Intelligent communication command and dispatch method and system based on artificial intelligence

Abstract

The invention provides an intelligent communication commanding and dispatching method and system based on artificial intelligence. The method comprises the following steps: acquiring perception data of a target scene, performing time extraction according to the perception data to generate a preliminary event, aligning the preliminary time according to a timestamp, and performing density clustering in space to obtain a plurality of candidate events; performing vector similarity matching on the standard structured event, determining a unit list which is successfully responded in a historical task, and generating a standard scheduling task package; constructing a scheduling heterogeneous graph according to the standard scheduling task packet in combination with a unit equipment set, a real-time resource state and communication link information; reasoning based on a graph neural network is carried out on the scheduling heterogeneous graph, and an optimal resource allocation scheme is deduced; and generating a control instruction set of a standard structure according to the optimal resource allocation scheme, and actually issuing and executing the control instruction set through a command and dispatch system.

Description

Technical Field

[0001] This invention belongs to the field of artificial intelligence, and in particular relates to an intelligent communication command and dispatch method and system based on artificial intelligence. Background Technology

[0002] With the continuous improvement of urban management intelligence, communication command and dispatch systems have become the fundamental capability support for key business scenarios such as urban emergency response, government event management, and public safety maintenance. In scenarios such as earthquake disasters, severe weather, public emergencies, and abnormal urban operations, government agencies and various enterprises and institutions need to respond quickly and work collaboratively to complete a series of command execution processes, including task resource allocation, personnel scheduling, equipment linkage, and information dissemination. These dispatch systems typically involve multi-source sensing (such as sensor data, video surveillance, and on-site voice), multi-role collaboration (such as multiple departments such as public security, fire protection, health, and transportation), and cross-network link data and command transmission, characterized by complex information flow, high difficulty in resource scheduling, and strong timeliness requirements for response.

[0003] Traditional communication command and dispatch systems primarily rely on fixed task plans and manual dispatch rules. Upon receiving an event alarm or sensor trigger signal, dispatchers assess the nature of the task based on experience, select an appropriate response unit and instruction template, and issue the command via static communication links (such as private network broadcasts, SMS push notifications, or dedicated dispatch terminals). While this approach is efficient in routine high-frequency processing, it suffers from significant shortcomings in complex environments characterized by multiple concurrent tasks, dynamic resource changes, and frequent link fluctuations. Specifically: First, existing dispatch strategies are often based on static rule matching, lacking intelligent judgment and real-time learning capabilities, and cannot automatically generate structured, highly adaptable dispatch task instructions based on the actual scenario. Second, dispatch resources (personnel, equipment, units) are heterogeneously distributed, and existing systems lack effective methods to determine their optimal matching relationships. Resource allocation often relies on subjective experience, leading to frequent problems such as resource mismatch and task delays. Third, the instability of communication links has not been systematically modeled and optimized. In particular, under conditions such as network congestion and channel loss, command failure may have catastrophic consequences. At present, there is a lack of systematic evaluation and control mechanisms for factors such as link status, path selection, and data retransmission.

[0004] Furthermore, while the industry has attempted to introduce artificial intelligence (AI) technology for assisted scheduling, most efforts focus on a single aspect, such as modular optimization of event classification, path planning, and speech recognition. A complete closed-loop system covering "task generation—communication issuance—task feedback" has not yet been formed, and there is a significant gap in the coordination between scheduling strategies and communication links. Existing research has failed to unify the modeling and linkage reasoning of task semantics, scheduling object characteristics, and communication network status, resulting in a lack of systematicity, stability, and adaptability in intelligent scheduling systems in practical applications. Therefore, how to construct a command and dispatch method and system with global awareness, multi-task dynamic reasoning capabilities, and the ability to intelligently adapt to changes in the communication environment is a key challenge that urgently needs to be addressed in the current technological field. Summary of the Invention

[0005] The purpose of this invention is to propose an intelligent communication command and dispatch method and system based on artificial intelligence, which solves the core bottleneck of the current system's fragmented functional modules and the inability to unify dispatch strategies and distribution paths, and significantly improves the actual command and dispatch efficiency and reliability of the system in highly dynamic scenarios such as urban emergency response and government coordination.

[0006] To achieve the above objectives, a first aspect of the present invention provides an intelligent communication command and dispatch method based on artificial intelligence, the method comprising:

[0007] The process involves acquiring perception data of the target scene, extracting time data to generate preliminary events, aligning these preliminary times with timestamps, and performing density clustering in space to obtain several candidate events. These candidate events are then fused and scored, and the candidate time with the highest score is selected to form a standard structured event. The standard structured event includes event type, cluster center coordinates, event time, and event source identifier.

[0008] Vector similarity matching is performed on the standard structured events to determine a list of units that have successfully responded in historical tasks, thus constructing a preliminary candidate response unit set. Then, the response score is calculated by combining the real-time resource status and geographical location of the current unit. The top k units with the highest scores are selected to form the final response unit set and the task action set corresponding to each unit, generating a standard scheduling task package. The standard scheduling task package includes a unique task identifier, a set of response units, a set of action instructions for each unit, an overall task priority, and a task completion time limit.

[0009] Based on the standard scheduling task package, and combined with the unit device set, real-time resource status and communication link information, a scheduling heterogeneity graph is constructed;

[0010] The optimal resource allocation scheme is inferred by performing graph neural network-based reasoning on the scheduling heterogeneous graph.

[0011] Based on the optimal resource allocation scheme, a standard set of control instructions is generated and executed through the command and dispatch system.

[0012] Furthermore, the perceived data includes video stream data, IoT device data stream, human voice or text reporting data, and historical scheduling task records;

[0013] Then, the step of generating preliminary events by extracting time from the perceived data includes:

[0014] The video stream data is processed by a lightweight object detection network to extract abnormal scene labels and corresponding confidence scores.

