Urban service agent scheduling method and system based on intention recognition

Abstract

The invention relates to the technical field of smart city services, and discloses a city service agent scheduling method and system based on intention recognition, and the method comprises the steps: collecting the multi-modal original request data of a city complex, and carrying out the preprocessing and feature extraction; identifying a multi-level intention, extracting an intention entity and a constraint condition, and evaluating an intention priority; constructing an intelligent agent capability model and a knowledge base, calculating a semantic matching degree between the intention and the intelligent agent, and generating an intelligent agent combination scheme according to a matching result; constructing a multi-objective optimization model and generating a scheduling scheme of a single intention and a composite intention; task issuing is carried out, execution abnormity is processed, fault-tolerant recovery is carried out, task completion quality is verified, and user feedback is collected; intention recognition optimization is carried out, scheduling strategy parameters are adjusted, and intelligent agent capability evaluation is updated; according to the method, the multi-dimensional intention of the user can be accurately understood, and multi-agent efficient collaboration and resource optimization configuration are realized.

Description

Technical Field

[0001] This invention relates to the field of smart city service technology, and more specifically, to a method and system for scheduling smart agents for urban services based on intent recognition. Background Technology

[0002] With the deepening of smart city construction, the demand for intelligent and precise urban services is becoming increasingly prominent. Urban complexes, as complex scenarios integrating multiple functions such as shopping, dining, entertainment, cultural exhibitions, and transportation hubs, receive a large number of visitors daily, generating massive service requests. To improve service efficiency and user experience, urban complexes deploy various types of intelligent service agents, including passenger flow analysis agents, parking guidance agents, shop recommendation agents, emergency evacuation agents, environmental monitoring agents, and facility maintenance agents. These agents have different service capabilities, response times, and resource consumption characteristics; how to rationally schedule them according to visitor needs has become a key technical issue.

[0003] When expressing service requests, visitors typically use natural language, and their true intentions are often hidden, diverse, and dynamic. For example, a visitor might say, "The air conditioning is too cold; I feel a bit chilly." This request could imply multiple service intentions, such as adjusting the environmental control system, changing their location, or requesting recommendations for warmer spaces. Traditional keyword-based or rule-based matching methods struggle to accurately understand users' deeper intentions, often leading to delayed service responses, wasted resources, and poor user experience. Furthermore, visitor service requests often involve multiple service dimensions and spatiotemporal constraints, requiring coordinated responses from multiple agents; a single agent cannot meet complex needs.

[0004] Current agent scheduling technologies suffer from the following shortcomings: They lack a single dimension for intent recognition, focusing primarily on simple intents within specific scenarios and failing to capture complex intents such as urgency, service quality expectations, time constraints, and personalized preferences in user requests; their application scenarios are highly limited, often targeting single scenarios like medical appointments and parking management, lacking a cross-domain general framework and unable to adapt to the diversity and complexity of urban services; their agent collaboration capabilities are insufficient, mainly focusing on scheduling single resources or service providers, lacking key technologies such as task decomposition, agent combination, and collaborative process design for complex tasks requiring multiple agents to collaborate; and they lack dynamic adaptability, relying heavily on static rules or historical data, lacking the ability to dynamically adjust and optimize for real-time changing environmental factors such as traffic conditions, agent load, and unexpected events.

[0005] Therefore, there is an urgent need for a city service agent scheduling method that can accurately understand users' multi-dimensional intentions, achieve efficient collaboration among multiple agents, dynamically optimize resource allocation, and ensure service reliability, so as to meet the growing demand for intelligent services in smart city construction. Summary of the Invention

[0006] This invention provides a method and system for scheduling urban service agents based on intent recognition, which solves the technical problems of strong application scenario limitations, insufficient agent collaboration capabilities, and lack of dynamic adaptability in related technologies.

[0007] This invention provides a method for scheduling urban service agents based on intent recognition, comprising: Collect multimodal raw request data from urban complexes and perform preprocessing and feature extraction to obtain personalized request feature representations; Based on personalized request feature representation, multi-level intents are identified and intent entities and constraints are extracted. Intent priorities are evaluated to obtain structured intent representations. Based on structured intent representation, an agent capability model and an agent capability knowledge base are constructed. The semantic matching degree between intent and agent is calculated. Based on the matching results, an agent combination scheme is generated, and a candidate set of multi-agent cooperation schemes is obtained. Based on the candidate set of multi-agent cooperative schemes, a multi-objective optimization model is constructed and scheduling schemes for single intents and composite intents are generated. Scheduling strategies and alternative schemes are formulated to obtain a complete scheduling scheme. Tasks are issued based on a complete scheduling scheme, task execution progress is monitored in real time, execution exceptions are handled and fault-tolerant recovery is implemented, task completion quality is verified and user feedback is collected, and task execution results are obtained. Based on the task execution results, intent recognition is optimized, scheduling strategy parameters are adjusted, and agent capability evaluation is updated to obtain a continuous optimization scheme.

[0008] In a preferred embodiment, the step of obtaining the personalized request feature representation includes: Collect text requests, voice requests, image requests, and dialogue log data, with each type of data accompanied by location and time information; Syntactic analysis and named entity recognition are performed on the text request, and a semantic representation is obtained using a pre-trained language model; The speech request is converted into text through speech recognition, and the prosodic features are extracted to determine the urgency of the need. Perform object detection to extract scene features from image requests; perform session segmentation on dialogue logs to identify intent evolution trajectories; Retrieve regional information based on the requested location coordinates and calculate the spatial relationship with surrounding facilities; obtain operational status information based on the requested timestamp; and construct a spatiotemporal context feature vector. User profiles are retrieved by visitor identifiers, and the weights of user profile features are calculated using an attention mechanism. The weighted fusion is then used to obtain a personalized request feature representation.

[0009] In a preferred embodiment, the step of obtaining the structured intent representation includes: A hierarchical intent classification system is constructed. The request feature representation is input into the classification model to predict the intent category according to the hierarchy, and the intent label and confidence score are obtained. For the classification result with a confidence score lower than the preset confidence threshold, the intent clarification mechanism is triggered. Analyze the subjective expressions in the request text, identify the implicit service requirements, establish a mapping rule base from subjective expressions to implicit intentions, and infer the implicit intentions through rule matching and semantic similarity calculation; Identify coordinating conjunctions and logical connectors, break down complex intents into multiple sub-intents, and analyze the dependencies between sub-intents; Entity extraction is performed on the request text and mapped to a normalized representation. Constraints are identified and converted into quantization ranges. The extracted entities are then filled into the parameter slots of the intent. Define the evaluation dimensions for intent priority, including urgency, scope of impact, time sensitivity, resource scarcity, and user importance, calculate a comprehensive priority score, and classify priority levels.

[0010] In a preferred embodiment, the step of obtaining a candidate set of multi-agent cooperative schemes includes: Define the core dimensions of agent capabilities, including service type, service capability boundaries, response time, resource consumption, and collaboration preferences. Fill in the capability parameters based on historical operation data and establish an agent capability knowledge base. Periodically collect the agent's running status, task queue, resource consumption, and performance indicators, and calculate the agent's load score; Extract the capability requirement features of the intent, retrieve the candidate agent set from the agent capability knowledge base, calculate the service type matching degree, capability level matching degree, response time matching degree, resource constraint matching degree, and collaboration matching degree, calculate the comprehensive matching degree, and filter out agents with matching degree lower than the preset matching degree threshold; It identifies composite intents, highly complex intents, and cross-domain intent scenarios that require multi-agent collaboration, decomposes composite intents into sub-intent sequences and analyzes dependencies, and uses combinatorial optimization methods to generate candidate schemes for agent combination.

[0011] In a preferred embodiment, the identification of complex intentions, highly complex intentions, and cross-domain intentions requiring multi-agent collaboration includes: For composite intents, they are decomposed into a sequence of sub-intents, the dependencies between sub-intents are analyzed, and an execution order graph is constructed. For highly complex intentions, a task splitting method is used to break them down into multiple parallel subtasks, and an agent is assigned to each subtask. For cross-domain intentions, retrieve cross-domain collaboration relationships from the knowledge base, select a collaboration mode, and instantiate it into a specific intelligent agent.

[0012] In a preferred embodiment, the step of obtaining the complete scheduling scheme includes: Define the objective function for scheduling optimization, including minimizing response time, maximizing service quality, maximizing resource utilization, and maximizing user satisfaction; define the constraints, including capacity constraints, resource constraints, time constraints, and collaboration constraints. For single-intention scenarios, select agents that meet the constraints, calculate the comprehensive scheduling score, fusion capability matching degree, load score, historical service quality, and response time prediction, and select the optimal agent to generate scheduling instructions; For complex intentions or collaborative situations, the feasibility of each collaborative scheme is verified and multi-objective optimization is evaluated. The optimal scheme is selected and scheduling instructions are generated for each agent. Define the conditions for triggering dynamic scheduling, and re-execute the scheduling decision process after triggering; pre-generate alternative scheduling schemes and establish the trigger conditions for alternative schemes.

[0013] In a preferred embodiment, the conditions for triggering dynamic scheduling include: Monitor the difference between the actual execution progress and the estimated progress. If the difference exceeds a preset deviation threshold, trigger a rescheduling. When the priority difference of a new intent is higher than the preset priority difference threshold, a preemption task is triggered, and the preempted task is re-added to the task queue. When the detection resource usage exceeds the capacity limit, the task will be reallocated. Monitor task failure rate, response time, and user complaints; switch to a backup agent when quality deteriorates.

