A closed-loop management system for handling public complaints on the 12345 platform

Abstract

This invention relates to the field of e-government information management technology, and discloses a closed-loop management system for handling requests on the 12345 platform. The system includes: extracting the acoustic urgency and semantic information density of the input voice stream and converting them into initial work order quantification values; calculating a spatiotemporal correlation score to generate merged work orders; subsequently outputting the probability distribution of the power and responsibility matching of candidate departments and iteratively optimizing to determine the optimal dispatch path; simultaneously, the system tracks the department processing rate to obtain an efficiency lag index, calculates the time-series change rate of the overall request density to generate an abnormal and sudden heat field, and finally aggregates the two to obtain a rectification priority value to generate a supervision letter. This invention effectively achieves dynamic load balancing and proactive anomaly warning for work orders across the entire domain.

Description

Technical Field

[0001] This invention relates to the field of e-government information management technology, and more specifically, to a closed-loop management system for handling requests on the 12345 platform. Background Technology

[0002] As the digital nerve center of modern urban governance, the 12345 government service platform faces massive concurrent non-standardized voice and text streams as its inputs. These inputs are often accompanied by anxiety, noise interference, and incomplete information regarding location and time. Especially under the stimulus of sudden public events (such as pipeline bursts or environmental pollution), pulse-like, homogeneous mass requests emerge within specific spatial grids, exhibiting strong non-linear aggregation and surge characteristics. Simultaneously, the platform's output execution network is an extremely complex hierarchical matrix grid, vertically divided into a tree-like structure of "city-district-street-community," and horizontally containing numerous overlapping edges of authority and responsibility. This ambiguous topology of authority and responsibility is highly susceptible to impedance misalignment when facing complex cross-departmental events. Currently, traditional request dispatch systems often rely heavily on static preset conditions (such as "if a certain keyword is present, it will be dispatched to a certain bureau," "if the distance exceeds 500 meters, it will not be merged") to forcibly bridge the supply and demand sides. This rigid judgment logic has significant flaws when dealing with the surge of requests due to high entropy and spatiotemporal chaos.

[0003] When faced with complex cross-domain events, static routing is extremely prone to deadlocks, causing work orders to be endlessly pushed back and forth between multiple departments, instantly exhausting the computing power of the grassroots level, and triggering a dynamic load avalanche. In addition, the existing system supervision mechanism relies heavily on post-event static indicators (such as "timeout red light"), lacking prior awareness of the dynamic congestion and backlog that is currently brewing at the grassroots nodes, making it difficult to detect execution damping, which can easily lead to the rapid spread of congestion at a single node, thereby paralyzing the entire flow loop. Summary of the Invention

[0004] This invention provides a closed-loop management system for handling requests on the 12345 platform, which solves the technical problems mentioned in the background.

[0005] This invention provides a closed-loop management system for handling complaints on the 12345 platform, the system being configured to perform the following steps: The acoustic urgency and semantic information density of the input voice stream are extracted through the 12345 platform, and the request is converted into an initial work order quantification value. The spatiotemporal correlation score is calculated based on the initial work order quantification value, and merged work orders are generated based on the spatiotemporal correlation score. Process the merged work orders and output the probability distribution of the matching rights and responsibilities of the candidate departments; Based on the probability distribution of the matching of rights and responsibilities, iterative optimization is performed to determine the optimal distribution path to the target department; Track the work order processing rate of the target department and calculate the efficiency lag index based on the tracking results; Calculate the temporal rate of change of the distribution density of demands across the entire region, and generate an abnormal sudden heat field based on the temporal rate of change; The efficiency lag index and the abnormal sudden heat field are aggregated to obtain the rectification priority value, and a supervision notice letter for the target department is generated based on the rectification priority value.

[0006] The beneficial effects of this invention are as follows: This invention effectively overcomes the routing deadlock and node congestion problems caused by rigid rule-based dispatching in traditional government work order systems. By transforming the non-standardized surge in demand flow into quantifiable momentum, and combining probabilistic cloud mapping and particle swarm optimization for global dynamic impedance, it achieves accurate distribution of work orders and balanced computing load across complex cross-domain departments. Simultaneously, relying on divergence calculation and deviation calculus monitoring, the system can sensitively capture clustered abnormal events at a very early stage and automatically issue targeted supervision letters, significantly improving the self-healing efficiency and dynamic response timeliness of urban governance workflows. Attached Figure Description

[0007] Figure 1 This is a module diagram of a closed-loop management system for handling requests on the 12345 platform according to the present invention. Detailed Implementation

[0008] The subject matter described herein will now be discussed with reference to exemplary embodiments. It should be understood that these embodiments are discussed only to enable those skilled in the art to better understand and implement the subject matter described herein, and changes may be made to the function and arrangement of the elements discussed without departing from the scope of this specification. Various processes or components may be omitted, substituted, or added as needed in the examples. Furthermore, features described in some examples may be combined in other examples.

[0009] like Figure 1 As shown, a closed-loop management system for handling complaints on the 12345 platform is configured to perform the following steps: The acoustic urgency and semantic information density of the input voice stream are extracted through the 12345 platform, and the request is converted into an initial work order quantification value. The spatiotemporal correlation score is calculated based on the initial work order quantification value, and merged work orders are generated based on the spatiotemporal correlation score. Process the merged work orders and output the probability distribution of the matching rights and responsibilities of the candidate departments; Based on the probability distribution of the matching of rights and responsibilities, iterative optimization is performed to determine the optimal distribution path to the target department; Track the work order processing rate of the target department and calculate the efficiency lag index based on the tracking results; Calculate the temporal rate of change of the distribution density of demands across the entire region, and generate an abnormal sudden heat field based on the temporal rate of change; The efficiency lag index and the abnormal sudden heat field are aggregated to obtain the rectification priority value, and a supervision notice letter for the target department is generated based on the rectification priority value.

[0010] Preferably, the step of extracting the acoustic urgency and semantic information density of the input voice stream through the 12345 platform and converting the request into an initial work order quantification value includes obtaining the initial work order quantification value using the following calculation formula:

[0011] In the formula, This represents the initial work order quantification value. This indicates the total duration of the input voice stream access. This indicates the current time frame of the scrolling time window. Indicates an instantaneous moment. Represents the integral element at an instant. This represents the semantic information density at a given instant. This indicates the acoustic urgency of the input speech stream at a given instant. This indicates the current scrolling time window and the time frame it represents. Average acoustic urgency background, This represents an exponential function with the natural constant as its base.

[0012] The input voice stream is the raw call voice data received through the 12345 platform's inbound call channel. It includes caller speech segments, pause segments, and environmental noise information, and can be collected through the 12345 platform's voice switch, recording server, session recording interface, and call recording files.

[0013] The total duration of voice stream access is the total duration of the input voice stream from the start of connection to the end of connection. It can be collected through call details, recording metadata, and switch session start and end timestamps.

[0014] The current rolling time window is a local time range used to statistically measure the average acoustic urgency background. The preferred value range is 30 to 90 seconds. This range can cover a relatively complete segment of the appeal, while avoiding an excessively long window that weakens the sensitivity of recognizing sudden emotional changes.

