A government affair work order intelligent distribution and efficiency closed loop management method and system

Abstract

The application discloses a kind of government affair work order intelligent distribution and efficiency closed loop management method and system, belong to government affair service intelligent processing technical field, the implementation of this method includes: through multimodal fusion classification model processing work order input data, through deep learning model realizes high-precision semantic understanding and entity extraction, generates structured work order entity information;Based on reinforcement learning dynamic decision engine, according to real-time department load state and geographic weight calculation optimal distribution path;Utilize the risk early warning model of bayesian network to monitor the progress of disposal, when the probability threshold triggers when overdue, automatically supervise and deal with.This application realizes the upgrading of the precision, real-time and closed loop of the whole link of work order processing through multimodal fusion classification, reinforcement learning distribution decision and bayesian risk intervention triple technical innovation, solves the problem of insufficient semantic understanding accuracy, lack of dynamic scheduling capability and poor real-time supervision in existing government affair work order system.

Description

Technical Field

[0001] This invention relates to the field of intelligent processing technology for government services, specifically a method and system for intelligent distribution and closed-loop management of government work orders. Background Technology

[0002] Currently, the government hotline receives over a million work orders daily, but manual sorting relies on experience-based matching, resulting in an error rate as high as 25% and inefficient cross-departmental processing. Furthermore, the lack of a real-time tracking mechanism leads to a processing overdue rate exceeding 30%, and citizen satisfaction is less than 60%. The existing system suffers from three major pain points: inefficient classification, static distribution, and lagging supervision, necessitating intelligent transformation.

[0003] Among existing technical solutions, the closest to the proposal in this application is likely a rule-based post-event statistical analysis work order processing system. This technology essentially generates departmental profiles through cluster analysis of work order data to aid in post-event optimization. However, this method has several significant drawbacks. First, both classification and profile generation are offline operations, and it lacks integrated dynamic resource scheduling algorithms, resulting in a lack of real-time decision-making capabilities and an inability to respond to work orders in real time, leading to a disconnect between scheduling and supervision. Second, such systems typically use fixed rule engines, whose rules cannot cover all dialects and vague descriptions, making them difficult to adapt to complex and ever-changing business scenarios. Finally, the entire process lacks a real-time interaction module, relying on manual report delivery to drive improvement, and while overdue work orders are discovered through post-event statistics, warnings cannot be issued during the processing phase. Summary of the Invention

[0004] The technical objective of this invention is to address the above-mentioned shortcomings by providing a method and system for intelligent distribution and efficiency closed-loop management of government work orders, which can achieve precise, real-time, and closed-loop upgrades across the entire work order processing chain.

[0005] The technical solution adopted by this invention to solve its technical problem is: A method for intelligent distribution and closed-loop efficiency management of government work orders, the implementation of which includes: (1) Process the work order input data through a multimodal fusion classification model, and achieve high-precision semantic understanding and entity extraction through a deep learning model to generate structured work order entity information; (2) Based on the reinforcement learning dynamic decision engine, the optimal distribution path is calculated according to the real-time department load status and geographical weight; (3) Use a Bayesian network risk warning model to monitor the progress of the disposal and automatically supervise when the overdue probability threshold is triggered.

[0006] This method addresses the technical bottlenecks in existing government work order systems, such as insufficient semantic understanding accuracy, lack of dynamic scheduling capabilities, and poor real-time monitoring. Through three technological innovations—multimodal fusion classification, reinforcement learning allocation decision-making, and Bayesian risk intervention—it achieves a more precise, real-time, and closed-loop upgrade across the entire work order processing chain.

[0007] Furthermore, in step (1), the speech and text dual-modal data are fused. Voice work orders are translated into text using an end-to-end dialect ASR model, and the acoustic model is trained using a region-specific phoneme set. The text work order extracts local features through TextCNN, concatenates them with BERT semantic vectors, and inputs them into the BiLSTM-CRF entity recognition layer; The output entity includes {location, event type, urgency level, and related department}.

[0008] Furthermore, step (1) is specifically implemented as follows: (1.1) Work order input and modal judgment: First, the system receives work order data pushed by the government service platform through an input interface. This data includes the following formats: Voice work order: Audio recorded by citizens through the hotline (e.g., MP3 format, 16kHz sampling rate). Text work order: Text content (such as UTF-8 encoded text) submitted by citizens through the APP / website. Then, modality is determined based on the input data format, and the processing branch is automatically routed: if the input is audio, the speech processing branch is started; if the input is text, the text processing branch is started. (1.2) Voice Work Order Processing Branch (Dialect ASR Translation): Audio preprocessing: Noise reduction and frame segmentation of the input audio, and extraction of MFCC (Mel frequency cepstral coefficients) feature vectors.

[0009] End-to-end dialect ASR translation: The specific structure of the classification model is as follows: Input layer: MFCC feature sequences; Convolutional layer: 2 convolutional layers to extract local spectral features; Recurrent layer: 3-layer bidirectional LSTM (Long Short-Term Memory network, 1024 hidden units per layer) to capture temporal dependencies; Output layer: CTC (Connection Temporal Classification) layer, outputting character probability distribution; The model is trained using dialect adaptation, employing a regional government speech database, and a hybrid loss function is used. ; in, =0.3, Responsible for aligning audio and text. Strengthen attention to dialect keywords; The model outputs translated text; (1.3) Text work order processing branch: If the work order is in the form of directly entered text, it will directly enter the subsequent semantic understanding module without translation; (1.4) Multimodal semantic understanding and entity extraction: Text feature extraction: TextCNN local feature extraction and BERT semantic encoding are used in parallel with dual channels to extract features from escaped text. TextCNN Local Feature Extraction: Input: The work order text after word segmentation. Embedding layer: word vector dimension 300 (Word2Vec model trained based on government affairs corpus). Convolutional layers: 3 types of convolutional kernels, each kernel outputs 64-dimensional features. Pooling layer: Max pooling, concatenating the outputs of all convolutional kernels to obtain a 192-dimensional feature vector; BERT semantic encoding: Context-aware text vectors are generated using a pre-trained BERT-base model, and the output corresponding to the [CLS] marker is taken as the overall text representation. The output of TextCNN is fused with the output of BERT.

[0010] Furthermore, in step (2), the optimal distribution path is calculated based on the real-time departmental load status and geographical weights, and the state space is defined as follows: ; The reward function is designed as follows: , in To respond to the delay, This is a predicted value for citizen satisfaction. Number of cross-departmental transfers; actions are output using the DQN algorithm. : This enables the distribution of work orders to the target departments.