[0014] In a preferred embodiment, the step of obtaining the task execution result includes: The task message is pushed to the message queue. The agent pulls the task message, parses it, checks its executability, and adds the task to the task queue and sends a confirmation message after the check is passed. The intelligent agent periodically reports its status, the monitoring center updates the task execution status, compares the actual execution progress with the estimated progress to calculate the execution deviation, and identifies execution timeouts, progress stagnation, resource anomalies, and quality anomalies. For agent failures, the faulty agent is isolated and the task is reassigned to a backup agent; In response to execution timeouts, if the timeout is less than the preset timeout threshold, the time window is extended; if the timeout is greater than or equal to the preset timeout threshold, task migration or rescheduling is triggered. After the intelligent agent completes the task, it reports the task result. The scheduling center verifies the result, and the verified task result is pushed to the visitor, triggering the collection of satisfaction evaluation.

[0015] In a preferred embodiment, the step of obtaining the continuously optimized solution includes: Cases with incorrect intent recognition were filtered from the feedback data, manually labeled, samples were selected using an active learning strategy, and the intent recognition model and intent knowledge graph were updated using an incremental training method. Analyze the weight configuration of each optimization objective in the scheduling objective function, adjust the dimension weights in the agent matching degree calculation, optimize the load score calculation method, and adjust the trigger threshold for dynamic scheduling; The actual service timeliness of the agent is statistically analyzed, and the deviation between the actual timeliness and the declared timeliness in the capability model is compared and the response timeliness parameters are updated to update the service capability boundary of the agent.

[0016] This invention provides an intent-based urban service agent scheduling system for executing the aforementioned intent-based urban service agent scheduling method, comprising: The data acquisition module is used to collect multimodal raw request data from urban complexes and perform preprocessing and feature extraction to obtain personalized request feature representations. The intent recognition module identifies multi-level intents and extracts intent entities and constraints based on personalized request feature representations, evaluates intent priorities, and obtains structured intent representations. The agent matching module constructs an agent capability model and an agent capability knowledge base based on structured intent representation, calculates the semantic matching degree between intent and agent, generates agent combination schemes based on matching results, and obtains a candidate set of multi-agent cooperation schemes. The scheduling decision module, based on the candidate set of multi-agent cooperative schemes, constructs a multi-objective optimization model and generates scheduling schemes for single intents and composite intents, formulates scheduling strategies and alternative schemes, and obtains a complete scheduling scheme. The task execution module issues tasks based on a complete scheduling scheme, monitors task execution progress in real time, handles execution exceptions and implements fault-tolerant recovery, verifies task completion quality and collects user feedback, and obtains task execution results. The continuous optimization module optimizes intent recognition based on task execution results, adjusts scheduling strategy parameters, updates agent capability assessment, and obtains a continuous optimization solution.

[0017] The beneficial effects of this invention are as follows: By employing multi-layered intent recognition, dynamic agent evaluation, and multi-objective optimization scheduling, this approach addresses issues such as inaccurate intent understanding, low collaborative efficiency, and unreasonable resource allocation in urban service scenarios. Regarding intent recognition, it integrates multimodal data, contextual analysis, and user profiles to accurately identify explicit and implicit intents, demonstrating particular strength in handling ambiguous expressions and complex needs. It processes complex intent structures through hierarchical classification and reasoning mechanisms. In agent scheduling, it establishes a multi-dimensional capability model and a dynamic evaluation mechanism to achieve precise matching of needs and capabilities. By dynamically selecting the optimal agent through real-time monitoring of load, quality, and resources, it automatically generates and optimizes execution plans for collaborative tasks using combinatorial reasoning and task decomposition algorithms.

[0018] In terms of resource optimization, a multi-objective optimization model is constructed to achieve reasonable allocation and load balancing. Dynamic scheduling and alternative solutions address real-time changes, and a priority evaluation mechanism ensures timely response to high-priority demands. A robust monitoring and fault-tolerance mechanism is designed to handle failures and timeouts promptly. Service continuity is ensured through task migration and solution switching, and the distributed framework supports parallel scheduling and asynchronous communication. Furthermore, a closed-loop optimization mechanism based on user feedback is established. By collecting evaluation and execution data, the model and strategies are continuously optimized, and proactive learning methods are used to quickly adapt to new intents and changing scenarios. This invention accurately understands visitor needs, efficiently schedules resources, optimizes system performance, and ensures reliability, providing intelligent, precise, and efficient solutions for smart city services. Attached Figure Description

[0019] Figure 1 This is a main flowchart of a city service intelligent agent scheduling method based on intent recognition according to the present invention; Figure 2 This is a detailed flowchart of a city service intelligent agent scheduling method based on intent recognition according to the present invention; Figure 3 This is a block diagram of an urban service intelligent agent scheduling system based on intent recognition according to the present invention. Detailed Implementation

[0020] 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, some features described in the examples may be combined in other examples.

[0021] At least one embodiment of the present invention discloses a method for scheduling urban service agents based on intent recognition, such as... Figures 1 to 2 As shown, it includes: Step 1: Collect multimodal raw request data of urban complexes and perform preprocessing and feature extraction to obtain personalized request feature representations; Specifically, the following steps are included: Step 1.1: Multi-channel access request and data collection; Based on mobile applications, smart service terminals, intelligent customer service robots, and image recognition-assisted functions deployed in urban complexes, a unified request access interface is used to collect visitor service requests. Mobile applications collect text content, geographic coordinates, request timestamps, and device identifiers; smart service terminals collect voice data, terminal location, and ambient noise levels; intelligent customer service robots collect voice, facial expressions, images, and conversation logs from multi-turn dialogues; and image recognition functions collect scene photos and shooting location and time information.

[0022] We obtain multimodal raw request data, including text request sets, voice request sets, image request sets, and dialogue log data. All types of data are accompanied by location, time, and device information.

[0023] Step 1.2, Multimodal data preprocessing and feature extraction; Based on the multimodal raw request data obtained in step 1.1, a targeted preprocessing method is used to extract the key features of each modality.

[0024] For text request data, word segmentation and dependency parsing are used to identify sentence structure. Named entity recognition is used to extract entities such as location, time, service type, and facility name. Keyword extraction algorithms are used to identify core words and calculate their weights. A pre-trained language model is used to map the text to the semantic space. The pre-trained model adopts a 12-layer transformer encoder architecture, with each layer containing 12 attention heads and a hidden layer dimension of 768. It is pre-trained on a Chinese corpus of 3 billion words, using a masked language model and a next sentence prediction task, and is trained for one million steps to obtain a distributed semantic representation.

[0025] The model was trained using the Adam optimizer with an initial learning rate of 1e-4 and a linear warm-up strategy of 10,000 steps, followed by cosine annealing to adjust the learning rate. The batch size was 256, the gradient accumulation steps were 4, and the maximum sequence length was 512. The loss function was a weighted sum of the masked language model loss and the next-sentence prediction loss, with a weight ratio of 1:1. On the validation set, the model achieved 72% accuracy in the masked language model and 89% accuracy in the next-sentence prediction.

[0026] After preprocessing the voice request data with denoising and volume normalization, an end-to-end speech recognition model is used to convert it into text and confidence scores. The model includes an acoustic feature extraction module, a six-layer bidirectional long short-term memory network encoder, an eight-head attention mechanism, and a two-layer decoder, and is trained for fifty rounds on a 1,000-hour Chinese speech dataset. Simultaneously, prosodic features such as speech rate, pitch, and pauses are extracted to identify the visitor's emotional state and the urgency of their needs.

[0027] For image request data, a deep convolutional neural network is used for object detection and scene recognition. The network uses a 50-layer residual network as its backbone, containing four residual block groups. After pre-training for 100 epochs on a dataset of one million images, it is fine-tuned for 20 epochs on a dataset of 50,000 composite scenes. The output includes object category, location, confidence score, and scene features, which are then converted into textual semantic features using an image description generation method.

[0028] For dialogue log data, we perform conversation segmentation to identify topic transition points, extract key information fragments, perform referential resolution to restore pronouns to entity expressions, and count dialogue rounds, response lengths, and semantic redundancy to assess the clarity of visitor expression.

[0029] The feature representation sets for each modality are obtained, including text semantic vectors, entity annotation sequences, speech prosody feature vectors, image scene feature vectors, and dialogue structure features.

[0030] Step 1.3, Location semantic enhancement and spatiotemporal context construction; Based on the feature representations and location-time information extracted in step 1.2, a location semantic enhancement method is used to enrich the contextual information of the request.

[0031] Based on the requested location coordinates, the system retrieves corresponding regional information from a spatial database constructed using a Geographic Information System (GIS), storing building outlines, floor plans, facility location coordinates, and road network topology. The system calculates the distance between the requested location and surrounding key facilities, employing nearest neighbor search and Dijkstra's shortest path algorithm to determine the actual walking distance and estimated walking time. Walking speed is dynamically adjusted based on pedestrian density: 1.2 meters per second in low-density areas, 0.9 meters per second in medium-density areas, and 0.6 meters per second in high-density areas. Spatial accessibility is categorized based on distance: near distance (less than 100 meters), medium distance (100 to 300 meters), and far distance (greater than 300 meters).

[0032] Based on the request timestamp, query the time database to obtain visitor traffic statistics, agent busyness, and historical service request hotspot distribution for the current time period, and determine whether the request time is during a holiday, weekend, or weekday.

[0033] A spatiotemporal context feature vector is constructed, encoding location and time semantic information into numerical features. This vector is then concatenated with the modal features obtained in step 1.2 to form an enhanced request feature representation. This yields a request feature representation fused with spatiotemporal context.

[0034] Step 1.4: User profile association and personalized feature fusion; User profiles are retrieved from the user profile database using the visitor's device identifier or membership account. User profiles include service preferences, behavioral habits, and service evaluation history. Service preferences include store type preference tags, price sensitivity levels, and service method preferences; behavioral habits include visit frequency, average stay duration, frequently visited area distribution, and active time period preferences; service evaluation history includes satisfaction ratings, written feedback, and response time sensitivity.

