Intelligent agent decision path optimization method fusing explicit topology and implicit semantics

Abstract

The invention relates to the technical field of artificial intelligence, in particular to an agent decision path optimization method fusing explicit topology and implicit semantics. The method comprises the following steps: acquiring multi-modal data; performing document analysis according to the multi-modal data to obtain document analysis data; performing knowledge injection according to the document analysis data to obtain knowledge injection data; carrying out implicit semantic detection according to the knowledge injection data to obtain implicit semantic data; constructing a user demand agent blueprint according to the implicit semantic data; constructing an agent workflow according to the agent blueprint required by the user; performing task execution scheduling according to the agent workflow to obtain task execution data; performing node explicit topology construction according to the task execution data to obtain node explicit topology data; and performing communication anomaly analysis based on the node explicit topology data to obtain node communication anomaly data. According to the invention, agent task scheduling and path optimization are realized based on an artificial intelligence technology, and the resource utilization rate of the system is improved.

Description

Technical Field

[0001] The present invention relates to the field of artificial intelligence technology, and particularly to an intelligent agent decision path optimization method that integrates explicit topology and implicit semantics. Background Art

[0002] In traditional intelligent agent decision path optimization, in terms of topological modeling, it often only relies on static network structures, fails to capture changes in communication link status in real time, and cannot dynamically reflect communication anomalies between nodes or real-time adjustments of the topological structure, resulting in slow response of path optimization in emergency scenarios; in terms of implicit semantic understanding, it generally relies on shallow semantic matching or keyword extraction methods, cannot effectively identify deep semantic dependency relationships in user requirements, and is difficult to achieve precise parsing of complex intentions, thus affecting the accuracy of task scheduling and resource binding of intelligent agents; in path optimization strategies, there is generally a lack of a multi-agent cooperation mechanism driven by reinforcement learning, resulting in the inability to balance global resource load and local task priorities during task scheduling, and prone to problems such as decision conflicts, unperceived and unoptimized resource bottlenecks, reducing the adaptability and robustness of the overall system in a dynamic environment; in the process of anomaly detection and path rearrangement, most use rule-based or static priority schemes, unable to achieve automatic perception and intelligent rearrangement of task conflicts, and difficult to meet the high-concurrency and high-dynamic intelligent agent cooperation requirements in large-scale distributed environments. Summary of the Invention

[0003] Based on this, it is necessary for the present invention to provide an intelligent agent decision path optimization method that integrates explicit topology and implicit semantics to solve at least one of the above technical problems.

[0004] To achieve the above object, an intelligent agent decision path optimization method that integrates explicit topology and implicit semantics includes the following steps: Step S1: Obtain multimodal data; perform document parsing according to the multimodal data to obtain document parsing data; perform knowledge injection according to the document parsing data to obtain knowledge injection data; Step S2: Perform implicit semantic detection according to the knowledge injection data to obtain implicit semantic data; construct an intelligent agent blueprint for user requirements according to the implicit semantic data; construct an intelligent agent workflow according to the intelligent agent blueprint for user requirements; perform task execution scheduling according to the intelligent agent workflow to obtain task execution data; Step S3: Construct an explicit node topology according to the task execution data to obtain explicit node topology data; perform communication anomaly analysis based on the explicit node topology data to obtain node communication anomaly data; identify abnormal communication links based on the node communication anomaly data to obtain abnormal communication link data; perform node resource scheduling based on the abnormal communication link data to obtain node resource scheduling data; Step S4: Construct a reinforcement learning decision engine based on the node resource scheduling data; perform multi-agent decision path optimization according to the reinforcement learning decision engine to obtain decision path optimization data.