[0011] Furthermore, the implementation of step (2) includes state awareness and data preparation, reward function calculation, decision generation, execution and feedback. (2.1) State Awareness and Data Preparation: The system perceives the current environmental state in real time. This state includes multiple key indicators, which are quantified into a multi-dimensional vector as input to the decision-making algorithm; including: Departmental real-time saturation calculation: Two key data points are retrieved in real time from the government data platform: the number of work orders currently pending in each department. And the department's maximum daily processing capacity pre-entered into the system. Saturation is calculated using the following formula: ; For example, if the Urban Management Bureau currently has 90 pending work orders, and its maximum daily processing capacity is 100, then its real-time saturation level is 90%. This is a very important indicator used to avoid assigning more work orders to departments that are already overloaded. Departmental Historical Processing Time Inquiry: Retrieve the average processing time (T_avg) for similar work orders from the historical work order database; for example, for work orders related to "sewer blockage," the water group's historical average processing time is 2 hours, while the municipal administration's is 1.5 hours. This indicator reflects the department's efficiency in handling specific issues; Geographic distance calculation: Using GIS (Geographic Information System) services, calculate the spherical distance between the latitude and longitude of the work order's location and the latitude and longitude of the center point of each department's jurisdiction. The unit is kilometers; this is to prioritize assigning work orders to the nearest department that can respond to the scene the fastest. Constructing a state vector: Combine all the above indicators into a structured state vector. , used to describe the global situation at the current time t; (2.2) Reward Function Calculation: The system calculates a reward score R for each potential department. The higher the score, the better the overall benefit of allocating the reward to that department. The reward function consists of three weighted components: Time-sensitive rewards: , in, This is the expected delay time. This is the maximum allowed delay time; These are weighting coefficients; Citizen satisfaction reward: , It is a predicted value, based on historical data, that predicts the citizen satisfaction score the department may receive after handling such work orders; These are the corresponding weighting coefficients; Collaboration efficiency penalty: , This indicates the number of times this work order is expected to require cross-departmental collaboration; It is the penalty coefficient; The final reward value R = timeliness reward + citizen satisfaction reward + collaborative efficiency penalty; (2.3) Decision generation: The final allocation decision is made using a deep Q-network (DQN), including the following process: Input: The state vector constructed in step (2.1) Input into the DQN model; Processing: The DQN model is a deep learning network that has internalized the rules of the reward function by learning from a large amount of historical allocation data and its results; the network processes state information internally and evaluates the potential long-term value of allocating to each department for the current state, which is called the Q value; Output: The DQN model outputs a final Q value for each selectable department, and the system automatically selects the department with the highest Q value as the allocation target for this work order; (2.4) Execution and feedback: After the decision is made, the system automatically dispatches the work order to the selected department through the government system's API (Application Programming Interface) and notifies the relevant person in charge; After a work order is processed, the system collects the actual processing time and the citizen's final satisfaction rating, comparing these actual results with the initial predictions. If the results are better than expected, the decision-making path is reinforced; if the results are worse than expected, the model's internal parameters are adjusted to avoid making similar decisions in the future. This closed loop of "decision-feedback-learning" enables the system to continuously evolve and become increasingly intelligent.

[0012] Furthermore, step (3) is implemented as follows: Constructing Bayesian network nodes: Prior probability Sourced from historical work order database; Overdue risk probability calculation: ; when In such cases, warning instructions are pushed to the responsible person via message queue, and the work order priority is increased.

[0013] Furthermore, step (3) specifically includes: real-time data collection and monitoring, dynamic prediction of overdue risks, early warning triggering and automatic supervision; (3.1) Real-time data acquisition and monitoring: The system performs real-time data perception throughout the entire process of handling all dispatched work orders, including multi-dimensional indicator monitoring and data stream processing. The multi-dimensional indicator monitoring includes: Response time: The system continuously calculates the ratio of the elapsed time of a work order to the standard processing time limit; for example, a work order with a standard processing time limit of 8 hours, after 2 hours, its... =2 / 8=0.25; the closer this ratio is to 1, the greater the time pressure. Processing progress updates: Does the system monitor whether the responsible department updates the processing log at key nodes? For example, do staff mark it as "received" within 1 hour, "verified on-site" within 3 hours, and "processed" within 6 hours? Prolonged "update silence" is a warning sign. Secondary interaction with citizens: The system integrates a public opinion monitoring module to capture secondary calls, complaints, or comments from citizens regarding the same event in real time; through NLP (Natural Language Processing) sentiment analysis technology, it automatically judges the positive, negative, or anxious level of citizens' emotions (for example, giving a negative emotion score from 0.0 to 1.0). Work order complexity (C): This is a quantifiable value assigned at the time of allocation, and is comprehensively evaluated based on factors including event type, number of departments requiring coordination, and required resources (e.g., simple problems). =0.3, cross-departmental joint law enforcement =0.9); The data stream processing involves encapsulating all of the dynamically changing data into messages, which are then pushed to the performance monitoring unit in real time via Kafka (a high-throughput distributed message queue) to ensure the immediacy of monitoring. (3.2) Dynamic prediction of overdue risk (Bayesian network inference): The Bayesian network probabilistic graphical model is used to comprehensively assess the overdue risk of work orders; including: Constructing a Bayesian network: A Bayesian network contains multiple nodes, representing various monitored metrics (such as response time). Work order complexity Public sentiment These nodes are connected by directed edges, the direction of which represents the probability dependency. Real-time probability calculation: When new monitoring data flows in, the system inputs it as evidence into the Bayesian network; the network performs probabilistic reasoning based on Bayes' theorem, dynamically calculating the posterior probability that the work order will eventually be overdue under the current known conditions (e.g., "the current progress is delayed by 30%" and "citizens are starting to complain"). Its probability value It is a value between 0 and 1 that quantifies the "risk level" of the current work order; (3.3) Early Warning Triggering and Automatic Supervision: After a risk is predicted, the system will not wait, but will immediately initiate an automated intervention process: Warning trigger condition: Set a risk threshold (e.g., 0.8), and trigger the warning when the calculated probability of exceeding the deadline is reached. If the work order is determined to be in a high-risk state, the early warning mechanism will be triggered immediately. Automatic supervision and execution includes: Message push: The system automatically generates a supervision message and accurately pushes it to the work terminal of the person in charge of the department through the government OA system, SMS or work DingTalk, WeChat; the message content includes the probability of high risk of overdue. Dynamic priority adjustment: At the same time, the system will automatically increase the priority of the work order in the department's task queue. The increase is linked to the risk value and follows the formula: New priority = Original priority × (1 + Overdue probability). For example, a work order with an original priority of 1 will have its new priority jump to 1.9 if the overdue probability is 0.9, thus placing it at the top of the department's to-do list and giving it priority in resource allocation. Closed-loop confirmation: After the supervision is issued, the system continues to monitor. The warning status will only be lifted when the responsible person handles and updates the status, and the system verifies that the work order has indeed been completed, thus forming a true "closed loop".

[0014] This invention also claims a government work order intelligent distribution and efficiency closed-loop management system, comprising: The multimodal work order classification module is used to fuse multimodal work order input data, achieve high-precision semantic understanding and entity extraction through deep learning models, and generate structured work order entity information. The work order dynamic distribution module is used to analyze multi-dimensional dynamic factors such as the handling capacity, historical performance, and geographical location of each department in real time through reinforcement learning algorithms, and calculate the optimal distribution path based on the real-time department load status and geographical weight. The efficiency closed-loop monitoring module is used to monitor the progress of work order processing in real time, proactively predict and intervene in potential overdue risks, and automatically supervise the process when the overdue probability threshold is triggered. This system can implement the methods described above.

[0015] The present invention also claims a device for intelligent distribution and efficiency closed-loop management of government work orders, comprising: at least one memory and at least one processor; The at least one memory is used to store a machine-readable program; The at least one processor is used to call the machine-readable program to implement the above method.

[0016] The present invention also claims a computer-readable medium storing computer instructions that, when executed by a processor, enable the implementation of the above-described method.

[0017] Compared with the prior art, the intelligent distribution and efficiency closed-loop management method and system for government work orders of the present invention have the following advantages: This invention achieves the following technical effects through a triple technological innovation: "multimodal work order classification → reinforcement learning dynamic distribution → Bayesian closed-loop supervision": 1. Significantly improved accuracy in work order processing: The multimodal fusion model (ASR+TextCNN) achieves accurate parsing of speech dialects and ambiguous descriptions, improving the classification accuracy from 57% to 92% of existing technologies, and reducing the error rate of semantically ambiguous work orders by 40%, thereby reducing cross-departmental misallocation at the source.