[0015] The current time of the scrolling time window is the time location corresponding to the current scrolling time window, which is used to determine the time position on which the background calculation is based.

[0016] An instantaneous moment is any specific point in time in the speech stream that is sampled or framed by the system, used to express the acoustic urgency and semantic information density at a certain moment.

[0017] The integral element at an instant is the smallest discrete time step used when integrating the speech stream over time. The preferred value range is 0.02 seconds to 0.1 seconds, which can balance the resolution of acoustic feature changes, system real-time performance, and computational resource consumption.

[0018] Semantic information density is the number or intensity of effective semantic units extracted from the content expressed by the caller per unit time. It is used to measure the concentration of event elements, location elements, object elements, and consequence elements contained in the appeal text within a certain period of time.

[0019] Acoustic urgency is a value representing the degree of urgency of a caller's voice by fusing multiple acoustic dimensions such as intensity, fundamental frequency fluctuation, speech rate, energy change, and emotional tension. It is used to reflect the caller's subjective sense of urgency at that moment.

[0020] The average acoustic urgency background is the average reference level of acoustic urgency within the current rolling time window, used to eliminate the bias caused by differences in the inherent vocal timbre of different speakers, differences in recording equipment, and differences in environmental noise on urgency determination.

[0021] The quotient is the relative amplification of the acoustic urgency at a given moment relative to the average acoustic urgency background, used to represent the degree of deviation of the current emotional urgency from the local normal level.

[0022] The exponential term is a nonlinear amplification result generated based on the degree of deviation of the quotient from the normal level. It is used to enhance the contribution of the current moment when the acoustic urgency is significantly higher than the background, and to reduce the contribution of the current moment when the acoustic urgency is lower than the background.

[0023] The product term is the instantaneous comprehensive contribution value formed by combining the semantic information density and the exponential term, which is used to simultaneously reflect the density of the appeal content and the urgency of the emotion.

[0024] The initial work order quantification value is the overall quantification result obtained by accumulating the instantaneous comprehensive contribution value of the entire voice stream during the call. It is used to characterize the comprehensive work order value of the request in terms of content intensity and emotional urgency.

[0025] In practice, the preprocessing of the input speech stream before it enters the calculation of acoustic urgency and semantic information density requires channel unification processing after the speech stream is input, which unifies the sampling rate to 8000 Hz or 16000 Hz and the bit depth to 16 bits. Then, echo cancellation, static noise estimation, spectral subtraction and noise reduction, and automatic gain control are performed. Then, speech activity detection is used to remove continuous silence segments and pure environmental noise segments, while retaining speech segments, breath segments, and short pause segments. For dual-channel recordings, the caller's main channel is extracted first and the operator's channel is used as an auxiliary correction signal.

[0026] In practice, regarding the division of the current rolling time window, instantaneous moment, and integral infinitesimal element of the instantaneous moment, a continuous time axis needs to be established starting from the moment the call is connected. An instantaneous moment is generated every 0.05 seconds, and each instantaneous moment corresponds to an analysis frame. The current rolling time window is formed by looking back 60 seconds from the current moment. When the total call length is less than 60 seconds, the already occurred call segment is taken as the window. Adjacent analysis frames can adopt an overlap ratio of half to three-quarters to improve the ability to capture short-term fluctuations. Boundary compensation is enabled 10 seconds before and 10 seconds after the call to avoid distortion of the average acoustic urgency background caused by window truncation.

[0027] In practice, the calculation of acoustic urgency requires extracting short-time energy, mean fundamental frequency, fundamental frequency jitter, speech rate, zero-crossing rate, spectral centroid, formant change rate, and emotional tension score from the analysis frame corresponding to each instant. Speaker-specific standardization and environmental noise compensation are then applied to these features. Short-time energy, fundamental frequency fluctuation, speech rate, and emotional tension are then weighted and fused as high-weight terms to output an acoustic urgency score in the range of 0 to 1. When three or more consecutive analysis frames simultaneously exhibit high intensity, high speech rate, and a significant rise in fundamental frequency, a burst enhancement coefficient can be added.

[0028] In practice, the calculation of semantic information density requires first feeding the speech segment into a Chinese speech recognition module to obtain a sentence-by-sentence transcribed text; then performing word segmentation, part-of-speech tagging, entity recognition, and event element extraction, identifying at least event elements, location elements, object elements, consequence elements, quantity elements, and time elements; assigning weights to different semantic units, with higher weights given to semantic units containing specific addresses, specific objects, and specific harmful consequences; and finally calculating the semantic information density based on the weighted total amount of effective semantic units per unit time.

[0029] In practice, the calculation of the average acoustic urgency background requires averaging the acoustic urgency of all valid speech frames within the current rolling time window, and using truncated mean or median smoothing compensation for extreme peaks; continuous silence segments, obvious background noise segments, and agent interruption segments are not included in the average acoustic urgency background; when the caller's emotions are naturally high or the equipment's pickup gain is too large, speaker baseline correction is used first to ensure that the average acoustic urgency background reflects the local normality of the call.

[0030] In practice, for the time integration of the product term over the total access time of the voice stream from zero, the product term of each analysis frame needs to be discretely accumulated in the order of instantaneous moments, and the integral element at each instantaneous moment is used as the time weight of the contribution of each frame. For frames identified as blank frames, severely noisy frames, and frames with low voice recognition confidence, the product term can be set to 0 or compensated by interpolation between two adjacent frames. When the call is transferred or access is resumed after a short interruption, the accumulated value is continuously accumulated under the same session identifier.

[0031] Preferably, the step of calculating the spatiotemporal relevance score based on the initial work order quantification value and generating a merged work order based on the spatiotemporal relevance score includes obtaining a first request and a second request from the requests, and deriving the spatiotemporal relevance score using the following calculation formula:

[0032] In the formula, This represents the spatiotemporal correlation score. These represent the numbers of the first claim and the second claim, respectively. This represents the initial work order quantification value corresponding to the first request. This represents the initial work order quantification value corresponding to the second request. This represents the spatial physical straight-line distance between the first claim and the second claim. This represents the absolute time difference between the first request and the second request for access. Indicates the current time of the global work order. This indicates that the current global work order is for a specific time period. The average velocity of spatiotemporal dynamic diffusion This represents the high-dimensional semantic cosine similarity between the first claim and the second claim.

[0033] The first request is one of the matching objects selected from the current set of requests to be analyzed on the 12345 platform. It is used to participate in the spatiotemporal correlation score calculation and can be collected through the 12345 platform's request acceptance database, work order pool, and call record table.

[0034] The second request is another object to be paired, which participates in the correlation calculation at the same time as the first request. It is used to determine whether the two requests belong to the same event or the same abnormal source. It can be collected through the 12345 platform's request acceptance database, work order pool, and call record table.

[0035] The numbers for the first and second requests are indexes used to identify the first and second requests, and can be collected through the 12345 platform work order primary key, acceptance serial number, and inbound session number.

[0036] The spatial physical straight-line distance is the actual straight-line distance between the corresponding geographical locations of the first and second demands, used to describe the degree of spatial proximity between the two demands.

[0037] The absolute time difference of access is the absolute difference between the first and second requests in terms of the platform access time. It is used to measure whether the two requests belong to the same event propagation process in terms of time.

