value as the upper limit for the "minor disturbance segment," thus forming a data-driven threshold interval. This interval, once constructed, can be fixed as model parameters for subsequent classification and judgment. It can also be dynamically updated in conjunction with real-time data when necessary to adapt to the evolving trends of service behavior patterns. This approach not only ensures the rationality and objectivity of the classification criteria but also enhances the model's adaptability in different environments.

[0107] The strategy execution module executes dynamic adjustment strategies for supply and demand matching response intensity corresponding to each category, based on the classification results.

[0108] In this embodiment, the strategy execution module executes dynamic adjustment strategies for supply and demand matching response intensity corresponding to each category based on the classification results, specifically as follows:

[0109] For the disturbed social perception data segments that are classified as pseudo-interference segments, the original resource allocation plan remains unchanged based on the service type corresponding to the data segment, and the behavioral characteristic patterns of the data segment are recorded in the interference behavior feature database for subsequent interference pattern comparison and identification.

[0110] For disturbed social perception data segments classified as pseudo-interference segments, their corresponding data anomaly assessment values ​​are low. This typically means that although the data segment exhibits some fluctuations in behavior, it does not create substantial resource pressure or supply-demand imbalance for actual public services. Therefore, no proactive adjustments to resource allocation are necessary. To enhance the system's ability to identify similar interference signals in the future, key feature information (such as user behavior trajectory sequences, keyword combinations, time distribution density, behavioral location coordinates, and behavioral path structures) can be extracted from the behavioral data segment using software. This information is then represented as a standardized behavioral pattern template using feature vector encoding and stored in an internal interference behavior feature library. This library serves as a baseline for subsequent judgments, supporting the system in performing interference feature comparison processes when new data arrives. When a new behavioral segment is found to be highly similar to historical interference patterns, it can be prioritized as a pseudo-interference segment, reducing unnecessary resource response operations. This approach reduces the probability of erroneous scheduling and strengthens the system's experience in identifying "non-service high-intensity event behavior patterns," thereby improving the stability and accuracy of the control system.

[0111] For the disturbed social perception data segments that are classified as lightly disturbed segments, resource scheduling adjustment operations with limited ranges are performed on the service types corresponding to the data segments. Specifically, upper and lower limits for resource allocation are set, and within these limits, the quantity of resources, scheduling frequency, and service node coverage are proportionally adjusted. At the same time, a behavior change trend monitoring program is launched to dynamically evaluate the results of resource allocation adjustment.

[0112] For the disturbed social perception data segments classified as "lightly disturbed," their anomaly level is in the intermediate range, meaning that the current behavioral signals may be partially affected by external non-service events, but the existence of genuine service demand cannot be ruled out. Therefore, a "controlled response" strategy should be adopted for resource allocation. In software implementation, a preset resource scheduling boundary strategy template can be invoked for the service type corresponding to this data segment, setting upper and lower thresholds for resource allocation (e.g., increasing resources by no more than 20% of the current amount, and maintaining basic support configuration at the minimum). Within this range, the system uses a proportional control algorithm to flexibly adjust the amount of resources, service scheduling frequency (e.g., vehicle frequency, window opening frequency), and the geographical coverage radius of service nodes. Simultaneously, the system will initiate a behavioral change trend monitoring program, periodically collecting a new round of social perception data input for this service type, analyzing whether the behavioral aggregation trend continues to strengthen, weaken, or stabilize, and comparing its effectiveness with the initial control actions in real time to dynamically adjust the current resource response magnitude. This approach avoids over-scheduling caused by misjudgment while retaining the system's appropriate response capability to potential service pressure, improving the robustness and adaptability of the control system.

[0113] For disturbed social perception data segments that are classified as heavily disturbed segments, the automatic resource scheduling process for the service type corresponding to the data segment is suspended, and a manually preset static resource allocation scheme is switched to perform fixed configuration distribution for service nodes, service time periods and service content. At the same time, the behavioral data input channel and feedback channel for this service type are suspended to prevent the current abnormal data from continuing to participate in the control decision.

[0114] When a segment of disturbed socially sensed data is classified as a heavily disturbed segment, it indicates that its anomaly level has significantly exceeded the system's trust threshold, and it is highly likely to be driven entirely by external non-service events. Continuing to use such data for automatic resource scheduling could easily lead to severe resource mismatch or even systemic service imbalance. Therefore, in the software implementation, the first step is to mark the automatic resource scheduling logic flow corresponding to the service type as "frozen" using a mapping table between service types and resource scheduling strategies, and prohibit the scheduling flow from being restarted by timed events or triggered by events. Subsequently, the system will call the default allocation template for this service type from a pre-set static resource allocation strategy library, loading resources into the scheduling plan table according to fixed rules in dimensions such as the number of nodes, service coverage area, service time period, and service items. To prevent continuous input of abnormal data from interfering with the decision-making logic, the input channel management program must temporarily disable the data access rights of this service type in the sensed data processing flow, and simultaneously disable the user feedback collection mechanism under this service type. This ensures that the entire control system enters a "static safe operation" mode when subjected to high-intensity interference, thereby maintaining the availability of basic services and waiting for the external disturbance to subside before resuming normal strategy execution.