[0018] 2. A qualitative leap in resource scheduling efficiency: The DQN reinforcement learning algorithm dynamically perceives departmental load (saturation, geographical weight, etc.), reducing the first response time of work orders from >30 minutes to <30 minutes, decreasing the cross-departmental workflow rate from 60% to 20%, and reducing administrative resource waste by 67%.

[0019] 3. Breakthrough in regulatory timeliness: Bayesian networks predict overdue risks in real time (an early warning is issued when the probability of multiple factors such as response time and citizen sentiment is greater than 0.8), reducing the overdue handling rate from 30% to less than 1%, reducing the need for manual follow-up by 90%, and achieving 100% closed-loop management of work orders within 24 hours. Attached Figure Description

[0020] Figure 1 This is a flowchart illustrating the intelligent distribution and efficiency closed-loop management method for government work orders provided in this embodiment of the invention. Figure 2 This is a flowchart illustrating the voice work order processing procedure provided in an embodiment of the present invention; Figure 3 This is a flowchart illustrating the dynamic work order distribution process provided in an embodiment of the present invention; Figure 4 This is a flowchart illustrating the performance closed-loop monitoring process provided in an embodiment of the present invention. Detailed Implementation

[0021] This invention provides a method for intelligent distribution and closed-loop efficiency management of government work orders. The implementation of this method includes: (1) Process the work order input data through a multimodal fusion classification model, and achieve high-precision semantic understanding and entity extraction through a deep learning model to generate structured work order entity information; (2) Based on the reinforcement learning dynamic decision engine, the optimal distribution path is calculated according to the real-time department load status and geographical weight; (3) Use a Bayesian network risk warning model to monitor the progress of the disposal and automatically supervise when the overdue probability threshold is triggered.

[0022] Step (1) involves fusing speech and text dual-modal data. Voice work orders are translated into text using an end-to-end dialect ASR model, and the acoustic model is trained using a region-specific phoneme set. The text work order extracts local features through TextCNN, concatenates them with BERT semantic vectors, and inputs them into the BiLSTM-CRF entity recognition layer; The output entity includes {location, event type, urgency level, and related department}.

[0023] In step (2), the optimal distribution path is calculated based on the real-time departmental load status and geographical weight. The state space is defined as follows: ; The reward function is designed as follows: , in To respond to the delay, This is a predicted value for citizen satisfaction. Number of cross-departmental transfers; actions are output using the DQN algorithm. : This enables the distribution of work orders to the target departments.

[0024] Step (3) is implemented as follows: Constructing Bayesian network nodes: Prior probability Sourced from historical work order database; Overdue risk probability calculation: ; when In such cases, warning instructions are pushed to the responsible person via message queue, and the work order priority is increased.

[0025] This method addresses the technical bottlenecks in existing government work order systems, such as insufficient semantic understanding accuracy, lack of dynamic scheduling capabilities, and poor real-time monitoring. Through three technological innovations—multimodal fusion classification, reinforcement learning allocation decision-making, and Bayesian risk intervention—it achieves a more precise, real-time, and closed-loop upgrade across the entire work order processing chain.

[0026] The specific implementation process of this method is as follows.

[0027] 1. Detailed description of the multimodal work order classification process.

[0028] This process aims to convert unstructured work orders (voice or text) submitted by citizens into structured work order entity information, providing accurate input for subsequent intelligent allocation. The core of the process lies in integrating voice and text dual-modal data, and using a deep learning model to achieve high-precision semantic understanding and entity extraction. The specific steps are as follows: 1.1 Work order input and modal judgment: First, the work order data pushed by the government platform is received through the input interface. Its format may be: (1) Voice work order: Audio recorded by citizens through the hotline (e.g., MP3 format, sampling rate 16kHz). (2) Text work order: Text content (such as UTF-8 encoded text) submitted by citizens through APP / website.

[0029] Then, modality is determined based on the input data format, and the processing branch is automatically routed: if the input is audio, the speech processing branch is started; if the input is text, the text processing branch is started.

[0030] 1.2 Voice Work Order Processing Branch (Dialect ASR Translation): 1.2.1 Audio preprocessing: Denoise the input audio (using spectral subtraction) and divide it into frames (frame length 25ms, frame shift 10ms), and extract the MFCC (Mel frequency cepstral coefficients) feature vector.

[0031] 1.2.2 End-to-end dialect ASR translation: The classification model uses an acoustic model based on the DeepSpeech2 architecture, the specific structure of which is as follows: Input layer: MFCC feature sequences; Convolutional layer: 2 convolutional layers (kernel size 5×5) to extract local spectral features; Recurrent layer: 3-layer bidirectional LSTM (Long Short-Term Memory network, 1024 hidden units per layer) to capture temporal dependencies; Output layer: CTC (Connection Temporal Classification) layer, outputting character probability distribution.

[0032] The model is trained using dialect adaptation, employing a regional government speech database, and a hybrid loss function is used. ; in, =0.3, Responsible for aligning audio and text. Strengthen attention to dialect keywords; The model outputs translated text.

[0033] 1.3 Text Work Order Processing Branch: If the work order is in the form of directly entered text, it will proceed directly to the subsequent semantic understanding module without translation.

[0034] 1.4 Multimodal semantic understanding and entity extraction: 1.4.1 Text Feature Extraction: TextCNN local feature extraction and BERT semantic encoding are used in parallel with dual channels to extract features from escaped text. (1) TextCNN local feature extraction: Input: The work order text after word segmentation; Embedding layer: word vector dimension 300 (Word2Vec model trained based on government affairs corpus); Convolutional layers: 3 types of convolutional kernels (sizes 2, 3, and 4), each kernel outputs 64-dimensional features; Pooling layer: Max pooling, concatenating all convolutional kernels to output a 192-dimensional feature vector.

[0035] (2) BERT semantic encoding: The pre-trained BERT-base model (Chinese version) is used to generate context-aware text vectors, and the output corresponding to the [CLS] marker is taken as the overall text representation (768-dimensional vector).

[0036] The TextCNN output (192-dimensional feature vector) and the BERT output (768-dimensional vector) are fused to form a feature fusion.

[0037] 2. The implementation process of dynamic work order distribution.

[0038] The core objective of this step is to intelligently allocate categorized structured work orders to the most suitable processing departments. The underlying technology lies in using reinforcement learning algorithms to analyze multi-dimensional dynamic factors such as the processing capacity, historical performance, and geographical location of each department in real time, calculating the optimal allocation path. This replaces the traditional static polling or manual allocation model, achieving load balancing and maximizing efficiency in work order processing.

[0039] The entire process can be divided into four core stages: state awareness and data preparation, reward function calculation, decision generation, and execution and feedback. The following is a detailed description: 2.1 State Awareness and Data Preparation: The system perceives the current environmental state in real time, which consists of a series of key indicators. These indicators are quantified into a multi-dimensional vector and used as input to the decision-making algorithm.

[0040] 2.1.1 Calculation of real-time saturation for each department: Data Acquisition: Two key data points are retrieved in real time from the government data platform: the number of work orders currently pending in each department. And the department's maximum daily processing capacity pre-entered into the system. ; Calculation method: Saturation is calculated using the following formula: .

[0041] For example, if the Urban Management Bureau currently has 90 pending work orders, and its maximum daily processing capacity is 100 orders, then its real-time saturation level is 90%. This is a very important indicator used to avoid assigning more work orders to departments that are already overloaded.