[0038] The current time of the work order in the entire domain is the system time point corresponding to the assessment of the overall diffusion situation. It is used to associate the spatiotemporal dynamic diffusion average speed with the current operating status of the platform. It can be collected through the platform system clock, streaming processing timestamp, and work order entry timestamp.

[0039] The average diffusion velocity in the spatiotemporal dynamics is the average diffusion velocity exhibited by the current work orders across the entire domain during spatial propagation and temporal evolution. It is used to convert the absolute time difference of access into an equivalent diffusion distance in space.

[0040] High-dimensional semantic cosine similarity is a similarity index obtained based on the angle relationship between the semantic vectors of two claims. It is used to represent the textual similarity between two claims in terms of event topic, object type, location, and handling of the claim.

[0041] The quality product term is the product of the initial work order quantification value corresponding to the first demand and the initial work order quantification value corresponding to the second demand, which is used to amplify the mutual attraction between high-intensity demands.

[0042] Time-converted distance utilizes the spatiotemporal dynamic diffusion average velocity to convert the absolute time difference of access into an equivalent distance in a spatial sense, thus incorporating the time difference into a unified spatiotemporal distance framework.

[0043] The sum of squares of spatiotemporal distance is a comprehensive separation measure that combines the physical straight-line distance and the time-converted distance on the same distance scale. It is used to measure the overall degree of dispersion of two claims in spatiotemporal space.

[0044] The similarity squared term is a semantic amplification factor that enhances the high-dimensional semantic cosine similarity, and is used to further improve the credibility of merging when the two claims are highly consistent in semantics.

[0045] The spatiotemporal relevance score is an association strength index obtained by combining the initial work order quantification value, spatial distance, temporal distance and semantic similarity. It is used to determine whether two requests should be regarded as the same event cluster.

[0046] Merged work orders are unified work order entities formed by combining multiple highly related requests based on spatiotemporal relevance scores, in order to avoid duplicate dispatching, duplicate processing, and duplicate supervision.

[0047] In practice, the selection and matching of the first and second requests should take the new access request as the central request and search the preset candidate pool for historical unmerged requests that are adjacent to its access time and geographical scope. The candidate pool should prioritize requests within the most recent 30 to 120 minutes and expand layer by layer according to the same district, the same street, and the same grid. For mixed scenarios of text and voice requests, they should be uniformly converted into structured request objects before entering the matching process.

[0048] In practice, to obtain the spatial physical straight-line distance, it is necessary to extract the address text, location point, administrative division and landmark name from the first and second requirements respectively; firstly, address standardization, place name disambiguation and geocoding are performed, and then they are uniformly converted into planar coordinates or latitude and longitude coordinates under the same map coordinate system; after the coordinates are completed, the straight-line distance between the two points is calculated; when the address only has the name of the community or the name of the road section, the location is completed by using the place name center point plus the linear projection of the road.

[0049] In practice, to obtain the absolute time difference of access, the unified access timestamp of the 12345 platform should be used as the time reference, and priority should be given to reading the work order acceptance time, call connection time, or text submission time; records written across systems should be uniformly converted to the same time zone and the same time format; and second-level errors caused by network jitter should be eliminated by aligning with the most recent sampling period.

[0050] In practice, to calculate the average velocity of spatiotemporal dynamic diffusion, it is necessary to extract the diffusion trajectory of similar event clusters on different spatial grids within the statistical window before the current time of the work order in the entire domain; calculate the average velocity of spatiotemporal dynamic diffusion of the current work order in the entire domain based on the average propagation distance and time difference from the earliest grid to the subsequent newly added grids of the event cluster; to avoid distorting the results by individual extreme events, quantile truncation and moving average updates can be used.

[0051] In practice, for the calculation of high-dimensional semantic cosine similarity, the first and second requests need to be converted into unified text representations that include event type, location entity, object entity, harmful consequences and handling requests, respectively; then input into the Chinese government affairs semantic encoding model to obtain equal-length vectors; after normalizing the length of the vectors, the similarity of the included angles is calculated, and the high-dimensional semantic cosine similarity in the interval of 0 to 1 is output.

[0052] In practice, for the merging threshold and merging decision of spatiotemporal correlation scores, it is necessary to first establish a threshold table according to the major event categories, and then dynamically adjust the threshold based on the diffusion strength of the current global work order at the current time. When the spatiotemporal correlation score is higher than the high threshold, it is directly merged. When it is between the high threshold and the low threshold, it enters the secondary verification process. The verification content includes at least address consistency, main event consistency and acceptance result conflict check.

[0053] In practice, when generating merged work orders, the earliest access time and the most complete information should be selected from the merged requests as the main request, and its main event title, first report code and main address should be inherited. Supplementary addresses, supplementary objects, supplementary photos and supplementary call information of other requests should be written into the supplementary evidence list. For requests that have not been dispatched synchronously by the responsible departments, the status should be locked first, and then the probability distribution of the matching of rights and responsibilities should be recalculated uniformly by the merged work orders.

[0054] Preferably, processing the merged work order and outputting the power and responsibility matching probability distribution of the candidate departments includes obtaining all candidate departments and using the following calculation formula to obtain the target power and responsibility matching probability degree of the target candidate department among the candidate departments:

[0055] In the formula, This represents the probability degree of target responsibility matching for the target candidate department. This refers to the target candidate department. This represents the deep feature vector of the merged work order. The embedding tensor represents the statutory function graph of the target candidate department. This represents the feature dimension scalar of the embedding space in which the statutory function graph embedding tensor resides. This represents the total number of all candidate departments. This refers to any one of the candidate departments among all the candidate departments. This represents the embedding tensor of the statutory function graph of any of the candidate departments. This represents an exponential function with the natural constant as its base.

[0056] All candidate departments are a set of departments that can handle the matter, selected based on the subject of the request, administrative level, statutory responsibilities, and business acceptance boundaries.

[0057] The deep feature vector of a merged work order is a multi-dimensional representation formed by jointly encoding the event type, location, object, severity, appeal direction, and historical handling semantics of the merged work order, and is used as input to the department matching model.

[0058] The embedding tensor of the statutory function graph of each candidate department is a structured representation obtained by mapping the statutory responsibilities, hierarchical relationships, collaborative relationships and prohibited boundaries of the department to a vector space, which is used to characterize the scope of authority, responsibility and capability of each department.

[0059] The feature dimension scalar of the embedding space in which the statutory function graph embedding tensor resides is a model dimension parameter used to determine the length of the department embedding representation. It is used to control the balance between semantic expressive power and computational complexity. The preferred value range is 128 to 512, which can better carry the semantic hierarchy of government responsibilities and facilitates real-time inference in actual platforms.

[0060] The target candidate department is the specific candidate department whose power and responsibility matching probability is currently being calculated, and is used to output the probability of that department's affiliation among all candidate departments.

[0061] The statutory function graph embedding tensor of the target candidate department is a target department-specific representation extracted from the set of responsibility embeddings of all candidate departments, and is used to match it with the deep feature vector of the merged work order.

[0062] The target inner product is the original similarity score obtained by matching the deep feature vector of the merged work order with the statutory function map embedding representation of the target candidate department. It is used to measure the degree of fit between the request and the target candidate department.