[0115] The regulation and feedback module acquires data on the operation of public services after implementing dynamic regulation strategies for the intensity of each supply and demand matching response. Based on the data on the operation of public services, it updates the classification results of the disturbed social perception data segments to enable subsequent classification judgments and adjustments to regulation strategies.

[0116] To implement the functionality of the "regulation and feedback module," the software can be designed with a continuously running subsystem for acquiring and monitoring operational data. This subsystem is used to obtain real-time operational status indicators for public services, such as service response latency, user request completion rate, and service node load rate. This data can be automatically collected through service platform logs, user feedback interfaces, or embedded operation and maintenance monitoring components. Then, using pre-defined data comparison logic, this operational data is compared with the baseline state before the regulation strategy took effect to determine the actual effect of the strategy. If the monitoring results show a significant improvement or stability in the operational status of the public service, it indicates that the current classification and regulation strategy are effectively matched; if the operational status does not meet expectations, the system marks the affected data segment as "pending review" and initiates an update judgment process.

[0117] During the update judgment process, the software calls upon historical classification records and current service performance data, using a rule engine or regression algorithm to determine whether the original classification contained misjudgments or miscalculations of interference levels. For example, a data segment originally marked as a "lightly disturbed segment" may be reclassified as a "pseudo-disturbed segment" if its service performance shows no significant abnormalities after adjustment; conversely, if the actual load deteriorates rapidly, it can be upgraded to a "heavily disturbed segment." This mechanism, by introducing service performance feedback, achieves a closed-loop update from "forward identification + classification" to "backward verification + adjustment." This not only improves the accuracy of classification judgments but also enhances the adaptability and dynamic correction capabilities of the adjustment strategy, avoiding the rigidity of strategies caused by static classification.

[0118] The above formulas are all dimensionless calculations. The formulas are derived from software simulations based on a large amount of collected data to obtain the most recent real-world results. The preset parameters in the formulas are set by those skilled in the art according to the actual situation.

[0119] The above embodiments can be implemented, in whole or in part, by software, hardware, firmware, or any other combination thereof. When implemented using software, the above embodiments can be implemented, in whole or in part, as a computer program product. The computer program product includes one or more computer instructions or computer programs. When the computer instructions or computer programs are loaded or executed on a computer, all or part of the processes or functions described in the embodiments of this application are generated. The computer can be a general-purpose computer, a special-purpose computer, a computer network, or other programmable device. The computer instructions can be stored in a computer-readable storage medium or transmitted from one computer-readable storage medium to another. For example, the computer instructions can be transmitted from one website, computer, server, or data center to another website, computer, server, or data center via wired or wireless (e.g., infrared, wireless, microwave, etc.) means. The computer-readable storage medium can be any available medium that a computer can access or a data storage device such as a server or data center that includes one or more sets of available media. The available medium can be a magnetic medium (e.g., floppy disk, hard disk, magnetic tape), an optical medium (e.g., DVD), or a semiconductor medium. The semiconductor medium can be a solid-state drive.

[0120] It should be understood that in the various embodiments of this application, the order of the above-mentioned processes does not imply the order of execution. The execution order of each process should be determined by its function and internal logic, and should not constitute any limitation on the implementation process of the embodiments of this application.

[0121] Those skilled in the art will recognize that the units and algorithm steps of the various examples described in conjunction with the embodiments disclosed herein can be implemented in electronic hardware, or a combination of computer software and electronic hardware. Whether these functions are implemented in hardware or software depends on the specific application and design constraints of the technical solution. Those skilled in the art can use different methods to implement the described functions for each specific application, but such implementation should not be considered beyond the scope of this application.

[0122] In the several embodiments provided in this application, it should be understood that the disclosed systems and methods can be implemented in other ways. For example, the embodiments described above are merely illustrative; for instance, the division of units is only a logical functional division, and in actual implementation, there may be other division methods. For example, multiple units or components may be combined or integrated into another system, or some features may be ignored or not executed. Furthermore, the coupling or direct coupling or communication connection shown or discussed may be an indirect coupling or communication connection through some interfaces, devices, or units, and may be electrical, mechanical, or other forms.

[0123] The units described as separate components may or may not be physically separate. The components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the units can be selected to achieve the purpose of this embodiment according to actual needs.

[0124] In addition, the functional units in the various embodiments of this application can be integrated into one processing unit, or each unit can exist physically separately, or two or more units can be integrated into one unit.