[0042] 2.1.2. Inquiry into the historical processing time of departments: Retrieve the average processing time (T_avg) for similar work orders from the historical work order database. For example, for work orders related to "sewer blockage," the historical average processing time for the water company is 2 hours, while for the municipal authority it is 1.5 hours. This metric reflects the department's efficiency in handling specific problems.

[0043] 2.1.3 Calculation of geographical distance: Using GIS (Geographic Information System) services, calculate the spherical distance between the latitude and longitude of the work order's location and the latitude and longitude of the center point of each department's jurisdiction. The unit is kilometers. This is to prioritize assigning work orders to the nearest department that can respond to the site the fastest.

[0044] 2.1.4 Constructing the state vector: All the above indicators are combined into a structured state vector. This is used to describe the global situation at the current time t. For example: ={Urban Management Bureau:[Saturation: 90%, Historical Timeframe: 4h, Distance: 3km], Water Group:[Saturation: 60%, Historical Timeframe: 2h, Distance: 1.5km],...}.

[0045] 2.2 Reward Function Calculation: The reward function is the brain of reinforcement learning; it defines what constitutes a "good" decision. The system calculates a reward score R for each potential department, with a higher score indicating better overall benefits for that department. The reward function consists of three weighted components: Time-sensitive rewards: , in, This is the expected delay time. This is the maximum allowed delay time. This section encourages the system to select the department that can process work orders the fastest. It is a weighting coefficient, indicating the degree of importance we attach to this item.

[0046] Citizen satisfaction reward: , This is a predicted value, based on historical data, that indicates the citizen satisfaction score a department is likely to receive after handling such work orders. This section encourages the system to select departments with better service quality and higher citizen ratings. These are the corresponding weighting coefficients.

[0047] Collaboration efficiency penalty: , This represents the estimated number of times this work order will require cross-departmental collaboration. This is a negative score, designed to penalize decisions that might cause the work order to flow back and forth between departments. By introducing negative scores, the system is encouraged to select departments that can complete the task independently or with minimal collaboration, significantly reducing the flow rate. It is the penalty coefficient.

[0048] The final reward value R = timeliness reward + citizen satisfaction reward + collaborative efficiency penalty.

[0049] 2.3 Decision Generation (DQN Decision): The final allocation decision is made using a deep Q-network (DQN).

[0050] Input: The state vector constructed in step 2.1 Input into the DQN model.

[0051] Processing: The DQN model is a deep learning network that has internalized the rules of the reward function by learning from a large amount of historical allocation data and its results. The network processes state information internally and evaluates the potential long-term value of allocating to each department for the current state, which is called the Q value.

[0052] Output: The DQN model outputs a final Q value for each selectable department. The system automatically selects the department with the highest Q value as the allocation target for this work order. For example, if the model calculates that the Q value for allocation to "Water Group" is 0.92, while the Q value for allocation to "Urban Management Bureau" is 0.73, then the system will select Water Group.

[0053] 2.4 Implementation and Feedback: Once the decision is made, the system automatically dispatches the work order to the selected department via the government system's API (Application Programming Interface) and notifies the relevant person in charge.

[0054] Feedback learning is key to the system's continuous optimization. After a work order is processed, the system collects the actual processing time and the citizen's final satisfaction rating, comparing these real results with the initial predictions. If the results are better than expected, the decision-making path is reinforced; if the results are worse than expected, the model's internal parameters are adjusted to avoid making similar decisions in the future. This closed loop of "decision-feedback-learning" enables the system to continuously evolve and become increasingly intelligent.

[0055] This dynamic distribution process is a complete, adaptive intelligent decision-making loop. It perceives the current situation through real-time data, defines the goal through the reward function, makes complex optimal decisions through the DQN network, and learns and optimizes itself through feedback results, thereby completely solving the technical defects of traditional distribution methods such as slow response, low efficiency, and uneven load.

[0056] 3. The process of implementing closed-loop performance monitoring.

[0057] This step is the "supervision and guarantee" stage to ensure that work orders are ultimately handled efficiently. Its technical goal is to change the traditional lagging supervision model of post-event statistics, and to proactively predict and intervene in the risk of potential delays by monitoring the progress of work order processing in real time, thereby forming a complete closed loop to ensure that every citizen's request is addressed and every issue receives a response.

[0058] The entire implementation process acts like an intelligent "early warning radar," continuously scanning the lifecycle of all pending work orders. It can be divided into three core stages: real-time data collection and monitoring, dynamic prediction of overdue risks, and early warning triggering and automatic supervision.

[0059] 3.1 Real-time data acquisition and monitoring: The system provides real-time data monitoring throughout the entire process of all dispatched work orders, which forms the basis for early warning and judgment. This includes multi-dimensional indicator monitoring and data stream processing.

[0060] 3.1.1 Multi-dimensional indicator monitoring: Response time ( The system continuously calculates the ratio of "time elapsed / standard time limit" for each work order. For example, a work order with a standard processing time limit of 8 hours, after 2 hours, its... =2 / 8=0.25; the closer this ratio is to 1, the greater the time pressure.

[0061] Processing progress updates: Does the system monitor whether the responsible department updates the processing log at key milestones? For example, do staff mark the log as "received" within 1 hour, "verified on-site" within 3 hours, and "processed" within 6 hours? Prolonged periods of "update silence" are a warning sign.

[0062] Secondary citizen interaction: The system integrates a public opinion monitoring module to capture secondary calls, complaints, or comments from citizens regarding the same event in real time. Through NLP (Natural Language Processing) sentiment analysis technology, it automatically determines the positive, negative, or anxious level of citizens' emotions (for example, giving a negative sentiment score from 0.0 to 1.0).

[0063] Work order complexity (C): This is a quantifiable value assigned at the time of allocation, and is comprehensively evaluated based on factors including event type, number of departments requiring coordination, and required resources (e.g., simple problems). =0.3, cross-departmental joint law enforcement =0.9).

[0064] 3.1.2 Data Stream Processing: All the dynamically changing data mentioned above are encapsulated into messages and pushed to the performance monitoring unit in real time via Kafka (a high-throughput distributed message queue) to ensure the timeliness of monitoring.

[0065] 3.2 Dynamic Prediction of Overdue Risk (Bayesian Network Inference): This is the core technology of the entire process. A Bayesian network, a probabilistic graphical model, is used to comprehensively assess the risk of work orders exceeding their deadlines.

[0066] 3.2.1 Constructing a Bayesian network: A Bayesian network contains multiple nodes, representing various monitored metrics (such as response time). Work order complexity Public sentiment These nodes are connected by directed edges, the directions of which represent probabilistic dependencies. The probability parameters in the network (prior probability and conditional probability) are derived by learning from massive amounts of historical work order data. For example, historical data shows that when... >0.8 (meaning the time taken has exceeded 80% of the standard time limit) and public sentiment When the threshold is >0.7 (strong negative emotions), the probability of exceeding the time limit is as high as 90%.

[0067] 3.2.1 Real-time Probability Calculation: When new monitoring data flows in, the system inputs it as evidence into the Bayesian network. The network performs probabilistic reasoning based on Bayes' theorem, dynamically calculating the posterior probability that the work order will eventually exceed its deadline under the current known conditions (e.g., "the current progress is delayed by 30%" and "citizens are starting to complain"). Its probability value It is a value between 0 and 1 that quantifies the "risk level" of the current work order.

[0068] 3.3 Warning Trigger and Automatic Supervision: After a risk is predicted, the system will not wait, but will immediately start an automated intervention process.