[0063] The target exponent term is a positive value obtained by normalizing the dimension of the target inner product and performing an exponential transformation. It is used to participate in the probability normalization calculation.

[0064] The total number of all candidate departments is the number of candidate departments currently included in the probability distribution calculation, used to determine the traversal range and normalization scale.

[0065] Any candidate department is any member of the set of all candidate departments, used to traverse all possible processing departments during the normalization process.

[0066] The statutory function graph embedding tensor of any candidate department is the corresponding responsibility embedding representation of any candidate department, which is used to generate the global exponent term and construct the normalized denominator.

[0067] The global index term is a positive value obtained by exponentially transforming the normalized matching score between any candidate department and the merged work order, and is used to reflect the relative competitive intensity of the candidate department.

[0068] The global index and term are the sum of the global index terms corresponding to all candidate departments, used to normalize the original matching strength into a probability distribution.

[0069] The target responsibility matching probability is the probability value of the target candidate department among all candidate departments, which is used to indicate the relative suitability of the department to undertake the merged work order.

[0070] The probability distribution of matching rights and responsibilities is a probability vector formed over all candidate departments.

[0071] In practice, to generate all candidate departments, an initial set of departments needs to be matched based on the event type, administrative level, geographical jurisdiction, and statutory responsibilities in the merged work order. Then, departments that are currently inactive, have no authority to accept cases, cross-jurisdictional departments, and departments that are obviously inconsistent with the event type are removed. For scenarios that require joint processing, multiple candidate departments can be retained at the same time, and the host attribute or co-host attribute can be recorded for each department.

[0072] In practice, the construction of the deep feature vector for merged work orders requires first extracting the event type, location, object, degree of harm, appeal direction, historical handling results, and attachment summary in a structured manner; then concatenating the structured fields with the original text summary into a unified input sequence; after inputting into the Chinese government affairs semantic model, sentence vectors and field-level vectors are obtained, and the deep feature vector of the merged work order is obtained through attention fusion.

[0073] In practice, the construction of the embedded tensor of the statutory function graph for each candidate department requires the establishment of a government affairs responsibility knowledge graph based on the departmental responsibility list, the three-fixed plan, the item list, the collaboration rules, the prohibited boundaries, and the superior-subordinate relationship; departments, items, jurisdictions, legal basis, and collaboration relationships are used as nodes and edges; then, the knowledge graph is mapped into a unified vector representation using the graph embedding training method; and each time there is a regulation adjustment, an item addition, or a jurisdiction change, it is synchronously retrained or incrementally updated.

[0074] In specific implementation, for the selection of feature dimension scalars in the embedding space where the statutory function graph embedding tensor is located, offline comparative experiments need to be conducted in the range of 128 to 512 at 128, 256, 384 and 512. The department matching accuracy, joint processing recall rate, real-time inference latency and memory usage are used as joint evaluation indicators. When the platform is centrally deployed at the municipal level and there are many departments, 256 or 384 is preferred. When the platform is lightly deployed at the district level, 128 or 256 is preferred.

[0075] In practice, for the calculation of the target inner product, target exponent, global exponent, and global exponent sum, it is necessary to check whether the length of the depth feature vector of the merged work order is consistent with the length of the embedding tensor of the statutory function map of each candidate department before the calculation; apply a negative infinity mask to departments without authority and departments across jurisdictions or remove them directly, so that they do not participate in the calculation of the global exponent sum; for excessively large target inner product values, subtract the maximum value of all candidates before performing exponent transformation to avoid numerical overflow.

[0076] In practice, for the output of the probability distribution of matching rights and responsibilities, it is necessary to output a probability vector of all candidate departments in descending order of probability, and at the same time provide suggestions for the preferred department, the secondary department, and joint handling. When the probability of matching the target rights and responsibilities of the preferred department is higher than the preset host threshold and the difference with the secondary department is large enough, it is directly determined to be handled by a single department. When the probabilities of multiple departments are close and the event itself has cross-functional attributes, a joint handling plan of host and co-host is generated.

[0077] Preferably, the step of iteratively optimizing based on the power-responsibility matching probability distribution to determine the optimal dispatch path to the target department includes obtaining all candidate departments and using the following calculation formula to derive the particle evolution adaptive inertia weight and the expected flow time fitness:

[0078]

[0079] In the formula, This represents the adaptive inertial weights of particle evolution. Indicates the current moment. This represents the standard deviation of the real-time backlog of work orders for all candidate departments. This represents the average number of backlogged work orders in real time for all candidate departments. This represents an exponential function with the natural constant as its base. This indicates the fitness of the expected circulation time. This represents the business weight distribution matrix. This represents the total number of all candidate departments. This refers to any one of the candidate departments among all the candidate departments. This indicates the business weight ratio corresponding to any of the candidate departments. This represents the real-time backlog of work orders for any of the candidate departments. This represents the average case closure throughput of any of the candidate departments. This represents the probability degree of matching rights and responsibilities for any candidate department obtained from the power and responsibility matching probability distribution.

[0080] The target department is the department that is ultimately determined or closely monitored based on the previous assignment calculations.

[0081] The standard deviation of the real-time backlog of work orders for all candidate departments is a statistic on the dispersion of the current backlog of work orders for each candidate department, which is used to reflect whether the current load on the platform is balanced.

[0082] The average number of backlogged work orders in real time for all candidate departments is a statistical measure of the average number of backlogged work orders in each candidate department, and is used as a benchmark for standard deviation normalization.

[0083] The ratio is the relative dispersion of the standard deviation of the real-time backlog of work orders for all candidate departments relative to the mean of the real-time backlog of work orders, and is used to measure the overall congestion imbalance of the system.

[0084] The particle evolution adaptive inertia weight is a dynamic weight used to adjust the degree to which the particle's historical motion is preserved during the optimization process, in order to establish a balance between global exploration and local convergence.

[0085] The current time is the system time point corresponding to the execution of dispatch path optimization. It is used to bind the load status and optimization process to the same running time and can be collected through the platform system clock, task scheduling timestamp, and process engine running timestamp.

[0086] The business weight ratio is the flow weight assigned to any candidate department, which is used to indicate the size of the work order flow share that candidate department undertakes in a round of work order dispatch decision.

[0087] The real-time backlog of work orders for any candidate department is the number of pending work orders that the department has not yet completed at the current moment. It is used to estimate the degree of queuing congestion and can be collected through the 12345 platform work order master database, department pending queue list and process engine status table.

[0088] The average case closure throughput of any candidate department is the number of work orders that the department can stably complete per unit time within the statistical period, which is used to characterize the department's actual processing capacity.

[0089] The expected queue clearing time is the estimated time required for any candidate department to process all current backlogged work orders, reflecting the waiting cost for that department after accepting a new work order.

[0090] Local workflow time is the single-department workflow cost obtained by combining the business weight ratio, expected queue clearing time, and probability of matching authority and responsibility. It is used to represent the local time cost of assigning a work order to that department.

[0091] The expected circulation time fitness is the overall optimization target value after the sum of the local circulation times of all candidate departments.

[0092] The business weight distribution matrix is ​​a matrix structure that uniformly expresses the business allocation weights of all candidate departments at one or more flow levels.

