A user order intelligent allocation method and system

Abstract

This application provides a method and system for intelligent allocation of user orders, relating to the field of intelligent scheduling. The method includes: acquiring geographical environment data, weather data, and historical order data to construct a three-dimensional spatiotemporal knowledge graph; mapping the three-dimensional spatiotemporal knowledge graph to a dynamic spiking neural network; collecting real-time detection data to generate a pulse sequence, and inputting the pulse sequence into the dynamic spiking neural network for pulse propagation, the path of which is jointly determined by the topological connection structure of the neuron cluster and the current weights of synaptic connections; updating the weights of synaptic connections based on a time-dependent synaptic plasticity mechanism, which dynamically adjusts the weight values ​​according to the time difference between presynaptic and postsynaptic pulses and the rider's real-time physiological state data; activating a pulse competition mechanism to match and arbitrate orders with riders, obtaining the matching arbitration result, and executing order allocation. This method solves the technical problems of missing three-dimensional spatiotemporal knowledge representation and weak dynamic adaptive weighting capabilities.

Description

Technical Field

[0001] This application relates to the field of intelligent scheduling, and in particular to a method and system for intelligent allocation of user orders. Background Technology

[0002] Currently, intelligent order allocation in the on-demand delivery field generally uses a core technology approach of 2D geographic information + rule engines + traditional operations research optimization or deep learning. The mainstream implementation method is to first collect basic geographic data such as latitude and longitude and road network through GIS, and then integrate environmental information such as real-time weather and traffic congestion index. Next, it integrates historical order data, rider trajectories, and user behavior data, and uses path planning algorithms such as Dijkstra's algorithm to calculate the geometrical similarity between riders and orders. This is combined with the Hungarian algorithm, KM algorithm, or improved ant colony algorithm to complete global matching. Some platforms further utilize reinforcement learning models for weight iteration and strategy optimization. This approach can cover on-demand fulfillment application scenarios such as food delivery, fresh produce delivery, and same-city errand running.

[0003] However, in actual operation, when order volume surges in a short period during peak sales periods, the existing algorithm suffers from computational complexity that increases exponentially with the number of orders and riders. This leads to significant overruns in single-round scheduling, system degradation, and a degeneration into greedy matching or random order assignment, resulting in uneven rider load and widespread order timeouts. Alternatively, in extreme environments such as heavy rain or road closures, the fixed-weight model still prioritizes short-term delivery, continuously assigning long-distance orders to nearby riders. This causes long-distance orders to occupy capacity for extended periods while nearby orders accumulate and go unfulfilled, leading to regional fulfillment collapse and preventing the overall order assignment from achieving the expected results. Summary of the Invention

[0004] This application provides a user order intelligent allocation method and system, which solves the technical problems of lack of three-dimensional spatiotemporal knowledge representation and weak dynamic adaptive weighting ability in the prior art.

[0005] To achieve the above objectives, this application adopts the following technical solution: Firstly, a method for intelligent allocation of user orders includes: acquiring geographical environment data, weather data, and historical order data; constructing a three-dimensional spatiotemporal knowledge graph, which includes a spatial topology layer, an environmental feature layer, and a semantic association feature layer; mapping the three-dimensional spatiotemporal knowledge graph to a dynamic spiking neural network, where the spatial topology layer maps to the topological connection structure of the neuron cluster, the dynamic environmental attributes in the environmental feature layer map to the input modulation signals of the neuron cluster, and the entity relationships in the semantic association layer map to the synaptic connections between the neuron clusters; collecting real-time detection data to generate a pulse sequence, and inputting the pulse sequence into the dynamic spiking neural network for pulse propagation, the path of pulse propagation being jointly determined by the topological connection structure of the neuron cluster and the current weights of the synaptic connections; updating the weights of the synaptic connections based on a time-dependent synaptic plasticity mechanism, which dynamically adjusts the weight values ​​according to the time difference between presynaptic and postsynaptic pulses and the rider's real-time physiological state data; activating a pulse competition mechanism to match and arbitrate orders with riders, obtaining the matching arbitration result, and executing order allocation.

[0006] In conjunction with the first aspect mentioned above, one possible implementation of the process of constructing a three-dimensional spatiotemporal knowledge graph includes: acquiring geographic information system data of the target city, dividing the urban space into micro-units, and assigning each micro-unit three-dimensional spatial coordinates including longitude, latitude, and floor height or altitude to form a spatial topology layer; collecting multi-source environmental data, aligning it spatiotemporally with the micro-units, and generating an environmental feature layer containing spatial location and dynamic environmental attributes, including real-time weather data, road congestion index, historical delivery success rate, and environmental characteristics of delivery points; collecting and integrating historical order data, rider trajectory data, and user behavior data, and extracting entity nodes and relational edges. Entity nodes include business district nodes, road segment nodes, merchant nodes, rider nodes, user nodes, delivery point nodes, weather area nodes, and congestion area nodes. Relational edges include spatial relationships representing spatial adjacency or connectivity, temporal relationships representing the temporal evolution of historical data, and semantic relationships representing the association between delivery point type and delivery difficulty. Entity nodes are attached to micro-units, relation edges are used as topological paths connecting different entity nodes, and environmental features in the environmental feature layer, such as merchant meal preparation timeliness features and regional historical congestion index, are attached to the corresponding entity nodes in the form of attribute vectors to generate a three-dimensional spatiotemporal knowledge graph.

[0007] In conjunction with the first aspect mentioned above, in one possible implementation, entity nodes include business district nodes, road segment nodes, merchant nodes, rider nodes, user nodes, delivery point nodes, weather area nodes, and congestion area nodes. The node-based attributes of entity nodes include environmental characteristics of the delivery area, merchant food preparation timeliness, and historical congestion index of the area. Relationship edges include spatial relationships representing spatial adjacency or connectivity, temporal relationships representing the evolution of historical data, and semantic relationships representing the association between delivery point type and delivery difficulty.

[0008] In conjunction with the first aspect mentioned above, in one possible implementation, the mapping construction process of the dynamic spiking neural network specifically includes: mapping entity nodes in the three-dimensional spatiotemporal knowledge graph to neuron clusters, each neuron cluster containing an input integrator, a membrane potential register, and a pulse emitter. Mapping relation edges in the three-dimensional spatiotemporal knowledge graph to synaptic connection arrays connecting two neuron clusters; initializing the weights of the synaptic connection arrays, each containing multiple parallel synaptic channels, and each synaptic channel including a weight register and a time-dependent synaptic plasticity learner; constructing a pulse routing protocol, which includes configuring a pulse routing table for each neuron cluster and a pulse gating unit for each synaptic connection array. The pulse routing table stores the physical address mappings of adjacent neuron clusters, and the pulse gating unit selects or blocks synaptic channels based on the membrane potential state of the presynaptic neuron cluster and external modulation signals. Constructing an asynchronous pulse computation engine, which includes an event queue, a parallel computing unit array, and a distributed clock synchronizer. The event queue receives and sorts pulse events, the parallel computing unit array performs membrane potential integration and pulse emission determination in parallel, and the distributed clock synchronizer maintains a global logical timestamp.

[0009] In conjunction with the first aspect mentioned above, in one possible implementation, the process of initializing the weights of the synaptic connection array specifically includes: calculating initial weight values ​​based on the reciprocal of the spatial distance corresponding to the relation edges, the historical cooperation success rate, and the semantic association strength, and storing them in a weight register. The time-dependent synaptic plasticity learner adjusts the weight values ​​in the weight register according to the time difference between the presynaptic pulse and the postsynaptic pulse.

[0010] In conjunction with the first aspect mentioned above, in one possible implementation, the process of updating the weights of synaptic connections based on a time-dependent synaptic plasticity mechanism specifically includes: acquiring real-time feedback pulses and rider physiological data generated during pulse propagation, and converting the physiological data into physiological constraint pulses. The real-time feedback pulses include performance result pulses and rider state update pulses, and the physiological data includes at least one of heart rate data, skin conductance response data, body temperature data, and motion acceleration data. A time-dependent synaptic plasticity learner is constructed, comprising a presynaptic time recorder, a postsynaptic time recorder, a weight adjustment factor calculation unit, a physiological constraint modulation unit, and a weight update execution unit. When a cluster of presynaptic neurons in any synaptic connection fires a pulse, the presynaptic time recorder records its firing timestamp. When a cluster of postsynaptic neurons in the same synaptic connection fires a pulse, the postsynaptic time recorder records its firing timestamp. The weight adjustment factor calculation unit calculates the time difference based on the firing timestamps and calculates the basic weight adjustment factor based on an asymmetric time window function. The physiological constraint modulation unit generates modulation coefficients based on physiological constraint pulses, which are used to reduce the weight adjustment amplitude of the corresponding synaptic connection when physiological data indicates that the rider is in a state of fatigue or stress. The weight update execution unit updates the weight of the synaptic connection based on the base weight adjustment factor, modulation coefficients, and the current weight value of the synaptic connection, combined with preset weight upper and lower limits.

[0011] In conjunction with the first aspect mentioned above, one possible implementation involves activating a pulse competition mechanism to match and arbitrate orders with riders, obtaining the matching arbitration result, and executing the order allocation process. Specifically, this includes: obtaining the current weight values ​​of synaptic connections and the current membrane potential values ​​of each neuron cluster; constructing a pulse competition array, which includes an order node competition pool, a rider node competition pool, a competition intensity calculation unit, a stability suppression input port, and a matching arbitration unit. Within the pulse competition array, the membrane potential values ​​of the order node neuron clusters are stored as the order competition intensity in the order competition pool, and the membrane potential values ​​of each rider node neuron cluster are stored as the rider competition intensity in the rider competition pool. All synaptic connections between order nodes and rider nodes are traversed, and the comprehensive competition intensity of each order-rider pair is calculated based on the current weight values ​​of the synaptic connections, the initial order competition intensity, and the initial rider competition intensity, generating a competition matrix. The stability suppression input port receives suppression signals and reduces the competition intensity of the corresponding order-rider connection. The suppression signals are generated based on the future stability index of the order-rider matching. An iterative competition suppression strategy is used to match and arbitrate the competition matrix, and the matching and arbitration results are sent to the scheduling and execution module. The scheduling and execution module then sends the order dispatch instruction to the corresponding rider terminal.

[0012] In conjunction with the first aspect mentioned above, one possible implementation includes, before order allocation, acquiring the pulse propagation path and pulse intensity during pulse propagation. The pulse propagation path is a sequence of intermediate nodes traversed by the pulse as it propagates from the order node neuron cluster to the rider node neuron cluster. The pulse intensity is a coefficient of amplitude attenuation recorded as the pulse travels along the propagation path. The sequence of intermediate nodes along the pulse propagation path is encoded as a path feature vector, which includes spatial resistance features, time delay features, and semantic matching features. Based on the pulse intensity and the attenuation process along the pulse propagation path, a path comprehensive impedance value is calculated, which is negatively correlated with the pulse intensity. Based on the path feature vector and the path comprehensive impedance value, a future stability index for each candidate order and rider match is calculated using a stability evaluation model. The future stability index includes the timeout probability and the rider load change trend. The future stability index is compared with a stability threshold to generate a suppression signal, which is used to reduce the activation intensity of the corresponding order and rider connection in the pulse competition mechanism.

[0013] Secondly, a user order intelligent allocation system is provided, comprising: a graph construction module, a pulse coding module, a brain-like computing core module, and a scheduling execution module. The graph construction module constructs a three-dimensional spatiotemporal knowledge graph, which includes a spatial topology layer, an environmental feature layer, and a semantic association layer. The pulse coding module encodes real-time data into pulse sequences. The brain-like computing core module includes a dynamic spiking neural network, mapped from the three-dimensional spatiotemporal knowledge graph. This network receives pulse sequences, performs pulse propagation, conducts online learning and weight adjustment based on a time-dependent synaptic plasticity mechanism, and outputs matching results through a pulse competition mechanism, which includes a stability inhibition input port. The scheduling execution module sends order dispatch instructions to rider terminals based on the matching results and collects the order execution results as feedback pulses to the brain-like computing core module.