[0069] 3.3.1 Warning Trigger Conditions: Set a risk threshold (e.g., 0.8). When the calculated probability of exceeding the deadline... If the work order is deemed to be in a high-risk state, an early warning mechanism will be immediately triggered.

[0070] 3.2.2 Automatic supervision and execution: Message push: The system automatically generates a supervision message and pushes it accurately to the work terminal of the person in charge of the department through the government OA system, SMS, or work DingTalk and WeChat; the message content is no longer a simple "please handle as soon as possible", but an intelligent prompt: "Work order XX has a high risk of being overdue due to [progress delay / negative feedback from citizens] (probability 92%), please pay attention immediately!"

[0071] Dynamic Priority Adjustment: Simultaneously, the system will automatically increase the priority of this work order in the department's task queue. The increase is linked to the risk value and follows the formula: New Priority = Original Priority × (1 + Probability of Overdue). For example, a work order with an original priority of 1, if its probability of overdue is 0.9, will have its new priority jump to 1.9, thus being placed at the top of the department's to-do list and receiving priority resource allocation.

[0072] Closed-loop confirmation: After the supervision is issued, the system continues to monitor. The warning status will only be lifted when the responsible person handles and updates the status, and the system verifies that the work order has indeed been completed, thus forming a true "closed loop".

[0073] In summary, the efficiency closed-loop monitoring process is a highly intelligent "prediction-intervention" system. It comprehensively perceives the situation through real-time data collection, uses Bayesian networks for scientific risk prediction, and ultimately achieves proactive intervention through an automated supervision mechanism, nipping problems in the bud. This completely overturns the traditional passive "post-event remediation" model, transforming "post-event statistics" into "pre-event early warning," and "manual reminders" into "automatic supervision," thereby ensuring the ultimate efficiency of government work order processing at the technical level.

[0074] The following case study, using a complete government service order processing example, illustrates the technical implementation process of this method in a real-world business scenario and its improvement compared to traditional solutions. The case study is adapted from a real event reported by a city's 12345 hotline, and the data has been anonymized.

[0075] Case background: Mr. Zhang, a citizen, complained in his local dialect: "The sewer at the intersection of XX Road and XX Road is blocked, and the water depth is more than 20 centimeters, making it impossible for pedestrians to pass. Please deal with it as soon as possible!"

[0076] Step S1: Multimodal work order classification: Step A1: Speech Recognition The dialect ASR model translates in real time and accurately outputs the text: "The sewer at the intersection of XX Road and XX Road is blocked, with water depth > 20cm, making passage difficult."

[0077] Step A2: Semantic Understanding: TextCNN+BERT model extracts key entities: { Location: Intersection of XX Road and XX Road "Event Type": "Municipal Facility Malfunction" Facility Name: "Sewer" Urgency level: "High" Related Departments: ["Water Group", "Municipal Management Office"] } The classification accuracy rate is 92%, avoiding the misjudgment of "water accumulation" as the responsibility of urban management.

[0078] Step S2: Reinforcement Learning Dynamic Distribution: Step B1: State Awareness (Real-time Data): Water Group: Saturation 65% (82 work orders pending processing); Municipal Administration Office: Saturation 40% (31 work orders pending); Location: The water group is 3km from the site, and the municipal management office is 1.5km from the site.

[0079] Step B2: DQN Decision: Reward function calculation: R = 0.6 * (1 - Delay) + 0.3 * Citizen Expectation - 0.1 * Number of Inter-departmental Tasks. The optimal solution is: dispatch to the Municipal Administration Office (overall timeliness score 0.92, higher than the Water Group's 0.73).

[0080] Step S3: Bayesian closed-loop supervision: Step C1: Risk Warning (Monitoring During Handling): Two hours after the response was initiated: Bayesian network detected "progress not updated + citizen called a second time"; Calculate the probability of exceeding the deadline: P(delay | progress delay, citizen anxiety) = 0.85; Automatic supervision: Triggers a Kafka message push to the person in charge of the municipal management office, "Work order XX needs to be processed quickly, high risk of being overdue!", and simultaneously increases the priority of the work order and allocates resources accordingly.

[0081] Outcome: The municipal management office cleared the blockage within 4 hours, and the citizens were "very satisfied".

[0082] This invention also provides an intelligent distribution and efficiency closed-loop management system for government work orders, comprising: The multimodal work order classification module is used to fuse multimodal work order input data, achieve high-precision semantic understanding and entity extraction through deep learning models, and generate structured work order entity information. The work order dynamic distribution module is used to analyze multi-dimensional dynamic factors such as the handling capacity, historical performance, and geographical location of each department in real time through reinforcement learning algorithms, and calculate the optimal distribution path based on the real-time department load status and geographical weight. The efficiency closed-loop monitoring module is used to monitor the progress of work order processing in real time, proactively predict and intervene in potential overdue risks, and automatically supervise the process when the overdue probability threshold is triggered. This system can realize the intelligent distribution and efficiency closed-loop management method for government work orders described in the above embodiments.

[0083] 1. Multimodal work order classification module.

[0084] This module aims to transform unstructured work orders (voice or text) submitted by citizens into structured work order entity information, providing accurate input for subsequent intelligent allocation. Its core lies in fusing voice and text dual-modal data, and using a deep learning model to achieve high-precision semantic understanding and entity extraction. The specific steps are as follows: 1.1 Work order input and modal judgment: First, the system receives work order data pushed by the government affairs platform through the input interface, which may be in the following form: (1) Voice work order: Audio recorded by citizens through the hotline (e.g., MP3 format, sampling rate 16kHz). (2) Text work order: Text content (such as UTF-8 encoded text) submitted by citizens through APP / website.

[0085] Then, the system performs modal judgment based on the input data format and automatically routes the processing branches: if the input is audio, the speech processing branch is started; if the input is text, the text processing branch is started.

[0086] 1.2 Voice Work Order Processing Branch (Dialect ASR Translation): 1.2.1 Audio preprocessing: Denoise the input audio (using spectral subtraction) and divide it into frames (frame length 25ms, frame shift 10ms), and extract the MFCC (Mel frequency cepstral coefficients) feature vector.

[0087] 1.2.2 End-to-end dialect ASR translation: The classification model uses an acoustic model based on the DeepSpeech2 architecture, the specific structure of which is as follows: Input layer: MFCC feature sequences; Convolutional layer: 2 convolutional layers (kernel size 5×5) to extract local spectral features; Recurrent layer: 3-layer bidirectional LSTM (Long Short-Term Memory network, 1024 hidden units per layer) to capture temporal dependencies; Output layer: CTC (Connection Temporal Classification) layer, outputting character probability distribution.

[0088] The model is trained using dialect adaptation, employing a regional government speech database, and a hybrid loss function is used. ; in, =0.3, Responsible for aligning audio and text. Strengthen attention to dialect keywords; The model outputs translated text.

[0089] 1.3 Text Work Order Processing Branch: If the work order is in the form of directly entered text, it will proceed directly to the subsequent semantic understanding module without translation.

[0090] 1.4 Multimodal semantic understanding and entity extraction: 1.4.1 Text Feature Extraction: TextCNN local feature extraction and BERT semantic encoding are used in parallel with dual channels to extract features from escaped text. (1) TextCNN local feature extraction: Input: The work order text after word segmentation; Embedding layer: word vector dimension 300 (Word2Vec model trained based on government affairs corpus); Convolutional layers: 3 types of convolutional kernels (sizes 2, 3, and 4), each kernel outputs 64-dimensional features; Pooling layer: Max pooling, concatenating all convolutional kernels to output a 192-dimensional feature vector.