[0035] For new visitors or those with incomplete profile information, a rule-based cold start strategy is employed for feature completion. The rule base contains over two hundred mapping rules, categorized into 40 rules for demographic features, 30 rules for time features, 35 rules for environmental features, 25 rules for price features, 30 rules for service methods, and 40 rules for special needs. The rule base is constructed through a combination of expert knowledge and data mining, with rules defined as those where the co-occurrence frequency of keywords and features exceeds 0.6. Keyword matching combines exact matching and semantic matching. Semantic matching is calculated using the cosine similarity of word vectors, with a similarity greater than 0.85 considered a successful match. For missing feature dimensions in the profile, a collaborative filtering method is used to calculate the similarity with historical users, selecting the top ten users with a similarity greater than 0.7, and extracting the average of their feature values ​​for completion.

[0036] User profile features are encoded into feature vectors, and an attention mechanism is used to calculate the importance weights of these features for request understanding. The attention mechanism maps request features and profile features to 128-dimensional query and key vectors through a linear transformation. The transpose of the query and key vectors is used to calculate the dot product similarity, which is then divided by a scaling factor (the square root of 128, approximately 11.31) to stabilize the gradient. The attention weight distribution is obtained through a softmax normalization function. The attention weights are multiplied by the value vectors of the profile features to obtain a weighted feature representation. Feature dimensions with attention weights greater than 0.3 are marked as strongly relevant features, while those with weights less than 0.1 are filtered out. A personalized request feature representation, including multimodal content features, spatiotemporal context features, and user profile features, is obtained through weighted fusion and serves as input data for intent recognition.

[0037] Step 2: Based on the personalized request feature representation, identify multi-level intents and extract intent entities and constraints, evaluate intent priorities, and obtain a structured intent representation; Specifically, the following steps are included: Step 2.1, Explicit intent extraction and category mapping; Based on the personalized request feature representation output in step 1, a city service intent classification system is constructed, dividing intents into three levels. Level 1 intents include six major categories: commercial services, facility services, environmental services, transportation services, emergency services, and information services. Level 2 intents are further subdivided under the Level 1 categories; for example, commercial services include subcategories such as shop search, product consultation, obtaining promotional information, and shopping guidance. Level 3 intents describe specific service actions and target objects.

[0038] A hierarchical intent classification model was trained, consisting of three classifier layers. The first classifier layer employs a bidirectional long short-term memory (LSTM) network architecture, comprising a word embedding layer, a bidirectional LSM network layer, and a fully connected classification layer. The word embedding layer maps words to 300-dimensional word vectors. The bidirectional LSM network layer has a hidden layer dimension of 256 in each direction, outputting a 512-dimensional sentence representation. The fully connected layer outputs the probability distributions of six primary intent categories. The training data consisted of 50,000 labeled samples, optimized using stochastic gradient descent with a learning rate of 0.002, a batch size of 32, and 10 training epochs. The second classifier layer trained six dedicated secondary classifiers for each primary category, with identical model structures. The training data consisted of 5,000 to 12,000 samples corresponding to the primary category. The third classifier layer used a lightweight model with a unidirectional LSM network hidden layer dimension of 128. The training data consisted of 800 to 3,000 samples, a learning rate of 0.001, a batch size of 16, and 8 training epochs.

[0039] During model inference, request features are input into the first-layer classifier to obtain the primary intent. Based on the primary intent, the corresponding second-layer classifier is selected to predict the secondary intent, and then the third-layer classifier is selected to predict the tertiary intent. The classification confidence thresholds are set to 0.7 for the first layer, 0.65 for the second layer, and 0.6 for the third layer. When the predicted probability is lower than the classification confidence threshold, an intent clarification mechanism is triggered, selecting the two candidate categories with the highest probabilities to generate clarification questions for the visitor. Clarification question generation uses a template-filling method, with predefined clarification question templates for each intent category. For example, when the predicted probabilities for business services and facility services are 0.68 and 0.65 respectively, the system generates the clarification question: "Do you want to inquire about shop information or do you need facility services?" After the visitor selects, the system uses the selected category as the final intent, updating the confidence level to 1.0. If the visitor does not respond within 30 seconds, the system defaults to selecting the category with the highest probability for further processing.

[0040] Obtain hierarchical labels for explicit intents, including first-level intent categories, second-level intent subcategories, third-level intent descriptions, and classification confidence levels for each level.

[0041] Step 2.2, Implicit Intent Reasoning and Complex Intent Recognition; Based on the explicit intent tags and personalized request feature representations identified in step 2.1, the subjective expressions and sentiment tendencies in the request text are analyzed to identify implicit service needs. A mapping rule base from subjective expressions to implicit intents is established, with rules summarized from domain knowledge and historical data, containing over two hundred mapping rules. An intent reasoning model is used to directly match common subjective expressions with the corresponding implicit intent candidate set through the rule base. For expressions not covered by the rule base, a semantic similarity calculation method is used to retrieve the semantically most similar intent node in the intent knowledge graph; a similarity greater than 0.85 is considered a successful match.

[0042] The system identifies the complex intent structure in requests, using syntactic analysis to identify coordinating conjunctions and logical connectors, breaking down the complex intent into multiple sub-intents. It then analyzes the dependencies between these sub-intents, determining whether there are sequential constraints or conditional dependencies. For each inferred implicit intent and sub-intent, the system calculates its association strength with the explicit intent and its probability of independence. The association strength is comprehensively evaluated using intent co-occurrence frequency and semantic similarity.

[0043] The resulting set of intents includes explicit intent labels, a list of implicit intent candidates, a sequence of sub-intents of a composite intent, and a dependency graph between the sub-intents.

[0044] Step 2.3, Intent Entity Extraction and Constraint Identification; Based on the intent set and personalized request feature representation obtained in step 2.2, entity extraction is performed on the request text to identify key entities related to the intent. Entity categories include location entities, time entities, object entities, attribute entities, and quantity entities. A named entity recognition model based on conditional random fields or bidirectional long short-term memory networks is adopted, with character or word sequences as input and entity labels for each location as output. The extracted entities are standardized to map colloquial expressions to normalized entity representations.

[0045] Identify the constraints in the request. Constraint categories include spatial constraints, temporal constraints, resource constraints, quality constraints, and preference constraints. Spatial constraints convert natural language expressions into quantified distances or regional ranges; temporal constraints map to specific points in time or time periods; resource constraints limit the types of resources that can be used; quality constraints describe service quality requirements; and preference constraints express the visitor's personalized preferences.

[0046] Establish the association between intents and entities, and populate the extracted entities into the parameter slots of the intents. Each intent category defines a set of parameter slots. Entities are populated into the corresponding slots by matching the entity category with the slot semantics, forming a structured intent representation. Unfilled slots are marked as missing parameters, and visitors can be asked to supplement them through a clarification dialogue.

[0047] A preliminary structured intent representation is obtained, including intent category labels, parameter slots and their values, a list of constraints, and the quantization range of the constraints.

[0048] Step 2.4, Prioritization of Intent and Determination of Urgency; Based on the preliminary structured intent representation and request context information obtained in step 2.3, the evaluation dimensions of intent priority are defined, including five dimensions: urgency, scope of impact, time sensitivity, resource scarcity, and user importance.

[0049] The urgency assessment employs a combination of rules and models. Emergency service intents are tagged with an urgency level, while other intent categories use an urgency prediction model to output an urgency score between 0 and 1. Impact scope is based on the location entities associated with the intent and spatial constraints, querying the population density and functional importance of the corresponding area, and calculating the average daily usage of facilities and the accessibility of alternative facilities. Time sensitivity is based on the intent's time constraints and current time period characteristics, extracting time entities and mapping them to a quantified time window width; a narrower time window results in a higher time sensitivity score. Resource scarcity is based on the agent type required by the intent and the current resource status, calculating the number of candidate agents capable of responding to the intent and the number of currently available agents. User importance is based on user profiles including membership level, spending amount, visit frequency, and complaint history.

[0050] The scores from the five dimensions are weighted and integrated to calculate a comprehensive priority score for the intent. Weighting coefficients are set based on the urban complex's operational strategy; urgency and scope of impact are typically assigned high weights, while the weights for time sensitivity and resource scarcity are dynamically adjusted based on the current system load. Priority levels are categorized based on the comprehensive priority score into four levels: urgent, high, medium, and low. A comprehensive priority score of 0.8 or higher qualifies as urgent, 0.6 to 0.8 as high, 0.3 to 0.6 as medium, and below 0.3 as low.

[0051] The complete structured intent representation is obtained, including intent category labels, parameter slots and their values, a list of constraints, priority sub-scores for five dimensions, a comprehensive priority score, and a priority level label.

[0052] Step 3: Based on the structured intent representation, construct the agent capability model and knowledge base, calculate the semantic matching degree between the intent and the agent, generate agent combination schemes according to the matching results, and obtain a candidate set of multi-agent cooperation schemes; Specifically, the following steps are included: Step 3.1, Construction of the intelligent agent capability model and maintenance of the knowledge base; Five core dimensions define the capabilities of an intelligent agent: Service type describes the set of intent categories that the agent can handle, and uses a hierarchical structure to identify the coverage; Service capability boundary defines the complexity of a single service, the upper limit of the number of concurrent requests, and the ability to handle special scenarios; Response timeliness includes preparation time, execution time, and feedback time; Resource consumption quantifies the computational resources, communication bandwidth, physical equipment, and human resources input; Collaboration preference identifies three role types: independent execution, leading coordination, or auxiliary cooperation.

[0053] Establish standardized capability description templates and populate capability parameters for each agent instance based on deployment configuration and historical operational data. Deployment configurations provide service type declarations, resource quotas, and interface specifications, while historical data statistics include actual response time, task completion rate, and average resource consumption. Compare declared capabilities with measured capabilities to identify discrepancies and update the capability model.

[0054] A knowledge base of agent capabilities is established using a graph database architecture. Agent instances are nodes containing capability description parameters, collaboration relationships are edges recording collaboration patterns and historical effects, and intent categories are also edges establishing a "can handle" relationship between the node and the agent, with edge weights reflecting the strength of the processing capability.