[0014] In conjunction with the second aspect mentioned above, in one possible implementation, the brain-like computing core module is integrated onto a dedicated neuromorphic computing chip. The dedicated neuromorphic computing chip adopts an event-driven architecture to process in parallel the membrane potential update and pulse firing events of all neuron groups in the dynamic spiking neural network.

[0015] This application provides a method and system for intelligent allocation of user orders. It constructs a three-dimensional spatiotemporal knowledge graph containing a spatial topology layer, an environmental feature layer, and a semantic association feature layer. This graph accurately depicts the spatial relationships and road network structure of riders, orders, and delivery areas, comprehensively captures dynamic attributes such as weather and congestion, and clearly outlines the semantic relationships between orders, riders, and areas. It achieves three-dimensional fusion representation of multi-source heterogeneous spatiotemporal data, solving the problem of missing three-dimensional spatiotemporal knowledge representation and laying a solid data foundation for accurate matching. Based on this, the graph is mapped layer by layer into a dynamic spiking neural network. Combined with real-time data, it generates and propagates spiking sequences. Relying on the time-dependent synaptic plasticity mechanism, it dynamically adjusts synaptic weights based on spiking time differences and rider physiological states, overcoming the problem of poor adaptability of fixed weights. This allows the scheduling strategy to be optimized in real time according to the environment and capacity status. Simultaneously, a spiking competition mechanism completes the arbitration of order and rider matching, balancing matching accuracy, delivery efficiency, and rider physiological safety. This significantly reduces order timeout rates, optimizes capacity allocation, and ensures stable operation under high concurrency and extreme environments, achieving efficient, stable, and user-friendly intelligent scheduling. Attached Figure Description

[0016] Figure 1 A flowchart illustrating a user order intelligent allocation method provided in an embodiment of this application; Figure 2 A flowchart illustrating the construction process of a three-dimensional spatiotemporal knowledge graph in a user order intelligent allocation method provided in this application embodiment; Figure 3 The present application provides a user order intelligent allocation method, which maps a static graph into a dynamic spiking neural network with dynamic computing capabilities, and provides a flowchart of the process of initializing network weights using prior knowledge. Figure 4 A flowchart illustrating the dynamic spiking neural network weight update process in a user order intelligent allocation method provided in this application embodiment; Figure 5 A flowchart illustrating the stability assessment process before activating the pulse competition mechanism for final matching arbitration in a user order intelligent allocation method provided in this application embodiment; Figure 6 The flowchart of the intelligent allocation method for user orders provided in this application embodiment is as follows: after completing pulse propagation, weight update and stability assessment, a specific competitive architecture and strategy are constructed to realize the final order matching arbitration process. Figure 7 This is a schematic diagram of the structure of a user order intelligent allocation system provided in an embodiment of this application. Detailed Implementation

[0017] To make the objectives, technical solutions, and advantages of this invention clearer, the technical solutions of the embodiments of this invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some, not all, of the embodiments of this invention. All other embodiments obtained by those skilled in the art based on the embodiments of this invention without creative effort are within the scope of protection of this invention.

[0018] Unless otherwise defined, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this invention pertains. The terminology used herein in the description of the invention is for the purpose of describing particular embodiments only and is not intended to be limiting of the invention.

[0019] Example 1: like Figure 1 As shown, this embodiment provides a method for intelligent allocation of user orders. This method constructs a three-dimensional spatiotemporal knowledge graph and maps it to a dynamic spiking neural network. It leverages the event-driven characteristics of brain-like computing to solve the computational bottleneck problem in high-concurrency scenarios, and combines a physiological data feedback mechanism to achieve humanized order scheduling. Specifically, it includes the following steps: Step S100: Obtain geographic environment data, weather data, and historical order data to construct a three-dimensional spatiotemporal knowledge graph.

[0020] In some implementations, geographic information system data of the target city is obtained and spatial indexing algorithms are used to divide the urban space into micro-units. Each micro-unit is assigned three-dimensional spatial coordinates of longitude, latitude and altitude to form a spatial topology layer, which is used to characterize the physical connectivity of the urban geographic space, such as the accessibility of the road network.

[0021] The system collects multi-source environmental data such as real-time weather data, road congestion index, and historical delivery success rate. Through a spatiotemporal alignment algorithm, each data point is matched with its corresponding micro-unit to generate an environmental feature layer containing spatial location and dynamic environmental attributes. This layer is used to attach dynamically changing attributes, such as real-time weather conditions and road congestion index.

[0022] Historical order data, rider trajectory data, and user behavior data are collected and integrated. From these, business district nodes, road segment nodes, merchant nodes, rider nodes, user nodes, delivery point nodes, weather area nodes, and congestion area nodes are extracted as entity nodes. Simultaneously, spatial relationships representing adjacency or connectivity, temporal relationships representing the evolution of historical data, and semantic relationships representing the correlation between delivery point type and delivery difficulty are extracted as relation edges, resulting in a semantic association feature layer. This layer is used to extract implicit logical relationships between entities, such as the semantic association between "hospital" and "high delivery difficulty." Then, the entity nodes are attached to the corresponding micro-units, the relation edges are used as the topological paths connecting different entity nodes, and the merchant's meal preparation timeliness characteristics and regional historical congestion index in the environmental feature layer are attached to the corresponding entity nodes in the form of attribute vectors, thereby generating a complete three-dimensional spatiotemporal knowledge graph.

[0023] Step S200: Map the three-dimensional spatiotemporal knowledge graph into a dynamic spiking neural network.

[0024] Among them, dynamic spiking neural networks are computational models that simulate the information processing mechanisms of biological nervous systems.

[0025] In some implementations, each entity node in the three-dimensional spatiotemporal knowledge graph is mapped to a cluster of neurons. Each cluster is configured with an input integrator to accumulate input pulses, a membrane potential register to record excitation states, and a pulse generator to generate output pulses when the membrane potential exceeds a preset threshold.

[0026] The relation edges in the knowledge graph are mapped to synaptic connection arrays connecting two clusters of neurons. These arrays contain multiple parallel synaptic channels, each with an independently configured weight register and a time-dependent synaptic plasticity learner.

[0027] Thus, the initial weight value can be calculated and stored in the weight register based on the reciprocal of the spatial distance corresponding to the relation edge, the historical cooperation success rate, and the semantic association strength. At the same time, a pulse routing table is configured for each neuron cluster to store the physical address mapping of adjacent clusters, and a pulse gating unit is configured for each synaptic connection array to select or block synaptic channels according to the membrane potential state of the presynaptic cluster and the external modulation signal.

[0028] Finally, an asynchronous pulse computing engine was constructed, comprising an event queue for receiving and sorting pulse events, a parallel computing unit array for parallel execution of membrane potential integration and pulse emission determination, and a distributed clock synchronizer for maintaining global logical timestamps, thus completing the mapping construction from static graphs to dynamic spiking neural networks.

[0029] Step S300: Collect real-time detection data to generate a pulse sequence, and input the pulse sequence into a dynamic pulse neural network for pulse propagation.

[0030] In some implementations, the pulse coding module monitors order generation events, rider location updates, and weather warnings in the business flow in real time. It calculates the initial pulse intensity based on the urgency of the order and encapsulates it together with the current timestamp into a pulse sequence, which can then be input into the corresponding neuron cluster in the dynamic spiking neural network.

[0031] Upon arrival of the pulse, the input integrator converts it into an increment of membrane potential and accumulates it in the membrane potential register. When the accumulated membrane potential exceeds a preset excitation threshold, the pulse generator generates an output pulse. At this point, the pulse queries the physical address of the adjacent clusters according to the pulse routing table of the current neuron cluster. Simultaneously, the pulse gating unit checks the membrane potential state of the presynaptic cluster and the external modulation signal to determine whether to select the synaptic channel. When the selected pulse is transmitted along the synaptic connection array to the next neuron cluster, it is multiplied by the current weight value in the weight register to achieve amplitude attenuation. The attenuated pulse continues to repeat the above integration, judgment, and propagation process in the next cluster until the pulse energy is attenuated to the point that it is insufficient to excite any neuron cluster or reach the target output node.

[0032] It's important to note that pulses tend to propagate along paths with higher synaptic weights, mimicking the "survival of the fittest" mechanism in biological neural networks. For example, a pulse from an order node will preferentially propagate to nearby rider nodes with a high historical success rate in cooperation, because the synaptic weights connecting these nodes are typically higher. This propagation mechanism based on topology and weights allows the system to quickly identify potentially high-quality matching solutions without traversing all possible combinations.

[0033] Step S400: Update the weights of synaptic connections based on the time-dependent synaptic plasticity mechanism.

[0034] Among them, the time-dependent synaptic plasticity mechanism is a weight adjustment method that simulates the learning rules of the biological brain. The presynaptic impulse refers to the output impulse emitted from the neuronal cluster at the signal sending end, and the postsynaptic impulse refers to the output impulse emitted by the neuronal cluster at the signal receiving end after receiving the presynaptic impulse.

[0035] In some implementations, this step endows the system with the ability to learn and adapt online. It allows time-dependent synaptic plasticity mechanisms to mimic the learning patterns of the biological brain: if the firing of a presynaptic neuron (such as an order node) leads to the excitation of a postsynaptic neuron (such as a rider node), the connection weight between the two is strengthened; conversely, it is weakened.

[0036] The system also incorporates riders' real-time physiological data (such as heart rate and fatigue level) as constraints for weight adjustment. When the system detects that a rider is fatigued or stressed, even if the rider is a good match for the order, the synaptic plasticity mechanism will inhibit the increase of weight and even actively reduce the connection strength, thereby avoiding over-assigning orders to fatigued riders. This mechanism not only improves the system's decision-making intelligence but also reflects humanistic care for delivery capacity, effectively preventing fulfillment risks caused by rider overwork.

[0037] Step S500: Activate the pulse competition mechanism to match and arbitrate orders with riders, obtain the matching and arbitration results, and execute order allocation.

[0038] Among them, the impulse competition mechanism is a matching arbitration method that simulates the lateral inhibition effect of the biological nervous system.

[0039] In some implementations, a pulse competition array is constructed, which includes an order node competition pool, a rider node competition pool, a competition intensity calculation unit, a stability suppression input port, and a matching arbitration unit.

[0040] The membrane potential value of the order node neuron cluster is stored in the order competition pool as the order competition intensity.

[0041] The membrane potential value of each rider node's neuron cluster is stored in the rider competition pool as the rider competition intensity.

[0042] The synaptic connections between all order nodes and rider nodes can be traversed. The competition intensity calculation unit calculates the comprehensive competition intensity of each order and rider pair based on the current weight value of the synaptic connection, the initial competition intensity of the order, and the initial competition intensity of the rider, generating a competition matrix.

[0043] Simultaneously, the stability suppression input port receives suppression signals generated based on future stability indicators matching orders and riders, thereby reducing the competitive intensity of the corresponding order and rider connection.

[0044] Finally, the matching arbitration unit adopts an iterative competition suppression strategy for the competition matrix. That is, it first finds the matching pair with the highest competition intensity and marks it as the winner. At the same time, it suppresses the connection strength between the order and other riders, and between the rider and other orders. Then, it repeats the above process in the remaining candidates until all orders or capacity are allocated. The matching arbitration result is sent to the scheduling execution module, which then sends the order dispatch instruction to the corresponding rider terminal.