[0091] (2) BERT semantic encoding: The pre-trained BERT-base model (Chinese version) is used to generate context-aware text vectors, and the output corresponding to the [CLS] marker is taken as the overall text representation (768-dimensional vector).

[0092] The TextCNN output (192-dimensional feature vector) and the BERT output (768-dimensional vector) are fused to form a feature fusion.

[0093] 2. Dynamic work order distribution module.

[0094] The core objective of this module is to intelligently allocate categorized structured work orders to the most suitable processing department. The underlying technology lies in using reinforcement learning algorithms to analyze multi-dimensional dynamic factors such as the processing capacity, historical performance, and geographical location of each department in real time, calculating the optimal allocation path. This replaces the traditional static polling or manual allocation method, achieving load balancing and maximizing efficiency in work order processing.

[0095] The entire process can be divided into four core stages: state awareness and data preparation, reward function calculation, decision generation, and execution and feedback. The following is a detailed description: 2.1 State Awareness and Data Preparation: The system perceives the current environmental state in real time, which consists of a series of key indicators. These indicators are quantified into a multi-dimensional vector and used as input to the decision-making algorithm.

[0096] 2.1.1 Calculation of real-time saturation for each department: Data Acquisition: The system pulls two key data points in real time from the government data platform: the number of work orders currently pending in each department. And the department's maximum daily processing capacity pre-entered into the system. ; Calculation method: Saturation is calculated using the following formula: .

[0097] For example, if the Urban Management Bureau currently has 90 pending work orders, and its maximum daily processing capacity is 100 orders, then its real-time saturation level is 90%. This is a very important indicator used to avoid assigning more work orders to departments that are already overloaded.

[0098] 2.1.2. Inquiry into the historical processing time of departments: The system retrieves the average processing time (T_avg) for similar work orders from the historical work order database. For example, for work orders related to "sewer blockage," the historical average processing time for the water company is 2 hours, while for the municipal authority it is 1.5 hours. This metric reflects the department's efficiency in handling specific problems.

[0099] 2.1.3 Calculation of geographical distance: The system uses GIS (Geographic Information System) services to calculate the spherical distance between the latitude and longitude of the work order's location and the latitude and longitude of the center point of each department's jurisdiction. The unit is kilometers. This is to prioritize assigning work orders to the nearest department that can respond to the site the fastest.

[0100] 2.1.4 Constructing the state vector: All the above indicators are combined into a structured state vector. This is used to describe the global situation at the current time t. For example: ={Urban Management Bureau:[Saturation: 90%, Historical Timeframe: 4h, Distance: 3km], Water Group:[Saturation: 60%, Historical Timeframe: 2h, Distance: 1.5km],...}.

[0101] 2.2 Reward Function Calculation: The reward function is the brain of reinforcement learning; it defines what constitutes a "good" decision. The system calculates a reward score R for each potential department, with a higher score indicating better overall benefits for that department. The reward function consists of three weighted components: Time-sensitive rewards: , in, This is the expected delay time. This is the maximum allowed delay time. This section encourages the system to select the department that can process work orders the fastest. It is a weighting coefficient, indicating the degree of importance we attach to this item.

[0102] Citizen satisfaction reward: , This is a predicted value, based on historical data, that indicates the citizen satisfaction score a department is likely to receive after handling such work orders. This section encourages the system to select departments with better service quality and higher citizen ratings. These are the corresponding weighting coefficients.

[0103] Collaboration efficiency penalty: , This represents the estimated number of times this work order will require cross-departmental collaboration. This is a negative score, designed to penalize decisions that might cause the work order to flow back and forth between departments. By introducing negative scores, the system is encouraged to select departments that can complete the task independently or with minimal collaboration, significantly reducing the flow rate. It is the penalty coefficient.

[0104] The final reward value R = timeliness reward + citizen satisfaction reward + collaborative efficiency penalty.

[0105] 2.3 Decision Generation (DQN Decision): The final allocation decision is made using a deep Q-network (DQN).

[0106] Input: The state vector constructed in step 2.1 Input into the DQN model.

[0107] Processing: The DQN model is a deep learning network that has internalized the rules of the reward function by learning from a large amount of historical allocation data and its results. The network processes state information internally and evaluates the potential long-term value of allocating to each department for the current state, which is called the Q value.

[0108] Output: The DQN model outputs a final Q value for each selectable department. The system automatically selects the department with the highest Q value as the allocation target for this work order. For example, if the model calculates that the Q value for allocation to "Water Group" is 0.92, while the Q value for allocation to "Urban Management Bureau" is 0.73, then the system will select Water Group.

[0109] 2.4 Implementation and Feedback: Once the decision is made, the system automatically dispatches the work order to the selected department via the government system's API (Application Programming Interface) and notifies the relevant person in charge.

[0110] Feedback learning is key to the system's continuous optimization. After a work order is processed, the system collects the actual processing time and the citizen's final satisfaction rating, comparing these real results with the initial predictions. If the results are better than expected, the decision-making path is reinforced; if the results are worse than expected, the model's internal parameters are adjusted to avoid making similar decisions in the future. This closed loop of "decision-feedback-learning" enables the system to continuously evolve and become increasingly intelligent.

[0111] This dynamic distribution process is a complete, adaptive intelligent decision-making loop. It perceives the current situation through real-time data, defines the goal through the reward function, makes complex optimal decisions through the DQN network, and learns and optimizes itself through feedback results, thereby completely solving the technical defects of traditional distribution methods such as slow response, low efficiency, and uneven load.

[0112] 3. Performance closed-loop monitoring module.

[0113] This module is the "supervision and guarantee" link to ensure that work orders are ultimately handled efficiently. Its technical goal is to change the traditional lagging supervision model of post-event statistics, and to proactively predict and intervene in the risk of potential delays by monitoring the progress of work order processing in real time, thereby forming a complete closed loop to ensure that every citizen's request is implemented and every issue receives a response.

[0114] The entire implementation process acts like an intelligent "early warning radar," continuously scanning the lifecycle of all pending work orders. It can be divided into three core stages: real-time data collection and monitoring, dynamic prediction of overdue risks, and early warning triggering and automatic supervision.

[0115] 3.1 Real-time data acquisition and monitoring: The system provides real-time data monitoring throughout the entire process of all dispatched work orders, which forms the basis for early warning and judgment. This includes multi-dimensional indicator monitoring and data stream processing.

[0116] 3.1.1 Multi-dimensional indicator monitoring: Response time ( The system continuously calculates the ratio of "time elapsed / standard time limit" for each work order. For example, a work order with a standard processing time limit of 8 hours, after 2 hours, its... =2 / 8=0.25; the closer this ratio is to 1, the greater the time pressure.

[0117] Processing progress updates: Does the system monitor whether the responsible department updates the processing log at key milestones? For example, do staff mark the log as "received" within 1 hour, "verified on-site" within 3 hours, and "processed" within 6 hours? Prolonged periods of "update silence" are a warning sign.

[0118] Secondary citizen interaction: The system integrates a public opinion monitoring module to capture secondary calls, complaints, or comments from citizens regarding the same event in real time. Through NLP (Natural Language Processing) sentiment analysis technology, it automatically determines the positive, negative, or anxious level of citizens' emotions (for example, giving a negative sentiment score from 0.0 to 1.0).

[0119] Work order complexity (C): This is a quantifiable value assigned at the time of allocation, and is comprehensively evaluated based on factors including event type, number of departments requiring coordination, and required resources (e.g., simple problems). =0.3, cross-departmental joint law enforcement =0.9).