[0055] Obtain an agent capability knowledge base, which includes agent capability descriptions, collaborative relationships between agents, and mappings of intents to agent capabilities.

[0056] Step 3.2, Real-time status monitoring and load assessment of the intelligent agent; Deploy an agent status monitoring agent to periodically collect status data. The collection period is dynamically adjusted according to the agent type; agents with high-frequency interactions use a second-level collection period, while agents with low-frequency services use a minute-level collection period. Status data includes four categories of information: running status, task queue, resource usage, and performance metrics.

[0057] The operational status includes four states: idle, busy, fault, and maintenance, determined by heartbeat detection and task feedback. The task queue records the list of pending and currently being processed tasks, and tracks queue length and average task waiting time. Resource consumption quantifies the agent's current computational resource usage, memory usage, network bandwidth usage, and external device status. Performance metrics evaluate task success rate, average response time, service satisfaction, and frequency of abnormal events.

[0058] Based on the collected state data, the agent's load score is calculated. The load score calculation formula is: Load score equals queue load weight multiplied by queue load rate, plus resource load weight multiplied by resource bottleneck index, plus time load weight multiplied by response time degradation coefficient. Queue load rate equals the current task queue length divided by the queue capacity limit, which is set to 10 to 25 depending on the agent type. Resource bottleneck index is the maximum value of CPU utilization and memory utilization. Response time degradation coefficient equals the current average response time divided by the baseline response time, which is the agent's average response time under light load conditions. The three weight coefficients are optimized using a grid search method: queue load weight 0.4, resource load weight 0.35, and time load weight 0.25. When the queue load rate, resource bottleneck index, or response time degradation coefficient exceeds 1, the value of that item is truncated to 1. If the calculated load score exceeds 1, it is truncated to 1; if it is less than 0, it is truncated to 0, ensuring that the load score value strictly falls within the range of 0 to 1.

[0059] The load score ranges from 0 to 1, and load levels are determined based on the load score: a load score less than 0.3 indicates light load, 0.3 to 0.7 indicates medium load, and a load score greater than or equal to 0.7 indicates high load. Scheduling decisions prioritize agents with load scores less than 0.5, and new tasks are no longer accepted when an agent's load score is greater than 0.8.

[0060] Obtain a real-time snapshot of the agent's status, including running status identifiers, task queue details, resource usage data, performance metric statistics, and load scores.

[0061] Step 3.3, Agent capability matching based on intent semantics; Extract the capability requirement features of the intent and transform the intent into a capability requirement vector. The capability requirement features include five dimensions: service type requirement, capability level requirement, response time requirement, resource constraints, and collaboration requirement. Service type requirement is directly derived from the category label of the intent; capability level requirement is evaluated based on the complexity of the intent and the number of constraints; response time requirement is determined based on the time sensitivity and priority score of the intent; resource constraints limit the resource consumption of the agent to not exceed the system's allocable resource quota; and collaboration requirement indicates whether the intent requires multi-agent collaboration.

[0062] Retrieve a set of candidate agents capable of handling the intent category from the agent capability knowledge base. Based on the intent category node, traverse the "capable of handling" relation edges in the knowledge base to obtain a list of connected agent nodes. If the intent is a composite intent, retrieve the agent sets corresponding to each sub-intent separately, and take the union of these sets as candidates.

[0063] For each candidate agent, its capability matching degree with the intent is calculated. The matching degree calculation employs a multi-dimensional similarity metric, including five dimensions: service type matching degree, capability level matching degree, response time matching degree, resource constraint matching degree, and collaboration matching degree. Service type matching degree is calculated through hierarchical matching between the intent category and the agent's service type, with a perfect match at level three. Capability level matching degree compares the agent's service capability boundaries with the intent's capability level requirements. Response time matching degree assesses whether the agent's response time meets the intent's time requirements. Resource constraint matching degree checks whether the agent's resource consumption is within the allocable quota. Collaboration matching degree assesses the consistency between the agent's collaboration preferences and the intent's collaboration requirements.

[0064] The matching scores from each dimension are weighted and fused to calculate the overall matching degree. The weighting reflects the importance of each dimension; service type and capability level matching degrees are typically given high weights, while the weight of response time matching degree is dynamically adjusted based on intent priority. A matching degree threshold is set to filter out agents with matching degrees below the threshold.

[0065] The matching results between intent and agent are obtained, including a list of candidate agents, a comprehensive matching score for each agent, and matching details by dimension.

[0066] Step 3.4, Evaluation of the combined capabilities of multi-agent collaboration; Identify intent scenarios that require multi-agent collaboration, including complex intents, highly complex intents, and cross-domain intents. Figure 3 Classes. Composite intents contain multiple sub-intents, each of which may require different types of agents to respond; high-complexity intents require a level of capability that exceeds the capability boundary of a single agent, requiring multiple agents to share the workload or provide complementary capabilities; cross-domain intents involve multiple service domains, requiring collaborative work between agents across domains.

[0067] For composite intents, they are decomposed into a sequence of sub-intents, with each sub-intent undergoing independent agent matching. Dependencies between sub-intents are analyzed, including three types: sequential dependency, conditional dependency, and data dependency. An execution order graph of the sub-intents is constructed based on these dependencies. Candidate agent combination schemes are generated, and a combinatorial optimization method is used to search for the optimal scheme. Optimization objectives include minimizing total completion time, minimizing resource consumption, maximizing combinatorial matching degree, and minimizing collaboration overhead.

[0068] For highly complex intents, a task partitioning method is used to break the intent down into multiple parallel subtasks, reducing the capability requirements of each subtask to the range that a single agent can handle. Agents are assigned to each subtask to ensure balanced load across agents.

[0069] For cross-domain intents, cross-domain collaboration relationships are retrieved from the knowledge base to identify agent composition patterns that enable collaborative work. Collaboration patterns define the roles, communication interfaces, and data exchange formats of the agents. A collaboration pattern matching the intent requirements is selected, and the roles within the pattern are instantiated as specific agents.

[0070] The feasibility and performance of each agent combination scheme are evaluated. Feasibility checks include agent availability verification, resource constraint satisfaction checks, and cooperation interface compatibility verification. Performance evaluation calculates the estimated completion time, total resource consumption, combination matching degree, and probability of successful cooperation for each scheme.

[0071] A candidate set of multi-agent cooperation schemes is obtained, including the agent combination, sub-task allocation, execution order, cooperation protocol and performance evaluation index of each scheme.

[0072] Step 4: Based on the candidate set of multi-agent cooperation schemes, construct a multi-objective optimization model and generate scheduling schemes for single intents and composite intents. Formulate dynamic scheduling strategies and alternative schemes to obtain a complete scheduling scheme. Specifically, the following steps are included: Step 4.1, Modeling and defining the scheduling optimization objectives and constraints; The objective function for scheduling optimization is defined, comprising four optimization goals: minimizing response time, maximizing service quality, maximizing resource utilization, and maximizing user satisfaction. The response time minimization goal aims to minimize the time from intent submission to service commencement, which includes task allocation latency, agent preparation time, and arrival time. The service quality maximization goal requires selecting agents with strong service capabilities and good historical performance, measured by agent capability matching and historical service quality scores. The resource utilization maximization goal aims to fully utilize system resources, measured by agent load balancing and resource occupancy. The user satisfaction maximization goal is comprehensively evaluated based on the degree of intent priority satisfaction, response time satisfaction, and service method preference satisfaction.

[0073] The constraints for scheduling decisions are defined into four categories: capability constraints, resource constraints, time constraints, and collaboration constraints. Capability constraints require that the assigned agent must possess the service capability to process the intent, the agent's service type must cover the intent category, and the capability boundary must satisfy the intent requirements. Resource constraints limit scheduling decisions to the total amount of resources available in the system, and the cumulative number of tasks assigned to each agent and the resource consumption cannot exceed the capacity limit. Time constraints require tasks to be completed within the intent's time window, and the sum of the agent's estimated response time and service duration cannot exceed the time window width. Collaboration constraints define the collaboration relationships and execution order in a multi-agent collaboration scheme; the execution order of sub-intents must satisfy dependencies.

[0074] The mathematical model of the scheduling optimization problem is obtained, including the expressions of four optimization objective functions, the mathematical descriptions of four types of constraints, and the definitions of decision variables.

[0075] Step 4.2, Single-Intent Single-Agent Scheduling Decision; Select a subset of agents that meet the capability and resource constraints from the candidate set of multi-agent cooperation schemes, exclude agents with insufficient capabilities, over-quota resources, faulty or under maintenance status, and retain only agents that are idle or have spare capacity.

[0076] For the selected subset of intelligent agents, a comprehensive scheduling score is calculated. The comprehensive score integrates four dimensions: capability matching degree, load score, historical service quality, and response time prediction. After normalizing the four scoring dimensions, a weighted fusion is performed, with the weight coefficients dynamically adjusted according to intent characteristics: high-priority or time-sensitive intents have increased response time weight, intents with high quality requirements have increased service quality weight, and intents with high system load periods have increased load score weight.

[0077] Sort agents by comprehensive score in descending order, and select the agent with the highest score that meets the time constraint. If the estimated completion time exceeds the time window, select the agent with the next highest score. If none of the agents can meet the time constraint, decide whether to accept the timeout risk or trigger the task queuing mechanism based on intent priority.

[0078] Generate scheduling instructions, including agent identifier, intent content, execution parameters, time requirements, and service location. Convert the structured representation of the intent into a task description that the agent can understand, set priority identifiers, calculate and provide estimated completion times.

[0079] A single-intention, single-agent scheduling scheme is obtained, including the selected agent, scheduling instructions, estimated completion time, and scoring criteria.

[0080] Step 4.3, Multi-agent collaborative scheduling decision-making based on composite intent; Feasibility verification is performed on each collaboration scheme, checking the availability of agents, resource constraints, compatibility of collaboration interfaces, and the rationality of the execution order of sub-intents. Infeasible schemes are eliminated, and the set of feasible schemes is retained.