[0125] The above description is merely a specific embodiment of this application, but the scope of protection of this application is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in this application should be included within the scope of protection of this application. Therefore, the scope of protection of this application should be determined by the scope of the claims.

Claims

1. A public service supply and demand matching intelligent control system based on social perception data, characterized in that, It includes a data construction module, a service mapping module, a disturbance identification module, an anomaly detection module, a strategy execution module, and a regulation feedback module; The data construction module collects social perception data from multiple sources, and after collection, it performs data cleaning, format unification and structuring to form a structured data record sequence; The service mapping module establishes a mapping relationship between the structured data record sequence and public service types. It classifies each data record through the association rules between behavior tags and service intentions, and generates a set of data records containing service category identifiers. The disturbance identification module performs time-series clustering analysis on the data record set to identify data segments affected by external non-service events and marks these data segments as disturbed socially perceived data segments. In the disturbance identification module, the data record set is divided into several data sequences arranged in chronological order according to the service category identifier. Clustering based on threshold division is performed on the change in the number of behaviors in each data sequence to identify data segments where the number of behaviors increases continuously and exceeds a preset growth threshold. Among them, data segments where the number of behaviors increases beyond the preset growth threshold but the path trajectory field does not contain the coordinate points in the corresponding service area coordinate set, and the keyword field in the data segment appears at a synchronously increasing frequency in external public information sources, are identified as data segments affected by external non-service events and are marked as disturbed social perception data segments. The anomaly detection module acquires the associated behavioral feature information of each disturbed social perception data segment, and after acquisition, evaluates the degree of data anomaly of each disturbed social perception data segment by constructing an anomaly degree generation model, and classifies it according to the evaluation results. The strategy execution module executes dynamic adjustment strategies for supply and demand matching response intensity corresponding to each category, based on the classification results. The regulation and feedback module acquires data on the operation of public services after implementing dynamic regulation strategies for the intensity of each supply and demand matching response. Based on the data on the operation of public services, it updates the classification results of the disturbed social perception data segments to enable subsequent classification judgments and adjustments to regulation strategies.

2. The intelligent regulation system for matching public service supply and demand based on social perception data according to claim 1, characterized in that, In the service mapping module, keyword information, behavior location path information, and time tag information are extracted from the structured data record sequence. Combined with the preset service semantic dictionary and public service scenario tag library, a mapping relationship between structured behavior information and public service type is established. Based on the established mapping relationship, the behavior tag in each data record is matched with the service intent according to rules. After matching, the data record is assigned a service category identifier to generate a set of data records containing service category identifiers.

3. The intelligent regulation system for matching public service supply and demand based on social perception data according to claim 1, characterized in that, In the anomaly detection module, the associated behavioral feature information of each disturbed social perception data segment is obtained and normalized. After normalization, a dimensionless feature vector is constructed, and an anomaly degree generation model is built based on the dimensionless feature vector to generate an anomaly degree evaluation value for each disturbed social perception data segment. The anomaly degree evaluation value of each disturbed social perception data segment is compared with a pre-set threshold range. The data anomaly degree of each disturbed social perception data segment is evaluated based on the comparison results, and each disturbed social perception data segment is divided into pseudo-disturbance segment, light disturbance segment, and heavy disturbance segment based on the evaluation results.

4. The intelligent regulation and control system for matching public service supply and demand based on social perception data according to claim 3, characterized in that, In the anomaly detection module, the associated behavioral characteristic information of each disturbed social perception data segment is obtained. The associated behavioral characteristic information includes four types of characteristic information, specifically: The percentage of user behavior records located within the public service facility area per unit time is specifically: the ratio between the number of user behavior records whose geographical coordinates fall within the public service facility area and the total number of records within that time unit in each preset time unit. The percentage of data records with a complete behavioral conversion path is specifically defined as the ratio between the number of data records that simultaneously include the four behavioral fields of query, click, navigation, and terminal access in the behavioral sequence and the total number of records within that time unit. The average growth rate of the frequency of high-frequency keywords in external public information sources is specifically: the average increase in the time-series change value of the frequency of keywords extracted from the disturbed social perception data segment in external information sources within the corresponding time period. The maximum difference between the rate of change of the number of behaviors in each behavior cluster unit after clustering is specifically: the maximum difference between the rate of change of the number of behaviors in several spatial behavior units obtained after clustering by inputting the geographic location field in the behavior record into the clustering process, within adjacent time windows.

5. The intelligent regulation system for matching public service supply and demand based on social perception data according to claim 4, characterized in that, The associated behavioral feature information of each disturbed social perception data segment is normalized, and a dimensionless feature vector is constructed after normalization, specifically as follows: The values ​​of each feature in the associated behavioral feature information are scaled relative to the maximum and minimum values ​​of each feature in the preset historical sample set, and the values ​​of each feature are mapped to a standardized numerical range between zero and one. After normalizing the various feature information, the proportion of user behavior records within the scope of public service facilities, the proportion of data records with complete behavior conversion paths, the average growth rate of the frequency of high-frequency keywords in external public information sources, and the maximum difference between the behavior quantity change rate of each behavior cluster unit after clustering are combined into a dimensionless feature vector in a preset order.