[0015] After the IoT device data is received via MQTT or HTTP protocol, preliminary event candidates are triggered based on a set threshold.

[0016] The artificial voice or text report data is converted into text through a speech-to-text network, then the TextRank is used to extract keywords, and the data is classified into a standard event type library.

[0017] Furthermore, the step of selecting the top k units by score to form the final response unit set and the corresponding task action set for each unit, and generating a standard scheduling task package, includes:

[0018] Obtain the response score; select the top k units of the response score to form the final response unit set; wherein the value of k is dynamically adjusted according to the event level corresponding to the event type;

[0019] Encode the event type into a vector representation;

[0020] Input the action reasoning network and introduce an action coupling regularization term into the action reasoning network to output the task action set corresponding to each unit;

[0021] Based on the final set of response units and the set of task actions corresponding to each unit, a standard scheduling task package is assembled.

[0022] Furthermore, the step of constructing a scheduling heterogeneity graph based on the standard scheduling task package, combined with the set of unit devices, real-time resource status, and communication link information, includes:

[0023] Obtain the standard scheduling task package;

[0024] A task node is generated for each task's unique identifier according to the standard scheduling task package;

[0025] Each response unit in the set of response units generates a unit node;

[0026] For each unit node, a device node is generated for each available device; the set of unit devices contains several available devices.

[0027] A communication node is generated for each available device associated with a communication gateway; the real-time resource status and communication link information includes several communication gateways;

[0028] Directed edges are constructed based on the task nodes, unit nodes, device nodes, and communication nodes, including: task nodes to unit nodes, unit nodes to device nodes, and device nodes to communication nodes.

[0029] By combining all task nodes, unit nodes, device nodes, communication nodes, and all directed edges, a scheduling heterogeneous graph is generated.

[0030] Furthermore, the edge weight w of the scheduling heterogeneous graph ij The structure is as follows:

[0031]

[0032] Where P represents the overall task priority; P max The ratio of the two values ​​represents the highest priority level, and is used to obtain a dimensionless urgency index.

[0033] LinkQual(n k ) represents the communication node n k Link quality score is calculated as LinkQual = NormalizedBandwidth - NormalizedPacketLoss, with a value range of [0,1]. Delay(n) is a normalized metric for link delay, ensuring it is dimensionless. k Real-time data collection, Delay max α1, α2, γ, δ1, δ2 are the maximum tolerable delay settings of the system; ActionAdapt(a) is the adaptability of action type a to the link requirements; TimePressure(D) is the task time limit pressure index; α1, α2, γ, δ1, δ2 are the weight adjustment coefficients respectively.

[0034] Furthermore, each edge of the scheduling heterogeneous graph is also provided with:

[0035] If the current unit node has multiple high-priority tasks attached, then the edge from the task to the unit will be penalized with a load penalty, and the priority will be reduced.

[0036] If the current number of connections of a communication node exceeds the system's preset number of connections, a delay penalty is added to the edge from the device to the communication node, reducing the transmission priority.

[0037] Furthermore, the step of performing graph neural network-based reasoning on the scheduling heterogeneous graph to infer the optimal resource allocation scheme includes:

[0038] Obtain the scheduling heterogeneity graph;

[0039] Different types of nodes in the scheduling heterogeneous graph are respectively processed through independent small feedforward neural networks to embed the original features into a unified low-dimensional space, thereby generating low-dimensional features.

[0040] The low-dimensional features are updated using a graph convolutional network to generate hidden representations for each layer of nodes after the update. In each convolutional operation of the graph convolutional network, different parameter matrices are used for different types of edges, and weighted aggregation is performed according to the edge weights to update the nodes.

[0041] After the convolution operation of the convolutional layer is completed, a global scheduling selector is introduced for each task node to aggregate the final states of all associated unit nodes and deduce the optimal resource selection set; wherein, the global scheduling selector is the Global Readout module.

[0042] Furthermore, the node representation update calculation is as follows:

[0043]

[0044] in, Let i be the hidden representation of node i at the +1l level. Let i be the hidden representation of node i at level l;

[0045] Let be the hidden representation of node j at level l; A set of edge types; w is the set of neighboring nodes connected to i via edge type r; ij c represents the edge weight from node i to node j. i,r This is a normalization constant; σ represents the weight matrix for different edge types and self-loop connections; σ is the activation function.

[0046] Furthermore, the step of generating a standard-structured control instruction set based on the optimal resource allocation scheme and actually issuing and executing it through the command and dispatch system includes:

[0047] Obtain the optimal resource allocation scheme and assemble the corresponding structured instructions;

[0048] After the instructions are encapsulated, they are uniformly distributed to the corresponding units through the system control bus;

[0049] Ultimately, we obtain the set of sent instructions corresponding to all scheduled task nodes, as well as the set of execution receipts returned in real time.

[0050] In another aspect of the invention, an intelligent communication command and dispatch system based on artificial intelligence is provided, the system comprising:

[0051] A multi-source data processing module is used to acquire perception data of the target scene, extract time based on the perception data to generate preliminary events, align the preliminary times by timestamps, and perform density clustering in space to obtain several candidate events; fuse and score the several candidate events, select the candidate time with the highest score to form a standard structured event; the standard structured event includes event type, cluster center coordinates, event time, and event source identifier;

[0052] The scheduling task generation module is used to perform vector similarity matching on the standard structured events, determine the list of units that have successfully responded in historical tasks, construct a preliminary candidate response unit set, and then calculate the response score by combining the real-time resource status and geographical location of the current unit. The top k units with the highest scores are selected to form the final response unit set and the task action set corresponding to each unit, generating a standard scheduling task package. The standard scheduling task package includes a unique task identifier, a set of response units, a set of action instructions for each unit, an overall task priority, and a task completion time limit.