[0081] The feasible solutions are evaluated using multi-objective optimization, calculating the performance of each solution across four optimization objectives. The response time objective calculates the total completion time of the solution; the longest time is taken for parallel sub-intents, and the time for serial sub-intents is the sum of their times, plus communication latency and coordination time. The service quality objective assesses the combination matching degree, using the geometric mean or minimum value to reflect the weakest link effect, and evaluating the probability of successful collaboration based on historical experience and agent stability. The resource utilization objective calculates the total resource consumption and load balancing degree; load balancing degree is measured by the variance of the load score. The user satisfaction objective assesses response time, service content completeness, and the friendliness of the collaboration process.

[0082] After normalizing the performance of each solution across the four objectives, a multi-objective optimization method is used to select the optimal solution. The selection of the multi-objective optimization method is determined based on the number of feasible solutions and intent characteristics. When the number of feasible solutions is greater than 10, Pareto optimality screening is first used to identify the non-dominated solution set, filtering out dominated solutions. When the intent priority is urgent or high, a weighted summation method is used, dynamically adjusting the weights according to the intent characteristics: the response time weight for urgent intents is set to 0.5, service quality weight to 0.3, resource utilization weight to 0.1, and user satisfaction weight to 0.1; the weights for high-priority intents are 0.35, 0.35, 0.15, and 0.15, respectively. The solution with the largest weighted sum is selected as the optimal solution. When the intent priority is medium or low, the ideal point method is used. The ideal point is defined as a vector composed of the optimal values ​​of each objective. The Euclidean distance between each solution and the ideal point is calculated, and the solution with the closest distance is selected.

[0083] After selecting the optimal solution, scheduling instructions are generated for each agent, including agent identifier, assigned sub-intent, execution order, collaboration information, and time requirements. Collaboration information includes roles, communication partners, and data exchange content. Trigger conditions between sub-intents are set to ensure smooth execution. A collaboration monitoring mechanism is established to track execution progress and collaboration status.

[0084] The main scheduling scheme is obtained, including the selected cooperation scheme, scheduling instructions, execution order, cooperation protocol definition and estimated completion time.

[0085] Step 4.4: Generation of dynamic scheduling strategies and alternative solutions; The system monitors operational status and environmental changes, including agent state changes, the arrival of new intents, resource supply fluctuations, and external environmental disturbances. Agent state changes include agent failures, sudden load increases, or task delays, which can cause the original scheduling scheme to fail or degrade performance. The arrival of new intents, especially high-priority intents, may require preemption of current resources or rescheduling of task queues.

[0086] The conditions for triggering dynamic scheduling are defined, including four categories: scheduling scheme execution deviation exceeding a threshold, preemption of high-priority intents, resource constraint violations, and service quality degradation. Scheduling scheme execution deviation monitors the difference between the agent's actual execution progress and the estimated progress; if the difference exceeds a set threshold of 20% to 30%, rescheduling is triggered. The high-priority intent preemption mechanism allows urgent intents to interrupt the execution of current low-priority tasks, determining whether the priority of the new intent is significantly higher than that of tasks in the current task queue; if the priority difference is more than two levels, preemption is triggered. Resource constraint violations detect whether system resource usage exceeds the capacity limit or whether the agent's load exceeds its carrying capacity. Service quality degradation monitors the agent's service quality indicators, such as increased task failure rate, significantly increased response time, or an increase in user complaints.

[0087] After dynamic scheduling is triggered, the scheduling decision process from steps 4.1 to 4.3 is re-executed to generate a new scheduling scheme based on the current system state and the remaining intent queue. The new scheme takes into account the execution progress of already assigned tasks to avoid repeatedly scheduling tasks that have been completed or are currently being executed.

[0088] Alternative scheduling schemes are pre-generated to supplement the primary scheme and serve as backups in case of failure. Alternative schemes select the agent or collaborative scheme with the second-best overall score, quickly switching to a backup scheme when the primary scheme fails or performs poorly. The number of backup schemes is determined based on intent priority and system load; high-priority intents generate multiple backup schemes to improve reliability. Triggering conditions for backup schemes are established, including primary scheme agent failure, primary scheme execution timeout, and primary scheme service quality not meeting standards. The monitoring system automatically activates a backup scheme when the triggering conditions are detected.

[0089] The complete scheduling scheme is obtained, including the main scheduling scheme, the set of alternative schemes, the dynamic scheduling strategy, and the definition of triggering conditions.

[0090] Step 5: Based on the complete scheduling scheme, tasks are issued, task execution progress is monitored in real time, execution exceptions are handled and fault-tolerant recovery is implemented, task completion quality is verified and user feedback is collected to obtain task execution results; Specifically, the following steps are included: Step 5.1: Issuance of scheduling instructions and task initialization; The task message structure is constructed, containing seven fields: task identifier, intent content, agent identifier, execution parameters, priority identifier, time requirement, and collaboration information. The task identifier is a unique number used for task tracking and status query. The intent content includes a structured representation of the intent converted into a format parsable by the agent. Execution parameters include parameter slot values, constraints, and service location coordinates. The priority identifier indicates the urgency of the task. The time requirement describes the task's time window and estimated completion time. For collaborative tasks, the collaboration information includes the collaboration roles, collaboration partners, and data exchange specifications.

[0091] Task messages are pushed to a message queue, which is partitioned by agent. Each agent subscribes to messages in its corresponding partition. The message queue guarantees reliable message delivery and ordering. Messages are cached in the queue when an agent is offline and received in order when the agent comes back online. Message priorities are set, with messages for higher-priority tasks being delivered first.

[0092] The agent retrieves task messages from the message queue and performs task parsing and initialization. It parses the intent content and execution parameters, and loads the required configuration files and resource data. It checks the task's executability, including whether the service type is within its capabilities, whether resources are sufficient, and whether the execution parameters are valid. If the task is unexecutable, the agent reports the rejection reason to the scheduling center, triggering a switch to an alternative solution.

[0093] After a task passes the executability check, the agent adds it to its internal task queue and arranges its execution order according to priority identifiers. The agent updates its state, adds the task identifier to the task list, increments the load count, and records the task reception time. A task acknowledgment message is sent to the scheduling center, containing the task identifier, reception time, and estimated start time. The scheduling center then updates the task status to "received."

[0094] The task issuance confirmation results are obtained, including the task reception status of each agent, the task queue update status, and the estimated start time.

[0095] Step 5.2, monitoring and tracking the task execution process; During task execution, the agent periodically reports its status, which includes task progress, current stage, location information, and resource consumption. Task progress is expressed as a percentage or stage identifier. The current stage describes the specific execution steps of the task. For agents that need to move, the location information reports the current coordinates and movement trajectory. Resource consumption reports the usage of computing, communication, and physical resources during task execution.

[0096] The monitoring center receives status reports from the agents and updates the task execution status database. The database records the complete lifecycle of a task, including creation time, distribution time, receipt time, start time, timestamps for each stage, and estimated completion time. The actual execution progress is compared with the estimated progress to calculate the execution deviation and deviation rate.

[0097] Set up anomaly detection rules to identify abnormal situations during task execution. Anomalies are categorized into four types: execution timeout, progress stagnation, resource anomalies, and quality anomalies. Execution timeout checks if the actual execution time exceeds a set multiple of the estimated time, triggering a timeout alarm to notify the scheduling center and visitors. Progress stagnation checks if the task progress has not been updated for an extended period; the monitoring center sends a heartbeat query to the agent, and if there is no response, the agent is considered faulty. Resource anomalies detect if the agent's resource consumption exceeds expectations or if resources are exhausted, triggering a resource adjustment process. Quality anomalies check whether the task output meets quality standards, determined through intermediate result verification or output result validation.

[0098] For collaborative tasks, monitor the collaboration progress and communication status between agents. Track the execution status of each sub-intention, identify the dependency fulfillment between sub-intentions, and determine whether subsequent sub-intentions can be triggered. Monitor data exchange between agents and detect communication delays or data format errors.

[0099] The system visualizes task execution status, displaying a task list, execution progress, agent distribution, and anomaly alerts on a monitoring panel. Administrators can view system operation in real time and manually intervene or adjust strategies for abnormal tasks.

[0100] Obtain real-time monitoring data for task execution, including task progress updates, execution deviation statistics, abnormal event records, and collaboration status tracking.

[0101] Step 5.3, Task execution exception handling and fault tolerance mechanism; In response to agent failures, a fault detection and isolation process is initiated. Fault detection is triggered by heartbeat timeouts, abnormal status reporting, or external system alarms to confirm that the agent has entered a fault state. The faulty agent is isolated from the available resource pool and marked as unschedulable. An impact assessment is performed on the tasks currently being executed by the faulty agent to determine whether the tasks are recoverable or require rescheduling. For tasks with low execution progress that can be quickly restarted, execution is attempted to continue after the faulty agent recovers. For tasks with high execution progress or time constraints, a task migration process is triggered to reassign the tasks to alternative agents.

[0102] The task migration process selects a candidate agent from the alternative options. If no candidate is pre-generated, the scheduling decision process is re-executed. Task status and execution progress data are passed to the candidate agent. Status data includes the task identifier, a list of completed sub-steps, intermediate result data, and remaining parameters. The candidate agent determines its current execution position based on the list of completed sub-steps, loads the intermediate result data, and continues execution from the next incomplete sub-step. If the task does not support resuming interrupted execution, the candidate agent executes the entire task flow from the beginning. The resuming interrupted execution capability of task types is predefined in the agent capability model. Task types that support resuming interrupted execution include data retrieval, path planning, and information aggregation. Task types that do not support resuming interrupted execution include real-time interaction and state-sensitive tasks. The task allocation record is updated, and the visitor agent is notified of the change and the new estimated completion time.

