Real-time task distribution system of Internet of Things based on AI scheduling

Abstract

The invention discloses an Internet of Things real-time task distribution system based on AI scheduling, and the system comprises a data collection module which is used for generating a task input vector, and constructing a node resource vector set; the semantic modeling module is used for inputting the task input vector into an improved StructBERT model to construct a task semantic vector; the intention analysis module is used for constructing a historical task intention vector library and a semantic conflict graph; the comparison judgment module is used for executing task conflict judgment and scheduling priority adjustment operation according to the task aggregation degree and the semantic difference vector; and the scheduling execution module is used for executing a scheduling calculation operation, generating a task allocation result and updating the historical task intention vector library and the semantic conflict graph. The method has the advantages of high task semantic understanding precision, high conflict judgment efficiency and high scheduling decision intelligent degree, and is suitable for a real-time task allocation scene in a multi-node heterogeneous Internet of Things environment.

Description

Technical Field

[0001] This invention relates to the field of IoT task management and intelligent scheduling technology, and in particular to an IoT real-time task allocation system based on AI scheduling. Background Technology

[0002] With the large-scale deployment of IoT terminal devices and the development of edge intelligence technology, the demand for real-time task scheduling in complex environments continues to grow. How to achieve efficient and stable task allocation under multi-node, multi-task concurrency has become an important research direction for IoT management systems. Existing IoT scheduling schemes mostly employ static rule matching or preset strategies based on resource load thresholds for task assignment, but they generally suffer from the following problems in dynamic scenarios:

[0003] Task requests are typically uploaded in the form of natural language or sensor events, with complex semantic expressions, unclear objectives, and ambiguous boundaries, making it difficult for traditional structured mapping methods to accurately extract task intent and execution conditions. There are multi-dimensional semantic relationships between tasks, such as overlapping objectives, conflicting conditions, and priority interference. Existing solutions often ignore the historical scheduling impact and semantic conflict accumulation between tasks, making it impossible to achieve dynamic identification and intelligent conflict resolution. Node resource status changes frequently, and the collection latency and evaluation methods are inconsistent. Traditional resource adaptation mechanisms cannot reflect the actual carrying capacity of nodes in real time, leading to scheduling decision mismatch. There is a lack of high-quality task feedback loops after scheduling, making it impossible to build a behavioral data foundation for continuous optimization of scheduling strategies.

[0004] Therefore, how to provide an AI-based real-time task allocation system for the Internet of Things is a problem that urgently needs to be solved by those skilled in the art. Summary of the Invention

[0005] One objective of this invention is to propose an AI-based real-time task allocation system for the Internet of Things (IoT). This invention constructs scheduling input information based on task input vectors and node resource vector sets. It then combines an improved StructBERT model incorporating task slot construction and divergence reinforcement structures to perform semantic structure modeling, generating task semantic vectors. Furthermore, it constructs a historical task intent vector library and a semantic conflict graph. By comparing the task semantic vectors with the historical task intent vector library, it calculates task aggregation degree and semantic difference vectors. Based on the aggregation degree range, it determines the task conflict state and scheduling priority. Finally, based on the task semantic vectors, scheduling priority values, and node resource vector sets, it performs scheduling calculations, outputting task allocation results and task execution instructions. It records task feedback and dynamically updates the historical task intent vector library and semantic conflict graph, achieving closed-loop modeling of tasks, resources, and feedback. This invention fully integrates task semantic modeling, historical scheduling behavior analysis, and node resource status evaluation mechanisms, detailing an intelligent task allocation process that dynamically optimizes semantic accuracy and scheduling priority. It possesses advantages such as refined semantic modeling, accurate conflict analysis, efficient scheduling decisions, and scalable feedback loops.

[0006] According to an embodiment of the present invention, an IoT real-time task allocation system based on AI scheduling includes: The data acquisition module is used to collect task request data uploaded by IoT terminals, perform format normalization processing, and generate task input vectors; it also collects the operating resource data of IoT nodes and constructs a set of node resource vectors. The semantic modeling module is used to input the task input vector into the improved StructBERT model, perform semantic structure modeling operations, and construct task semantic vectors; the improved StructBERT model introduces task slot construction structure and branching reinforcement structure; The intent analysis module is used to construct a historical task intent vector library and a semantic conflict graph. The historical task intent vector library stores task semantic vectors and scheduling execution result data. The semantic conflict graph represents the target overlap relationship, execution condition conflict relationship, and priority interference relationship between tasks. The comparison and judgment module is used to compare the task semantic vector with the historical task intent vector library, calculate the semantic structure similarity, target overlap and condition conflict value, and generate task aggregation degree and semantic difference vector based on the comparison results; perform task conflict judgment and scheduling priority adjustment operations according to the task aggregation degree and semantic difference vector; if the task aggregation degree is less than the first preset threshold, semantic clarification data is generated; if the task aggregation degree is between the first preset threshold and the second preset threshold, the scheduling priority value is adjusted; if the task aggregation degree is greater than the second preset threshold, it is marked as a task that can be directly scheduled. The scheduling and execution module is used to perform scheduling calculation operations based on task semantic vectors, scheduling priority values ​​and node resource vector sets, generate task allocation results, send task execution instructions to target nodes, record task execution feedback data, and update the historical task intent vector library and semantic conflict graph.

[0007] Optionally, modules can be integrated using the following methods: S1. Collect task request data uploaded by IoT terminals, perform format normalization processing, and generate task input vectors; S2. Collect the operational resource data of IoT nodes and construct a set of node resource vectors; S3. Input the task input vector into the improved StructBERT model, perform semantic structure modeling operations, and construct the task semantic vector; the improved StructBERT model introduces a task slot construction structure and a branching reinforcement structure; S4. Construct a historical task intent vector library and a semantic conflict graph; the historical task intent vector library stores task semantic vectors and scheduling execution result data; the semantic conflict graph represents the target overlap relationship, execution condition conflict relationship and priority interference relationship between tasks. S5. Perform a comparison operation between the task semantic vector and the historical task intent vector library, calculate the semantic structure similarity, target overlap and condition conflict value, and generate task aggregation degree and semantic difference vector based on the comparison results. S6. Perform task conflict determination and scheduling priority adjustment operations based on task aggregation degree and semantic difference vector; if the task aggregation degree is less than the first preset threshold, generate semantic clarification data; if the task aggregation degree is between the first preset threshold and the second preset threshold, adjust the scheduling priority value; if the task aggregation degree is greater than the second preset threshold, mark it as a task that can be directly scheduled. S7. Based on the task semantic vector, scheduling priority value and node resource vector set, perform scheduling calculation operation, generate task allocation result, send task execution instruction to target node, record task execution feedback data, and update historical task intent vector library and semantic conflict graph.