[0053] The heterogeneous graph construction module is used to construct a scheduling heterogeneous graph based on the standard scheduling task package, combined with the set of unit devices, real-time resource status and communication link information;

[0054] The graph neural network reasoning module is used to perform graph neural network-based reasoning on the scheduling heterogeneous graph to infer the optimal resource allocation scheme.

[0055] The instruction generation and issuance module is used to generate a standard set of control instructions based on the optimal resource allocation scheme, and to actually issue and execute them through the command and dispatch system.

[0056] The beneficial technical effects of the present invention are at least as follows:

[0057] To address the problems of insufficient intelligent task generation, poor resource matching adaptability, unstable communication links, and fragmented intelligent capabilities in existing communication command and dispatch systems, this invention proposes an intelligent communication command and dispatch method and system based on artificial intelligence. This invention constructs a unified task representation structure, integrating information such as sensor data, historical plans, and unit resource status to achieve automatic generation and structured expression of dispatch instructions. Furthermore, it establishes a multi-dimensional linkage model between task objects and the communication network, automatically analyzing the feasibility and stability of the communication link while generating dispatch tasks, and dynamically constructing the optimal task execution path and distribution strategy. This method enables collaborative optimization of tasks and communication, possesses scenario adaptation, autonomous reasoning, and fault-tolerant scheduling capabilities, and solves the core bottlenecks of fragmented functional modules and the inability to unify scheduling strategies and distribution paths in current systems. It significantly improves the actual command and dispatch efficiency and reliability of the system in highly dynamic scenarios such as urban emergency response and government coordination. Attached Figure Description

[0058] The present invention will be further described with reference to the accompanying drawings, but the embodiments in the drawings do not constitute any limitation on the present invention. For those skilled in the art, other drawings can be obtained based on the following drawings without creative effort.

[0059] Figure 1 This is a flowchart of an intelligent communication command and dispatch method based on artificial intelligence, according to an embodiment of the present invention.

[0060] Figure 2 This is a framework diagram of an intelligent communication command and dispatch system based on artificial intelligence, according to an embodiment of the present invention. Detailed Implementation

[0061] Embodiments of the present invention are described in detail below. Examples of these embodiments are shown in the accompanying drawings, wherein the same or similar reference numerals denote the same or similar elements or elements having the same or similar functions throughout. The embodiments described below with reference to the accompanying drawings are exemplary and are only used to explain the present invention, and should not be construed as limiting the present invention.

[0062] like Figure 1 As shown in the embodiment of the present invention, the intelligent communication command and dispatch method based on artificial intelligence includes the following steps S1-S5:

[0063] S1. Obtain perception data of the target scene, extract time based on the perception data to generate preliminary events, align the preliminary times by timestamp, and perform density clustering in space to obtain several candidate events; fuse and score the several candidate events, select the candidate time with the highest score to form a standard structured event; the standard structured event includes event type, cluster center coordinates, event time, and event source identifier.

[0064] Specifically, the purpose of this step is to process, filter, and fuse multi-source heterogeneous perception data from real-world scenarios to generate structured event information that can serve as the basis for inference in intelligent scheduling tasks. In real-world urban emergency communication and command scenarios, events may originate from unusual situations captured by cameras, abnormal indicators detected by sensors, reports from citizens, or even potential dispatch needs inferred from historical mission data. Therefore, it is essential to design a multi-source fusion event extraction method oriented towards dispatch decision-making, going beyond mere conventional perception to focus on generating the "task triggering conditions" required for subsequent dispatch decisions.

[0065] This step receives the following four types of raw sensor data:

[0066] V t Video stream data, such as real-time RTSP video streams from urban surveillance systems;

[0067] S t Data streams from IoT devices, such as alarm data for smoke, temperature, vibration, and power anomalies;

[0068] H t : Data reported via human voice or text, from citizen reporting hotlines or emergency dispatch apps;

[0069] Historical scheduling task records include task frequency, time, location, and unit response status.

[0070] For example, a security camera in a certain area captured images of thick smoke rising (V). t The surrounding smoke sensor alarm (S) t A citizen called to report that "the factory is on fire" (H). t Furthermore, a search of historical records revealed that there had been two fire alarms in this area within the past month. This information needs to be processed in a unified manner to ultimately form a clear event description that is valuable for scheduling decisions.

[0071] Specifically include:

[0072] First, modal decoding and time synchronization are required for different types of data. Video stream (V t The data is processed using a lightweight object detection network (YOLOv7-tiny, a 32-layer convolutional network with an input size of 416x416) to extract abnormal scene labels and corresponding confidence scores. IoT device data (S...) t After receiving the data via MQTT or HTTP protocol, it will be processed according to a set threshold (e.g., smoke concentration greater than 7.5 mg / m³). 3 Triggering initial event candidates. Manual input (H)t The text is converted into text using a speech-to-text module (such as a DeepSpeech2 network with a 5-layer bidirectional GRU), and then keywords are extracted using TextRank and classified into a standard event type library.

[0073] All extracted preliminary events are aligned by timestamp (time difference less than 2 minutes is considered synchronous) and spatially (e.g., latitude and longitude difference less than 300 meters) density clustered (DBSCAN algorithm, ∈ = 0.003 degrees, minimum number of samples = 3). Each cluster center represents an event that may need to be scheduled.

[0074] Then, for each candidate event e k A fusion score is calculated, taking into account the confidence level and historical frequency of different modalities. The specific scoring formula is as follows:

[0075]

[0076] Where M is the modality set, including video, IoT sensors, text reports, and historical data; α m Modal confidence weights (e.g., α) video =0.4, α iot =0.3, α text =0.2, α history =0.1), set by the platform based on experience values; Conf m (e k ) is the confidence score for event recognition within each modality (e.g., softmax output for object detection, normalized score for outliers in IoT); λ is the historical record regularization factor, which enhances the importance of the location of historical high-frequency events in the final decision. It is an indicator function that indicates that if the event type has appeared in the history of the task, points will be awarded.