[0093] The optimal dispatch path is a departmental workflow scheme that minimizes the expected workflow time adaptability. It is used to guide work orders to be optimally dispatched from the current node to the target department or collaborating departments.

[0094] In practice, to calculate the standard deviation of the real-time backlog of work orders for all candidate departments and the average of the real-time backlog of work orders for all candidate departments, it is necessary to read the current number of pending work orders for all candidate departments at a unified refresh cycle, with a refresh cycle preferably ranging from 1 minute to 5 minutes. After reading the results, abnormal work orders that have been withdrawn, suspended, or repeatedly returned should be removed. Then, the average of the real-time backlog of work orders for all candidate departments and the standard deviation of the real-time backlog of work orders for all candidate departments should be calculated.

[0095] In specific implementation, for the algorithm process of particle evolution adaptive inertia weight participating in optimization iteration, it is necessary to define a particle as a candidate solution of a set of business weight ratio distribution matrices; initialize the particle position, velocity, individual optimal value and global optimal value; in each round of iteration, first update the particle evolution adaptive inertia weight according to the standard deviation and mean of the real-time backlog of work orders of all candidate departments, and then update the particle velocity and position by combining the individual optimal term and the group optimal term; stop in advance when the improvement of the expected flow time fitness is insufficient for several consecutive rounds.

[0096] In practice, to address the constraints of the business weight ratio and the business weight ratio distribution matrix, a weight vector needs to be set for each work order to be dispatched, with each column corresponding to a candidate department; the weight of each column should not be less than 0, and the total weight of the same work order in the same row should be equal to 1; a minimum weight lower limit should be set for the lead department, and an upper limit should be set for the cooperating departments to prevent the responsibility from being unclear due to the average distribution among all departments.

[0097] In practice, to calculate the average case closure throughput of any candidate department, it is necessary to select the completion records of similar matters within the most recent 7 to 30 days, and to separately calculate the number of cases completed per unit time according to working days and holidays; to truncate abnormal peaks caused by centralized cancellation, historical supplementation, and batch archiving in the system; and then to use a moving average or exponential smoothing to obtain the average case closure throughput of any candidate department.

[0098] In practice, to solve for the expected processing time adaptability, it is necessary to calculate the local processing time for each candidate department based on the business weight ratio, expected queue clearing time, and probability of matching authority and responsibility, and then sum them to obtain the overall cost; at the same time, hard constraint penalty items are superimposed. The hard constraints include at least the legality of the jurisdiction, the legality of the authority, the upper limit of the number of co-organizers, and the uniqueness of the host; when a candidate department violates the hard constraints, its corresponding candidate solution is directly judged as invalid or given a high penalty.

[0099] In practice, for the output of the optimal dispatch path, the optimized business weight distribution matrix needs to be decoded into a specific flow link, at least outputting the first dispatching department, the cooperating department, whether it needs to be pushed to the next higher level department, and whether it needs to be copied to the local unit simultaneously; when the weight of the first-choice department is higher than the single dispatch threshold, a single-department optimal dispatch path is formed; when the weights of two or more departments reach the coordination threshold, a multi-department optimal dispatch path is formed, consisting of the host and cooperating departments.

[0100] Preferably, the step of tracking the work order processing rate of the target department and calculating the efficiency lag index based on the tracking results includes deriving the efficiency lag index using the following calculation formula:

[0101] In the formula, This indicates the performance lag index. Indicates the target department. Indicates the instantaneous tracking moment. This represents the real-time backlog change rate of the target department. This represents the average work order processing rate across the entire domain. This indicates the work order processing rate of the target department. A constant representing extremely small positive numbers to prevent the denominator from being zero. This represents the natural logarithm function.

[0102] The instantaneous tracking moment is the specific point in time when the system samples or refreshes the operating status of the target department.

[0103] The real-time backlog queue change rate is the rate at which the pending work order queue of a target department increases or decreases within a unit of time, reflecting the changing trend of the current congestion status of the target department.

[0104] The overall average work order processing rate is the average completion speed of all departments under the same statistical caliber, and is used as a reference baseline to measure whether the target department is lagging behind the overall level.

[0105] Work order processing rate is the number of work orders that a target department actually processes or completes within a unit of time, reflecting the department's current actual execution capability.

[0106] The minimal positive number to prevent zero value in the denominator is a protective constant set to avoid instability of the denominator caused by the work order processing rate being zero or close to zero. The preferred value range is 0.01 to 0.1. This range can avoid zero value in the denominator and numerical explosion, and will not significantly distort the difference between normal processing rates.

[0107] The corrected work order processing rate is a robust rate value obtained by adding the work order processing rate to the protection constant, which is used to avoid zero denominators and extreme amplification.

[0108] The rate ratio is the ratio of the average work order processing rate across the entire region to the corrected work order processing rate. It is used to measure the relative speed between the target department and the overall level.

[0109] The logarithmic deviation term is the deviation obtained by performing a logarithmic transformation on the rate ratio. It is used to compress extreme rate differences while retaining information about the direction of speed.

[0110] The efficiency lag index is a comprehensive indicator that couples the real-time backlog queue change rate with the logarithmic deviation term. It is used to simultaneously reflect the backlog deterioration trend and the degree of deviation in processing efficiency of the target department.

[0111] In practice, for the statistics of instantaneous tracking time, work order processing rate and overall average work order processing rate, the status of each department needs to be refreshed at a fixed sampling period, with the sampling period preferably ranging from 1 minute to 5 minutes; the work order processing rate is obtained by dividing the actual number of completed work orders in the most recent sliding window by the window duration; the overall average work order processing rate is obtained by dividing the total number of completed work orders in the same window of all departments by the number of departments or the total effective working time.

[0112] In practice, to calculate the change rate of the real-time backlog queue, it is necessary to record the difference in the number of pending work orders between two adjacent instantaneous tracking moments, and then divide it by the time interval between the two sampling moments. To reduce jitter, a moving average or median filter can be used for the most recent 3 to 5 sampling points. When the sampling interval is abnormally long or there are missing points, time alignment and interpolation compensation are performed first.

[0113] In practice, for setting the constant to prevent zero value in the denominator for extremely small positive numbers, the candidate range should be one-thousandth to one-hundredth of the average work order processing rate across the entire domain, and then the final value should be selected in combination with the historical minimum non-zero work order processing rate. When the overall processing rate of the platform increases or decreases significantly, it can be recalibrated monthly or quarterly. For ultra-low frequency items, the constant to prevent zero value in the denominator for extremely small positive numbers can be appropriately increased to avoid abnormal amplification when the work order processing rate is close to zero.

[0114] In practical implementation, the interpretation of the logarithmic deviation term and the efficiency lag index should be as follows: when the work order processing rate of the target department is lower than the average work order processing rate of the entire region, the logarithmic deviation term should be positive, indicating that there is efficiency lag; when the work order processing rate of the target department is higher than the average work order processing rate of the entire region, the logarithmic deviation term should be negative, indicating that efficiency is leading. Upper and lower cutoff boundaries can be set for extreme large and small values ​​to prevent a single sampling point from causing distortion of the efficiency lag index.