[0008] Optionally, S1 specifically includes: Collect task request data uploaded by IoT terminals; the task request data includes text command data, speech recognition and transcription data, sensor event data, and image trigger data. The text command data and speech recognition-written data are processed in a standardized manner to unify them into a structured language expression format; Sensor event data and image trigger data are processed to extract event labels, generating event representation information and trigger context information; Vector encoding is performed on the text expression format and time tag information to construct a task expression vector with a unified dimension; The task representation vector is normalized and padded with timestamps to generate a task input vector; the task input vector includes action semantics, object semantics and conditional semantics.

[0009] Optionally, S2 specifically includes: Collect operational resource data of IoT nodes; the operational resource data includes computing power data, response latency data, network connection status data, and operational load status data; The computing power data is standardized to generate node computing performance indicators; the response latency data is processed by sliding time window statistical processing to generate node average response time and fluctuation amplitude indicators; the network connection status data is processed by status labeling to generate node connection availability indicators; the running load status data is processed by interval quantization to generate node load level indicators; and the node computing performance indicators, average response time, connection availability indicators and load level indicators are concatenated into vectors to construct node resource vectors. Perform dimension normalization on all node resource vectors to generate a set of node resource vectors.

[0010] Optionally, S3 specifically includes: The task input vector is fed into the improved StructBERT model in the order of action semantics, object semantics, and conditional semantics. Structural position labeling is performed on the action semantics, object semantics, and conditional semantics of the task input vector to form a structural position labeling sequence; Establish a correspondence between the structure location marker sequence and the task input vector to construct a structured task representation; Perform context association modeling on the structured task representation to generate a sequence of context semantic representations; Structural boundary recognition processing is performed on the context semantic representation sequence to identify action boundaries, object boundaries, and condition boundaries, forming a slot label sequence; The slot label sequence is processed by vector concatenation according to the arrangement of action slot boundaries, object slot boundaries and condition slot boundaries to generate task semantic vector; The task input vector is compared with the preset comparison samples to calculate the difference, generate a divergence score, and perform reinforcement training based on the divergence score during the training process.

[0011] Optionally, the improved StructBERT model introduces a task slot construction structure and a branching reinforcement structure, specifically: The task slot construction structure performs semantic component segmentation on the task input vector, constructing slot placeholder information from action semantics, object semantics, and condition semantics respectively; it then performs concatenation processing on the slot placeholder information and structural position embedding information to generate a structured slot embedding sequence; it performs context association modeling on the structured slot embedding sequence to generate a slot context vector; it performs boundary recognition processing on the slot context vector to output action slot boundaries, object slot boundaries, and condition slot boundaries; and it constructs a slot structure representation from the action slot boundaries, object slot boundaries, and condition slot boundaries according to a set concatenation order and performs fusion processing with the original context representation to generate a slot enhanced expression vector. The divergence enhancement structure performs difference-sensitive feature extraction on the slot enhancement expression vector to generate a difference initial vector; it then performs a contrast construction process on the difference initial vector and the original embedding sequence of the task input vector to generate a difference contrast representation; it performs a dual-channel attention distribution calculation operation on the difference contrast representation to generate a difference attention vector; it performs weight enhancement processing on the difference attention vector to generate a difference enhancement representation; it performs a fusion process with the difference enhancement representation and the slot enhancement expression vector to generate a semantic divergence representation vector; it performs a divergence degree calculation operation on the semantic divergence representation vector to generate divergence score information, and updates the parameter update path in the semantic structure modeling process based on the divergence score information to output the task semantic vector.

[0012] Optionally, S4 specifically includes: Collect task semantic vectors and scheduling execution result data, and establish task semantic record entries according to task identifiers; Perform vector storage processing on the task semantic record entries, and write the task semantic vector and scheduling execution result data into the historical task intent vector library in sequence; Perform vector validation processing on the historical task intent vector library and establish an index structure based on timestamp information, task identification information and semantic vector dimension information; Semantic similarity calculation is performed on all task semantic vectors to construct a set of semantic structure relationships; task target extraction is performed on the scheduling execution result data to generate a set of task target labels; target overlap is calculated on the task target label set to construct task target overlap relationships; condition conflict analysis is performed on the conditional semantics in the scheduling execution result data and task input vectors to construct execution condition conflict relationships; interference is calculated on the priority records in the scheduling execution result data to construct priority interference relationships. By establishing a set of graph nodes and a set of graph edges according to the identification order of task pairs, the overlapping relationships of task objectives, conflicting relationships of execution conditions, and interference relationships are constructed, and a semantic conflict graph is built. The semantic conflict graph is subjected to structural consistency verification. Graph edges that do not meet the structural constraints are removed to generate a verified semantic conflict graph.

[0013] Optionally, S5 specifically includes: Perform a comparison process between the task semantic vector and all semantic vectors in the historical task intent vector library, and construct a comparison sequence according to the vector dimension order; Semantic structure similarity calculation is performed on the control sequence, and semantic structure similarity values ​​are generated based on the correspondence between action slot boundaries, object slot boundaries and condition slot boundaries; The target label extraction process is performed on the scheduling execution result data in the historical task intent vector library to generate a target label set, and then the label is compared with the target-related components in the task semantic vector to generate a target overlap value. Perform condition conflict analysis and processing on the condition records in the scheduling execution result data and the condition slot information in the task semantic vector to generate condition conflict values; The semantic structure similarity value, target overlap value, and condition conflict value are jointly measured according to the set weight order to generate an association measurement sequence. Perform aggregation calculations on the associated metric sequences to generate the task aggregation degree; Perform difference vector construction processing on semantic structure similarity value, target overlap value and condition conflict value, and generate semantic difference vectors by arranging the difference components in the order of action difference, object difference and condition difference.