[0045] It should be noted that after pulse propagation and weight updates, there may be multiple potential matching pairs in the network. The pulse competition mechanism can resolve resource conflicts, and its core logic is "winner-takes-all". The system compares the overall activation intensity of each potential matching path (determined by multiple factors such as pulse intensity, propagation path impedance, and weights), and the order and rider connection pair with the highest activation intensity will win the competition.

[0046] Based on the above technical solutions, a three-dimensional spatiotemporal knowledge graph is constructed to integrate geographical environment, weather, and historical order data into structured prior knowledge, solving the representation bottleneck problem of existing technologies that struggle to integrate multi-source heterogeneous spatiotemporal information for joint decision-making. Building upon this, the three-dimensional spatiotemporal knowledge graph is mapped to a dynamic spiking neural network. An event-driven architecture ensures that the computational load is linearly related to the number of active spiking events, rather than cubically related to the total number of orders, thus resolving the system degradation and failure problem caused by the explosion of computational complexity under high concurrency. Simultaneously, real-time data is collected to generate spiking sequences, which are then input into the network for spiking propagation. This allows spiking events to naturally converge along paths with higher weights, achieving parallel optimization and avoiding the computational overhead of traditional algorithms that traverse all matching combinations. Furthermore, based on a time-dependent synaptic plasticity mechanism, weights are dynamically updated according to spiking time differences and rider physiological states. This enables the network to automatically suppress the connection strength between distant orders and nearby riders and enhance regional mobility connections in extreme environments such as heavy rain, solving the rigid coupling problem of fixed weights in dynamic environments. Finally, the pulse competition mechanism is activated to perform matching arbitration through the side inhibition principle. While ensuring global balanced allocation, it outputs the optimal dispatch instruction, thereby achieving millisecond-level real-time response in high-concurrency scenarios and adaptive and stable scheduling in extreme environments.

[0047] Example 2: This embodiment is a detailed description of step S100 in Embodiment 1. In constructing the three-dimensional spatiotemporal knowledge graph, this embodiment deeply integrates physical space, dynamic environment, and business semantics through a layered construction approach, specifically including the following steps: Step S101: Obtain geographic information system data of the target city, divide the urban space into micro-units, and assign each micro-unit three-dimensional spatial coordinates including longitude, latitude, and floor height or altitude to form a spatial topology layer.

[0048] The micro-unit is the smallest computational grid that divides urban space according to geographic information system data, and is used to carry the basic physical attributes of the delivery environment. The spatial topology layer is a three-dimensional network structure composed of micro-units through spatial adjacency or connectivity, and is used to describe the actual spatial connections between riders, orders, and delivery points.

[0049] In some implementations, geographic information system (GIS) data of the target city is acquired, and spatial indexing algorithms are used to divide the urban planar area into multiple macro-grids. Within these macro-grids, micro-units can then be further subdivided based on building density and road structure, and each micro-unit is assigned longitude and latitude coordinates.

[0050] Simultaneously, using Building Information Modeling (BIM) or Digital Elevation Modeling (DEM), the average floor height or ground elevation of the main buildings within the unit is obtained to form three-dimensional spatial coordinates. Then, all micro-units are established with spatially adjacent or connected edges according to actual road connectivity and internal building passage relationships, constructing a spatial topology layer that reflects the three-dimensional spatial structure of the city, providing a spatial benchmark for the subsequent mounting of environmental features and physical nodes.

[0051] It should be noted that the granularity of the micro-unit division can be dynamically adjusted according to the type of urban area. A finer granularity is used in densely populated commercial areas to improve spatial resolution, while a coarser granularity is used in suburban areas to reduce computational complexity. The floor height data of the three-dimensional spatial coordinates can be obtained through the building's 3D model or LiDAR point cloud data, which is used to distinguish the different spatial characteristics of ground delivery and in-building delivery.

[0052] Step S102: Collect multi-source environmental data, align it with micro-units in time and space, and generate an environmental feature layer that includes spatial location and dynamic environmental attributes.

[0053] Multi-source environmental data refers to real-time information related to the delivery environment collected from different data sources, including real-time weather data, road congestion index, historical delivery success rate, and environmental characteristics of delivery points.

[0054] In some implementations, real-time weather data is collected from meteorological service platforms; road congestion indices are obtained from traffic management systems; historical delivery success rates for each delivery point at different times are statistically analyzed from historical order databases; and environmental characteristics of delivery points are obtained through on-site data collection or user feedback.

[0055] This allows various types of data to be located to their corresponding microscopic units according to their spatial coordinates: Weather data is spatially interpolated and allocated according to regional range; The road congestion index is mapped to the micro-units covered by road segments; Historical delivery success rates are aggregated and statistically analyzed based on the micro-unit to which the delivery point belongs.

[0056] This allows for timestamp alignment of multi-source data within each micro-unit, combining different types of environmental data collected at the same time into an environmental attribute vector for that unit.

[0057] Finally, an environmental feature record containing spatial location identifiers and dynamic environmental attributes is generated for each micro-unit. The environmental feature records of all micro-units together constitute the environmental feature layer.

[0058] Step S103: Collect and integrate historical order data, rider trajectory data, and user behavior data, and extract entity nodes and relationship edges.

[0059] Historical order data refers to the historical delivery order information recorded by the system, including fields such as the origin and destination of the order, the order time, the promised delivery time, the actual delivery time, the order type, and the delivery result.

[0060] Rider trajectory data is a sequence of location points recorded in chronological order during the delivery process, reflecting the rider's movement path and speed changes.

[0061] User behavior data is a record of user actions during the delivery service process, including interactions such as placing orders, expediting orders, canceling orders, and providing feedback.

[0062] Physical nodes include business district nodes, road segment nodes, merchant nodes, rider nodes, user nodes, delivery point nodes, weather area nodes, and congestion area nodes.

[0063] The node-based attributes of physical nodes include environmental characteristics of the delivery area, the timeliness of the merchant's food preparation, and the historical congestion index of the region.

[0064] Relationship edges include spatial relationships that represent spatial adjacency or connectivity, temporal relationships that represent the evolution of historical data, and semantic relationships that represent the association between delivery point type and delivery difficulty.

[0065] In some implementations, historical order data is collected from the business database. From this data, the originating merchant node, target delivery point node, executing rider node, and time-series information such as order placement time and delivery time are extracted. Rider trajectory data is also collected, and path analysis is performed on each rider's historical location sequence to extract the sequence of road segment nodes traversed by the rider and the delivery point nodes where they stopped, thus constructing the access relationships between rider nodes, road segment nodes, and delivery point nodes.

[0066] Simultaneously, the association between user nodes, merchant nodes, and delivery point nodes is extracted from user order records; negative feedback relationships between user nodes and specific order nodes or rider nodes are extracted from order reminder and cancellation records.

[0067] This allows for the extraction of business district nodes based on business district boundaries; road segment nodes based on road networks; delivery point nodes based on historical delivery popularity; weather area nodes based on weather data coverage; and congestion area nodes based on congestion data distribution.

[0068] Finally, the extracted entity nodes are attached to the corresponding micro-units according to their spatial location, and relationship edges are established based on the spatial connectivity, temporal sequence, and semantic similarity between the nodes, forming the initial graph structure of entity nodes and relationship edges.

[0069] Step S104: Mount the entity nodes to the micro-unit, use the relation edges as the topological paths connecting different entity nodes, and mount the environmental features in the environmental feature layer, such as the merchant's meal preparation timeliness features and the regional historical congestion index, to the corresponding entity nodes in the form of attribute vectors to generate a three-dimensional spatiotemporal knowledge graph.

[0070] Among them, mounting refers to associating an entity node with a corresponding micro-unit according to its spatial coordinate position, so that the entity node can obtain spatial positioning attributes.

[0071] In some implementations, based on the spatial coordinates of the entity nodes, the spatial range of all micro-units is traversed, and each entity node is matched to the micro-unit to which its coordinates belong, establishing the membership relationship between the entity node and the micro-unit. Then, based on the relationship edges, according to the spatial connectivity, temporal order, and semantic association strength between entity nodes, the relationship edges are organized into topological paths connecting different entity nodes, forming a network connectivity structure between entity nodes.

[0072] The environmental attribute vector corresponding to each micro-unit is obtained from the environmental feature layer. Merchant meal preparation timeliness features are extracted and attached to the attribute fields of the corresponding merchant nodes. Historical congestion indices for the region are extracted and attached to the attribute fields of the corresponding road segment nodes and congested area nodes. Finally, the entity nodes with attached attribute vectors are integrated with the organized topological path relationship edges to form a three-dimensional spatiotemporal knowledge graph containing spatial positioning information, environmental attribute information, and entity association information.

[0073] It should be noted that using attribute vector mounting allows abstract semantic information to be quantified into specific numerical features. For example: For the "hospital" delivery point node, the system uses historical data to calculate its average delivery time, elevator waiting time, access control pass rate, and other data to generate an attribute vector containing features such as "high cost of going upstairs" and "low probability of access control pass rate", and attaches it to the node.

[0074] For the "Congested Area" node, the system attaches attributes such as real-time congestion index and historical peak-hour congestion probability to it. For the "Merchant" node, it attaches attributes such as the merchant's food preparation timeliness and the region's historical congestion index.

[0075] Based on the above technical solution, a three-dimensional spatiotemporal knowledge graph is constructed, comprising a spatial topology layer, an environmental feature layer, and a semantic association feature layer. The spatial topology layer accurately depicts the spatial location relationships and road network topology between riders, orders, and delivery points by dividing urban space into micro-units with three-dimensional coordinates. The environmental feature layer comprehensively captures dynamic environmental factors affecting delivery efficiency by spatiotemporally aligning dynamic attributes such as real-time weather and congestion indices with micro-units. Simultaneously, the semantic association feature layer extracts entity nodes and relational edges from historical orders, rider trajectories, and user behavior data, clearly outlining the implicit logical relationships between entities such as order type, rider fulfillment ability, and regional delivery popularity. By attaching entity nodes to micro-units and environmental features to entity nodes in the form of attribute vectors, the three layers collaboratively achieve a leap from two-dimensional planar modeling to three-dimensional spatial fusion. This provides comprehensive and accurate structured data support for subsequent precise matching of orders and riders, fundamentally breaking through the representation bottleneck of existing technologies that cannot simultaneously consider multi-dimensional information (spatial, environmental, and semantic) for joint decision-making.

[0076] Example 3: This embodiment is a detailed description of step S200 in Embodiment 1. After constructing the three-dimensional spatiotemporal knowledge graph, this embodiment elaborates on how to map the static graph into a dynamic spiking neural network with dynamic computing capabilities, and how to initialize network weights using prior knowledge, specifically including the following steps: Step S201: Map the entity nodes in the three-dimensional spatiotemporal knowledge graph to a cluster of neurons.

[0077] Among them, the neuron cluster is a computational unit that simulates the function of a biological neuron group. It is composed of multiple homogeneous or heterogeneous neuron models and is used to represent the comprehensive state and behavior pattern of a certain entity node in the knowledge graph.

[0078] In some implementations, each entity node in the three-dimensional spatiotemporal knowledge graph is mapped to a cluster of neurons. This allows an input integrator to receive pulse signals from other neuron clusters, which are then weighted according to synaptic weights and accumulated in a membrane potential register. The membrane potential register continuously records the current membrane potential value, which decays naturally over time via a leakage mechanism. When the value in the membrane potential register exceeds a preset activation threshold, a pulse generator is triggered, generating a standardized output pulse. This pulse carries the identifier and timestamp of the current neuron cluster and is ready to be transmitted to neighboring neuron clusters.