[0103] In case of execution timeout, initiate the timeout handling process. Analyze the cause of the timeout, including slow agent processing, external dependency delays, or underestimated task complexity. If the timeout is within a tolerable range and the agent is still executing normally, extend the task's time window, update the estimated completion time, and notify the visitor. If the timeout is significant or there is a risk of failure, trigger task migration or rescheduling.

[0104] To address progress stalling, first attempt to wake up or restart the agent. Send a wake-up command to the agent and check if the stalling is caused by entering a hibernation or waiting state. If waking up fails, attempt to restart the agent service to clean up any potential resource deadlocks or state anomalies. If restarting still fails to resolve the issue, it is considered a fault, and task migration is triggered.

[0105] In response to resource anomalies, initiate a resource adjustment process. If resource consumption exceeds expectations but remains within system capacity, dynamically scale up the agent by allocating additional computing, memory, or bandwidth resources. If resource exhaustion leads to service failure, release resources held by other tasks to prioritize the completion of the current task. If resource adjustments fail to resolve the issue, migrate the task to an agent with sufficient resources.

[0106] For quality anomalies, initiate a quality remediation process. For automatically remediable quality issues, such as incorrect data format or parameter configuration, the system automatically corrects and re-executes the task. For quality issues requiring manual intervention, trigger a manual review process where experts analyze the cause of the problem and provide a remediation plan. If the quality fails to meet standards, the task is marked as failed, the visitor is notified, and an alternative solution is provided.

[0107] For anomalies in collaborative tasks, initiate a collaborative repair process. If a sub-intention fails, assess its impact on subsequent sub-intentions. If it's a non-critical dependency, skip the failed sub-intention and continue. If it's a critical dependency, the entire collaborative task fails, triggering a rescheduling or notifying the visitor that the task cannot be completed. For communication anomalies between agents, check network connectivity and protocol compatibility, fix communication issues, and retry data exchange.

[0108] Record all abnormal events and their handling processes to establish an abnormal event database. The database includes the abnormal type, occurrence time, involved agent, cause, handling method, and result. Statistically analyze the frequency and impact of abnormal events to identify system weaknesses and high-risk scenarios.

[0109] Obtain the results of exception handling and the execution record of the fault tolerance mechanism, including fault recovery operations, task migration records, resource adjustment logs and quality repair plans.

[0110] Step 5.4: Task completion confirmation and result feedback; After completing a task, the agent reports the task result, which includes the task identifier, completion time, execution summary, service output, and resource consumption statistics. The scheduling center receives the task result and verifies it. Verification includes checking output integrity, format compliance, and content rationality. Tasks that pass verification are marked as successfully completed; those that fail trigger quality anomaly handling. Successfully completed tasks are updated to "Completed," and their completion time and execution time are recorded. The deviation between the actual execution time and the estimated time is calculated, and this deviation data is used to optimize the time prediction model.

[0111] The task results are pushed to visitors via various methods, including app notifications, SMS, and voice announcements, based on visitor preferences. The push notification includes a summary of the service results, a link to detailed information, and an entry point for satisfaction rating. For collaborative tasks, the execution results of each sub-intention are aggregated to generate an overall service report, which is then pushed to visitors.

[0112] Trigger the satisfaction rating collection process and guide visitors to provide feedback after receiving the results. Rating dimensions include responsiveness, service quality, agent attitude, and overall satisfaction. Rating methods include star ratings, written reviews, and feedback. Implement an incentive mechanism to reward visitors who complete the rating with points or coupons, thereby increasing participation.

[0113] Collect visitor review data and store it in a feedback database. The database records review content, review time, task identifier, and visitor identifier, establishing a correlation between reviews and task execution. Perform sentiment analysis on the review text to identify positive, negative, and neutral reviews, and extract the specific reasons for visitor dissatisfaction.

[0114] Key performance indicators for task completion include completion rate, average completion time, resource utilization rate, and average satisfaction rate. The completion rate is the ratio of successfully completed tasks to the total number of tasks; the average completion time is the average of the actual completion times for all tasks; the resource utilization rate is the ratio of task-consuming resources to the total system resources; and the average satisfaction rate is the average of all evaluation scores.

[0115] Obtain complete task execution results, including task execution status, exception handling records, performance data, and user feedback.

[0116] Step 6: Based on the task execution results, optimize intent recognition, adjust scheduling strategy parameters, update agent capability evaluation, and obtain a continuous optimization scheme. Specifically, the following steps are included: Step 6.1, Incremental optimization of the intent recognition model; Cases with incorrect intent identification or low confidence levels were filtered from the feedback data. Incorrect cases included incorrect intent category judgment, omission of implicit intent, and incorrect parameter extraction. Intent comprehension biases were identified by comparing visitor reviews with actual service results.

[0117] Error cases are manually labeled with the correct intent category, implicit intent, and parameter values. This manual labeling is performed by domain experts to ensure label quality and consistency. The labeled data is added to the training sample set for incremental training of the model.

[0118] An active learning strategy is employed to select the most valuable samples for labeling, thereby reducing labeling costs. Active learning selects samples based on the uncertainty of the model's predictions. This uncertainty is measured by the entropy or variance of the predicted probability, with samples of high entropy or high variance being prioritized for labeling.

[0119] Incremental training is used to update the intent recognition model, retaining the model's learning from historical data and fine-tuning the model parameters on new samples. Incremental training avoids the high computational cost of training from scratch and prevents catastrophic forgetting. After training, the model performance is evaluated on the validation set to confirm improved accuracy without significant performance degradation.

[0120] The intent knowledge graph is updated by adding newly discovered intent categories and intent relationships. New intent categories arise from novel user requests or new needs arising from changing scenarios. Intent relationships are mined through co-occurrence analysis and semantic similarity calculations to enrich the knowledge graph structure.

[0121] The optimization scheme for the intent recognition model is obtained, including the model parameter update strategy, knowledge graph expansion content, and training sample supplementation plan.

[0122] Step 6.2, adaptive adjustment of scheduling strategy parameters; Analyze the weight configuration of each optimization objective in the scheduling objective function and evaluate whether the current weights align with operational goals. Analyze the actual performance of each objective using task completion data, with average response time, service quality score, resource utilization, and user satisfaction as evaluation criteria. If the performance of a particular objective significantly falls short of expectations, increase its weight to strengthen optimization efforts.

[0123] Adjust the dimensional weights in the agent matching degree calculation, dynamically optimizing them based on the contribution of each dimension to scheduling success. Identify the matching dimensions with the greatest impact on user satisfaction through correlation and regression analysis. Increase the weight of highly influential dimensions, such as response time matching degree, to prioritize ensuring response speed. Appropriately reduce the weight of less influential dimensions.

[0124] The load score calculation method is optimized by adjusting the scoring formula based on the agent's actual load capacity. Service quality changes for each agent under different load scores are statistically analyzed to identify load thresholds and performance inflection points. A non-linear function is introduced into the load score, causing the score to rise rapidly when the load approaches the threshold, thus preventing over-allocation that could lead to performance crashes.

[0125] The trigger threshold for dynamic scheduling was adjusted to balance system stability and response sensitivity. A threshold that is too low leads to frequent rescheduling triggers, increasing system overhead and instability; a threshold that is too high results in delayed responses to anomalies, impacting service quality. The optimal threshold range was determined through simulation experiments and historical data analysis. Specifically, one thousand task execution cases were extracted from historical data, including three scenarios: normal execution, minor deviation, and severe deviation. The distribution characteristics of execution deviation were statistically analyzed for each scenario: the mean deviation for normal execution was 5%, with a standard deviation of 3%; the mean deviation for minor deviation was 18%, with a standard deviation of 5%; and the mean deviation for severe deviation was 45%, with a standard deviation of 12%. Receiver operation characteristic curve analysis was used, with the number of rescheduling attempts and the service quality degradation rate as evaluation indicators, to plot performance curves under different thresholds. The threshold corresponding to the inflection point of the curve was selected as the optimal value. Experimental results showed that the optimal threshold range was 22% to 28%, and the system defaulted to using 25% as the trigger threshold.

[0126] Optimize the number of alternative solutions generated and the selection strategy, dynamically configuring them based on intent priority and system load. Increase the number of alternative solutions for high-priority intents to improve fault tolerance; pre-generate more alternative solutions during low-load periods and reduce them during high-load periods to save resources.

[0127] The optimization schemes for scheduling strategies are obtained, including target weight adjustment suggestions, matching weight optimization strategies, load scoring formula improvement schemes, trigger threshold adjustment ranges, and alternative strategy configuration rules.

[0128] Step 6.3, Dynamic update of the agent capability model; The actual service timeliness of the intelligent agent is statistically analyzed, including the distribution of average response time, service duration, and completion time. The actual timeliness is compared with the declared timeliness in the capability model to identify discrepancies. Sources of discrepancies include changes in agent performance, load impact, and external environmental factors. The response timeliness parameters in the capability model are updated based on recent data to reflect the true level.

[0129] Evaluate the service quality stability of agents by statistically analyzing time series data on task success rate, anomaly frequency, and user satisfaction. Identify service quality fluctuation trends and abnormal periods; agents with large fluctuations are marked as unstable and their priority is reduced during scheduling. Agents with continuously declining service quality trigger maintenance checks to investigate performance issues.

[0130] Update the service capability boundaries of the agent, adjusting them based on the actual task complexity and concurrency. Capability boundaries include the upper limit of complexity for a single service, the upper limit of the number of concurrent tasks, and the ability to handle special scenarios. If the agent frequently handles tasks exceeding the declared boundaries and performs well, increase its capability boundary assessment. If the agent still encounters problems with tasks within the boundaries, decrease the capability boundary rating.

[0131] Adjust the resource consumption model of the intelligent agent and revise the estimated values ​​based on actual resource usage data. Resource consumption includes the use of computing resources, communication bandwidth, and physical equipment. An accurate resource consumption model is helpful for resource planning and cost control.