[0115] In practice, for the alarm classification of performance lag indicators, three threshold ranges of mild, moderate and severe need to be set according to historical distribution. When the performance lag indicator is in the mild range for several consecutive sampling periods, a prompt is triggered. When it enters the moderate range, a departmental warning is initiated. When it enters the severe range, it is added to the supervision candidate pool. At the same time, it is combined with the abnormal and sudden heat field to determine whether it is necessary to upgrade the handling.

[0116] Preferably, the step of calculating the time-series change rate of the global demand distribution density and generating an abnormal sudden heat field based on the time-series change rate includes obtaining the abnormal sudden heat field using the following calculation formula:

[0117] In the formula, This refers to the aforementioned abnormal sudden thermal field. Represents the target space grid coordinates. Indicates time, This represents the overall demand distribution density. This represents taking the second-order partial derivative with respect to time. This represents the time-series rate of change of the distribution density of the global demands. This represents the variance of the normal distribution across the entire region. This represents the mean square background of the normal distribution across the entire region.

[0118] The target spatial grid coordinates are the location identifiers of a grid unit after the urban governance space is discretized. They are used to locate the spatial areas where abnormal demands gather. The three-level coding of district / county number, street number and basic grid number is preferred. The side length of the basic grid is preferably 250 meters, and the preferred value range is 200 meters to 500 meters. This scale can cover the scope of community governance events and avoid the hotspots being diluted due to the grid being too large.

[0119] Time is a unified time variable used to describe the evolution of the distribution density of demands across the entire domain, and is used to express the dynamic trend of demand density changing over time.

[0120] The density of demands across the entire city is the degree of demand aggregation obtained by statistical analysis per unit area or per unit grid within each spatial grid. It is used to reflect the strength of the spatial distribution of demands across the entire city.

[0121] The temporal rate of change of global demand distribution density is an acceleration measure of the change of global demand distribution density over time, used to identify the difference between ordinary fluctuations and sudden surges.

[0122] The global normal distribution variance is a statistical measure of the dispersion of the distribution of demands in each grid under historical normal conditions, and is used to characterize the historical background level of spatial heterogeneity.

[0123] The mean square background of the global normal distribution is the mean square reference value of the distribution intensity under historical normal conditions, which is used to normalize the variance of the global normal distribution.

[0124] The spatial variation coefficient is a normalized fluctuation coefficient composed of the variance of the global normal distribution and the mean square background of the global normal distribution. It is used to reflect the degree of prior imbalance in the current urban spatial distribution.

[0125] The abnormal outbreak heat field is a spatial heat result formed by coupling the temporal change rate of the distribution density of demands across the entire region with the spatial variation coefficient term. It is used to represent the intensity distribution of an abnormal demand outbreak at a certain time and place.

[0126] In practice, for the division of target spatial grid coordinates, a unified spatial grid system needs to be established first according to the urban governance map, with the basic grid side length preferably set to 200 to 500 meters; then each grid is coded into a three-level code consisting of district / county number, street number, and basic grid number; for special areas such as crossing rivers, elevated roads, and large industrial parks, thematic grid patches can be overlaid.

[0127] In practice, to calculate the distribution density of demands across the entire region, it is necessary to count the number of demands corresponding to the grid coordinates of each target space in each sampling period, and normalize them by combining grid area, population weight or facility density weight; demands from voice, web pages, applications and offline transcriptions are uniformly entered into the database and counted according to the same standard; duplicate demands can be processed by reducing their weight as they have been merged and marked.

[0128] In practice, to achieve the time-series change rate of the global demand distribution density, it is necessary to record the global demand distribution density at continuous time intervals with a fixed sampling interval; use the central difference method or the three-point difference method to approximate the second-order time change; and perform a moving average or low-pass filtering on the original density sequence before calculation to remove occasional spike noise.

[0129] In practice, to establish the baselines for the variance of the normal distribution and the mean square background of the normal distribution across the entire region, it is necessary to select historical periods without major emergencies within the past 30 to 180 days as normal samples; establish stratified sample databases according to weekdays, holidays, daytime and nighttime; calculate the dispersion and mean square reference value of the distribution of demands across the entire region for each time slice in the sample database, and then continuously update the variance of the normal distribution and the mean square background of the normal distribution across the entire region.

[0130] In practice, to determine and connect abnormal sudden heat fields, it is necessary to first classify the abnormal sudden heat field values ​​of each grid into thresholds, and regard grids below the basic threshold as background noise; then perform connected domain clustering on adjacent high-value grids to identify the range, peak value and duration of continuous heat areas; when the same heat area lasts for more than a preset time or spans multiple administrative units, it is marked as a key abnormal heat area.

[0131] Preferably, the step of aggregating the performance lag indicator with the abnormal sudden heat field to obtain a rectification priority value, and generating a supervision notice for the target department based on the rectification priority value, includes obtaining the initial monitoring time and the current time corresponding to the performance lag indicator, and using the following calculation formula to obtain the rectification priority value:

[0132] In the formula, This indicates the rectification priority value. Indicates the target department. This indicates the initial monitoring time corresponding to the performance lag indicator. This indicates the current time at which the data was obtained. This indicates the assessment time for a sudden heat field. Denotes the moment of the integral infinitesimal element. Denotes an integral infinitesimal element. The performance lag index represents the target department at the integral infinitesimal time. This represents the norm of the abnormal sudden heat field associated with the target department. This represents the historical average heat field background under conditions of no sudden events across the entire region. Represents the natural constant. This represents the natural logarithm function.

[0133] The initial monitoring time is the starting point at which the system begins to continuously track the performance lag of a specific target department. This can be collected through monitoring task creation records, rule trigger logs, and departmental alarm status tables.

[0134] The integral infinitesimal moment is a discrete sampling time point used when integrating the performance lag index over time, and is used to represent each cumulative node within the monitoring period.

[0135] The integral infinitesimal is the smallest discrete time step when integrating the execution time of the performance lag indicator. The preferred value range is 1 minute to 5 minutes, which can take into account the actual rhythm of monitoring sensitivity, computational overhead and departmental state fluctuations.

[0136] The assessment time for a sudden heat field is the time point at which the system reads and assesses the heat status of the target department.

[0137] The cumulative lag indicator is the cumulative result of the performance lag indicators of the target department at each time point within the monitoring period, used to reflect the total amount of continuous lag of the department over a period of time.

[0138] The norm of the abnormal sudden heat field associated with the target department is the result of compressing the overall intensity of the abnormal sudden heat field within the jurisdiction of the target department or the scope of the matter into a single scalar, which is used to characterize the intensity of external sudden pressure faced by the department.

[0139] The historical average heat field baseline under no-emergency conditions is a heat reference baseline obtained from long-term statistics during the normal operation of the city, used to measure whether the current emergency heat is significantly higher than the normal level.

[0140] The heat ratio is the amplification of the norm of the abnormal sudden heat field associated with the target department relative to the historical average heat field background, and is used to characterize the degree of abnormality of the current external sudden pressure.

[0141] The summation term is the logarithmic input obtained by adding the normal protection term to the heat ratio.

[0142] The sudden penalty coefficient is the penalty intensity obtained by performing a logarithmic transformation on the summation term. It is used to smoothly superimpose external sudden pressures onto the rectification priority.