[0014] Optionally, S6 specifically includes: Perform threshold comparison processing on the task aggregation degree and construct a task aggregation degree comparison sequence; The task aggregation degree values ​​in the task aggregation degree comparison sequence are compared with a first preset threshold to determine their size relationship, generating a first determination result; the task aggregation degree values ​​in the task aggregation degree comparison sequence are compared with a second preset threshold to determine their size relationship, generating a second determination result; joint state analysis processing is performed on the first determination result and the second determination result to generate state labeling information; If the task aggregation degree corresponding to the state label information is less than the first preset threshold, then perform difference extraction processing on the semantic difference vector to extract action difference components, object difference components and condition difference components, construct a clarification expression sequence, and perform format construction processing on the clarification expression sequence to generate semantic clarification data; If the task aggregation degree corresponding to the status flag information is between the first preset threshold and the second preset threshold, then priority adjustment processing is performed on the scheduling priority value. An adjustment reference sequence is generated according to the order of action difference components, object difference components and condition difference components. Weight mapping processing is performed on the adjustment reference sequence to generate priority increase and decrease information. Numerical synthesis operation is performed on the priority increase and decrease information to generate the updated scheduling priority value. If the task aggregation degree corresponding to the status flag information is greater than the second preset threshold, then the task identifier is processed for schedulable marking, and a directly schedulable marking record is generated.

[0015] Optionally, S7 specifically includes: Collect task semantic vectors, scheduling priority values, and node resource vector sets to construct a set of inputs for scheduling calculations; Based on the scheduling calculation input set, a scheduling calculation operation is performed to generate a task allocation result. The scheduling calculation operation includes resource adaptability calculation, semantic matching degree calculation and priority sorting operation. Based on the task allocation results, a task execution instruction is generated and sent to the corresponding target node; Collect task execution feedback data from the target node, including task completion status, resource usage changes, and execution time information; Based on the task semantic vector and the task execution feedback data, perform a historical task intent vector library update operation and store the updated vector in the historical task intent vector library; The semantic distribution relationships in the historical task intent vector library are analyzed to identify conflict relationships between tasks and construct a semantic conflict graph.

[0016] The beneficial effects of this invention are: This invention constructs a structural system comprising a data acquisition module, a semantic modeling module, an intent analysis module, a comparison and judgment module, and a scheduling execution module. It combines task input vectors and node resource vector sets to construct a scheduling computation input set. An improved StructBERT model with task slot construction and divergence reinforcement structures is introduced to process the position marking and context modeling of action semantics, object semantics, and conditional semantics, thereby improving the slot boundary recognition accuracy and divergence expression capability of task semantic vectors. During task scheduling, a task aggregation degree and semantic difference vector are generated through a joint measurement of semantic structure similarity, target overlap, and conditional conflict value. This invention performs conflict determination and scheduling priority adjustment operations based on task aggregation degree. During task allocation, it integrates resource adaptability calculation, semantic matching degree calculation, and priority sorting operations to generate task allocation results and task execution instructions. It also collects task execution feedback data and updates the historical task intent vector library and semantic conflict map to achieve closed-loop evolution of semantic scheduling. This invention is designed for real-time task allocation scenarios in the Internet of Things. While maintaining semantic integrity, it effectively avoids multi-task conflict interference, improves scheduling matching accuracy, strengthens the consistency of task expression and dynamic priority adjustment capabilities, and has the beneficial effects of accurate semantic recognition, refined conflict determination, efficient scheduling calculation, and closed-loop feedback modeling. Attached Figure Description

[0017] The accompanying drawings are provided to further illustrate the invention and form part of the specification. They are used in conjunction with embodiments of the invention to explain the invention and do not constitute a limitation thereof. In the drawings:

[0018] Figure 1 This is a schematic diagram of an IoT real-time task allocation system based on AI scheduling proposed in this invention; Figure 2 This is a flowchart of an IoT real-time task allocation system based on AI scheduling proposed in this invention; Figure 3 This is a schematic diagram of the structure of the improved StructBERT model proposed in this invention. Detailed Implementation

[0019] The present invention will now be described in further detail with reference to the accompanying drawings. These drawings are simplified schematic diagrams, illustrating only the basic structure of the invention, and therefore only show the components relevant to the invention.

[0020] refer to Figure 1 An AI-based real-time task allocation system for the Internet of Things (IoT) includes: The data acquisition module is used to collect task request data uploaded by IoT terminals, perform format normalization processing, and generate task input vectors; it also collects the operating resource data of IoT nodes and constructs a set of node resource vectors. The semantic modeling module is used to input the task input vector into the improved StructBERT model, perform semantic structure modeling operations, and construct task semantic vectors; the improved StructBERT model introduces task slot construction structure and branching reinforcement structure; The intent analysis module is used to construct a historical task intent vector library and a semantic conflict graph. The historical task intent vector library stores task semantic vectors and scheduling execution result data. The semantic conflict graph represents the target overlap relationship, execution condition conflict relationship, and priority interference relationship between tasks. The comparison and judgment module is used to compare the task semantic vector with the historical task intent vector library, calculate the semantic structure similarity, target overlap and condition conflict value, and generate task aggregation degree and semantic difference vector based on the comparison results; perform task conflict judgment and scheduling priority adjustment operations according to the task aggregation degree and semantic difference vector; if the task aggregation degree is less than the first preset threshold, semantic clarification data is generated; if the task aggregation degree is between the first preset threshold and the second preset threshold, the scheduling priority value is adjusted; if the task aggregation degree is greater than the second preset threshold, it is marked as a task that can be directly scheduled. The scheduling and execution module is used to perform scheduling calculation operations based on task semantic vectors, scheduling priority values ​​and node resource vector sets, generate task allocation results, send task execution instructions to target nodes, record task execution feedback data, and update the historical task intent vector library and semantic conflict graph.

[0021] In this embodiment, the data acquisition module collects task request data uploaded by IoT terminals and performs format normalization processing to generate a structurally unified task input vector, thereby improving the input consistency of the subsequent semantic modeling process. Simultaneously, it collects operational resource data from IoT nodes, constructing a multi-dimensional node resource vector set to provide comprehensive resource status support for scheduling computation. The semantic modeling module inputs the task input vector into an improved StructBERT model that incorporates a task slot construction structure and a divergence enhancement structure, performing semantic structure modeling operations to construct a task semantic vector with slot boundary information and difference enhancement capabilities, thereby enhancing the accuracy and distinguishability of semantic representation. The intent analysis module constructs a historical task intent vector library and a semantic conflict graph, recording task semantic vectors and corresponding scheduling execution result data, establishing a semantic association structure between task objectives, execution conditions, and priorities, further supporting task conflict identification and scheduling reference. The comparison and judgment module performs a multi-dimensional comparison operation between the task semantic vector and the historical task intent vector library, calculating semantic structure similarity, objective overlap, and condition conflict values, generating task aggregation degree and semantic difference vectors, and determining task scheduling feasibility and priority adjustment strategies based on the aggregation degree, thereby improving the intelligence and robustness of scheduling allocation. The scheduling and execution module performs scheduling calculations based on task semantic vectors, scheduling priority values, and node resource vector sets, generates task allocation results, sends task execution instructions to target nodes, records task execution feedback data, and further updates the historical task intent vector library and semantic conflict graph. This achieves dynamic closed-loop updates of task scheduling and system semantic knowledge, thereby improving the real-time performance, accuracy, and adaptability of IoT task scheduling.