[0132] Update the agents' collaboration preferences and history, recording successful agent combinations and collaboration patterns. Collaboration history data is used to generate collaboration schemes, prioritizing combinations with good historical collaboration results. Combinations that fail or conflict are marked as incompatible to avoid repeated scheduling.

[0133] The updated agent capability model is obtained, and the model parameters are consistent with the actual operating level, providing accurate capability information and performance prediction for scheduling decisions.

[0134] This step outputs a complete continuous optimization scheme, including an intent recognition model optimization strategy, scheduling strategy parameter adjustment suggestions, agent capability model update rules, and a continuous improvement mechanism. This scheme guides the iterative upgrade and performance improvement of the system, realizing a feedback-based continuous optimization closed loop.

[0135] An intent-based urban service agent scheduling system is provided to execute the aforementioned intent-based urban service agent scheduling method, such as... Figure 3 As shown, it includes: The data acquisition module is used to collect multimodal raw request data from urban complexes and perform preprocessing and feature extraction to obtain personalized request feature representations. The intent recognition module identifies multi-level intents and extracts intent entities and constraints based on personalized request feature representations, evaluates intent priorities, and obtains structured intent representations. The agent matching module constructs an agent capability model and an agent capability knowledge base based on structured intent representation, calculates the semantic matching degree between intent and agent, generates agent combination schemes based on matching results, and obtains a candidate set of multi-agent cooperation schemes. The scheduling decision module, based on the candidate set of multi-agent cooperative schemes, constructs a multi-objective optimization model and generates scheduling schemes for single intents and composite intents, formulates scheduling strategies and alternative schemes, and obtains a complete scheduling scheme. The task execution module issues tasks based on a complete scheduling scheme, monitors task execution progress in real time, handles execution exceptions and implements fault-tolerant recovery, verifies task completion quality and collects user feedback, and obtains task execution results. The continuous optimization module optimizes intent recognition based on task execution results, adjusts scheduling strategy parameters, updates agent capability assessment, and obtains a continuous optimization solution.

[0136] In one embodiment of the present invention, a specific example is provided: A 30-day field test was conducted in a large urban complex. During the test, 50 instances of seven types of intelligent agents were deployed, including eight agents for passenger flow analysis, six for parking guidance, ten for shop recommendation, five for emergency evacuation, 12 for environmental monitoring, six for facility maintenance, and three for security monitoring. The test area covered approximately 150,000 square meters, encompassing commercial shopping, dining and entertainment, cultural exhibitions, and transportation hubs, receiving an average of approximately 80,000 visitors per day.

[0137] During the test, a total of 12,580 visitor service requests were processed, and complete data on intent recognition, agent scheduling, and task execution were collected. Test data was collected using a distributed logging system, with all service requests, intent recognition results, scheduling decisions, and task execution processes recorded in real time to the log database. Log entries included timestamps, request identifiers, visitor identifiers, intent recognition results, scheduling schemes, execution status, and user feedback. Approximately 1.5 million log entries were collected during the test, and after data cleaning, 12,580 valid entries were retained.

[0138] The statistical analysis methods are defined as follows: Recognition accuracy is calculated through manual sampling verification. Two thousand eight hundred and twenty samples are randomly selected proportionally from each intent category. Three annotators independently annotate the correct intents, and a majority vote is used to determine the final correct intent. Recognition accuracy equals the number of correctly identified samples divided by the total number of samples. Scheduling success rate is defined as the number of successfully completed tasks divided by the total number of tasks. The criteria for successful task completion include three conditions: the agent completes the service and reports the task result, the visitor confirms receipt of the result, and there are no serious abnormal events during task execution. Average response time is defined as the time interval from request submission to the agent starting the service. User satisfaction is the arithmetic mean of the five-star user ratings. During the test period, over eight thousand valid reviews were received, representing a review rate of approximately 65%.

[0139] Table 1 shows the statistics of recognition accuracy and scheduling success rate for different intent categories: Table 1: Performance statistics of intent recognition and scheduling;

[0140] To verify the effectiveness of the method of this invention, a control experiment was set up using a traditional rule-based scheduling method. The control method uses fixed intent keyword matching for intent recognition and employs a round-robin scheduling strategy to allocate agents, without considering load balancing or dynamic adjustment. On the same test data, the control method achieved an average recognition accuracy of 78.3%, a scheduling success rate of 87.6%, an average response time of 18.5 seconds, and a user satisfaction score of 3.8. In comparison, the method of this invention improves the recognition accuracy by 15.3 percentage points, the scheduling success rate by 8.6 percentage points, shortens the average response time by 41.6%, and increases user satisfaction by 0.8 points. As shown in Table 1, the method of this invention achieves an accuracy rate of over 89% for all types of intent recognition, a scheduling success rate exceeding 93%, an average response time controlled within 11 seconds, and a user satisfaction score of 4.6 out of 5. Emergency service intents exhibit the highest recognition accuracy and scheduling success rate, and the shortest response time, demonstrating the effectiveness of the priority mechanism.

[0141] Table 2 shows the comparison data of the execution efficiency of intelligent agent collaborative tasks: Table 2: Comparison of execution efficiency for multi-agent collaborative tasks;

[0142] As can be seen from Table 2, although the completion time of multi-agent collaborative tasks is slightly longer than that of single-agent tasks, the resource utilization rate is improved by 13 to 20 percentage points and the user satisfaction is improved by 0.3 to 0.6 points, indicating that the collaborative scheduling mechanism can effectively handle complex needs and improve service quality.

[0143] The following typical examples illustrate the complete execution process of the method of the present invention.

[0144] During the test, at 2:30 PM on a weekend afternoon, a visitor, Ms. Zhang, submitted a service request through the mobile application: "I want to find a restaurant suitable for taking children, preferably with children's play facilities, convenient parking, and not too far away. I am currently on the 3rd floor of Building B." This request is a typical composite intent scenario, involving three sub-intents: shop recommendation, parking guidance, and route planning.

[0145] Multimodal intent acquisition and preprocessing were performed. The system received a text request and automatically obtained Ms. Zhang's location coordinates as the third floor of Building B, with the request submitted at 2:30 PM on a weekend afternoon. Text preprocessing extracted keywords such as "restaurant," "children," "playground facilities," "parking," "convenient," "walking," and "far." Entity recognition extracted the restaurant type entity "restaurant," the attribute entity "children's playground facilities," and the spatial constraint entity "don't walk too far." Location semantic enhancement identified the third floor of Building B as a commercial shopping area with a medium pedestrian density at the current time, indicating a peak dining period on a weekend afternoon. User profile queries showed that Ms. Zhang was a member user with five visits in the past thirty days, frequently visiting the commercial shopping and dining areas, thus identifying her as a user with a preference for family-friendly facilities.

[0146] Intent identification and classification were performed. The intent classification model determined the primary intent to be "commercial services" (0.89 confidence level), the secondary intent to be "shop search" (0.85 confidence level), and the tertiary intent to be "restaurant recommendation" (0.82 confidence level). Implicit intent reasoning identified "convenient parking" triggering parking guidance intent and "don't walk too far" triggering route planning intent, forming a composite intent containing three sub-intents. Entity extraction determined the restaurant type to be "family-friendly restaurant," the attribute requirement to be "children's play facilities," and the spatial constraint to be "moderate walking distance," with a walking distance threshold set at 500 meters or a walking time of less than 10 minutes. Priority evaluation calculated a comprehensive priority score of 0.58, classifying it as a medium priority.

[0147] The process begins with matching the capabilities of the agents. Three candidate agents are retrieved from the agent capability database: SR-05 (store recommendation, capability matching score 0.92, load score 0.60); PG-03 (parking guidance, capability matching score 0.88, load score 0.75); and NAV-02 (navigation, capability matching score 0.85, load score 0.50). Sub-intention dependencies are analyzed: restaurant recommendation is the primary task and must be completed first; parking guidance and route planning are auxiliary tasks that depend on the restaurant recommendation results. A collaborative solution is generated: the store recommendation agent leads, with the parking guidance and navigation agents assisting. The execution order is sequential: restaurant recommendation is executed first, followed by parking guidance and route planning in parallel. The estimated completion time is four minutes.

[0148] Execute scheduling decisions. Based on a multi-objective optimization model, the collaboration scheme is evaluated. The estimated completion time for the response time objective is four minutes, meeting the time requirement. The service quality objective has a combination matching degree of 0.88, the resource utilization objective shows relatively balanced load, and the user satisfaction objective has an estimated satisfaction score of 4.5. A main scheduling scheme is generated, and scheduling instructions are issued. The task identifier is REQ-20260215-143008, the priority is medium, and the collaboration protocol defines the data exchange format as JSON.

[0149] The task is assigned and monitored step by step. Three intelligent agents collaborate to execute the task. The shop recommendation agent SR-05 begins execution at 2:30:15 PM, searching the restaurant database within the complex. The filtering criteria are: restaurant type containing the tags "family-friendly" or "children's," equipped with children's play facilities, and within a 500-meter walking distance of Ms. Zhang's current location. Five candidate restaurants are found, with the top three being "Happy Family Restaurant," "Children's Fun Restaurant," and "Family Amusement Park Restaurant." The shop recommendation agent completes the search at 2:32 PM, taking one minute and forty-five seconds, and sends the information of the top three restaurants to the parking guidance agent and the navigation agent. The parking guidance agent PG-03 queries parking availability near the three restaurants, completing the query at 2:33 PM, taking fifty-five seconds. The navigation agent NAV-02 calculates the walking route from Ms. Zhang's current location to the three restaurants, completing the calculation at 2:33 PM, taking fifty-five seconds. The system integrates the execution results of the three agents, selects the Happy Family Restaurant with the highest comprehensive score as the recommendation result, generates a recommendation report and pushes it to Ms. Zhang. The entire service process takes four minutes.