6. The intelligent regulation system for matching public service supply and demand based on social perception data according to claim 5, characterized in that, In the anomaly detection module, an anomaly severity generation model is constructed based on dimensionless feature vectors to generate anomaly severity assessment values ​​for each disturbed socially perceived data segment, specifically: The dimensionless feature vector of each disturbed social perception data segment is mapped to the labeled feature vector in the historical sample set. Each sample in the historical sample set includes a set of dimensionless feature vectors and their corresponding public service response status labels. An anomaly level generation model is constructed using supervised learning. The model consists of an input layer, a cross-feature construction layer, and a weighted fusion layer. The input layer receives a dimensionless feature vector. The cross-feature construction layer performs feature combination operations on features of each dimension to generate several interaction terms. The weighted fusion layer weights each interaction term based on the feature combination weights learned during the training phase and outputs a single evaluation value representing the degree of data anomaly. During the training phase, a loss function is constructed by minimizing the error between the model output value and the sample label, and the model parameters are iteratively optimized using backpropagation. After the model training is completed, the dimensionless feature vector of any newly received disturbed social perception data segment is input into the anomaly degree generation model to obtain the corresponding anomaly degree evaluation value.

7. The intelligent regulation system for matching public service supply and demand based on social perception data according to claim 6, characterized in that, An anomaly level generation model is constructed using supervised learning. This model consists of an input layer, a cross-feature construction layer, and a weighted fusion layer, where: The input layer is used to receive the dimensionless feature vector after normalization. The dimensionless feature vector includes the proportion of user behavior records within the scope of public service facilities per unit time, the proportion of data records with complete behavior conversion paths, the average growth rate of the frequency of high-frequency keywords in external public information sources, and the maximum difference between the behavior quantity change rates of each behavior cluster unit after clustering. The cross-feature construction layer performs pairwise combinations on the feature information of each dimension in the dimensionless feature vector to form feature interaction terms, which are used to capture the mutual influence relationship between various behavioral features. The weighted fusion layer performs a weighted summation on all feature interaction terms based on the feature combination weight parameters determined during the training phase, and outputs a single evaluation value representing the degree of data anomaly of each disturbed socially perceived data segment.

8. The intelligent regulation system for matching public service supply and demand based on social perception data according to claim 7, characterized in that, The anomaly assessment value of each disturbed social sensing data segment is compared with a pre-set threshold range. Based on the comparison results, the degree of data anomaly of each disturbed social sensing data segment is assessed, and based on the assessment results, each disturbed social sensing data segment is divided into pseudo-disturbance segments, slightly disturbed segments, and heavily disturbed segments, specifically: When the anomaly assessment value is less than the minimum value of the preset threshold range, the data anomaly level of the disturbed social perception data segment is determined to be low, and it is then classified as a false disturbance segment. When the abnormality assessment value is greater than or equal to the minimum value of the preset threshold interval and less than or equal to the maximum value of the preset threshold interval, the abnormality of the disturbed social perception data segment is determined to be moderate, and it is classified as a slightly disturbed segment. When the anomaly assessment value is greater than the maximum value of the preset threshold range, the data anomaly of the disturbed social perception data segment is determined to be severe, and it is then classified as a heavily disturbed segment.

9. The intelligent regulation system for matching public service supply and demand based on social perception data according to claim 8, characterized in that, In the strategy execution module, based on the classification results, dynamic adjustment strategies for supply and demand matching response intensity corresponding to each category are executed, specifically as follows: For the disturbed social perception data segments that are classified as pseudo-interference segments, the original resource allocation plan remains unchanged based on the service type corresponding to the data segment, and the behavioral characteristic patterns of the data segment are recorded in the interference behavior feature database for subsequent interference pattern comparison and identification. For the disturbed social perception data segments that are classified as lightly disturbed segments, resource scheduling adjustment operations with limited ranges are performed on the service types corresponding to the data segments. Specifically, upper and lower limits for resource allocation are set, and within these limits, the quantity of resources, scheduling frequency, and service node coverage are proportionally adjusted. At the same time, a behavior change trend monitoring program is launched to dynamically evaluate the results of resource allocation adjustment. For disturbed social perception data segments that are classified as heavily disturbed segments, the automatic resource scheduling process for the service type corresponding to the data segment is suspended, and a manually preset static resource allocation scheme is switched to perform fixed configuration distribution for service nodes, service time periods and service content. At the same time, the behavioral data input channel and feedback channel for this service type are suspended to prevent the current abnormal data from continuing to participate in the control decision.

Images