[0079] After the pulse is delivered, the membrane potential register is reset to the resting potential value according to the reset mechanism, and waits for the next integration cycle.

[0080] Throughout the process, the input integrator, membrane potential register, and pulse emitter form a closed loop, jointly simulating the complete information processing chain of a biological neuron receiving, integrating, and firing pulses.

[0081] It is important to note that the neuron cluster is not a simple scalar numerical storage unit, but rather a logical component that simulates the function of biological neurons. The input integrator receives and accumulates pulse signals transmitted from other neuron clusters, simulating the dendritic function of biological neurons. The membrane potential register temporarily stores the current excitation state of the neuron; its value changes with the accumulation of input pulses, simulating the changes in cell body potential in biological neurons. The pulse generator monitors the value of the membrane potential register; when the value exceeds a preset excitation threshold, it generates and sends out a pulse signal, simulating the axonal function of biological neurons.

[0082] For example, when the membrane potential of a cluster of neurons in an order node gradually increases and exceeds a threshold due to pulse input from surrounding rider nodes, the cluster of neurons is "activated," meaning that the order has the potential to be assigned. This mapping mechanism transforms static graph nodes into dynamic computational agents with the ability to evolve states and transmit information.

[0083] Step S202: Map the relation edges in the three-dimensional spatiotemporal knowledge graph to synaptic connection arrays connecting two neuron clusters, and initialize the weights of the synaptic connection arrays.

[0084] Among them, the synaptic connection array is a multi-channel parallel connection structure that connects two clusters of neurons, with each channel independently responsible for transmitting a specific type of association.

[0085] In some implementations, a corresponding number of parallel synaptic channels are established between two neuron clusters based on the type and number of relation edges in the three-dimensional spatiotemporal knowledge graph, and each channel is independently associated with a semantic relation.

[0086] For each channel, three indicators are obtained: spatial distance, historical cooperation success rate, and semantic association strength. These three indicators are then combined into an initial weight value using a weighted calculation formula. The calculated initial weight value is then stored in the weight register of that channel as the baseline attenuation coefficient for pulse transmission. Simultaneously, a time-dependent synaptic plasticity learner is configured for each channel. This learner continuously monitors the arrival time difference between the anterior and posterior synaptic pulses. When the time difference is positive and its absolute value is small, a positive adjustment factor is generated to enhance the weight; when the time difference is negative, a negative adjustment factor is generated to weaken the weight. The adjusted weight value is then written back to the weight register.

[0087] Step S203: Construct a pulse routing protocol, which includes configuring a pulse routing table for each neuron cluster and configuring a pulse gating unit for each synaptic connection array.

[0088] In some implementations, a pulse routing table is constructed for each neuron cluster. All adjacent nodes in the three-dimensional spatiotemporal knowledge graph that have a relationship edge with the entity node are traversed. The unique identifier, physical memory address or computational core number of each adjacent node, as well as the index list of all parallel synaptic channels between them are recorded in the routing table.

[0089] At the same time, each synaptic connection array is equipped with an independent pulse gating unit. This unit has two input ports: one is connected to the membrane potential register of the presynaptic cluster to read the current excitation state, and the other is connected to an external modulation signal bus to receive environmental constraints such as weather and congestion.

[0090] When the pulse routing protocol receives an output pulse from a neuron cluster, it first queries the pulse routing table to obtain the address information of the target neighboring cluster and the corresponding synaptic channel index. Then, it sends the pulse to the corresponding pulse gating unit. The gating unit checks whether the presynaptic membrane potential is higher than the activity threshold and whether the external modulation signal allows transmission. Only when both conditions are met will the pulse be selected to the synaptic channel for subsequent transmission.

[0091] It's important to note that the pulse routing table, similar to the routing table in an internet router, records the physical address mapping between the current neuron cluster and adjacent neuron clusters, ensuring that pulse signals can be accurately transmitted along the topology defined by the graph. It also serves as a setting for the pulse gating unit, acting as a "gate" for pulse transmission. The gating unit not only detects whether the presynaptic neuron has fired a pulse but also detects external modulation signals (such as environmental features like severe weather or road closures). Only when all conditions are met does the gating unit select the synaptic channel, allowing the pulse to pass. This enables an "event-driven" computing model: computational resources are only consumed when an event occurs (pulse firing) and conditions allow (gating open). For connections without pulses or with closed gating, the system does not perform computation, thus greatly reducing unnecessary computation and solving the computational resource bottleneck problem in high-concurrency scenarios.

[0092] Step S204: Construct an asynchronous pulse computing engine, which includes an event queue, a parallel computing unit array, and a distributed clock synchronizer.

[0093] In some implementations, an event queue is first constructed as a unified entry point for pulse events. This queue uses a priority queue structure to sort all received pulse events by timestamp, ensuring that earlier events are processed first. Then, a parallel computing unit array is initialized, and all neurons in the dynamic spiking neural network are distributed to each computing unit according to the principle of computational load balancing. Each computing unit independently runs the membrane potential integrator and the pulse firing decision unit.

[0094] When a pulse event pops out of the event queue, the event is routed to the corresponding parallel computing unit according to the identifier of the target cluster. The computing unit reads the current membrane potential value of the cluster, accumulates it according to the pulse intensity, and then determines whether it exceeds the firing threshold.

[0095] The distributed clock synchronizer maintains a monotonically increasing logical timestamp globally. Before processing a pulse event, each computing unit requests the current timestamp from the synchronizer. After processing, it adds the current timestamp to the newly generated pulse event and pushes it back to the event queue, forming a complete asynchronous computing closed loop.

[0096] It should be noted that the parallel computing unit array extracts events from the event queue and processes the membrane potential integrals and impulse firing determination of different neuron clusters in parallel. Since the computations of different neurons are independent, this architecture naturally supports large-scale parallel computing. The distributed clock synchronizer is responsible for maintaining a global logical timestamp, ensuring the correct causal order of impulse events in the network.

[0097] For example, during peak order periods, tens of thousands of neurons may be active simultaneously. The asynchronous pulse computing engine can use a parallel computing unit array to process the state updates of these neurons at the same time without waiting for all neurons to complete synchronously, thus ensuring the system's response speed within milliseconds.

[0098] Step S205: Calculate the initial weight value based on the reciprocal of the spatial distance corresponding to the relation edge, the historical cooperation success rate, and the semantic association strength, and store it in the weight register.

[0099] In some implementations, the spatial coordinates of two entity nodes connected by relational edges are obtained from a three-dimensional spatiotemporal knowledge graph, the Euclidean distance or road network distance between them is calculated, and the reciprocal of the two distances is taken as the reciprocal feature value of the spatial distance.

[0100] The historical order database is used to count the total number of times the two entity nodes corresponding to the relationship edge co-occur in historical deliveries, as well as the number of times they were successfully fulfilled. The historical cooperation success rate is calculated as the ratio of the number of successful fulfillments to the total number of co-occurrences. If the number of co-occurrences is zero, it is set as the baseline value by default.

[0101] Using pre-trained entity embedding vectors in the knowledge graph, the cosine similarity between two entity node vectors is calculated as the semantic association strength.

[0102] The three feature values ​​of spatial distance reciprocal, historical cooperation success rate and semantic association strength are fused according to preset weight coefficients using a weighted summation formula to calculate the initial weight value, and the value is written into the weight register of the corresponding synaptic channel.

[0103] Step S206: The time-dependent synaptic plasticity learner adjusts the weight values ​​in the weight register according to the time difference between the presynaptic pulse and the postsynaptic pulse.

[0104] In some implementations, an independent time-dependent synaptic plasticity learner is configured for each synaptic channel. This learner internally maintains a presynaptic time recorder and a postsynaptic time recorder, which store the timestamp of the most recent pulse emission, respectively.

[0105] When the learner detects a cluster of presynaptic neurons firing a pulse, it stores the current global timestamp in the presynaptic time recorder and waits for a postsynaptic response.

[0106] When a cluster of postsynaptic neurons in the same synaptic channel is detected firing a pulse, the current timestamp is stored in the postsynaptic time recorder and the weight adjustment calculation is immediately triggered.

[0107] At this point, the learner reads the timestamps from the previous and next recorders to calculate the time difference. The time difference is then input into an asymmetric time window function to generate a basic weight adjustment factor. If the time difference is positive and the absolute value is small, the adjustment factor is positive and the value is large, indicating that the weight should be increased. If the time difference is negative, the adjustment factor is negative, indicating that the weight should be weakened.

[0108] Finally, the adjustment factor is added to the current weight value and constrained within the preset upper and lower weight limits. The updated weight value is then written to the weight register to complete the weight adjustment.

[0109] It's important to note that after initialization, the network weights are not fixed but possess online learning capabilities. The time-dependent synaptic plasticity learner monitors the time difference between the firing pulses of neurons at both ends of a connection. If a presynaptic neuron (e.g., an order node) fires a pulse, and a postsynaptic neuron (e.g., a rider node) fires a pulse shortly afterward (meaning a successful match), the weight of that synapse is increased according to the STDP rule; conversely, if the time difference is too large or there is no response, the weight is decreased. This ensures that the network can continuously optimize connection strength based on real-time feedback, achieving adaptive evolution.

[0110] Based on the above technical solution, by mapping entity nodes of the three-dimensional spatiotemporal knowledge graph to neuron clusters and relation edges to synaptic connection arrays and initializing weights, a structured prior knowledge foundation is constructed for the dynamic spiking neural network, thus solving the representation bottleneck problem of existing technologies that struggle to integrate multi-source heterogeneous spatiotemporal information for joint decision-making. On this basis, a spiking routing protocol and an asynchronous spiking computation engine are constructed, ensuring that spiking propagates only along active paths and that computational complexity is linearly related to the number of active nodes. Combined with weight initialization based on the reciprocal of spatial distance, historical cooperation success rate, and semantic association strength, this addresses the system degradation and failure problem caused by the exponential increase in computational complexity of traditional algorithms under high concurrency. Furthermore, a time-dependent synaptic plasticity learner dynamically adjusts weights according to spiking time differences, enabling the network to automatically suppress the connection strength between distant orders and nearby riders and enhance regional mobility connections in dynamic environments such as extreme weather, solving the rigid coupling failure problem of fixed weights during sudden environmental changes. The three components work together to achieve the dual technical effects of millisecond-level real-time response in high-concurrency scenarios and adaptive stable scheduling in extreme environments.

[0111] Example 4: This embodiment is a detailed description of step S400 in Embodiment 1, and a further development of step S206 in Embodiment 3. During the operation of the dynamic spiking neural network, the weight updates are not static, but dynamically adjusted based on real-time feedback and physiological constraints. Specifically, it includes the following steps: Step S401: Obtain the real-time feedback pulses and rider physiological data generated during pulse propagation, and convert the physiological data into physiological constraint pulses.

[0112] The real-time feedback pulses are supervisory signals collected and encoded from the actual business execution results by the dynamic spiking neural network after executing order allocation decisions. The fulfillment result pulses are generated based on whether the order was delivered on time, canceled, or timed out, indicating the success or failure of the matching decision. For example, when the rider delivers on time, the system generates a positive feedback pulse; when a timeout or cancellation occurs, the system generates a negative feedback pulse. Rider status update pulses are generated based on real-time reported status change events such as order acceptance, pickup, and delivery by the rider, dynamically reflecting changes in the availability of transportation resources, such as riders going online, offline, or changing location. More importantly, this embodiment introduces rider physiological data as a constraint. This physiological data can be collected in real-time through smart wearable devices worn by the rider (such as smart bracelets or smart helmets), including heart rate data, skin conductance data, body temperature data, and motion acceleration data. These raw physiological signals are converted into physiological constraint pulses to modulate the learning intensity of the network. For example, an elevated heart rate usually indicates that the rider is under high-intensity exercise or stress, while dramatic fluctuations in skin conductance data may indicate emotional stress. By incorporating these physiological characteristics into the network input, the system can perceive the rider's "latent state," thereby reflecting humanistic care and risk control in order dispatch decisions.