[0077] For example, regarding the aforementioned factory fire incident, e k If a smoke detection confidence level of 90% is achieved in video recognition, smoke alarms of 85% are achieved in IoT, the keyword "fire" is extracted from text, and similar cases exist in historical records, then the final fusion score is very high, and the system can determine that it is a core event that requires immediate dispatch and response.

[0078] For each event cluster, select the highest score e. * To form standard structured events

[0079] in:

[0080] Et The event type (e.g., "industrial fire") is obtained through modality fusion classification.

[0081] L t These are the cluster center coordinates (GPS coordinates), obtained by averaging the positions of IoT devices and cameras;

[0082] T t It is the event time; take the source timestamp with the highest confidence.

[0083] R t It is the event source identifier (such as camera number, sensor ID, report ID, etc.).

[0084] The data sources and generation logic for these fields are clearly explained in the above steps to ensure that the data is traceable and verifiable, and meets the data governance requirements for urban emergency management.

[0085] S2. Perform vector similarity matching on the standard structured events to determine the list of units that have successfully responded in historical tasks, so as to construct a preliminary candidate response unit set. Then, combine the real-time resource status of the current unit and the unit's geographical location to calculate the response score. Select the top k units with the highest scores to form the final response unit set and the task action set corresponding to each unit, and generate a standard scheduling task package. The standard scheduling task package includes a unique task identifier, a response unit set, an action instruction set for each unit, an overall task priority, and a task completion time limit.

[0086] Specifically, in traditional solutions, task generation typically relies on static templates or rule engines, which are difficult to adjust flexibly based on real-time resource status, historical scheduling experience, and dynamic spatial environment, leading to scheduling failures or resource waste. To address this, this step innovatively proposes a unit scoring mechanism based on the fusion of historical, geographical, and resource three-dimensional features, and an action generation mechanism based on semantic reasoning and action relationship regularization. This achieves dual optimization of scheduling rationality and real-time performance, fully meeting the "variable, complex, and time-sensitive" requirements of urban emergency dispatch scenarios.

[0087] Specifically, it includes:

[0088] First, the system searches the historical task database to filter out past tasks related to E. t The set of event records with the closest type. Using event embedding vectors (encoding event text descriptions into 128-dimensional vectors via a pre-trained small Text2Vec encoder), the current E... t Perform vector similarity matching with historical events and select the top N. h Historical tasks (generally N) h =5).

[0089] Then, based on the list of units that have successfully responded in historical tasks, a preliminary set of candidate response units is constructed. Simultaneously, combining the real-time resource status of the current unit (from the city's IoT resource management platform) and the unit's geographical location (preloaded in the local cache), the response score for each unit u is calculated, and the following patent-customized scoring formula is proposed:

[0090]

[0091] The variables are defined as follows:

[0092] Score u The score is based on the response.

[0093] HistFreq u (E t ): The unit u historically processed E t The frequency of successful responses to similar events is statistically analyzed from a historical task database.

[0094] NormDist u (L t ): Unit u and event location L t The normalized geographic distance is calculated using the spherical distance formula and then linearly normalized to the [0,1] interval.

[0095] Avail u The resource availability score per unit u (e.g., the percentage of idle devices) is derived from real-time IoT device status monitoring.

[0096] Load u The unit u has carried the workload in the past 6 hours, normalized to [0,1], for load suppression;

[0097] LinkStable u The stability score of the communication link between unit u and the command and dispatch platform (a high score is given for a low packet loss rate) is derived from the communication link monitoring module.

[0098] β1,β2,β3,γ,η: Weighting coefficients, which are dynamically adjusted by the system according to the task type. For example, for fire-related events, geographical proximity is emphasized, while for terrorist events, communication stability is emphasized.

[0099] Compared to traditional simple scoring, this formula is the first to use communication link stability as a formal scoring factor, which reflects the core idea of ​​"scheduling decision-communication stability linkage" in this patent and lays the foundation for link modeling in subsequent heterogeneous graph modeling.

[0100] Based on the scoring formula described above, the system selects the top k units by score to form the final response unit set. The value of k can be determined based on E.t The corresponding event level is dynamically adjusted (e.g., k=5 for Level 1 response, k=2 for Level 3 response).

[0101] Next, for each unit The system needs to generate a corresponding set of task actions. This patent employs a reasoning mechanism based on an event-action relationship matrix and incorporates a regularization term for action rationality coupling. The specific process is as follows:

[0102] Event type E t Encode as a vector representation;

[0103] Input action reasoning network (two-layer MLP + softmax output action score, 64 hidden units, each layer followed by ReLU activation);

[0104] An action coupling regularization term is introduced to control the reasonable combination of different actions. The regularization objective is as follows:

[0105]

[0106] in:

[0107] Standard action classification cross-entropy loss;

[0108] Rel(a i ,a j ): The score of the reasonable relationship between actions, extracted from the action knowledge graph (e.g., "firefighting" and "evacuation" are highly correlated, while "environmental detection" and "patient transfer" are lowly correlated).

[0109] δ: Regularization coefficient, the weight of the constraint on the rationality of the control action (generally set to 0.2 to 0.4).

[0110] By introducing action rationality regularization terms, the problem of "fragmented and disordered scheduling actions" is systematically solved, making the generated task action combinations not only reasonable but also more coherent in the execution chain, thereby improving the overall execution success rate of the scheduling system.

[0111] Finally, based on the event urgency level (e.g., E) t Matching response time limit criteria), timestamp T t Assign an overall priority P and a completion time limit D to the entire task.