[0150] Execution feedback optimization. After dining at the restaurant at 3:15 PM, Ms. Zhang submitted a five-star review via the mobile application, stating, "The recommendations were very accurate, my child had a great time, and parking was very convenient." The system optimized based on the execution data: confirming accurate recognition of the composite intent, the combination of the shop recommendation agent SR-05, the parking guidance agent PG-03, and the navigation agent NAV-02 was marked as the preferred solution; collaborative relationship edges were added to the agent capability knowledge base, with an edge weight set to 0.9; the response time rating of the shop recommendation agent SR-05 was updated, with an actual response time of one minute and forty-five seconds, better than the estimated two minutes; the user profile was updated, confirming the accuracy of Ms. Zhang's preference weight for family-friendly facilities.

[0151] The embodiments of the present invention have been described above. However, the embodiments are 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 more equivalent embodiments under the guidance of the present embodiments, and all of them are within the protection scope of the present embodiments.

Claims

1. A method for scheduling urban service agents based on intent recognition, characterized in that, Includes the following steps: Collect multimodal raw request data from urban complexes and perform preprocessing and feature extraction to obtain personalized request feature representations; Based on personalized request feature representation, multi-level intents are identified and intent entities and constraints are extracted. Intent priorities are evaluated to obtain structured intent representations. Based on structured intent representation, an agent capability model and an agent capability knowledge base are constructed. The semantic matching degree between intent and agent is calculated. Based on the matching results, an agent combination scheme is generated, and a candidate set of multi-agent cooperation schemes is obtained. Based on the candidate set of multi-agent cooperative schemes, a multi-objective optimization model is constructed and scheduling schemes for single intents and composite intents are generated. Scheduling strategies and alternative schemes are formulated to obtain a complete scheduling scheme. Tasks are issued based on a complete scheduling scheme, task execution progress is monitored in real time, execution exceptions are handled and fault-tolerant recovery is implemented, task completion quality is verified and user feedback is collected, and task execution results are obtained. Based on the task execution results, intent recognition is optimized, scheduling strategy parameters are adjusted, and agent capability evaluation is updated to obtain a continuous optimization scheme.

2. The urban service agent scheduling method based on intent recognition according to claim 1, characterized in that, The step of obtaining the personalized request feature representation includes: Collect text requests, voice requests, image requests, and dialogue log data, with each type of data accompanied by location and time information; Syntactic analysis and named entity recognition are performed on the text request, and a semantic representation is obtained using a pre-trained language model; The speech request is converted into text through speech recognition, and the prosodic features are extracted to determine the urgency of the need. Perform object detection to extract scene features from image requests; perform session segmentation on dialogue logs to identify intent evolution trajectories; Retrieve regional information based on the requested location coordinates and calculate the spatial relationship with surrounding facilities; obtain operational status information based on the requested timestamp; and construct a spatiotemporal context feature vector. User profiles are retrieved by visitor identifiers, and the weights of user profile features are calculated using an attention mechanism. The weighted fusion is then used to obtain a personalized request feature representation.

3. The urban service agent scheduling method based on intent recognition according to claim 1, characterized in that, The steps for obtaining the structured intent representation include: A hierarchical intent classification system is constructed. The request feature representation is input into the classification model to predict the intent category according to the hierarchy, and the intent label and confidence score are obtained. For the classification result with a confidence score lower than the preset confidence threshold, the intent clarification mechanism is triggered. Analyze the subjective expressions in the request text, identify the implicit service requirements, establish a mapping rule base from subjective expressions to implicit intentions, and infer the implicit intentions through rule matching and semantic similarity calculation; Identify coordinating conjunctions and logical connectors, break down complex intents into multiple sub-intents, and analyze the dependencies between sub-intents; Entity extraction is performed on the request text and mapped to a normalized representation. Constraints are identified and converted into quantization ranges. The extracted entities are then filled into the parameter slots of the intent. Define the evaluation dimensions for intent priority, including urgency, scope of impact, time sensitivity, resource scarcity, and user importance, calculate a comprehensive priority score, and classify priority levels.

4. The urban service agent scheduling method based on intent recognition according to claim 1, characterized in that, The steps for obtaining a candidate set of multi-agent cooperative schemes include: Define the core dimensions of agent capabilities, including service type, service capability boundaries, response time, resource consumption, and collaboration preferences. Fill in the capability parameters based on historical operation data and establish an agent capability knowledge base. Periodically collect the agent's running status, task queue, resource consumption, and performance indicators, and calculate the agent's load score; Extract the capability requirement features of the intent, retrieve the candidate agent set from the agent capability knowledge base, calculate the service type matching degree, capability level matching degree, response time matching degree, resource constraint matching degree, and collaboration matching degree, calculate the comprehensive matching degree, and filter out agents with matching degree lower than the preset matching degree threshold; It identifies composite intents, highly complex intents, and cross-domain intent scenarios that require multi-agent collaboration, decomposes composite intents into sub-intent sequences and analyzes dependencies, and uses combinatorial optimization methods to generate candidate schemes for agent combination.

5. The urban service agent scheduling method based on intent recognition according to claim 4, characterized in that, The identification scenarios requiring multi-agent collaboration, including complex intentions, highly complex intentions, and cross-domain intentions, include: For composite intents, they are decomposed into a sequence of sub-intents, the dependencies between sub-intents are analyzed, and an execution order graph is constructed. For highly complex intentions, a task splitting method is used to break them down into multiple parallel subtasks, and an agent is assigned to each subtask. For cross-domain intentions, retrieve cross-domain collaboration relationships from the knowledge base, select a collaboration mode, and instantiate it into a specific intelligent agent.

6. The urban service agent scheduling method based on intent recognition according to claim 1, characterized in that, The steps to obtain the complete scheduling scheme include: Define the objective function for scheduling optimization, including minimizing response time, maximizing service quality, maximizing resource utilization, and maximizing user satisfaction; define the constraints, including capacity constraints, resource constraints, time constraints, and collaboration constraints. For single-intention scenarios, select agents that meet the constraints, calculate the comprehensive scheduling score, fusion capability matching degree, load score, historical service quality, and response time prediction, and select the optimal agent to generate scheduling instructions; For complex intentions or collaborative situations, the feasibility of each collaborative scheme is verified and multi-objective optimization is evaluated. The optimal scheme is selected and scheduling instructions are generated for each agent. Define the conditions for triggering dynamic scheduling, and re-execute the scheduling decision process after triggering; pre-generate alternative scheduling schemes and establish the trigger conditions for alternative schemes.

7. The urban service agent scheduling method based on intent recognition according to claim 6, characterized in that, The conditions for triggering dynamic scheduling are defined as follows: Monitor the difference between the actual execution progress and the estimated progress. If the difference exceeds a preset deviation threshold, trigger a rescheduling. When the priority difference of a new intent is higher than the preset priority difference threshold, a preemption task is triggered, and the preempted task is re-added to the task queue. When the detection resource usage exceeds the capacity limit, the task will be reallocated. Monitor task failure rate, response time, and user complaints; switch to a backup agent when quality deteriorates.

8. The urban service agent scheduling method based on intent recognition according to claim 1, characterized in that, The steps for obtaining the task execution result include: The task message is pushed to the message queue. The agent pulls the task message, parses it, checks its executability, and adds the task to the task queue and sends a confirmation message after the check is passed. The intelligent agent periodically reports its status, the monitoring center updates the task execution status, compares the actual execution progress with the estimated progress to calculate the execution deviation, and identifies execution timeouts, progress stagnation, resource anomalies, and quality anomalies. For agent failures, the faulty agent is isolated and the task is reassigned to a backup agent; In response to execution timeouts, if the timeout is less than the preset timeout threshold, the time window is extended; if the timeout is greater than or equal to the preset timeout threshold, task migration or rescheduling is triggered. After the intelligent agent completes the task, it reports the task result. The scheduling center verifies the result, and the verified task result is pushed to the visitor, triggering the collection of satisfaction evaluation.

9. The urban service agent scheduling method based on intent recognition according to claim 1, characterized in that, The steps for obtaining the continuously optimized solution include: Cases with incorrect intent recognition were filtered from the feedback data, manually labeled, samples were selected using an active learning strategy, and the intent recognition model and intent knowledge graph were updated using an incremental training method. Analyze the weight configuration of each optimization objective in the scheduling objective function, adjust the dimension weights in the agent matching degree calculation, optimize the load score calculation method, and adjust the trigger threshold for dynamic scheduling; The actual service timeliness of the agent is statistically analyzed, and the deviation between the actual timeliness and the declared timeliness in the capability model is compared and the response timeliness parameters are updated to update the service capability boundary of the agent.

10. A city service intelligent agent scheduling system based on intent recognition, characterized in that, A method for scheduling urban service agents based on intent recognition as described in any one of claims 1-9 includes: The data acquisition module is used to collect multimodal raw request data from urban complexes and perform preprocessing and feature extraction to obtain personalized request feature representations. The intent recognition module identifies multi-level intents and extracts intent entities and constraints based on personalized request feature representations, evaluates intent priorities, and obtains structured intent representations. The agent matching module constructs an agent capability model and an agent capability knowledge base based on structured intent representation, calculates the semantic matching degree between intent and agent, generates agent combination schemes based on matching results, and obtains a candidate set of multi-agent cooperation schemes. The scheduling decision module, based on the candidate set of multi-agent cooperative schemes, constructs a multi-objective optimization model and generates scheduling schemes for single intents and composite intents, formulates scheduling strategies and alternative schemes, and obtains a complete scheduling scheme. The task execution module issues tasks based on a complete scheduling scheme, monitors task execution progress in real time, handles execution exceptions and implements fault-tolerant recovery, verifies task completion quality and collects user feedback, and obtains task execution results. The continuous optimization module optimizes intent recognition based on task execution results, adjusts scheduling strategy parameters, updates agent capability assessment, and obtains a continuous optimization solution.

Images