[0113] In some implementations, biosensors integrated into rider smart bracelets or smart helmets are used to collect heart rate, skin conductance, body temperature, and motion acceleration data in real time at a fixed sampling rate, and the raw signals are filtered to remove noise and outliers.

[0114] The preprocessed physiological data is input into a preset physiological state assessment model. This model compares real-time heart rate with an individual resting heart rate threshold obtained from historical data statistics, calculates the short-term fluctuation variance of skin conductance response, and analyzes the cumulative energy value of exercise acceleration, outputting a fatigue stress coefficient between 0 and 1. This coefficient is then converted into a physiological constraint pulse with a specific timestamp and intensity using a pulse coding function. That is, when the fatigue stress coefficient exceeds a preset warning threshold, a high-intensity physiological constraint pulse is generated; otherwise, a low-intensity or zero-value pulse is generated.

[0115] Therefore, these physiological constraint pulses are injected into the synaptic plasticity learner connected to the corresponding rider node. During the weight update phase, they are multiplied by the basic weight adjustment factor as modulation coefficients. This mathematically achieves the reduction of the weight adjustment amplitude of the corresponding synaptic connection when the rider is in a state of fatigue or stress, thus completing the physiological constraint on the network learning process.

[0116] Step S402: Construct a time-dependent synaptic plasticity learner.

[0117] It should be noted that the synaptic plasticity learner is the core component of the Dynamic Spiking Neural Network (DSPN) that enables it to possess "intelligence" and "adaptive" capabilities. It simulates the mechanism by which the connection strength between neurons in the biological brain changes with activity experience. The presynaptic and postsynaptic time recorders are used to capture the timing information of pulse transmission. The weight adjustment factor calculation unit is responsible for calculating the basic learning intensity based on the timing information. Simultaneously, the physiological constraint modulation unit introduces physiological constraints to correct the learning intensity. Finally, the weight update execution unit completes the weight read and write operations, thus forming a complete closed loop of perception, calculation, correction, and execution.

[0118] Step S403: When any presynaptic neuron cluster connected by a synapse fires a pulse, the presynaptic time recorder records its firing timestamp; when the postsynaptic neuron cluster connected by the same synapse fires a pulse, the postsynaptic time recorder records its firing timestamp.

[0119] Among them, the time-dependent synaptic plasticity learner is a mathematical computation module that simulates the Hebbian learning rule in the biological brain. It is used to dynamically adjust the connection weights based on the time difference between the firing pulses of neurons before and after the synapse.

[0120] In some implementations, an independent time-dependent synaptic plasticity learner is instantiated for each parallel synaptic channel in each synaptic connection array. This learner internally maintains a presynaptic time recorder and a postsynaptic time recorder.

[0121] When the system detects that the presynaptic neuron cluster is firing a pulse, the learner immediately reads the global logical timestamp and stores it in the presynaptic time recorder; When a cluster of postsynaptic neurons in the same synaptic channel is detected firing a pulse, the current timestamp is read and stored in the postsynaptic time recorder, and the weight adjustment calculation process is triggered.

[0122] Meanwhile, the weight adjustment factor calculation unit extracts timestamps from the two recorders, calculates the time difference, and substitutes the difference into a predefined asymmetric time window function. This function outputs a larger positive adjustment factor to simulate synaptic enhancement when the time difference is positive and the absolute value is small, and outputs a negative adjustment factor to simulate synaptic inhibition when the time difference is negative.

[0123] Meanwhile, the physiological constraint modulation unit receives physiological constraint pulses converted from rider physiological data and maps the pulse intensity to a modulation coefficient between 0 and 1.

[0124] The weight update execution unit obtains the basic weight adjustment factor and modulation coefficient, multiplies the two to obtain the final adjustment amount, adds it to the weight value in the current weight register, and compares the result with the preset upper and lower weight limits. If it exceeds the range, it is truncated to the boundary value. Finally, the updated weight value is written back to the weight register to complete a complete synaptic plasticity learning process.

[0125] For example, when the order node (presynapse) fires a pulse first, and then the rider node (postsynapse) is activated and fires a pulse, this "cause and effect" temporal relationship indicates that there is a strong causal relationship between the two, and the connection should be strengthened.

[0126] Conversely, if the rider node is already in an excited state before the order node issues its pulse, or if the time interval between the two is too long, it indicates a weak correlation. Accurate timestamp recording provides the data foundation for subsequent calculations.

[0127] Step S404: The weight adjustment factor calculation unit calculates the time difference based on the issuance timestamp and calculates the basic weight adjustment factor based on the asymmetric time window function.

[0128] In some implementations, the asymmetric time window function is the mathematical core of the STDP mechanism. This function defines the non-linear relationship between the time difference and the weight adjustment magnitude. The specific logic is as follows: If the time difference is positive (i.e., the presynaptic pulse precedes the postsynaptic pulse), the function outputs a positive value, indicating that the synaptic connection should be strengthened, and the smaller the time difference, the greater the strengthening.

[0129] If the time difference is negative (i.e., the postsynaptic pulse precedes the presynaptic pulse), the function outputs a negative value, indicating that the synaptic connection should be weakened.

[0130] This asymmetry mimics the "Hebb's learning" rule of biological neurons, where "neurons that fire together are linked together." Therefore, through this function, the network can autonomously learn which order and rider combinations are closely correlated in time, thereby gradually optimizing the network topology.

[0131] Step S405: The physiological constraint modulation unit generates modulation coefficients based on the physiological constraint pulse.

[0132] Among them, the physiological constraint modulation unit is an interface module in the time-dependent synaptic plasticity learner, which is used to convert the rider's physiological state into a modulated signal that can be intervened for weight updates.

[0133] In some implementations, a physiological constraint pulse generated for a specific rider node is received. This pulse carries a pulse intensity value, which is preset to be between 0 and 1, where 0 indicates that the rider's physiological state is completely normal and 1 indicates that the rider is in a state of extreme fatigue or stress.

[0134] Therefore, a piecewise linear mapping mechanism is adopted: When the pulse intensity is lower than the first preset threshold T1, the modulation coefficient output is 1, indicating that no suppression is performed; When the pulse intensity is higher than the second preset threshold T2, the modulation coefficient output is 0, indicating that the weight adjustment is completely suppressed; When the pulse intensity is between T1 and T2, the modulation coefficient is calculated according to the linear interpolation formula k=1-(pulse intensity-T1) / (T2-T1) to achieve a smooth transition.

[0135] After completing the calculation, the modulation unit outputs the modulation coefficient to the weight update execution unit. This coefficient will be multiplied by the basic weight adjustment factor in the subsequent weight update process, thereby mathematically scaling the weight adjustment range.

[0136] Throughout the calculation process, the greater the intensity of the physiological constraint pulse, the smaller the generated modulation coefficient, and the more significantly the weight adjustment amplitude of the corresponding synaptic connection is suppressed. In this way, the physiological mechanism of the inhibition of neuronal activity in an organism under fatigue state is simulated in the computer system.

[0137] Step S406: The weight update execution unit updates the weight of the synaptic connection based on the basic weight adjustment factor, modulation coefficient, and the current weight value of the synaptic connection, combined with the preset upper and lower weight limits.

[0138] In some implementations, the basic weight adjustment factor is obtained. The modulation coefficient k is obtained from the physiological constraint modulation unit, and the current weight value is read from the weight register of the current synaptic channel. .

[0139] The execution unit multiplies the base weight adjustment factor by the modulation coefficient. The final weight adjustment amount is calculated. The final adjustment amount will then be added to the current weight value. To obtain temporary weight values .

[0140] After completing the addition operation, the execution unit will add the temporary weight value to the preset weight limit. and weight lower bound Comparison: If temporary weight value Greater than the weight limit Then set the updated weight value to ; If temporary weight value Less than the lower weight limit Then set the updated weight value to ; If temporary weight value Due to the weight cap and weight lower bound In between, directly As the updated weight value.

[0141] Finally, the execution unit writes the weight values ​​after boundary constraint processing into the weight register of the synaptic channel, completing a full weight update process. This accurately simulates the dynamic change mechanism of biological synaptic connection strength, which is regulated by activity experience and constrained by physiological state, in the computer system.

[0142] Based on the above technical solution, by acquiring real-time feedback pulses and rider physiological data and converting them into physiological constraint pulses, a supervisory signal and physiological constraint source can be provided for subsequent weight adjustment. A time-dependent synaptic plasticity learner is then constructed, utilizing its internal presynaptic and postsynaptic time recorders to accurately capture the pulse delivery timing. The weight adjustment factor calculation unit then calculates the basic weight adjustment factor based on the time difference and asymmetric time window function. The physiological constraint modulation unit generates modulation coefficients based on the physiological constraint pulses. Finally, the weight update execution unit multiplies the basic adjustment factor and modulation coefficients, combining them with preset upper and lower weight limits to complete the weight update. This entire mechanism solves the technical problem in existing technologies where fixed weights and the inability to perceive the true state of transport capacity lead to decision failure in extreme environments. On the one hand, it allows the network to autonomously learn which spatiotemporal paths are more efficient from the fulfillment results using the pulse timing-dependent STDP mechanism, achieving dynamic adaptive evolution of weights. On the other hand, it scales the weight adjustment range through the modulation coefficients generated by the physiological constraint pulses, actively suppressing the growth of related connection weights when riders are fatigued or stressed, preventing the system from over-assigning orders to high-risk transport capacity. This allows business feedback and physiological state to be co-encoded as constraints for weight updates, enabling the model to maintain its ability to continuously learn efficient matching in dynamic environments, while also automatically implementing protective inhibition when capacity approaches its limit. This achieves the dual benefits of intelligent adaptive weight adjustment and safe and humanized capacity scheduling.

[0143] Example 5: This embodiment provides a detailed description of the preliminary steps of step S500 in Embodiment 1. Before activating the pulse competition mechanism for final matching arbitration, this embodiment introduces a stability assessment step to address the potential problem of high immediate matching accuracy but high future performance risk, serving as a "pre-emptive defense" for the competition mechanism. Specifically, it includes the following steps: Step S501: Obtain the pulse propagation path and pulse intensity during pulse propagation.

[0144] In a dynamic spiking neural network, the pulse propagation path refers to the sequence of nodes a pulse passes through from the source neuron cluster (e.g., the order node), through a series of intermediate neuron clusters (e.g., merchant nodes, road segment nodes, congestion area nodes), and finally to the target neuron cluster (e.g., the rider node). Pulse intensity is a dynamic coefficient that records the amplitude of the pulse as it propagates along the aforementioned path, gradually decreasing as it passes through synaptic connections and nodes.

[0145] In some implementations, an event-driven architecture is used to track the complete propagation trajectory of a pulse in a dynamic spiking neural network. When an order node's neuron cluster generates a pulse due to a new order input, a unique identifier is assigned to the pulse and its initial intensity is recorded. Subsequently, according to the pulse routing table of the neuron cluster, it is directed to an adjacent neuron cluster. When passing through the synaptic connection array, its intensity is multiplied by the current weight value of that connection, achieving the first decay.

[0146] When a pulse reaches an intermediate node, it triggers the integration of the membrane potential of the neuron cluster at that node. If the membrane potential exceeds the threshold, the node itself will fire a new pulse. At this time, the node's identifier is recorded as part of the path.