[0120] 3.1.2 Data Stream Processing: All the dynamically changing data mentioned above are encapsulated into messages and pushed to the performance monitoring unit in real time via Kafka (a high-throughput distributed message queue) to ensure the timeliness of monitoring.

[0121] 3.2 Dynamic Prediction of Overdue Risk (Bayesian Network Inference): This is the core technology of the entire process. A Bayesian network, a probabilistic graphical model, is used to comprehensively assess the risk of work orders exceeding their deadlines.

[0122] 3.2.1 Constructing a Bayesian network: A Bayesian network contains multiple nodes, representing various monitored metrics (such as response time). Work order complexity Public sentiment These nodes are connected by directed edges, the directions of which represent probabilistic dependencies. The probability parameters in the network (prior probability and conditional probability) are derived by learning from massive amounts of historical work order data. For example, historical data shows that when... >0.8 (meaning the time taken has exceeded 80% of the standard time limit) and public sentiment When the threshold is >0.7 (strong negative emotions), the probability of exceeding the time limit is as high as 90%.

[0123] 3.2.1 Real-time Probability Calculation: When new monitoring data flows in, the system inputs it as evidence into the Bayesian network. The network performs probabilistic reasoning based on Bayes' theorem, dynamically calculating the posterior probability that the work order will eventually exceed its deadline under the current known conditions (e.g., "the current progress is delayed by 30%" and "citizens are starting to complain"). Its probability value It is a value between 0 and 1 that quantifies the "risk level" of the current work order.

[0124] 3.3 Warning Trigger and Automatic Supervision: After a risk is predicted, the system will not wait, but will immediately start an automated intervention process.

[0125] 3.3.1 Warning Trigger Conditions: Set a risk threshold (e.g., 0.8). When the calculated probability of exceeding the deadline... If the work order is deemed to be in a high-risk state, an early warning mechanism will be immediately triggered.

[0126] 3.2.2 Automatic supervision and execution: Message push: The system automatically generates a supervision message and pushes it accurately to the work terminal of the person in charge of the department through the government OA system, SMS, or work DingTalk and WeChat; the message content is no longer a simple "please handle as soon as possible", but an intelligent prompt: "Work order XX has a high risk of being overdue due to [progress delay / negative feedback from citizens] (probability 92%), please pay attention immediately!"

[0127] Dynamic Priority Adjustment: Simultaneously, the system will automatically increase the priority of this work order in the department's task queue. The increase is linked to the risk value and follows the formula: New Priority = Original Priority × (1 + Probability of Overdue). For example, a work order with an original priority of 1, if its probability of overdue is 0.9, will have its new priority jump to 1.9, thus being placed at the top of the department's to-do list and receiving priority resource allocation.

[0128] Closed-loop confirmation: After the supervision is issued, the system continues to monitor. The warning status will only be lifted when the responsible person handles and updates the status, and the system verifies that the work order has indeed been completed, thus forming a true "closed loop".

[0129] This invention also provides a device for intelligent distribution and performance closed-loop management of government work orders, comprising: at least one memory and at least one processor; The at least one memory is used to store a machine-readable program; The at least one processor is used to call the machine-readable program to implement the intelligent distribution and efficiency closed-loop management method for government work orders described in the above embodiments.

[0130] This invention also provides a computer-readable medium storing computer instructions. When executed by a processor, the computer instructions cause the processor to perform the intelligent distribution and efficiency closed-loop management method for government work orders described in the above embodiments. Specifically, a system or device equipped with a storage medium storing software program code that implements the functions of any of the embodiments described above can be provided, enabling the computer (or CPU or MPU) of the system or device to read and execute the program code stored in the storage medium.

[0131] In this case, the program code read from the storage medium can itself implement the function of any of the above embodiments, and therefore the program code and the storage medium storing the program code constitute part of the present invention.

[0132] Storage media embodiments for providing program code include floppy disks, hard disks, magneto-optical disks, optical disks (such as CD-ROM, CD-R, CD-RW, DVD-ROM, DVD-RAM, DVD-RW, DVD+RW), magnetic tapes, non-volatile memory cards, and ROMs. Alternatively, program code can be downloaded from a server computer via a communication network.

[0133] Furthermore, it should be clear that not only can the program code read by the computer be executed, but also the operating system or other components operating on the computer can be instructed based on the program code to perform some or all of the actual operations, thereby realizing the function of any of the embodiments described above.

[0134] Furthermore, it is understood that the program code read from the storage medium is written to the memory set in the expansion board inserted into the computer or to the memory set in the expansion unit connected to the computer. Then, based on the instructions of the program code, the CPU or other components installed on the expansion board or expansion unit execute some and all of the actual operations, thereby realizing the function of any of the embodiments described above.

[0135] The present invention has been shown and described in detail above with reference to the accompanying drawings and preferred embodiments. However, the present invention is not limited to these disclosed embodiments. Based on the above embodiments, those skilled in the art will know that more embodiments of the present invention can be obtained by combining the code review methods in the different embodiments. These embodiments are also within the protection scope of the present invention.

Claims

1. A method for intelligent distribution and closed-loop efficiency management of government work orders, characterized in that, The implementation of this method includes: (1) Process the work order input data through a multimodal fusion classification model, and achieve high-precision semantic understanding and entity extraction through a deep learning model to generate structured work order entity information; (2) Based on the reinforcement learning dynamic decision engine, the optimal distribution path is calculated according to the real-time department load status and geographical weight; (3) Use a Bayesian network risk warning model to monitor the progress of the disposal and automatically supervise when the overdue probability threshold is triggered.

2. The method for intelligent distribution and closed-loop efficiency management of government work orders according to claim 1, characterized in that, Step (1) involves fusing speech and text dual-modal data. Voice work orders are translated into text using an end-to-end dialect ASR model, and the acoustic model is trained using a region-specific phoneme set. The text work order extracts local features through TextCNN, concatenates them with BERT semantic vectors, and inputs them into the BiLSTM-CRF entity recognition layer; The output entity includes {location, event type, urgency level, and related department}.

3. A method for intelligent distribution and closed-loop management of government work orders according to claim 1 or 2, characterized in that, The specific implementation method of step (1) is as follows: (1.1) Work order input and modal judgment: First, the system receives work order data pushed by the government service platform through an input interface. This data includes the following formats: Voice work orders: Audio recordings made by citizens via hotline calls; Text work orders: Text content submitted by citizens through the APP / website; Then, modality is determined based on the input data format, and the processing branch is automatically routed: if the input is audio, the speech processing branch is started; if the input is text, the text processing branch is started. (1.2) Voice work order processing branch: Audio preprocessing: Denoising and framing the input audio, and extracting MFCC feature vectors; End-to-end dialect ASR translation: The classification model structure is as follows: Input layer: MFCC feature sequences; Convolutional layer: 2 convolutional layers to extract local spectral features; Recurrent layer: 3-layer bidirectional LSTM to capture temporal dependencies; Output layer: CTC layer, outputs character probability distribution; The model is trained using dialect adaptation, employing a regional government speech database, and a hybrid loss function is used. ； in, =0.3, Responsible for aligning audio and text. Strengthen attention to dialect keywords; The model outputs translated text; (1.3) Text work order processing branch: If the work order is in the form of directly entered text, it will proceed directly to the subsequent semantic understanding module without translation. (1.4) Multimodal semantic understanding and entity extraction: Text feature extraction: TextCNN local feature extraction and BERT semantic encoding are used in parallel with dual channels to extract features from escaped text. TextCNN Local Feature Extraction: Input: The work order text after word segmentation. Embedding layer: word vector dimension 300, Convolutional layers: 3 types of convolutional kernels, each kernel outputs 64-dimensional features. Pooling layer: Max pooling, concatenating all convolutional kernels to obtain a 192-dimensional feature vector; BERT semantic encoding: Context-aware text vectors are generated using a pre-trained BERT-base model, and the output corresponding to the [CLS] marker is taken as the overall text representation. The output of TextCNN is fused with the output of BERT.