[0112] Finally, assemble the standard scheduling task package:

[0113]

[0114] Wherein, TID: a unique identifier for the task, used to generate task nodes; A set of response units, with each unit generating a unit node; Action instruction set for each unit, used for action adaptation of node features and edge weights; P: overall task priority, used for calculating the urgency of edge weights; D: task completion time limit, used for time limit pressure regularization of task node features and edge weights.

[0115] S3. Based on the standard scheduling task package, and combined with the unit device set, real-time resource status and communication link information, construct a scheduling heterogeneous graph.

[0116] Specifically, the goal of this step is to build upon the scheduling task package generated in the previous step. By combining real-time resource status and communication link information, a scheduling heterogeneous graph G=(V,E) is constructed to support subsequent intelligent reasoning. In traditional scheduling systems, resource scheduling only stays at the device or unit level, failing to effectively integrate dynamic environmental factors such as communication link quality, task urgency, and unit load. This results in insufficient stability and timeliness of scheduling decisions in complex urban emergency environments.

[0117] This step introduces a joint edge weighting mechanism of urgency and link quality, along with action adaptation regularization and time-pressure regularization, to enable the heterogeneous graph to truly reflect the dynamic changes in the scheduling environment, ensuring that the intelligent inference stage can output a more accurate and stable scheduling scheme.

[0118] This step uses the complete scheduling task package output in the previous step. In addition, the following dynamic information is retrieved in real time: TID: Unique identifier for the task, used to generate task nodes; A set of response units, with each unit generating a unit node; Each unit's action instruction set is used for node feature and edge weight action adaptation; P: overall task priority, used for edge weight urgency calculation; D: task completion time limit, used for task node feature and edge weight time limit pressure regularization; S: unit device set, derived from the real-time resource platform, with each device generating a device node; The set of communication nodes and link status information (such as bandwidth, latency, and packet loss rate) are used to generate communication nodes and calculate link quality. All inputs will be actually used in subsequent steps to ensure a complete and closed information chain.

[0119] Specifically include:

[0120] First, based on Generate a task node for each task TID. Each response unit Generate a unit node, whose characteristics include the unit's historical response frequency, current idle rate, and action type encoding. And geolocation encoding (such as latitude and longitude vectors or discretized raster IDs).

[0121] Subsequently, the available equipment under each unit was... Generate device nodes. Device node characteristics include device health (inverse indicator of failure rate in the past hour), remaining energy ratio (such as the proportion of electric vehicle battery power), and action execution capability tags (such as whether it supports high-definition video transmission).

[0122] Finally, a communication gateway is associated with each available device. Communication nodes are generated, and their characteristics include real-time link bandwidth (Mbps), packet loss rate (%), and latency (ms). The data comes from the communication monitoring subsystem.

[0123] After constructing the node set V, define three types of directed edges:

[0124] Task node to unit node (TID→u) i ): Indicates the allocation of scheduling instructions;

[0125] Unit node to device node (u i →d j ): Indicates the mobilization of specific equipment by a unit;

[0126] Device node to communication node (d j →n k ): Indicates the device's upload link.

[0127] To accurately represent the complexity of the scheduling environment, the edge weight design adopts the joint weighting mechanism innovatively proposed in this patent, with the specific formula as follows:

[0128]

[0129] ActionAdapt(a)+δ2·TimePressure(D)(6);

[0130] The definitions and sources of the variables are as follows:

[0131] P: Overall task priority, from P max The ratio of the two values, representing the highest priority level (set to 5), yields a dimensionless urgency index.

[0132] LinkQual(n k ): Communication node n k Link quality score is calculated as LinkQual = NormalizedBandwidth - NormalizedPacketLoss, with a value range of [0,1].

[0133] Delay(n) is a normalized metric for link delay, ensuring it is dimensionless. k Real-time data collection, Delay max Set the maximum tolerable delay value for the system (e.g., 500ms);

[0134] ActionAdapt(a): The degree to which action type a adapts to link requirements (e.g., high-definition video backhaul actions rely more on high-bandwidth links, while rescue actions rely more on low-latency links), which is derived from the action link requirement mapping table.

[0135] TimePressure(D): Task time limit pressure index, calculated after normalization based on the remaining time of D. The shorter the time, the greater the pressure.

[0136] α1, α2, γ, δ1, δ2: Weight adjustment coefficients, which can be dynamically set according to the task type. For example, in a fire scenario, the weight of α1 is increased to emphasize urgency, and in a traffic management scenario, the weight of α2 is increased to emphasize link stability.

[0137] Understandably, this edge weight design greatly enhances the sensitivity and adaptability of the scheduling graph to changes in the complex urban environment, enabling subsequent graph neural network inference to comprehensively weigh paths based on task urgency, link status, and action characteristics, and output a scheduling scheme that better meets actual needs.

[0138] Furthermore, to further improve the actual performance of graph construction, dynamic regularization was designed for each edge:

[0139] Unit load regularization: If a unit node has multiple high-priority tasks attached, then apply a load penalty to the edge from the task to the unit, reducing its priority.

[0140] Link congestion regularization: If communication node n k If the current number of connections exceeds the system preset (e.g., 20 connections), a latency penalty is added to the edge from the device to the communication node, reducing the transmission priority.

[0141] This fine-grained dynamic regularization mechanism effectively avoids overall scheduling failures caused by single-point congestion, improving the overall robustness and reliability of the system.

[0142] The final output is a heterogeneous scheduling graph G = (V, E), where:

[0143] The node set V comprehensively covers four types of entities: tasks, units, devices, and communications.

[0144] The edge set E connects each node according to the real physical scheduling process, and the weight of each edge dynamically integrates the urgency of the task, the quality of the link, the action requirements and the time limit pressure.

[0145] S4. Perform graph neural network-based reasoning on the scheduling heterogeneous graph to deduce the optimal resource allocation scheme.