[0147] This process iterates until the pulse intensity decays below the excitation threshold or reaches the rider node. Simultaneously, the sequence of all traversed node identifiers (i.e., the propagation path) and the pulse intensity value after decay at each stage are recorded, thus obtaining complete dynamic propagation information from the order to each candidate rider.

[0148] For example, a pulse from an order node might pass through merchant nodes, congested road node, and community entrance node before finally reaching the rider node; this sequence is the pulse propagation path. The pulse intensity reflects the energy loss of the signal during transmission. Since the pulse loses energy at each node or edge due to synaptic weight attenuation or node threshold filtering, the pulse intensity is actually a sequence of attenuation coefficients. Obtaining this path and intensity information is equivalent to recording the "trajectory" and "energy state" of the matching scheme in the graph, providing raw data for subsequent stability assessment.

[0149] Step S502: Encode the sequence of intermediate nodes of the pulse propagation path into a path feature vector.

[0150] In some implementations, an intermediate node sequence is extracted from the recorded pulse propagation path, removing the source and target nodes. For each node in the sequence, its pre-stored attribute vector is read from a 3D spatiotemporal knowledge graph. For road segment nodes, their average travel time and congestion fluctuation coefficient are extracted to form spatial resistance characteristic components. For merchant nodes, extract the distribution of meal preparation time by time period to form a time delay feature component; For delivery point nodes, extract their building type labels (such as hospitals, schools, and residential areas) to form semantic matching feature components.

[0151] Finally, the feature components of each node are weighted and aggregated according to the path order to form a fixed-length path feature vector. The aggregation weight decays exponentially according to the position of the node in the path, and the node closer to the target node is given a higher weight.

[0152] For example, if the path contains a "congested road segment node", the feature value is higher; if the path contains a "building with elevators node", the feature value is lower.

[0153] The time delay feature is used to quantify the time consumption risk of a path. For example, if the path contains "merchant nodes with slow food preparation", then the feature value will be larger.

[0154] Semantic matching features can be used to quantify the fit between order requirements and rider attributes. For example, if an order is for "fragile items" and the route includes a "new rider node", the semantic matching feature value may be low.

[0155] This encoding method can map complex heterogeneous information such as road conditions, merchant characteristics, and rider capabilities into a unified feature vector, enabling the stability assessment model to examine the feasibility of matching schemes from multiple perspectives.

[0156] Step S503: Based on the pulse intensity and the attenuation process along the pulse propagation path, calculate the path composite impedance value.

[0157] In some implementations, a complete pulse intensity decay sequence (including the real-time intensity value of the pulse after it starts from the source node, passes through each synaptic connection and each cluster of intermediate neurons) is extracted from the recorded pulse propagation events. The total attenuation factor of the pulse intensity is calculated by dividing the initial pulse intensity by the pulse intensity when it finally reaches the target node, which reflects the total attenuation effect of the path on the pulse.

[0158] The total attenuation factor is mapped to the path impedance value through a negative correlation conversion function, specifically using logarithmic normalization: impedance value = 1 - log(total attenuation factor) / log(preset maximum attenuation factor). The impedance value is constrained between 0 and 1, with a larger value indicating greater path resistance. This conversion process simulates the cumulative loss mechanism of signal transmission in biological neural systems. A higher attenuation factor indicates more impedance nodes or lower synaptic weights in the path, corresponding to a higher path impedance value.

[0159] Step S504: Based on the path feature vector and the path comprehensive impedance value, calculate the future stability index of each candidate order and rider matching through the stability evaluation model.

[0160] Among them, the stability assessment model is a pre-trained machine learning model used to predict the risk level of the matching scheme in the future performance process based on the path feature vector and the path comprehensive impedance value.

[0161] In some implementations, the encoded path feature vector is concatenated with the path composite impedance value to form a complete input feature vector. This vector is then input into a pre-built stability assessment model, which uses a gradient boosting decision tree structure to output two values ​​in parallel: timeout probability (between 0 and 1) and rider load change trend (positive for load increase, negative for load decrease). These two values ​​together constitute the future stability index.

[0162] The training process of the stability assessment model is as follows: A large number of historical matching samples are extracted from the historical order database. Each sample contains the path feature vector, the comprehensive path impedance value, and the final fulfillment result (whether there was a timeout, and the change in subsequent orders accepted by the rider) when a successful match is achieved. The path feature vector and the comprehensive path impedance value are used as input features, and timeout status and rider load changes are used as supervision labels. The model is trained by minimizing the cross-entropy loss function to learn the nonlinear mapping relationship between path features and fulfillment risk. After training, the model is deployed to the online service.

[0163] For example, even if a rider is currently close to a merchant and has a high instantaneous pulse intensity, the path feature vector indicates that the rider is about to enter a congested area, and the overall path impedance value shows that the subsequent path resistance is high. Based on this, the model calculates that the rider has a high probability of exceeding the timeout limit in the next 30 minutes, and the rider's load change trend indicates that he may be fatigued due to taking too many orders. This predictive ability allows the system to identify potential risk scenarios before matching occurs.

[0164] Step S505: Compare the future stability index with the stability threshold to generate a suppression signal.

[0165] In some implementations, two components of the stability assessment model's output are obtained: the timeout probability and the rider load change trend. The timeout probability threshold is obtained by using the 80th quantile of historical order timeout rates; the load change threshold is obtained by using the upper limit of the normal fluctuation range of rider load. These two components can then be compared with preset stability thresholds. If the timeout probability exceeds a threshold or the rider load change trend exceeds a threshold, the matching scheme is deemed to have a fulfillment risk, and a suppression signal is generated. This signal contains two key parameters: the suppressed order and rider connection identifier, and the suppression strength coefficient. The suppression strength coefficient is dynamically calculated based on the degree to which the threshold is exceeded; the more the threshold is exceeded, the stronger the suppression. The calculation formula is suppression coefficient = min(1, timeout probability / threshold + load change / threshold), with the result constrained between 0 and 1.

[0166] The generated suppression signal is sent to the stability suppression input port of the pulse competition array, and multiplied by the original comprehensive competition strength of the connection during subsequent competition strength calculation to achieve proactive weight reduction for risk matching.

[0167] It should be noted that this inhibitory signal does not directly cancel the match, but rather acts as a negative regulatory factor input into the subsequent impulse competition mechanism. Its mechanism of action is similar to the "inhibitory postsynaptic potential" in biological nervous systems, reducing the activation intensity of the neuronal connection.

[0168] Based on the above technical solution, by acquiring the pulse propagation path and intensity and encoding them as path feature vectors, the complex topological structure experienced by the pulse in the three-dimensional spatiotemporal knowledge graph can be quantified into spatial resistance, time delay, and semantic matching features, thus solving the representation bottleneck of joint decision-making for multi-source heterogeneous spatiotemporal information. Furthermore, the comprehensive path impedance value is calculated based on pulse attenuation, and the dynamic resistance of the entire link is quantified using the cumulative effect of amplitude attenuation, compensating for the deficiency that static distance cannot reflect real-time impedance. Then, through a stability assessment model, future stability indicators are calculated, fusing multi-dimensional features into predictable risk values ​​such as timeout probability and load change trends, realizing the ability to learn risk patterns from historical experience. Finally, the indicators are compared with a stability threshold to generate a suppression signal, actively reducing the weight of high-risk matches in the pulse competition mechanism, simulating the lateral inhibition mechanism of the biological nervous system. This forms a closed loop of path tracking, feature quantification, risk prediction, and active suppression, solving the defect of existing decision-making that only focuses on current advantages while ignoring future performance risks, achieving an upgrade from instant matching to stable prediction, and improving the timeout controllability of order allocation and the global robustness of capacity scheduling.

[0169] Example 6: This embodiment is a detailed description of step S500 in Embodiment 1, and also a specific application of the stability evaluation results in Embodiment 5. After completing pulse propagation, weight update, and stability evaluation, the system achieves final order matching arbitration by constructing a specific competitive architecture and strategy. Specifically, it includes the following steps: Step S601: Obtain the current weight values ​​of synaptic connections and the current membrane potential values ​​of each neuron cluster, and construct a pulse competition array.

[0170] Among them, the pulse competition array is a matching arbitration logic structure in dynamic spiking neural networks used to simulate the lateral inhibition effect of biological nervous systems.

[0171] In some implementations, the spiking competition array serves as the "execution center" for the output decisions of the dynamic spiking neural network. It is not a physically independent hardware array, but rather a logical computation area partitioned within the neuromorphic computing core module. The order node competition pool and the rider node competition pool are used to temporarily store the state information of all currently unassigned order neuron clusters and available rider neuron clusters, respectively. The competition intensity calculation unit is responsible for quantifying the competitiveness of each potential matching pair. The stability inhibition input port, as an external intervention interface, receives the inhibition signal generated in Example 5. Simultaneously, the matching arbitration unit is responsible for outputting the final instruction based on the competition results. This array-based design enables the system to process the competition relationships of hundreds or thousands of order and rider pairs in parallel, avoiding the latency caused by traditional serial computation.

[0172] Step S602: Within the pulse competition array, the membrane potential value of the order node neuron cluster is stored in the order competition pool as the order competition intensity, and the membrane potential value of each rider node neuron cluster is stored in the rider competition pool as the rider competition intensity.

[0173] In some implementations, the value currently stored in the membrane potential register of each order node's neuron cluster is read from the dynamic spiking neural network, and this value is directly used as the order competition intensity, and then stored in the order competition pool according to the order node identifier.

[0174] At the same time, the value currently stored in the membrane potential register of each rider node's neuron cluster is read, and this value is directly used as the rider competition intensity and stored in the rider competition pool according to the rider node identifier.

[0175] During the storage process, the system maintains the mapping relationship between the order competition pool and the rider competition pool, ensuring that the competition intensity of each order node and rider node corresponds to its index position when participating in the subsequent calculation of the overall competition intensity. This transforms the continuous membrane potential state values ​​of the neuron clusters in the dynamic spiking neural network into static initial intensity values ​​in the discrete competition pool, providing input data for the subsequent competition intensity calculation unit to traverse the synaptic connections between all orders and riders and calculate the overall competition intensity.

[0176] It's important to note that membrane potential values ​​reflect the "excitation level" of a neuronal cluster. For order nodes, a higher membrane potential indicates a more urgent order (e.g., impending timeout) or lower environmental impedance. For rider nodes, a higher membrane potential indicates a lighter load and better rider condition. Storing these analog values ​​in a competition pool is equivalent to mapping the urgency of orders and availability of delivery capacity in the physical world to initial energy values ​​in a competition space. This achieves the transformation from continuous neurodynamic states to discrete competition variables.

[0177] Step S603: Traverse all synaptic connections between order nodes and rider nodes. Based on the current weight value of the synaptic connection, the initial intensity of order competition, and the initial intensity of rider competition, calculate the comprehensive competition intensity of each order and rider pair and generate a competition matrix.

[0178] The competition matrix is ​​a two-dimensional array structure composed of the overall competitive intensity of all order and rider pairs. Its rows correspond to order nodes and its columns correspond to rider nodes. Each element in the matrix represents the matching competitiveness of that order and rider combination.

[0179] In some implementations, the initial order competition strength of each order node is read from the order competition pool; the initial rider competition strength of each rider node is read from the rider competition pool; and the current weight values ​​of the synaptic connections between each order node and each rider node are obtained from the dynamic spiking neural network.

[0180] Iterate through all combinations of order nodes and rider nodes. For each pair of order i and rider j, use the competition intensity calculation unit to multiply the initial competition intensity of order i, the initial competition intensity of rider j, and the current weight value of the synaptic connection between them to obtain the comprehensive competition intensity of the order and rider pair.