[0143] The rectification priority value is a comprehensive priority result formed by coupling the cumulative lag indicator item and the sudden penalty coefficient item, which is used to determine whether the supervision action is triggered and the intensity of the triggering.

[0144] The AI ​​big data model is a preset text generation model used to automatically generate the main text of a supervision notice based on the rectification priority value, department information, event summary and supervision rules. The optimal range of parameter size is 7 billion to 30 billion, which can achieve a balance between the quality of Chinese government writing, factual binding ability, deployment cost and proofreading efficiency.

[0145] It should be noted that the training samples for the large-scale artificial intelligence model consist of structured and unstructured text from all historically completed supervised events handled by the 12345 government service platform. The core samples cover original official supervision notices from different types of urban governance events, different supervision levels, and responsible departments at different administrative levels; core event summaries of corresponding merged work orders; details of the requests; legal authority and responsibility information of the responsible departments; monitoring data on performance lag indicators; correlation data of abnormal and sudden hot topics; and complete records of historical rectification and feedback processes. Additionally, the training samples include the departmental "three-fixed" plans, lists of powers and responsibilities, and legally mandated official document formats from various government information disclosure departments. Standardized formats and government supervision management norms serve as supplementary training samples, with a total sample size of no less than 100,000 entries. These samples are evenly distributed and stratified according to event type, administrative level, supervision level, and regional attributes to ensure comprehensive coverage of the 12345 platform's supervision needs across all scenarios. The sample labels, corresponding one-to-one with the training samples, consist of the full text of officially effective supervision notices that have been manually reviewed and approved by government departments. Each sample is also simultaneously labeled with event type, supervision level, responsible department's authority and responsibility boundaries, document format compliance, and factual accuracy. This labeling system strictly adheres to the legal framework for government supervision documents. The standards and business management rules of the 12345 platform are aligned to ensure complete alignment between the tags and the business logic and control requirements of the supervision scenarios. The large-scale model training uses a multi-task joint weighted loss function. The main loss function is the cross-entropy loss function, used to constrain the fluency, semantic coherence, and content completeness of the supervision notification letter's text generation. Auxiliary loss functions include a factual consistency loss function, a document format compliance loss function, and a supervision level matching loss function. The factual consistency loss function constrains the generated content against the input work order factual elements, departmental authority and responsibility information, performance monitoring data, and other relevant factors. The perfect matching of constant heat field data eliminates factual deviations and misalignment of rights and responsibilities. The document format compliance loss function is used to constrain the generated text to strictly comply with the legal format specifications and writing requirements of government supervision letters. The supervision level matching loss function is used to constrain the generated rectification requirements, processing time limits, supervision intensity and input rectification priority values ​​to correspond accurately. The overall loss function is the weighted sum of each loss item. The weights of each item are dynamically adjusted according to the business priority of the government supervision scenario to ensure that the supervision notice letter generated by the large model not only complies with the legal document specifications, but also adapts to the control logic and business needs of the 12345 platform's closed-loop handling of requests.

[0146] The supervision notice is a formal supervision document automatically generated by the system for the target department. It is used to convey abnormal situations, rectification requirements, processing time limits, responsibility boundaries, and feedback requirements.

[0147] In practice, the selection and updating of the initial monitoring time, current time, and integral microelement should be carried out as follows: when the target department enters the early warning pool for the first time or undertakes a key work order, the time should be recorded as the initial monitoring time; then the current time should be refreshed according to a fixed monitoring cycle, and the monitoring cycle should be kept consistent with the integral microelement; if the department leaves the monitoring pool for more than a preset time and then re-enters, a new initial monitoring time should be generated.

[0148] In practice, the calculation of the cumulative lag index requires reading the efficiency lag index of each sampling point in the monitoring period in the order of integral elements, multiplying the value of each sampling point by the corresponding time step, and then summing them up. For gaps caused by monitoring interruption, system maintenance, or sampling loss, linear interpolation between two adjacent points or short-term delay based on the most recent valid value can be used. If the interruption time exceeds the preset threshold, the current round of integration is disconnected and restarted after recovery.

[0149] In practice, the calculation of the norm of the abnormal sudden heat field associated with the target department requires first determining the set of heat zones associated with the department based on the jurisdiction boundaries, matter acceptance boundaries, and scope of collaborative responsibilities of the target department; then extracting the abnormal sudden heat field values ​​of the corresponding grids, and compressing them into a single scalar using the square root of the sum of squares, weighted absolute value summation, or a mixed norm of maximum value plus mean; for jointly handled matters, the heat zone intensity can be weighted according to the responsibility weight of the lead agent and the responsibility weight of the co-agent.

[0150] In practice, to establish the baseline of the historical average heat field under the condition of no emergencies in the whole region, it is necessary to select time periods without red alerts, major public opinion events and batch supervision records from the historical sample database as samples without emergencies; to calculate the average intensity of the abnormal emergency heat field at each moment in the sample, and to maintain the baseline according to season, time period and urban functional area respectively; and to recalibrate when the urban governance structure or the scale of demands changes significantly.

[0151] In practice, regarding the classification and action mapping of rectification priority values, it is necessary to first divide the rectification priority values ​​into observation level, early warning level, supervision level, and urgent supervision level based on the historical supervision results. The observation level only generates internal prompts, the early warning level pushes to the department head, and the urgent supervision level directly triggers a supervision notice and simultaneously sends a copy to the superior authority. When the rectification priority value rises continuously or stays at a high level for a long time, the frequency of supervision and the feedback time limit requirements are automatically increased.

[0152] In practice, to generate a supervisory notice using the AI-powered large model, the following elements need to be assembled into a structured input: priority value for rectification, name of the target department, event summary, hot zone range, reasons for delays in efficiency, processing time limit, responsibility boundary, previous feedback, and current supervisory requirements. Then, a special government document instruction template is loaded, requiring the output of title, body, rectification requirements, completion time limit, feedback path, and signature elements. After generation, the notice undergoes factual constraint verification, sensitive word verification, date verification, and responsibility boundary verification. Finally, the rules engine determines whether to issue the notice automatically or require manual review.

[0153] The embodiments of this example have been described above. However, this example is not limited to the specific implementation methods described above. The specific implementation methods described above are merely illustrative and not restrictive. Those skilled in the art can make many other forms based on the guidance of this example, and all of them are within the protection scope of this example.

Claims

1. A closed-loop management system for handling complaints on the 12345 platform, characterized in that, The system is configured to perform the following steps: The acoustic urgency and semantic information density of the input voice stream are extracted through the 12345 platform, and the request is converted into an initial work order quantification value. The spatiotemporal correlation score is calculated based on the initial work order quantification value, and merged work orders are generated based on the spatiotemporal correlation score. Process the merged work orders and output the probability distribution of the matching rights and responsibilities of the candidate departments; Based on the probability distribution of the matching of rights and responsibilities, iterative optimization is performed to determine the optimal distribution path to the target department; Track the work order processing rate of the target department and calculate the efficiency lag index based on the tracking results; Calculate the temporal rate of change of the distribution density of demands across the entire region, and generate an abnormal sudden heat field based on the temporal rate of change; The efficiency lag index and the abnormal sudden heat field are aggregated to obtain the rectification priority value, and a supervision notice letter for the target department is generated based on the rectification priority value.