[0146] Specifically, the goal of this step is to infer the optimal scheduling path and resource allocation scheme based on the heterogeneous scheduling graph G=(V,E) constructed in the previous step, by designing a graph neural network (GNN) inference mechanism for urban intelligent communication command and dispatch scenarios. Traditional GNN-based methods mostly focus on node classification or link prediction, which is difficult to meet the requirements of simultaneous modeling and inference of three complex factors in urban emergency command and dispatch: "task priority," "dynamic changes in resource load," and "communication link fluctuations." This patented solution proposes a scheduling inference network guided by dynamic perception of task urgency and a dual regularization optimization mechanism for resource load and link stability, effectively improving the real-time performance, feasibility, and overall system execution success rate of scheduling inference.

[0147] Specifically, it includes:

[0148] First, the input graph G undergoes a type-aware feature encoding process. For different types of nodes (task node TID, unit node u, device node d, communication node n), the original features are embedded into a unified low-dimensional space (uniform dimension 64) through independent small feedforward neural networks (each network is a 2-layer MLP with 64 hidden units per layer and ReLU activation).

[0149] This categorical feature encoding ensures that the model can correctly distinguish the semantics of different nodes during subsequent inference, avoiding information interference caused by node heterogeneity.

[0150] Subsequently, a graph convolutional network is used to update the node representations. In each convolutional operation, different parameter matrices are used for different edge types (task-unit, unit-device, device-communication), while simultaneously applying the comprehensive edge weights w modeled earlier. ij Perform weighted aggregation. The node update formula is:

[0151]

[0152] in:

[0153] The hidden representation of node i at level l;

[0154] Edge type set (task-unit, unit-device, device-communication);

[0155] The set of neighboring nodes connected by edge type r and i;

[0156] w ij The edge weight from node i to node j is derived from step three.

[0157] c i,r : Normalization constant, which can be set as the number of neighbors;

[0158] Weight matrices for different edge types and self-loop connections;

[0159] σ: Activation function (ReLU).

[0160] Understandably, this aggregation process ensures that during information flow, the influence of neighboring nodes can be dynamically adjusted based on scenario-specific factors such as task urgency and link stability, effectively improving the inference model's adaptability to real urban emergency environments.

[0161] After all convolutional layers, a global scheduling selector (GlobalReadout module) is introduced for each task node TID to aggregate the final states of all associated unit nodes and infer the optimal resource selection set. To improve the timely response capability to high-priority tasks during the inference stage, this patent introduces a dynamic task urgency-guided regularization term into the inference loss and combines it with a resource load balancing mechanism to form a joint optimization objective:

[0162]

[0163] in:

[0164] The core allocation loss (such as cross-entropy in multi-label binary classification) measures the difference between the inferred unit / device selection and the true optimal solution.

[0165] The urgency dynamic regularization term penalizes the response delay of reasoning results for high-priority tasks and is defined as the weighted loss of the difference between the reasoning selection score and the task priority.

[0166] The resource load balancing regularization term penalizes instances of excessive load on a unit / device during the inference process, specifically through a secondary penalty for the difference between the current load of the selected node and the system average load.

[0167] λ1, λ2: Regularization weight coefficients, usually set between [0.1, 0.3], and adjusted according to the scheduling scenario.

[0168] It is important to note that, unlike existing methods that only optimize inference accuracy, this solution... and Joint constraints enable the inference model not only to pursue optimal accuracy but also to naturally prioritize high-urgency tasks, while ensuring overall system load balance, which greatly improves the stability and response efficiency of the scheduling system in complex urban environments.

[0169] Finally, during actual reasoning, the system selects the optimal unit u and corresponding device d for each task TID from high to low based on the reasoned score ranking results, and then assigns the optimal unit u and corresponding device d according to the action set. The system assigns specific actions to be executed, ultimately generating a resource matching scheme. These inference results directly drive the city's emergency dispatch and command platform to perform on-site dispatch operations, ensuring timely task response, efficient resource utilization, and stable and reliable communication links.

[0170] S5. Based on the optimal resource allocation scheme, generate a standard control instruction set and issue it for execution through the command and dispatch system.

[0171] Specifically, the core objective of this step is to standardize and generate a set of control instructions with a standard structure based on the resource allocation scheme output from step four. And through the actual issuance and execution via the command and dispatch system, the closed loop of the city's intelligent communication command and dispatch system from perception to decision-making to execution is truly completed.

[0172] This step follows all the inference output from step four and includes: Task ID (TID): a unique identifier for the entire task; and the set of selected units. Unit corresponding equipment set (S) * ); Action instructions for each device (a); Link health score (real-time bandwidth, latency, packet loss rate).

[0173] Specifically, it includes:

[0174] First, for each task TID, according to the inferred unit u i and equipment d j Assemble the corresponding structured instructions The instructions use a standard field format, including: Task ID (TID); Unit ID (u i Device ID (d) j ); Action command code (a j The system includes: unified mapping encoding; necessary action parameters (such as fire extinguishing water volume and traffic control range, defined according to action type); additional link health information (current main link ID, packet loss rate, bandwidth information); redundant link activation flag (dynamically determined based on packet loss rate); and verification fields (CRC32 or Hash digest, to prevent tampering and mistransmission).

[0175] The specific generation operation is represented by the following standard assembly function:

[0176]

[0177] in:

[0178] Assemble(·) is the standard instruction encapsulation process;

[0179] LinkQual(d j Extracted from link health score;

[0180] Param(a j According to action a j Additional parameters are dynamically added as needed.

[0181] Furthermore, the redundancy strategy activation logic is as follows:

[0182]

[0183] (Where θ is the system-defined threshold, typically a 3% packet loss rate.)