4. The method for intelligent distribution and closed-loop management of government work orders according to claim 1, characterized in that, In step (2), the optimal distribution path is calculated based on the real-time departmental load status and geographical weight. The state space is defined as follows: ； The reward function is designed as follows: ， in To respond to the delay, This is a predicted value for citizen satisfaction. Number of cross-departmental transfers; actions are output using the DQN algorithm. : This enables the distribution of work orders to the target departments.

5. A method for intelligent distribution and closed-loop management of government work orders according to claim 1 or 4, characterized in that, The implementation of step (2) includes state awareness and data preparation, reward function calculation, decision generation, execution and feedback. (2.1) State Awareness and Data Preparation: The system perceives the current environmental state in real time. This state includes multiple key indicators, which are quantified into a multi-dimensional vector as input to the decision-making algorithm; including: Departmental real-time saturation calculation: Two key data points are retrieved in real time from the government data platform: the number of work orders currently pending in each department. And the department's maximum daily processing capacity pre-entered into the system. Saturation is calculated using the following formula: ； Departmental historical processing time query: Retrieve the average time taken for this department to process similar work orders from the historical work order database; Geographic distance calculation: Using GIS services, calculate the spherical distance between the latitude and longitude of the work order's location and the latitude and longitude of the center point of each department's jurisdiction. Constructing a state vector: Combine all the above indicators into a structured state vector. , used to describe the global situation at the current time t; (2.2) Reward Function Calculation: The system calculates a reward score R for each potential department. The higher the score, the better the overall benefit of allocating the reward to that department. The reward function consists of three weighted components: Time-sensitive rewards: ， in, This is the expected delay time. This is the maximum allowed delay time; These are weighting coefficients; Citizen satisfaction reward: ， It is a predicted value, based on historical data, that predicts the citizen satisfaction score the department may receive after handling such work orders; These are the corresponding weighting coefficients; Collaboration efficiency penalty: ， This indicates the number of times this work order is expected to require cross-departmental collaboration; It is the penalty coefficient; The final reward value R = timeliness reward + citizen satisfaction reward + collaborative efficiency penalty; (2.3) Decision generation: The final allocation decision is made using a deep Q-network (DQN), including the following process: Input: The state vector constructed in step (2.1) Input into the DQN model; Processing: The DQN model is a deep learning network that processes state information internally and evaluates the potential long-term value of allocating to each department for the current state, which is called the Q-value; Output: The DQN model outputs a final Q value for each selectable department, and the system automatically selects the department with the highest Q value as the allocation target for this work order; (2.4) Execution and feedback: After the decision is made, the system automatically dispatches the work order to the selected department through the API of the government system and notifies the relevant person in charge; After a work order is processed, the system collects the actual processing time and the citizen's final satisfaction rating, and compares these actual results with the initial predictions. If the results are better than expected, the decision-making path is reinforced; if the results are worse than expected, the internal parameters of the model are adjusted to avoid making similar decisions in the future.

6. The method for intelligent distribution and closed-loop efficiency management of government work orders according to claim 1, characterized in that, Step (3) is implemented as follows: Constructing Bayesian network nodes: Prior probability Sourced from historical work order database; Overdue risk probability calculation: ； when In such cases, warning instructions are pushed to the responsible person via message queue, and the work order priority is increased.

7. A method for intelligent distribution and closed-loop management of government work orders according to claim 1 or 5, characterized in that, The specific implementation steps of step (3) include: real-time data collection and monitoring, dynamic prediction of overdue risk, early warning triggering and automatic supervision; (3.1) Real-time data acquisition and monitoring: The system performs real-time data perception throughout the entire process of handling all dispatched work orders, including multi-dimensional indicator monitoring and data stream processing. The multi-dimensional indicator monitoring includes: Response time: The system continuously calculates the ratio of the time elapsed for a work order to the standard time limit; the closer this ratio is to 1, the greater the time pressure. Processing progress update: Whether the responsible department for system monitoring updates the processing log at key stages; Secondary interaction with citizens: The system integrates a public opinion monitoring module to capture secondary calls, complaints, or comments from citizens regarding the same event in real time; and automatically judges the positive, negative, or anxious level of citizens' emotions through NLP sentiment analysis technology. Work order complexity: This is a quantitative value assigned at the time of allocation, which is comprehensively evaluated based on multiple factors including event type, number of departments that need to be coordinated, and required resources; The data stream processing involves encapsulating the dynamically changing data into messages and pushing them to the performance monitoring unit in real time via Kafka. (3.2) Dynamic prediction of overdue risk: Bayesian networks are used to comprehensively assess the overdue risk of work orders; including: Constructing a Bayesian network: A Bayesian network contains multiple nodes, representing various monitored metrics; these nodes are connected by directed edges, the direction of which represents probabilistic dependencies. Real-time probability calculation: When new monitoring data flows in, the system inputs it as evidence into the Bayesian network; the network performs probabilistic reasoning based on Bayes' theorem, dynamically calculating the posterior probability that the work order will eventually expire under the current known conditions. Its probability value It is a value between 0 and 1, which quantifies the risk level of the current work order; (3.3) Early Warning Triggering and Automatic Supervision: After a risk is predicted, the system immediately initiates an automated intervention process: Warning trigger conditions: Set a risk threshold. When the calculated probability of overdue payment is greater than the set threshold, the work order is determined to be in a high-risk state and the warning mechanism is triggered immediately. Automatic supervision and execution includes: Message push: The system automatically generates a supervision message and accurately pushes it to the work terminal of the person in charge of the department; the message content includes the probability of high risk of overdue payment; Dynamic priority adjustment: At the same time, the system will automatically increase the priority of this work order in the department task queue. The increase is linked to the risk value and follows the formula: New priority = Original priority × (1 + Probability of overdue). Closed-loop confirmation: After the supervision order is issued, the system continues to monitor. The warning status will only be lifted when the responsible person handles and updates the status, and the system verifies that the work order has indeed been completed.

8. A government work order intelligent distribution and efficiency closed-loop management system, characterized in that, include: The multimodal work order classification module is used to fuse multimodal work order input data, achieve high-precision semantic understanding and entity extraction through deep learning models, and generate structured work order entity information. The work order dynamic distribution module is used to analyze the multi-dimensional dynamic factors of each department in real time through reinforcement learning algorithms, and calculate the optimal distribution path based on the real-time department load status and geographical weight. The efficiency closed-loop monitoring module is used to monitor the progress of work order processing in real time, proactively predict and intervene in potential overdue risks, and automatically supervise the process when the overdue probability threshold is triggered. The system is capable of implementing the method described in any one of claims 1 to 7.

9. A device for intelligent distribution and closed-loop management of government work orders, characterized in that, include: At least one memory and at least one processor; The at least one memory is used to store a machine-readable program; The at least one processor is configured to invoke the machine-readable program to implement the method according to any one of claims 1 to 7.

10. A computer-readable medium, characterized in that, The computer-readable medium stores computer instructions that, when executed by a processor, enable the implementation of the method described in any one of claims 1 to 7.

Images