[0022] refer to Figure 2-3 In this embodiment, the modules are connected through the following method: S1. Collect task request data uploaded by IoT terminals, perform format normalization processing, and generate task input vectors; S2. Collect the operational resource data of IoT nodes and construct a set of node resource vectors; S3. Input the task input vector into the improved StructBERT model, perform semantic structure modeling operations, and construct the task semantic vector; the improved StructBERT model introduces a task slot construction structure and a branching reinforcement structure; S4. Construct a historical task intent vector library and a semantic conflict graph; the historical task intent vector library stores task semantic vectors and scheduling execution result data; the semantic conflict graph represents the target overlap relationship, execution condition conflict relationship and priority interference relationship between tasks. S5. Perform a comparison operation between the task semantic vector and the historical task intent vector library, calculate the semantic structure similarity, target overlap and condition conflict value, and generate task aggregation degree and semantic difference vector based on the comparison results. S6. Perform task conflict determination and scheduling priority adjustment operations based on task aggregation degree and semantic difference vector; if the task aggregation degree is less than the first preset threshold, generate semantic clarification data; if the task aggregation degree is between the first preset threshold and the second preset threshold, adjust the scheduling priority value; if the task aggregation degree is greater than the second preset threshold, mark it as a task that can be directly scheduled. S7. Based on the task semantic vector, scheduling priority value and node resource vector set, perform scheduling calculation operation, generate task allocation result, send task execution instruction to target node, record task execution feedback data, and update historical task intent vector library and semantic conflict graph.

[0023] In this embodiment, S1 specifically refers to: Collect task request data uploaded by IoT terminals; the task request data includes text command data, speech recognition and transcription data, sensor event data, and image trigger data. The text command data and speech recognition-written data are processed in a standardized manner to unify them into a structured language expression format; Sensor event data and image trigger data are processed to extract event labels, generating event representation information and trigger context information; Vector encoding is performed on the text expression format and time tag information to construct a task expression vector with a unified dimension; The task representation vector is normalized and padded with timestamps to generate a task input vector; the task input vector includes action semantics, object semantics and conditional semantics.

[0024] In this embodiment, S2 specifically refers to: The system collects operational resource data of IoT nodes, including computing power data, response latency data, network connection status data, and operational load status data. For computing power data, it performs structural standardization processing, constructing a linear mapping function based on preset maximum and minimum computing power to map the computing power values ​​of each node to the [0,1] interval, generating node computing performance indicators. For response latency data, it performs sliding time window statistical processing, setting a fixed window length and step size, extracting the average response latency and standard deviation of each window from the continuous time series, generating node average response time indicators and response fluctuation amplitude indicators respectively. For network connection status data, it performs status labeling processing, dividing network connection status values ​​into three states: stable connection, intermittent connection, and interrupted connection, assigning fixed scalar values ​​to each, generating node connection availability indicators. For operational load status data, it performs interval quantization processing, dividing node CPU and memory utilization into several load intervals, mapping each interval to three levels: low load, medium load, and high load, each corresponding to a discrete numerical label, generating node load level indicators. The node computing performance metrics, average response time metrics, response fluctuation metrics, connection availability metrics, and load level metrics are concatenated into vectors in a unified order to construct a five-dimensional resource vector for a single node. All node resource vectors are then combined sequentially according to their node identifiers to form an unnormalized set of node resource vectors. Each dimension of this set undergoes max-min normalization, using the minimum and maximum values ​​to construct a linear normalization function, ensuring that the values ​​of each dimension are distributed within the [0, 1] interval, thus generating a normalized set of node resource vectors. This set of node resource vectors serves as the input feature for subsequent scheduling priority calculations and task allocation weight adjustments, improving the matching degree of scheduled task execution and the level of resource load balancing.

[0025] In this embodiment, S3 specifically refers to: The task input vector is fed into the improved StructBERT model in the order of action semantics, object semantics, and conditional semantics. Structural position labeling is performed on the action, object, and conditional semantics of the task input vector, constructing a structural position label sequence by assigning fixed position labels to different semantic components. A one-to-one correspondence is established between the structural position label sequence and the task input vector, mapping action semantic segments to action position labels, object semantic segments to object position labels, and conditional semantic segments to conditional position labels, thus constructing a structured task representation. Context association modeling is then performed on the structured task representation, establishing cross-segment dependencies between action semantics, object semantics, and conditional semantics through a multi-head semantic association mechanism to generate a contextual semantic representation. The sequence is processed by performing structural boundary recognition on the context semantic representation sequence. Action boundaries, object boundaries, and condition boundaries are identified through structural boundary judgment rules to form a slot label sequence. The slot label sequence is then processed by vector concatenation according to the arrangement of action slot boundaries, object slot boundaries, and condition slot boundaries. Slot-level representations are constructed by continuously concatenating semantic units within the same slot to generate task semantic vectors. The task input vector is compared with preset comparison samples by performing difference calculation. A divergence score is generated by calculating the degree of deviation between the two in the action semantic dimension, object semantic dimension, and condition semantic dimension. During training, reinforcement training is performed based on the divergence score to enhance the model's sensitivity to fine-grained differences in task semantics by increasing the weight of high-divergence samples.

[0026] This implementation method enables action semantics, object semantics, and conditional semantics to maintain stable expression within the structural boundaries, allows for sufficient modeling of dependencies across semantic segments, and enhances the expressive power of subtle differences between different tasks, thereby improving the discriminability and stability of task semantic vectors in scheduling comparison, conflict determination, and priority adjustment.