[0181] The calculated overall competition intensity is then stored in the corresponding matrix positions according to the order index and rider index, thus constructing a complete competition matrix. The number of rows in this competition matrix represents the total number of orders to be assigned, and the number of columns represents the total number of available riders. Each element in the matrix reflects the initial advantage of the corresponding matching combination in subsequent competition arbitration.

[0182] For example, if an order has high competition intensity (urgent) and the synaptic weight connecting to a rider is high (close distance, good historical cooperation), but the rider's competition intensity is low (current load is full), then the calculated overall competition intensity will be lowered. The system iterates through all possible combinations to generate an N×M competition matrix, which fully represents the global matching situation at the current moment.

[0183] Step S604: The stability suppression input port is used to receive suppression signals and reduce the competition intensity of the corresponding order and rider connection.

[0184] In some implementations, the stability suppression input port receives a suppression signal from an external source. This signal carries the identifiers of the suppressed order node and rider node, as well as a suppression strength coefficient. Based on these identifiers, the corresponding element in the competition matrix is ​​located. The overall competition strength of this element is then multiplied by the suppression strength coefficient, and the product replaces the original overall competition strength, thereby reducing the competition strength for that order and rider connection.

[0185] The suppression strength coefficient ranges from 0 to 1. When the suppression strength coefficient is 0, it means that the match is completely suppressed. When the suppression strength coefficient is 1, it means that no suppression is performed.

[0186] This operation simulates the mechanism of inhibitory postsynaptic potentials in biological neural systems, reducing the activation intensity of specific neural connections through external negative input, thereby decreasing the probability of winning high-risk matches in subsequent match arbitration.

[0187] Step S605: The competition matrix is ​​matched and arbitrated using an iterative competition suppression strategy. The matching and arbitration results are then sent to the scheduling and execution module, which in turn sends a dispatch instruction to the corresponding rider terminal.

[0188] The iterative competition suppression strategy is a matching arbitration algorithm that simulates the lateral inhibition effect of biological neural networks. It selects the matching pair with the highest overall competition strength through multiple iterations, and suppresses other matching pairs related to the winning pair after each iteration until all orders or riders are allocated. This simulates the "lateral inhibition" mechanism in biological neural networks.

[0189] In some implementations, the matching arbitration unit obtains a competition matrix adjusted for stability suppression, which records the overall competition strength of all order-rider pairs. This allows for the execution of an iterative competition suppression strategy. Find the element with the highest overall competition intensity in the current competition matrix, mark the order node and rider node corresponding to that element as a winning match, and record them in the matching arbitration result set.

[0190] Next, remove the row containing the winning match and the column containing the rider from the competition matrix. At the same time, suppress the competition intensity of other matching pairs related to that order or rider, simulating the winner-take-all effect of lateral inhibition in biological neural systems. Repeat the above iterative process of finding the maximum value, marking the winner, and removing rows and columns until the competition matrix is ​​empty or all orders or riders have been assigned.

[0191] Finally, the generated matching arbitration result is sent to the scheduling execution module, which then sends a dispatch instruction to the corresponding rider terminal based on the result.

[0192] It's important to note that in the competition matrix, each element is not only related to its own strength but also inhibited by surrounding elements. We can first identify the matching pair with the highest competition strength, mark it as the winner, and simultaneously suppress the connection strength between that order and other riders, and between that rider and other orders, simulating a "winner-takes-all" effect. Then, we repeat this process in the remaining candidate set until all orders or capacity are allocated. This ensures the balance of the global allocation and avoids overall imbalance caused by local optima.

[0193] Based on the above technical solution, by constructing a pulse competition array and reading the current weight values ​​of synaptic connections and the current membrane potential values ​​of each neuron cluster, the membrane potential values ​​can be stored as order competition intensity and rider competition intensity in the order competition pool and rider competition pool, respectively. Then, by traversing all synaptic connections between orders and riders, the initial order competition intensity, the initial rider competition intensity, and the current synaptic weight values ​​are multiplied to generate a comprehensive competition intensity to construct a competition matrix. This quantifies the dynamic state of the neural network into a directly comparable competition intensity, solving the bottleneck problem of existing technologies in integrating multi-source spatiotemporal information and capacity status into a unified decision indicator. Furthermore, by receiving inhibition signals at the stability inhibition input port and multiplicatively reducing the weights of corresponding elements in the competition matrix, the inhibitory postsynaptic potential mechanism of biological nervous systems is simulated. This allows external stability assessment results to directly intervene in the competition intensity of high-risk matches, addressing the decision-making deficiency of existing technologies that only focus on current matching advantages while ignoring future fulfillment risks. Finally, an iterative competition suppression strategy is adopted for matching arbitration of the competition matrix. By selecting the globally optimal match in each round and removing the corresponding row and column, the winner-takes-all effect of side suppression is simulated. A conflict-free matching arbitration result is generated and the scheduling execution module sends the dispatch instruction. This process reduces the exponential computational complexity of the traditional combinatorial optimization algorithm O(N³) to event-driven computation that is linearly related to the number of active nodes. This solves the problem of system degradation and failure caused by the explosion of computational complexity in high-concurrency scenarios and achieves the dual technical effects of millisecond-level real-time response under high concurrency and adaptive and stable scheduling under extreme environments.

[0194] Example 7: This embodiment provides a user order intelligent allocation system, which is applied to the user order intelligent allocation methods described in the above embodiments. For example... Figure 7 As shown, this system, through its modular architecture design, organically combines 3D spatiotemporal modeling, spiking neural computation, and business scheduling execution, achieving an end-to-end closed loop from data perception to decision output. Specifically, it includes: a graph construction module, a spiking coding module, a brain-like computing core module, and a scheduling execution module.

[0195] The graph construction module obtains raw data from external data sources (such as GIS servers, meteorological bureau open interfaces, and business databases) through API interfaces or data synchronization tools. Internally, it integrates spatial indexing algorithms (such as R-tree indexes) and entity relationship extraction models to transform the raw data into structured graph data. The module's output is a graph instance stored in a graph database (such as Neo4j) or distributed memory, containing entity nodes (such as merchants, riders, and orders) and their attribute vectors, relation edges, and their topological connections. Simultaneously, the graph construction module updates the graph content periodically or under specific event triggers, such as refreshing the road congestion index every minute or writing newly generated order nodes in real time, ensuring that the knowledge base upon which subsequent calculation modules rely is real-time and accurate.

[0196] The pulse coding module monitors the event stream in the business flow in real time, including new order generation, rider location GPS drift, and weather warnings. For each event, the coding module converts it into a pulse signal with timestamps and intensity according to preset coding rules. For example, for a new order event, the coding module calculates the initial pulse intensity based on its urgency (such as the proximity of the promised delivery time); for rider location updates, it generates a continuous pulse sequence based on the rider's speed and direction. These pulse sequences are encapsulated into standard data packets and transmitted to the neuromorphic computing core module via a high-speed message bus (such as Kafka or Flink). This encoding method directly transforms complex business state changes into electrical signals that can be processed by neural networks, providing standardized input for subsequent neuromorphic computing.

[0197] The core module of neuromorphic computing includes a dynamic spiking neural network, which is mapped from a three-dimensional spatiotemporal knowledge graph. It is used to receive spiking sequences, perform spiking propagation, conduct online learning and weight adjustment based on time-dependent synaptic plasticity mechanisms, and output matching results through a spiking competition mechanism, which includes a stability inhibition input port.

[0198] Specifically, the neuromorphic computing core module is the "brain" of the system, undertaking the core matching and decision-making tasks. During the system initialization phase, this module reads the three-dimensional spatiotemporal knowledge graph generated by the graph construction module and constructs the corresponding dynamic spiking neural network structure in memory or dedicated hardware.

[0199] During operation, this module receives pulse sequences from the pulse coding module, driving the network's dynamic evolution. The cluster of neurons within the network updates membrane potentials based on the input pulses, transmits signals through synaptic connections, and adjusts connection weights according to the STDP mechanism. Ultimately, the optimal order-rider match is selected through a pulse competition mechanism. The matching result output by this module is no longer a complex feature vector, but a direct decision instruction. Furthermore, the stability inhibition input port, a key interface of this module, receives inhibition signals from an external evaluation module (such as the stability evaluation model in Example 5) to intervene in the competition process, thereby ensuring the robustness of the decision.

[0200] After the neuromorphic computing core module outputs the matching result, the scheduling and execution module pushes the order dispatch instruction to the corresponding rider's smart terminal (such as a mobile APP or smart helmet) via a mobile communication network (such as 4G / 5G). At this time, the instruction includes order details and suggested pickup routes. Simultaneously, this module is also responsible for closed-loop feedback: first, it monitors the rider's response status (order acceptance, order rejection, or no response after timeout) and subsequent fulfillment status (pickup, delivery). Then, after the execution result is collected, it is converted into a feedback pulse signal and sent back to the neuromorphic computing core module. For example, if the rider successfully delivers, a positive feedback pulse is generated, strengthening the weight of the relevant synaptic connections; if a timeout occurs, a negative feedback pulse is generated. This feedback mechanism allows the system to continuously adjust its decision-making model based on actual execution results, achieving adaptive evolution.

[0201] As a preferred implementation, the neuromorphic computing core module is integrated onto a dedicated neuromorphic computing chip. This chip employs an event-driven architecture to process in parallel the membrane potential updates and pulse firing events of all neuronal groups in the dynamic spiking neural network. Specifically, the dedicated neuromorphic computing chip (such as one based on an in-memory computing architecture) is optimized for the sparsity and event-driven characteristics of spiking neural networks. Under this event-driven architecture, the internal computing units are only activated upon receiving a pulse event; neurons in a quiescent state do not consume computing resources or power. This architecture is highly compatible with the operating mechanism of dynamic spiking neural networks.

[0202] Example 8: To verify the performance of the user order intelligent allocation method and system provided by this invention in real-world complex scenarios, this embodiment selects two highly challenging typical scenarios in the instant delivery business and provides a detailed explanation of the specific operation process and beneficial effects of the technical solution.

[0203] Scenario 1: A surge in orders during peak sales season.

[0204] During a major annual promotional event, order volume in the target city surged to five times its off-peak level within a short period. Under the existing technical architecture, the computational complexity of traditional combinatorial optimization algorithms (such as the Hungarian algorithm or the KM algorithm) increases cubically with the number of orders N. When the number of orders and riders reached thousands, the computation time for a single round of scheduling jumped dramatically from 80ms during off-peak hours to over 500ms, exceeding the system's 200ms service level agreement threshold. This forced the system to trigger a degradation protection mechanism, reverting to simple greedy matching or random order assignment, severely damaging delivery efficiency and user experience.

[0205] When applying the system of this embodiment, the neuromorphic computing core module is integrated on a dedicated neuromorphic computing chip, employing an event-driven architecture. When massive order pulse sequences are input into the dynamic spiking neural network, the system does not perform traversal calculations on all "order-rider" pairs, but only performs membrane potential integration and pulse firing determination on the cluster of active neurons that receive the pulse signal. Although the total number of network nodes is huge, only a small portion of the neurons are truly in an active competitive state during any scheduling cycle. Therefore, the computational load of the system is linearly related to the number of active pulses, rather than cubically related to the total number of nodes. This ensures that even with a surge in order volume, the single-round pulse propagation convergence time of the system remains stably within a certain controllable threshold, without triggering any degradation mechanism. Thus, by leveraging the event-driven characteristics of SNNs, the "computational explosion" problem in high-concurrency scenarios is effectively solved, ensuring the stable execution of multi-objective optimization strategies during peak periods and avoiding capacity efficiency losses due to system degradation.

[0206] Scenario 2: Fulfillment collapse defense scenario under extreme rainstorm weather.