[0184] Furthermore, after instruction encapsulation is completed, all instructions are uniformly distributed through the system control bus. The instruction is sent to the corresponding unit's scheduling and receiving module. Each instruction includes an execution acknowledgment (ACK) mechanism to ensure execution reliability.

[0185] To ensure clarity, here is a complete practical example:

[0186]

[0187] Furthermore, the sending process is as follows:

[0188] Command Center will Distribute the data in batches to all units;

[0189] Upon receiving the task, each unit should immediately register it in its internal task system.

[0190] After receiving the instruction, the device executes the corresponding action and sends back confirmation of execution.

[0191] The command system records the sending and execution status of each instruction in real time, forming a closed-loop data chain.

[0192] The final output is the set of sent instructions corresponding to all scheduled task TIDs. And the set of execution receipts (ACK / NACK) returned in real time.

[0193] All success or failure statuses will be entered into the scheduling log module for subsequent execution monitoring.

[0194] like Figure 2 As shown, another embodiment of the present invention provides an intelligent communication command and dispatch system based on artificial intelligence, the system comprising:

[0195] The multi-source data processing module 301 is used to acquire perception data of the target scene, extract time based on the perception data to generate preliminary events, align the preliminary times by timestamps, and perform density clustering in space to obtain several candidate events; fuse and score the several candidate events, select the candidate time with the highest score to form a standard structured event; the standard structured event includes event type, cluster center coordinates, event time, and event source identifier;

[0196] The scheduling task generation module 302 is used to perform vector similarity matching on the standard structured events, determine the list of units that have successfully responded in historical tasks, construct a preliminary candidate response unit set, and then calculate the response score by combining the real-time resource status and geographical location of the current unit. The top k units with the highest scores are selected to form the final response unit set and the task action set corresponding to each unit, generating a standard scheduling task package. The standard scheduling task package includes a unique task identifier, a set of response units, a set of action instructions for each unit, an overall task priority, and a task completion time limit.

[0197] The heterogeneous graph construction module 303 is used to construct a scheduling heterogeneous graph based on the standard scheduling task package, combined with the unit device set, real-time resource status and communication link information;

[0198] Graph neural network reasoning module 304 is used to perform graph neural network-based reasoning on the scheduling heterogeneous graph to infer the optimal resource allocation scheme.

[0199] The instruction generation and issuance module 305 is used to generate a standard structure set of control instructions based on the optimal resource allocation scheme, and to actually issue and execute them through the command and dispatch system.

[0200] In addition, for technical details not described in detail in this embodiment, please refer to the parameter operation method provided in any embodiment of the present invention, which will not be repeated here.

[0201] It should be noted that, in this document, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or system that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or system. Unless otherwise specified, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or system that includes that element.

[0202] The sequence numbers of the above embodiments of the present invention are for descriptive purposes only and do not represent the superiority or inferiority of the embodiments.

[0203] Through the above description of the embodiments, those skilled in the art can clearly understand that the methods of the above embodiments can be implemented by means of software plus necessary general-purpose hardware platforms. Of course, they can also be implemented by hardware, but in many cases the former is a better implementation method. Based on this understanding, the technical solution of the present invention, in essence, or the part that contributes to the prior art, can be embodied in the form of a software product. This computer software product is stored in a storage medium (such as read-only memory / random access memory, magnetic disk, optical disk) and includes several instructions to cause a terminal device (which may be a mobile phone, computer, server, air conditioner, or network device, etc.) to execute the methods described in the various embodiments of the present invention.

[0204] The above are merely preferred embodiments of the present invention and do not limit the scope of the patent. Any equivalent structural or procedural transformations made based on the description and drawings of the present invention, or direct or indirect applications in other related technical fields, are similarly included within the scope of patent protection of the present invention.

Claims

1. An intelligent communication command and dispatch method based on artificial intelligence, characterized in that, The method includes: The process involves acquiring perception data of the target scene, extracting time data to generate preliminary events, aligning these preliminary times with timestamps, and performing density clustering in space to obtain several candidate events. These candidate events are then fused and scored, and the candidate time with the highest score is selected to form a standard structured event. The standard structured event includes event type, cluster center coordinates, event time, and event source identifier. Vector similarity matching is performed on the standard structured events to determine a list of units that have successfully responded in historical tasks, thus constructing a preliminary candidate response unit set. Then, the response score is calculated by combining the real-time resource status and geographical location of the current unit. The top k units with the highest scores are selected to form the final response unit set and the task action set corresponding to each unit, generating a standard scheduling task package. The standard scheduling task package includes a unique task identifier, a set of response units, a set of action instructions for each unit, an overall task priority, and a task completion time limit. Based on the standard scheduling task package, and combined with the unit device set, real-time resource status and communication link information, a scheduling heterogeneity graph is constructed; The optimal resource allocation scheme is inferred by performing graph neural network-based reasoning on the scheduling heterogeneous graph. Based on the optimal resource allocation scheme, a standard set of control instructions is generated and executed through the command and dispatch system.

2. The intelligent communication command and dispatch method based on artificial intelligence according to claim 1, characterized in that, The sensing data includes video stream data, IoT device data stream, human voice or text reporting data, and historical scheduling task records. Then, the step of generating preliminary events by extracting time from the perceived data includes: The video stream data is processed by a lightweight object detection network to extract abnormal scene labels and corresponding confidence scores. After the IoT device data is received via MQTT or HTTP protocol, preliminary event candidates are triggered based on a set threshold. The artificial voice or text report data is converted into text through a speech-to-text network, then the TextRank is used to extract keywords, and the data is classified into a standard event type library.