[0027] In this embodiment, the improved StructBERT model introduces a task slot construction structure and a branching reinforcement structure, specifically: The task input vector is subjected to semantic component segmentation, where fragments corresponding to action semantics, object semantics, and condition semantics are labeled as independent semantic components to construct slot occupancy information. The slot occupancy information and structural position embedding information are then concatenated, aligning the slot category labels and position encoding vectors dimensionally and concatenating them into a unified structure to generate a structured slot embedding sequence. Context association modeling is performed on the structured slot embedding sequence, capturing the interaction relationships between different slot fragments based on a multi-layer self-attention mechanism to generate a slot context vector. Boundary recognition is performed on the slot context vector, extracting action slot boundaries, object slot boundaries, and condition slot boundaries based on continuity features and relative positional deviations. These boundaries are then concatenated according to a set concatenation order to construct a slot structure representation. Finally, the slot structure representation is fused with the original context representation, fusing semantic structure features and context semantic information through residual connections and gating mechanisms to generate an enhanced slot expression vector. The process involves extracting difference-sensitive features from the slot-enhanced representation vector, using a perturbation amplification mechanism to extract local features responsive to small semantic changes, and generating an initial difference vector. This initial difference vector is then compared to the original embedding sequence of the task input vector, with difference vectors calculated position-by-position using sequence alignment to construct a difference comparison representation. A dual-channel attention distribution calculation is performed on this comparison representation, generating difference attention vectors for both the difference information channel and the structural information channel. Weight enhancement is applied to the difference attention vector, mapping attention weights to difference significance scores and normalizing and scaling to generate a enhanced difference representation. The enhanced difference representation and the slot-enhanced representation vector are then fused, using a weighted average and channel fusion method to generate a semantic divergence representation vector. Divergence calculation is performed on the semantic divergence representation vector, measuring the distance between the semantic divergence vector and the target difference vector in the training set and outputting a divergence score. Based on the divergence score, the parameter update path in the semantic structure modeling process is updated, introducing directional weight shifts through divergence guidance signals to optimize the model training strategy and output the task semantic vector.

[0028] This implementation enables the slot structure representation of action semantics, object semantics, and conditional semantics to have boundary stability and context dependence, amplifies subtle semantic differences under the attention mechanism, and makes the training process have an explicit response to semantic bias, thereby improving the expression accuracy of task semantic vectors and the model's ability to identify discrepancies in task conflicts.

[0029] In this embodiment, S4 specifically refers to: Collect task semantic vectors and scheduling execution result data, and combine the two based on the unique task identifier to generate task semantic record entries; Perform vector storage processing on the task semantic record entries, and write the task semantic vector and scheduling execution result data into the historical task intent vector library in chronological order; Perform validation processing on the historical task intent vector library, and establish a joint index structure for each task semantic record based on timestamp information, task identifier information and semantic vector dimension information; Semantic similarity calculation is performed on all task semantic vectors in sequence according to task identifiers. The structural similarity between each pair of task semantic vectors is calculated using a similarity metric based on slot structure alignment, and a set of semantic structural relationships is constructed. The task target extraction process is performed on the scheduling execution result data to extract key operation target fields from the execution results and form a standardized target label set. Compare the label pairs in the task target label set, and use the label intersection and union ratio method to calculate the target overlap degree between each pair of task targets, and construct the task target overlap relationship; Perform condition semantic comparison processing on the execution condition field recorded in the scheduling execution result data and the condition semantic field in the task input vector, construct a semantic comparison matrix, determine whether there is an execution logic conflict based on the comparison matrix, and generate execution condition conflict relationship; Perform time series difference analysis on the priority records in the scheduling execution results data and the priority of the same node in the historical task records. Construct interference influence factors based on the priority adjustment magnitude and frequency to generate priority interference relationships. The semantic structure relationship set, the task goal overlap relationship, the execution condition conflict relationship and the priority interference relationship are constructed according to the identification order of the task pairs to construct the graph node set and the graph edge set. The task semantic vector is used as the graph node, and the goal overlap, condition conflict and priority interference relationship between tasks is used as the graph edge to construct the semantic conflict graph. Perform structural consistency verification on the semantic conflict graph, verify whether there are redundant edges, isolated edges and logically disconnected paths in the connection relationship between nodes, perform a removal operation on graph edges that do not meet the structural connection constraints, update the graph topology, and generate a verified semantic conflict graph.

[0030] In this embodiment, S5 specifically refers to: A set of vector pairs is constructed by comparing the task semantic vector with all semantic vectors in the historical task intent vector library, and arranged in a unified vector dimension order to form a task comparison sequence. For each pair of vectors in the task comparison sequence, perform semantic structure similarity calculation. Construct an alignment matrix based on the correspondence between action slot boundaries, object slot boundaries and condition slot boundaries. Count the number of matching boundary pairs and divide by the maximum boundary length to generate the semantic structure similarity value between each pair of vectors. Perform target label extraction processing on the scheduling execution result data associated in the historical task intent vector library, extract standardized target entity identifiers from the target field, and construct a target label set; Perform target entity recognition processing on the target semantic components related to the action slot in the current task semantic vector to generate the current task target entity set, compare it one by one with the historical task target label set, count the number of entity intersections and unions, and calculate the target overlap value according to the intersection-union ratio method. Perform field comparison and logical matching operations on the condition record information in the scheduling execution result data and the condition slot information in the current task semantic vector to determine whether there are resource mutual exclusion, time conflict or constraint overlap, count the number of conflicting fields and the total number of condition fields, and calculate the conflict ratio as the condition conflict value. The semantic structure similarity value, target overlap value, and condition conflict value are weighted and summed according to a preset weighting ratio to generate a joint association score for each pair of vectors, forming an association measurement sequence. Perform aggregation calculation on the correlation metric sequence, calculate the mean of all correlation scores, use it as the average degree of aggregation between the current task and historical tasks, and output the task aggregation degree. The semantic structure similarity value, target overlap value, and condition conflict value are processed by structural difference extraction. The residual values ​​of action slot difference, object slot difference, and condition slot difference are calculated. The three types of residual values ​​are concatenated into a complete semantic difference vector according to the order of action difference, object difference, and condition difference.