[0207] During sudden heavy rainstorms, road congestion indices rise, and rider speeds decrease. Current technology, employing a fixed-weight model trained on historical data, still prioritizes "user timeliness" above all else. This leads the system to tend to assign long-distance orders to riders closest to merchants, attempting to shorten pickup time. However, when a rider is occupied by long-distance orders for an extended period, nearby orders pile up due to lack of response, leading to user requests for cancellations and creating a negative spiral of long-distance orders consuming capacity, nearby orders going unanswered, and riders running empty, ultimately causing regional fulfillment collapse.

[0208] When applying the method of this embodiment, the dynamic spiking neural network achieves adaptive adjustment through the time-dependent synaptic plasticity mechanism (STDP). In the early stages of a rainstorm, the system may still issue pulses according to the existing weights, but when the matching of far-distance orders and near-distance riders frequently leads to timeouts (i.e., the fulfillment result pulse is negative feedback), the STDP learner automatically reduces the synaptic weight of such connections based on the time difference between the preceding and following synaptic pulses.

[0209] Meanwhile, the "congestion area nodes" in the 3D spatiotemporal knowledge graph are equipped with a high congestion index attribute, which is injected into the network as a high-impedance environmental characteristic modulation signal, further suppressing the intensity of pulses passing through the area. After a short period of online learning (approximately a few minutes), the weight distribution within the network changes: the weight of long-distance connections is significantly suppressed, while the weight of connections involving nearby orders, neighboring riders, and rider collaboration is enhanced due to high fulfillment success rates. This allows the system's decision-making preference to automatically switch from time-sensitive priority to regional liquidity priority, prioritizing the order processing capacity of local areas.

[0210] The above are merely specific embodiments of the present invention, but the scope of protection of the present invention is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the technical scope disclosed in the present invention should be included within the scope of protection of the present invention. Therefore, the scope of protection of the present invention should be determined by the scope of the claims.

Claims

1. A method for intelligent allocation of user orders, characterized in that, include: Geographic environment data, weather data, and historical order data are acquired to construct a three-dimensional spatiotemporal knowledge graph, which includes a spatial topology layer, an environmental feature layer, and a semantic association feature layer. The three-dimensional spatiotemporal knowledge graph is mapped to a dynamic spiking neural network, wherein the spatial topology layer is mapped to the topological connection structure of the neuron cluster, the dynamic environmental attributes in the environmental feature layer are mapped to the input modulation signal of the neuron cluster, and the entity association relationship in the semantic association layer is mapped to the synaptic connection between the neuron cluster. Real-time detection data is collected to generate a pulse sequence, and the pulse sequence is input into the dynamic spiking neural network for pulse propagation. The path of pulse propagation is determined by the topological connection structure of the neuron cluster and the current weight of the synaptic connection. The weights of the synaptic connections are updated based on a time-dependent synaptic plasticity mechanism, which dynamically adjusts the weight values ​​according to the time difference between presynaptic and postsynaptic pulses and the rider's real-time physiological state data. The pulse competition mechanism is activated to match and arbitrate orders with riders, obtain the matching and arbitration results, and execute order allocation.

2. The intelligent user order allocation method according to claim 1, characterized in that, The process of constructing a three-dimensional spatiotemporal knowledge graph specifically includes: Geographic information system data of the target city is acquired, the urban space is divided into micro-units, and each micro-unit is assigned three-dimensional spatial coordinates including longitude, latitude and floor height or altitude to form a spatial topology layer. Collect multi-source environmental data and align it with the micro-unit in time and space to generate an environmental feature layer containing spatial location and dynamic environmental attributes. The multi-source environmental data includes real-time weather data, road congestion index, historical delivery success rate and delivery point environmental characteristics. Collect and integrate historical order data, rider trajectory data, and user behavior data, and extract entity nodes and relational edges. The entity nodes include business district nodes, road segment nodes, merchant nodes, rider nodes, user nodes, delivery point nodes, weather area nodes, and congestion area nodes. The relational edges include spatial relationships that represent spatial adjacency or connectivity, temporal relationships that represent the temporal evolution of historical data, and semantic relationships that represent the association between delivery point type and delivery difficulty. The entity nodes are attached to the micro-unit, the relation edges are used as topological paths connecting different entity nodes, and the environmental features in the environmental feature layer, such as the merchant's meal preparation timeliness and the regional historical congestion index, are attached to the corresponding entity nodes in the form of attribute vectors to generate a three-dimensional spatiotemporal knowledge graph.

3. The intelligent user order allocation method according to claim 2, characterized in that, The physical nodes include business district nodes, road segment nodes, merchant nodes, rider nodes, user nodes, delivery point nodes, weather area nodes, and congestion area nodes; The node-based attributes of the entity nodes include environmental characteristics of the delivery area, the merchant's food preparation timeliness, and the region's historical congestion index. The relation edges include spatial relations that represent spatial adjacency or connectivity, temporal relations that represent the temporal evolution of historical data, and semantic relations that represent the association between delivery point type and delivery difficulty.

4. The intelligent user order allocation method according to claim 1, characterized in that, The mapping construction process of the dynamic spiking neural network specifically includes: The entity nodes in the three-dimensional spatiotemporal knowledge graph are mapped to neuron clusters, and the neuron clusters include an input integrator, a membrane potential register, and a pulse generator. The relation edges in the three-dimensional spatiotemporal knowledge graph are mapped to synaptic connection arrays connecting two neuron clusters. The weights of the synaptic connection arrays are initialized. The synaptic connection arrays contain multiple parallel synaptic channels, and each synaptic channel includes a weight register and a time-dependent synaptic plasticity learner. A pulse routing protocol is constructed, which includes configuring a pulse routing table for each neuron cluster and configuring a pulse gating unit for each synaptic connection array. The pulse routing table stores the physical address mapping of adjacent neuron clusters, and the pulse gating unit selects or blocks the synaptic channel according to the membrane potential state of the presynaptic neuron cluster and the external modulation signal. An asynchronous pulse calculation engine is constructed, which includes an event queue, a parallel computing unit array, and a distributed clock synchronizer. The event queue is used to receive and sort pulse events, the parallel computing unit array is used to perform membrane potential integration and pulse firing determination in parallel, and the distributed clock synchronizer is used to maintain a global logical timestamp.

5. The intelligent user order allocation method according to claim 4, characterized in that, The process of initializing the weights of the synaptic connection array specifically includes: Based on the reciprocal of the spatial distance corresponding to the relation edge, the historical cooperation success rate, and the semantic association strength, the initial weight value is calculated and stored in the weight register; The time-dependent synaptic plasticity learner adjusts the weight values ​​in the weight register based on the time difference between the presynaptic pulse and the postsynaptic pulse.

6. The intelligent allocation method for user orders according to claim 1, characterized in that, The process of updating the weights of the synaptic connections based on a time-dependent synaptic plasticity mechanism specifically includes: The system acquires real-time feedback pulses and rider physiological data generated during pulse propagation, and converts the physiological data into physiological constraint pulses. The real-time feedback pulses include performance result pulses and rider status update pulses. The physiological data includes at least one of heart rate data, skin conductance response data, body temperature data, and motion acceleration data. A time-dependent synaptic plasticity learner is constructed, comprising a presynaptic time recorder, a postsynaptic time recorder, a weight adjustment factor calculation unit, a physiological constraint modulation unit, and a weight update execution unit. When any presynaptic neuron cluster connected by a synapse fires a pulse, the presynaptic time recorder records its firing timestamp; when a postsynaptic neuron cluster connected by the same synapse fires a pulse, the postsynaptic time recorder records its firing timestamp. The weight adjustment factor calculation unit calculates the time difference based on the issuance timestamp and calculates the basic weight adjustment factor based on the asymmetric time window function. The physiological constraint modulation unit generates modulation coefficients based on the physiological constraint pulses, which are used to reduce the weight adjustment amplitude of the corresponding synaptic connection when the physiological data indicates that the rider is in a state of fatigue or stress. The weight update execution unit updates the weight of the synaptic connection based on the basic weight adjustment factor, the modulation coefficient, and the current weight value of the synaptic connection, combined with a preset upper and lower weight limit.

7. The intelligent user order allocation method according to claim 1, characterized in that, The activation pulse competition mechanism, which matches and arbitrates orders and riders, obtains the matching and arbitration results, and executes order allocation, specifically includes: The current weight values ​​of synaptic connections and the current membrane potential values ​​of each neuron cluster are obtained, and a pulse competition array is constructed. The pulse competition array includes an order node competition pool, a rider node competition pool, a competition intensity calculation unit, a stability inhibition input port, and a matching arbitration unit. Within the pulse competition array, the membrane potential value of the order node neuron cluster is stored in the order competition pool as the order competition intensity, and the membrane potential value of each rider node neuron cluster is stored in the rider competition pool as the rider competition intensity. Iterate through all synaptic connections between order nodes and rider nodes, and calculate the overall competition intensity of each order and rider pair based on the current weight value of the synaptic connection, the initial competition intensity of the order, and the initial competition intensity of the rider, and generate a competition matrix. The stability suppression input port is used to receive suppression signals and reduce the competition intensity of the corresponding order and rider connection. The suppression signal is generated based on the future stability index of order and rider matching. An iterative competition suppression strategy is used to match and arbitrate the competition matrix, and the matching and arbitration result is sent to the scheduling and execution module. The scheduling and execution module then sends a dispatch instruction to the corresponding rider terminal.

8. The intelligent user order allocation method according to claim 1, characterized in that, Before executing the order allocation, the process also includes obtaining the pulse propagation path and pulse intensity in the pulse propagation. The pulse propagation path is a sequence of intermediate nodes that the pulse passes through when it propagates from the order node neuron cluster to the rider node neuron cluster. The pulse intensity is a coefficient of amplitude attenuation that the pulse passes along the propagation path. The sequence of intermediate nodes in the pulse propagation path is encoded into a path feature vector, which includes spatial resistance features, time delay features, and semantic matching features. Based on the pulse intensity and the attenuation process along the pulse propagation path, the path composite impedance value is calculated, and the path composite impedance value is negatively correlated with the pulse intensity. Based on the path feature vector and the path comprehensive impedance value, the future stability index of each candidate order and rider matching is calculated through a stability evaluation model. The future stability index includes the timeout probability and the rider load change trend. The future stability index is compared with the stability threshold to generate a suppression signal, which is used to reduce the activation intensity of the corresponding order and rider connection in the pulse competition mechanism.

9. A user order intelligent allocation system, characterized in that, The user order intelligent allocation method according to any one of claims 1-8 is provided, wherein the user order intelligent allocation system specifically includes: a graph construction module, a pulse coding module, a brain-like computing core module, and a scheduling execution module; The graph construction module is used to construct a three-dimensional spatiotemporal knowledge graph, which includes a spatial topology layer, an environmental feature layer, and a semantic association layer. The pulse coding module is used to encode real-time data into a pulse sequence; The neuromorphic computing core module includes a dynamic spiking neural network, which is mapped from the three-dimensional spatiotemporal knowledge graph. It is used to receive the spiking sequence, perform spiking propagation, perform online learning and weight adjustment based on the time-dependent synaptic plasticity mechanism, and output matching results through a spiking competition mechanism, wherein the spiking competition mechanism includes a stability inhibition input port. The scheduling execution module is used to send a dispatch instruction to the rider terminal according to the matching result, and collect the dispatch execution result as a feedback pulse to return to the brain-like computing core module.

10. A user order intelligent allocation system according to claim 9, characterized in that, The neuromorphic computing core module is integrated on a dedicated neuromorphic computing chip, which adopts an event-driven architecture to process in parallel the membrane potential update and pulse firing events of all neuron groups in the dynamic spiking neural network.

Images