2. The closed-loop management system for handling complaints on the 12345 platform according to claim 1, characterized in that, The process of extracting the acoustic urgency and semantic information density of the input voice stream through the 12345 platform and converting the request into an initial work order quantification value includes: Obtain the total access duration of the input voice stream; Extract the acoustic urgency of the input speech stream at each instant, and obtain the semantic information density at each instant; Determine the average acoustic urgency background within the current rolling time window; Divide the acoustic urgency at the instant by the average acoustic urgency background to obtain a quotient. Use the difference between the quotient and one as the exponent of the natural constant to calculate the exponent term. The semantic information density at the instant is multiplied by the exponential term to obtain the product term; The product term is integrated over time from zero to the total duration of the voice stream access to obtain the initial work order quantization value.

3. The closed-loop management system for handling complaints on the 12345 platform according to claim 1, characterized in that, The step of calculating the spatiotemporal correlation score based on the initial work order quantification value, and generating merged work orders based on the spatiotemporal correlation score, includes: Obtain the first request and the second request from the requests, and obtain the initial work order quantification value corresponding to the first request and the initial work order quantification value corresponding to the second request; Obtain the spatial physical straight-line distance and the absolute time difference of access between the first request and the second request; Obtain the average spatiotemporal diffusion rate of the current global work orders; Obtain the high-dimensional semantic cosine similarity between the first claim and the second claim; Multiply the initial work order quantification value corresponding to the first request by the initial work order quantification value corresponding to the second request to obtain the quality product term; Multiply the average velocity of the spatiotemporal dynamic diffusion by the absolute time difference of access to obtain the time-converted distance; Add the square of the physical straight-line distance in space to the square of the time-converted distance to obtain the sum of squares of the spacetime distance; The high-dimensional semantic cosine similarity is squared to obtain the similarity squared term. The spatiotemporal correlation score is calculated by dividing the mass product term by the spatiotemporal distance sum of squares term and multiplying the resulting quotient by the similarity squares term. The first request and the second request are merged based on the spatiotemporal correlation score to generate the merged work order.

4. The closed-loop management system for handling complaints on the 12345 platform according to claim 1, characterized in that, The process of processing the merged work order and outputting the probability distribution of the matching rights and responsibilities of the candidate departments includes: Obtain all candidate departments, and obtain the deep feature vector of the merged work order and the statutory function map embedding tensor of each candidate department; Obtain the feature dimension scalar of the embedding space in which the legal function graph embedding tensor resides; For the target candidate department among all the candidate departments, the inner product operation is performed between the depth feature vector of the merged work order and the legal function map embedding tensor of the target candidate department to obtain the target inner product value; The target inner product value is divided by the square root of the feature dimension scalar, and the resulting quotient is used as the exponent of the natural constant to calculate the target exponent term. Traverse all the candidate departments, calculate the inner product of the depth feature vector of the merged work order and the embedding tensor of the legal function map of the candidate department, divide it by the square root and use it as the natural constant exponent to obtain the global index term corresponding to each candidate department, and add all the global index terms to obtain the global index sum term; Divide the target index term by the global index term to output the target responsibility matching probability degree of the responsibility matching probability distribution belonging to the target candidate department.

5. A closed-loop management system for handling complaints on the 12345 platform according to claim 1, characterized in that, The step of iteratively optimizing based on the probability distribution of the matching rights and responsibilities to determine the optimal distribution path to the target department includes: Obtain the standard deviation of the real-time backlog of work orders for all candidate departments and the mean of the real-time backlog of work orders for all candidate departments; The ratio is obtained by dividing the standard deviation of the real-time backlog of work orders of all candidate departments by the mean of the real-time backlog of work orders of all candidate departments. The negative of the ratio is used as the exponent of the natural constant to calculate the particle evolution adaptive inertia weight. For any candidate department among all the candidate departments, obtain the business weight ratio corresponding to the candidate department, the real-time backlog of work orders of the candidate department, and the average case closure throughput of the candidate department; Obtain the power and responsibility matching probability degree of any candidate department from the power and responsibility matching probability distribution; Divide the real-time backlog of work orders in any candidate department by the average case closure throughput of any candidate department to obtain the expected queue clearing time. Multiply the business weight ratio corresponding to any candidate department, the expected queue clearing time, and the reciprocal of the power-responsibility matching probability of any candidate department to obtain the local flow time corresponding to any candidate department. The local processing time corresponding to any one of the candidate departments is summed to calculate the expected processing time fitness. The optimization iteration is performed using the particle evolution adaptive inertia weight and the expected flow time fitness to obtain the business weight ratio distribution matrix with the minimum expected flow time fitness, and the business weight ratio distribution matrix is ​​used as the optimal dispatch path to the target department.

6. A closed-loop management system for handling complaints on the 12345 platform according to claim 1, characterized in that, The tracking of the work order processing rate of the target department, and the calculation of the efficiency lag index based on the tracking results, includes: Obtain the real-time backlog queue change rate of the target department; Obtain the work order processing rate of the target department and the average work order processing rate across the entire domain, obtained from the tracking. The corrected work order processing rate is obtained by adding the work order processing rate of the target department to the constant of the smallest positive number to prevent the denominator from being zero. Divide the global average work order processing rate by the corrected work order processing rate to obtain the rate ratio. Performing a natural logarithmic operation on the rate ratio yields the logarithmic deviation term; The efficiency lag index is calculated by multiplying the real-time backlog queue change rate of the target department by the logarithmic deviation term.

7. A closed-loop management system for handling complaints on the 12345 platform according to claim 1, characterized in that, The calculation of the time-series change rate of the global demand distribution density, and the generation of anomaly burst heat field based on the time-series change rate, includes: Obtain the overall demand distribution density; The time-series change rate of the global demand distribution density is calculated by performing a second-order partial derivative operation on the global demand distribution density. Obtain the variance of the global normal distribution and the mean square background of the global normal distribution; Divide the variance of the global normal distribution by the mean square background of the global normal distribution to obtain the spatial variation coefficient term; The abnormal sudden heat field is calculated by multiplying the temporal rate of change of the global demand distribution density with the spatial variation coefficient term.

8. A closed-loop management system for handling complaints on the 12345 platform according to claim 1, characterized in that, The step of aggregating the performance lag index with the abnormal sudden heat field to obtain a rectification priority value, and generating a supervisory notice for the target department based on the rectification priority value, includes: Obtain the initial monitoring time and the current time corresponding to the performance lag indicator; The cumulative lag indicator item is obtained by performing time integration on the performance lag indicator corresponding to the target department from the initial monitoring time to the current time. Obtain the abnormal sudden heat field norm of the abnormal sudden heat field associated with the target department; Obtain the historical average heat field background under conditions of no sudden events across the entire domain; Divide the abnormal sudden heat field norm by the historical average heat field background to obtain the heat ratio; Adding the natural constant to the heat ratio yields the summation term; Performing the natural logarithm operation on the summation term yields the burst penalty coefficient term; The cumulative lag indicator is multiplied by the sudden penalty coefficient to calculate the rectification priority value; The rectification priority value is input into the artificial intelligence big data model for instruction-driven automatic generation of the supervision notice letter for the target department.

Images