3. The intelligent communication command and dispatch method based on artificial intelligence according to claim 1, characterized in that, The process of selecting the top k units by score to form the final response unit set and the corresponding task action set for each unit generates a standard scheduling task package, including: Obtain the response score; select the top k units of the response score to form the final response unit set; wherein the value of k is dynamically adjusted according to the event level corresponding to the event type; Encode the event type into a vector representation; Input the action reasoning network and introduce an action coupling regularization term into the action reasoning network to output the task action set corresponding to each unit; Based on the final set of response units and the set of task actions corresponding to each unit, a standard scheduling task package is assembled.

4. The intelligent communication command and dispatch method based on artificial intelligence according to claim 1, characterized in that, The step of constructing a scheduling heterogeneity graph based on the standard scheduling task package, combined with the unit device set, real-time resource status, and communication link information, includes: Obtain the standard scheduling task package; A task node is generated for each task's unique identifier according to the standard scheduling task package; Each response unit in the set of response units generates a unit node; For each unit node, a device node is generated for each available device; the set of unit devices contains several available devices. A communication node is generated for each available device associated with a communication gateway; the real-time resource status and communication link information includes several communication gateways; Directed edges are constructed based on the task nodes, unit nodes, device nodes, and communication nodes, including: task nodes to unit nodes, unit nodes to device nodes, and device nodes to communication nodes. By combining all task nodes, unit nodes, device nodes, communication nodes, and all directed edges, a scheduling heterogeneous graph is generated.

5. The intelligent communication command and dispatch method based on artificial intelligence according to claim 4, characterized in that, The edge weight w of the scheduling heterogeneous graph ij The structure is as follows: Where P represents the overall task priority; P max The ratio of the two values ​​represents the highest priority level, yielding a dimensionless urgency index; LinkQual(n) k ) represents the communication node n k Link quality score is calculated as LinkQual = NormalizedBandwidth - NormalizedPacketLoss, with a value range of [0,1]. Delay(n) is a normalized metric for link delay, ensuring it is dimensionless. k Real-time data collection, Delay max α1, α2, γ, δ1, δ2 are the maximum tolerable delay settings of the system; ActionAdapt(a) is the adaptability of action type a to the link requirements; TimePressure(D) is the task time limit pressure index; α1, α2, γ, δ1, δ2 are the weight adjustment coefficients respectively.

6. The intelligent communication command and dispatch method based on artificial intelligence according to claim 4, characterized in that, Each edge of the scheduling heterogeneous graph is also configured with: If the current unit node has multiple high-priority tasks attached, then the edge from the task to the unit will be penalized with a load penalty, and the priority will be reduced. If the current number of connections of a communication node exceeds the system's preset number of connections, a latency penalty is added to the edge from the device to the communication node, reducing the transmission priority.

7. The intelligent communication command and dispatch method based on artificial intelligence according to claim 5, characterized in that, The step of performing graph neural network-based reasoning on the scheduling heterogeneous graph to infer the optimal resource allocation scheme includes: Obtain the scheduling heterogeneity graph; Different types of nodes in the scheduling heterogeneous graph are respectively processed through independent small feedforward neural networks to embed the original features into a unified low-dimensional space, thereby generating low-dimensional features. The low-dimensional features are updated using a graph convolutional network to generate hidden representations for each layer of nodes after the update. In each convolutional operation of the graph convolutional network, different parameter matrices are used for different types of edges, and weighted aggregation is performed according to the edge weights to update the nodes. After the convolution operation of the convolutional layer is completed, a global scheduling selector is introduced for each task node to aggregate the final states of all related unit nodes and deduce the optimal resource selection set; wherein, the global scheduling selector is the GlobalReadout module.

8. The intelligent communication command and dispatch method based on artificial intelligence according to claim 7, characterized in that, The node represents the update calculation as follows: in, Let i be the hidden representation of node i at the +1l level. Let i be the hidden representation of node i at level l; Let be the hidden representation of node j at level l; A set of edge types; w is the set of neighboring nodes connected to i via edge type r; ij c represents the edge weight from node i to node j. i,r This is the normalization constant; σ represents the weight matrix for different edge types and self-loop connections; σ is the activation function.

9. The intelligent communication command and dispatch method based on artificial intelligence according to claim 1, characterized in that, The step of generating a standard-structured control instruction set based on the optimal resource allocation scheme and actually issuing and executing it through the command and dispatch system includes: Obtain the optimal resource allocation scheme and assemble the corresponding structured instructions; After the instructions are encapsulated, they are uniformly distributed to the corresponding units through the system control bus; Ultimately, we obtain the set of sent instructions corresponding to all scheduled task nodes, as well as the set of execution receipts returned in real time.

10. An intelligent communication command and dispatch system based on artificial intelligence, characterized in that, The system includes: A multi-source data processing module is used to acquire perception data of the target scene, extract time based on the perception data to generate preliminary events, align the preliminary times by timestamps, and perform density clustering in space to obtain several candidate events; fuse and score the several candidate events, select the candidate time with the highest score to form a standard structured event; the standard structured event includes event type, cluster center coordinates, event time, and event source identifier; The scheduling task generation module is used to perform vector similarity matching on the standard structured events, determine the list of units that have successfully responded in historical tasks, construct a preliminary candidate response unit set, and then calculate the response score by combining the real-time resource status and geographical location of the current unit. The top k units with the highest scores are selected to form the final response unit set and the task action set corresponding to each unit, generating a standard scheduling task package. The standard scheduling task package includes a unique task identifier, a set of response units, a set of action instructions for each unit, an overall task priority, and a task completion time limit. The heterogeneous graph construction module is used to construct a scheduling heterogeneous graph based on the standard scheduling task package, combined with the set of unit devices, real-time resource status and communication link information; The graph neural network reasoning module is used to perform graph neural network-based reasoning on the scheduling heterogeneous graph to infer the optimal resource allocation scheme. The instruction generation and issuance module is used to generate a standard set of control instructions based on the optimal resource allocation scheme, and to actually issue and execute them through the command and dispatch system.

Images