[0031] In this embodiment, S6 specifically refers to: Perform threshold comparison processing on the task aggregation degree and construct a task aggregation degree comparison sequence; A "less than" judgment operation is performed on the task aggregation degree values ​​in the task aggregation degree comparison sequence compared to a first preset threshold, and a first judgment result is output. A "less than" judgment operation is performed on the task aggregation degree values ​​in the task aggregation degree comparison sequence compared to a second preset threshold, and a second judgment result is output. Joint state analysis processing is performed on the first and second judgment results to generate state labeling information, which is used to distinguish between semantic clarification state, priority adjustment state, and direct scheduling state. If the task aggregation degree state corresponding to the state labeling information is less than the first preset threshold, component extraction processing is performed on the semantic difference vector to extract action difference components, object difference components, and condition difference components, respectively. The action difference components, object difference components, and condition difference components are concatenated into a clarification expression sequence. Language normalization and expression format construction processing are performed on the clarification expression sequence to generate semantic clarification data and perform structured encapsulation. If the task aggregation degree corresponding to the status marker information is between the first preset threshold and the second preset threshold, then the scheduling priority value is numerically adjusted; the action difference component, object difference component, and condition difference component are combined into an adjustment reference sequence, and each component in the adjustment reference sequence is assigned a corresponding weight parameter, and a weight mapping operation is performed to generate priority increase / decrease information; the priority increase / decrease information and the original scheduling priority value are numerically synthesized, and the updated scheduling priority value is output; if the task aggregation degree corresponding to the status marker information is greater than the second preset threshold, then the current task identifier is status-marked, a directly schedulable marker record is generated and written into the task scheduling status set.

[0032] In this embodiment, S7 specifically refers to: The system collects task semantic vectors, scheduling priority values, and node resource vector sets. A joint index is constructed based on task identifiers, priority identifiers, and node identifiers to establish a scheduling calculation input set. Semantic matching degree calculation is performed on the task semantic vectors and node resource vector sets in the scheduling calculation input set, calculating the semantic matching degree based on the correspondence between action slots, object slots, condition slots, and node resource dimensions. Resource adaptability calculation is performed on the node resource vector sets in the scheduling calculation input set, constructing a resource evaluation function based on node load level, average response time, and connection availability indicators, and outputting a resource adaptability score. Scheduling priority values ​​are sorted, generating a priority ranking sequence from high to low. The semantic matching degree, resource adaptability score, and priority ranking sequence are then combined according to a set fusion rule to perform a multi-factor scheduling score calculation, outputting the scheduling score result, and generating task allocation results based on the scheduling score result. Based on the task allocation results, task execution instructions are constructed. The task identifier, target node identifier, and scheduling score value in each task allocation record are integrated to generate a set of task execution instructions in a standard format. The set of task execution instructions is sent to the target node, and the instruction sending status is processed by an acknowledgment mechanism to generate a task issuance acknowledgment status. The system collects task execution feedback data from target nodes, classifies task completion status, and distinguishes between success and failure. It performs vector difference calculation on resource usage change information and outputs resource change vectors. It analyzes execution time distribution to generate time interval statistics. The system encapsulates task execution feedback data in a standard format, binds it to the corresponding task semantic vector, performs vector combination processing, and generates historical task intent update entries. Finally, it writes these historical task intent update entries into the historical task intent vector library and updates the corresponding task identifier index. Joint analysis and processing are performed on the task semantic vectors and task execution feedback data in the historical task intent vector library to calculate the semantic similarity and resource conflict frequency between task pairs. Conflict relationship tags are generated for task pairs with high semantic similarity and resource conflict records, and priority interference tags are generated for task pairs with priority execution delays. Conflict relationship tags and interference relationship tags are used to establish graph nodes and graph edges according to the task pair identifiers to construct a semantic conflict graph. The constructed semantic conflict graph is subjected to structural consistency verification, and graph edges that do not meet the conflict definition rules are deleted to generate an updated semantic conflict graph.

[0033] Example 1: To verify the feasibility of this invention in practice, it was applied to a virtual power plant platform project under a comprehensive energy service company. This platform aggregates various distributed resources, including industrial and commercial distributed photovoltaic power, electrochemical energy storage, electric vehicle charging and discharging stations, and adjustable air conditioning loads, with the goal of achieving aggregated scheduling of adjustable capacity within the region, market transaction response, and cost optimization.

[0034] Before implementation, the project faced several technical bottlenecks: First, the resource distribution structure was complex and the nodes had poor correlation, resulting in inefficient scheduling. Second, the traditional rule aggregation method failed to dynamically capture the collaborative potential between resources, causing a certain degree of "mismatched aggregation". Third, the resource contribution fluctuated greatly, the strategy update was lagging, and the platform's operating revenue fluctuated significantly.

[0035] To address the aforementioned issues, the resource aggregation method based on graph embedding learning proposed in this invention is deployed in the "intelligent aggregation platform" of this platform. It is mainly used for resource graph construction, graph embedding vector generation, collaborative potential assessment and aggregation strategy optimization. The underlying system integrates the graph modeling rules, embedding expression mechanism, dynamic aggregation algorithm and feedback update process disclosed in this invention.

[0036] The entire experiment lasted three months. The platform selected 180 aggregateable resource nodes to participate in the comparative test group, which was divided into a traditional system aggregation group (control group) and an invention system graph embedding aggregation group (experimental group). All resources were kept consistent in basic configuration, access protocol, and operating conditions, with only the aggregation method differing, in order to compare core indicators such as resource utilization efficiency, aggregation response capability, transaction success rate, and platform economic benefits.

[0037] During the three-month testing period, the platform regularly collected various operational data, resulting in the following table: Table 1. Comparison of Key Indicators in the Resource Aggregation Process

[0038] As shown in Table 1, the system of this invention exhibits significant advantages in several key aspects. Firstly, regarding response hit rate, the system constructs semantic and behavioral feature graphs between nodes, enabling the aggregation unit to possess stronger matching capabilities, thereby increasing the hit rate from 73.4% to 91.2%. Secondly, the average aggregation synergy index also increases from 0.61 to 0.85, indicating a more accurate match between the actual synergistic capabilities of the nodes and improving overall response efficiency. Regarding aggregation time, the system of this invention controls the time to within 4.1 seconds through graph embedding pre-computation and a fast matching mechanism, improving aggregation efficiency. The number of mismatched aggregations is significantly reduced, reflecting that graph learning methods can more accurately identify resource matching relationships, effectively solving the "resource selection error" problem in traditional rule-based methods. Furthermore, the policy update frequency is reduced by 47%, and the response bias is significantly decreased, indicating that the graph embedding model has a certain degree of adaptability, contributing to stable system operation.

[0039] In addition, to further verify the platform's overall economic benefits and transaction execution effectiveness, the system simultaneously compiled the following economic and operational indicators, summarized below: Table 2 Comparison of Platform Operation Economic Indicators

[0040] Table 2 shows that the system of this invention has brought about a significant improvement in economic benefits. Within three months, the platform's average monthly electricity market revenue increased by 42.1%, mainly due to the accurate capture of resource contribution potential by graph embedding learning, enabling the platform to effectively concentrate resources with regulation capabilities during high-price periods. Simultaneously, the load peak-shaving cost on the user side decreased by 33%, reflecting that the aggregation strategy can complete power regulation tasks at a lower cost. Furthermore, the transaction success rate increased from 68.7% to 85.3%, significantly improving the platform's operational stability. Particularly noteworthy is the nearly 70% decrease in the proportion of ineffective strategies, the reduction of average daily scheduling operating costs by over 30%, and the significant improvement in platform penalty control, demonstrating the graph embedding strategy's dual optimization capabilities for real-time performance and accuracy.

[0041] Based on actual operating scenarios, this invention constructs a graph structure relationship of multiple types of nodes, extracts topological, state and historical behavior features, and generates aggregation vectors by using graph embedding representation learning methods. It establishes collaborative aggregation paths among multiple resources, replacing the traditional static policy aggregation methods based on rules or thresholds, and completely solves the problems of fuzzy resource relationship identification, inaccurate collaboration degree evaluation and lagging policy adjustment.

[0042] This embodiment fully verifies the application value of the method proposed in the present invention in a real virtual power plant platform. It not only solves the problems of insufficient collaboration, frequent strategy adjustments and large response deviations under the traditional aggregation method, but also achieves higher resource utilization efficiency and platform profitability, and has significant promotion value and commercialization potential.

[0043] The above description is only a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any equivalent substitutions or modifications made by those skilled in the art within the scope of the technology disclosed in the present invention, based on the technical solution and inventive concept of the present invention, should be covered within the scope of protection of the present invention.

Claims

1. An IoT real-time task allocation system based on AI scheduling, characterized in that, include: The data acquisition module is used to collect task request data uploaded by IoT terminals, perform format normalization processing, and generate task input vectors. Collect operational resource data of IoT nodes and construct a set of node resource vectors; The semantic modeling module is used to input the task input vector into the improved StructBERT model, perform semantic structure modeling operations, and construct task semantic vectors; the improved StructBERT model introduces task slot construction structure and branching reinforcement structure; The intent analysis module is used to construct a historical task intent vector library and a semantic conflict graph. The historical task intent vector library stores task semantic vectors and scheduling execution result data. The semantic conflict graph represents the target overlap relationship, execution condition conflict relationship, and priority interference relationship between tasks. The comparison and judgment module is used to compare the task semantic vector with the historical task intent vector library, calculate the semantic structure similarity, target overlap and condition conflict value, and generate task aggregation degree and semantic difference vector based on the comparison results; perform task conflict judgment and scheduling priority adjustment operations according to the task aggregation degree and semantic difference vector; if the task aggregation degree is less than the first preset threshold, semantic clarification data is generated; if the task aggregation degree is between the first preset threshold and the second preset threshold, the scheduling priority value is adjusted; if the task aggregation degree is greater than the second preset threshold, it is marked as a task that can be directly scheduled. The scheduling and execution module is used to perform scheduling calculation operations based on task semantic vectors, scheduling priority values ​​and node resource vector sets, generate task allocation results, send task execution instructions to target nodes, record task execution feedback data, and update the historical task intent vector library and semantic conflict graph.

2. The IoT real-time task allocation system based on AI scheduling according to claim 1, characterized in that, The modules are connected in the following way: S1. Collect task request data uploaded by IoT terminals, perform format normalization processing, and generate task input vectors; S2. Collect the operational resource data of IoT nodes and construct a set of node resource vectors; S3. Input the task input vector into the improved StructBERT model, perform semantic structure modeling operations, and construct the task semantic vector; the improved StructBERT model introduces a task slot construction structure and a branching reinforcement structure; S4. Construct a historical task intent vector library and a semantic conflict graph; the historical task intent vector library stores task semantic vectors and scheduling execution result data; the semantic conflict graph represents the target overlap relationship, execution condition conflict relationship and priority interference relationship between tasks. S5. Perform a comparison operation between the task semantic vector and the historical task intent vector library, calculate the semantic structure similarity, target overlap and condition conflict value, and generate task aggregation degree and semantic difference vector based on the comparison results. S6. Perform task conflict determination and scheduling priority adjustment operations based on task aggregation degree and semantic difference vector; if the task aggregation degree is less than the first preset threshold, generate semantic clarification data; if the task aggregation degree is between the first preset threshold and the second preset threshold, adjust the scheduling priority value; if the task aggregation degree is greater than the second preset threshold, mark it as a task that can be directly scheduled. S7. Based on the task semantic vector, scheduling priority value and node resource vector set, perform scheduling calculation operation, generate task allocation result, send task execution instruction to target node, record task execution feedback data, and update historical task intent vector library and semantic conflict graph.

3. The IoT real-time task allocation system based on AI scheduling according to claim 2, characterized in that, Specifically, S1 is: Collect task request data uploaded by IoT terminals; the task request data includes text command data, speech recognition and transcription data, sensor event data, and image trigger data. The text command data and speech recognition-written data are processed in a standardized manner to unify them into a structured language expression format; Sensor event data and image trigger data are processed to extract event labels, generating event representation information and trigger context information; Vector encoding is performed on the text expression format and time tag information to construct a task expression vector with a unified dimension; The task representation vector is normalized and padded with timestamps to generate a task input vector; the task input vector includes action semantics, object semantics and conditional semantics.

4. The IoT real-time task allocation system based on AI scheduling according to claim 2, characterized in that, Specifically, S2 is: Collect operational resource data of IoT nodes; the operational resource data includes computing power data, response latency data, network connection status data, and operational load status data; Standardize the computing power data to generate node computing performance metrics; Perform sliding time window statistical processing on response latency data to generate node average response time and fluctuation amplitude indicators; perform status labeling processing on network connection status data to generate node connection availability indicators; perform interval quantization processing on runtime load status data to generate node load level indicators; and concatenate the node computing performance indicators, average response time, connection availability indicators, and load level indicators into a vector to construct a node resource vector. Perform dimension normalization on all node resource vectors to generate a set of node resource vectors.

5. The IoT real-time task allocation system based on AI scheduling according to claim 2, characterized in that, Specifically, S3 is: The task input vector is fed into the improved StructBERT model in the order of action semantics, object semantics, and conditional semantics. Structural position labeling is performed on the action semantics, object semantics, and conditional semantics of the task input vector to form a structural position labeling sequence; Establish a correspondence between the structure location marker sequence and the task input vector to construct a structured task representation; Perform context association modeling on the structured task representation to generate a sequence of context semantic representations; Structural boundary recognition processing is performed on the context semantic representation sequence to identify action boundaries, object boundaries, and condition boundaries, forming a slot label sequence; The slot label sequence is processed by vector concatenation according to the arrangement of action slot boundaries, object slot boundaries and condition slot boundaries to generate task semantic vector; The task input vector is compared with the preset comparison samples to calculate the difference, generate a divergence score, and perform reinforcement training based on the divergence score during the training process.

6. The IoT real-time task allocation system based on AI scheduling according to claim 5, characterized in that, The improved StructBERT model introduces a task slot construction structure and a branching reinforcement structure, specifically: The task slot construction structure performs semantic component segmentation processing on the task input vector, constructing slot occupancy information by dividing action semantics, object semantics, and conditional semantics respectively; and performs splicing processing on the slot occupancy information and structural position embedding information to generate a structured slot embedding sequence. Perform context association modeling on the structured slot embedding sequence to generate slot context vectors; perform boundary recognition processing on the slot context vectors to output action slot boundaries, object slot boundaries, and condition slot boundaries; construct slot structure representations by combining action slot boundaries, object slot boundaries, and condition slot boundaries according to a set concatenation order and perform fusion processing with the original context representations to generate slot enhanced expression vectors; The divergence enhancement structure performs difference-sensitive feature extraction on the slot enhancement expression vector to generate a difference initial vector; it then performs a contrast construction process on the difference initial vector and the original embedding sequence of the task input vector to generate a difference contrast representation; it performs a dual-channel attention distribution calculation operation on the difference contrast representation to generate a difference attention vector; it performs weight enhancement processing on the difference attention vector to generate a difference enhancement representation; it performs a fusion process with the difference enhancement representation and the slot enhancement expression vector to generate a semantic divergence representation vector; it performs a divergence degree calculation operation on the semantic divergence representation vector to generate divergence score information, and updates the parameter update path in the semantic structure modeling process based on the divergence score information to output the task semantic vector.

7. The IoT real-time task allocation system based on AI scheduling according to claim 2, characterized in that, Specifically, S4 is: Collect task semantic vectors and scheduling execution result data, and establish task semantic record entries according to task identifiers; Perform vector storage processing on the task semantic record entries, and write the task semantic vector and scheduling execution result data into the historical task intent vector library in sequence; Perform vector validation processing on the historical task intent vector library and establish an index structure based on timestamp information, task identification information and semantic vector dimension information; Perform semantic similarity calculation on the semantic vectors of all tasks to construct a set of semantic structural relationships; The task target extraction process is performed on the scheduling execution result data to generate a task target label set; the target overlap degree calculation process is performed on the task target label set to construct the task target overlap relationship; Analyze and process the conflict between the scheduling execution result data and the conditional semantics execution conditions in the task input vector, and construct the execution condition conflict relationship; Perform interference calculation on the priority records in the scheduling execution result data to construct priority interference relationships; By establishing a set of graph nodes and a set of graph edges according to the identification order of task pairs, the overlapping relationships of task objectives, conflicting relationships of execution conditions, and interference relationships are constructed, and a semantic conflict graph is built. The semantic conflict graph is subjected to structural consistency verification. Graph edges that do not meet the structural constraints are removed to generate a verified semantic conflict graph.

8. The IoT real-time task allocation system based on AI scheduling according to claim 2, characterized in that, Specifically, S5 is: Perform a comparison process between the task semantic vector and all semantic vectors in the historical task intent vector library, and construct a comparison sequence according to the vector dimension order; Semantic structure similarity calculation is performed on the control sequence, and semantic structure similarity values ​​are generated based on the correspondence between action slot boundaries, object slot boundaries and condition slot boundaries; The target label extraction process is performed on the scheduling execution result data in the historical task intent vector library to generate a target label set, and then the label is compared with the target-related components in the task semantic vector to generate a target overlap value. Perform condition conflict analysis and processing on the condition records in the scheduling execution result data and the condition slot information in the task semantic vector to generate condition conflict values; The semantic structure similarity value, target overlap value, and condition conflict value are jointly measured according to the set weight order to generate an association measurement sequence. Perform aggregation calculations on the associated metric sequences to generate the task aggregation degree; Perform difference vector construction processing on semantic structure similarity value, target overlap value and condition conflict value, and generate semantic difference vectors by arranging the difference components in the order of action difference, object difference and condition difference.

9. The IoT real-time task allocation system based on AI scheduling according to claim 2, characterized in that, Specifically, S6 is: Perform threshold comparison processing on the task aggregation degree and construct a task aggregation degree comparison sequence; Perform a size relationship judgment operation between the task aggregation degree values ​​in the task aggregation degree comparison sequence and the first preset threshold to generate the first judgment result; The task aggregation degree values ​​in the task aggregation degree comparison sequence are compared with the second preset threshold to determine their size relationship, generating a second judgment result; joint state analysis processing is performed on the first judgment result and the second judgment result to generate state labeling information; If the task aggregation degree corresponding to the state label information is less than the first preset threshold, then perform difference extraction processing on the semantic difference vector to extract action difference components, object difference components and condition difference components, construct a clarification expression sequence, and perform format construction processing on the clarification expression sequence to generate semantic clarification data; If the task aggregation degree corresponding to the status flag information is between the first preset threshold and the second preset threshold, then priority adjustment processing is performed on the scheduling priority value. An adjustment reference sequence is generated according to the order of action difference components, object difference components and condition difference components. Weight mapping processing is performed on the adjustment reference sequence to generate priority increase and decrease information. Numerical synthesis operation is performed on the priority increase and decrease information to generate the updated scheduling priority value. If the task aggregation degree corresponding to the status flag information is greater than the second preset threshold, then the task identifier is processed for schedulable marking, and a directly schedulable marking record is generated.

10. The IoT real-time task allocation system based on AI scheduling according to claim 2, characterized in that, Specifically, S7 is: Collect task semantic vectors, scheduling priority values, and node resource vector sets to construct a set of inputs for scheduling calculations; Based on the scheduling calculation input set, a scheduling calculation operation is performed to generate a task allocation result. The scheduling calculation operation includes resource adaptability calculation, semantic matching degree calculation and priority sorting operation. Based on the task allocation results, a task execution instruction is generated and sent to the corresponding target node; Collect task execution feedback data from the target node, including task completion status, resource usage changes, and execution time information; Based on the task semantic vector and the task execution feedback data, perform a historical task intent vector library update operation and store the updated vector in the historical task intent vector library; The semantic distribution relationships in the historical task intent vector library are analyzed to identify conflict relationships between tasks and construct a semantic conflict graph.

Images