[0005] Through acquiring multi-modal data and conducting in-depth document parsing and knowledge injection, the present invention realizes the structured expression and semantic enrichment of information, effectively improves the ability to understand complex data, enables subsequent implicit semantic detection to capture the deep semantic dependencies and logical relationships in user requirements, thereby accurately constructing the intelligent agent blueprint of user requirements and reasonably planning the intelligent agent workflow, ensuring the scientificity and efficiency of task execution scheduling. The explicit node topology constructed based on task execution data can dynamically reflect the real-time state of the network structure and communication links, timely identify and analyze communication anomalies and abnormal links, enhancing the sensitivity and response speed to network anomalies. Node resource scheduling realizes dynamic adjustment for the impact of abnormal links and resource bottlenecks, optimizes resource allocation, avoids node overload or resource idleness, and effectively improves the system's load balancing ability. The reinforcement learning decision engine constructed using node resource scheduling data introduces a multi-agent cooperation mechanism, which can make dynamic trade-offs under global and local resource constraints, achieving the unified consideration of long-term macro goals and short-term task priorities, significantly reducing scheduling conflicts and resource bottlenecks, and enhancing the adaptability and robustness of the system in complex dynamic environments. During the decision path optimization process, through the real-time inference and adjustment of the reinforcement learning model, the path planning becomes more flexible and accurate, automatically senses task conflicts and performs priority sorting and path rearrangement, realizing the intelligent management of scheduling in high-concurrency, multi-task, and multi-agent environments. BRIEF DESCRIPTION OF THE DRAWINGS

[0006] Other features, objects, and advantages of the present invention will become more apparent from the following detailed description of non-restrictive embodiments with reference to the accompanying drawings: Figure 1 It is a schematic flowchart of the steps of an intelligent agent decision path optimization method that combines explicit topology and implicit semantics according to the present invention; Figure 2 It is a detailed schematic flowchart of step S1 in the present invention; Figure 3 It is a detailed schematic flowchart of step S16 in the present invention; The implementation, functional features, and advantages of the present invention will be further described with reference to the embodiments and the accompanying drawings. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0007] The technical method of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are part of the embodiments of the present invention, rather than all of them. All other embodiments obtained by those skilled in the art based on the embodiments of the present invention without creative work belong to the scope of protection of the present invention.

[0008] In addition, the accompanying drawings are only schematic diagrams of the present invention and are not necessarily drawn to scale. The same reference numerals in the drawings denote the same or similar parts, and thus repeated descriptions thereof will be omitted. Some of the block diagrams shown in the drawings are functional entities and do not necessarily correspond to physically or logically independent entities. The functional entities can be implemented in software form, or in one or more hardware modules or integrated circuits, or in different networks and / or processor methods and / or microcontroller methods.

[0009] It should be understood that although terms such as "first", "second", etc. may be used herein to describe various units, these units should not be limited by these terms. These terms are only used to distinguish one unit from another. For example, without departing from the scope of the exemplary embodiments, the first unit can be called the second unit, and similarly the second unit can be called the first unit. The term "and / or" used herein includes any and all combinations of one or more of the listed related items.

[0010] To achieve the above object, please refer to Figures 1 to 3 , the present invention provides an intelligent agent decision path optimization method that integrates explicit topology and implicit semantics. The method includes the following steps: Step S1: Obtain multimodal data; perform document parsing according to the multimodal data to obtain document parsing data; perform knowledge injection according to the document parsing data to obtain knowledge injection data; In this example, an image acquisition module (such as an Intel RealSense D435 camera) and a text data acquisition system are used to synchronously capture image data and corresponding text task description data from an industrial process monitoring scenario. The image data resolution is set to 1920×1080, with a sampling frequency of 10 frames per second. The text data is obtained from the process control log system and uniformly formatted using UTF-8 encoding. The collected multimodal data is input into a data preprocessing module. The image portion is processed using an image denoising and color normalization algorithm based on OpenCV 4.8. The text portion is processed using an NLTK tokenizer for word segmentation and stop word removal, and the Porter algorithm is used for word stemming. After preprocessing, the semantic segmentation module is used to segment the text regions, setting the average block length to 120 characters, and each block is divided into logical paragraphs. Tesseract OCR is then used to identify text regions in the image data to obtain the coordinates of the text region bounding boxes. The YOLOv5 model detects objects in the image and extracts the location information (x, y, w, h) and category label for each object. A preliminary image-text association is then established based on the spatial relationship between the object and the text within the image. During the vectorization phase, BERT (bert-base-uncased) encoding is used for the text, converting each semantic block into a 768-dimensional vector. For the image, a ResNet-50 feature extraction network is used to obtain a 2048-dimensional feature vector for the target region. The image-text features are fed into a multimodal alignment module, where a method based on Canonical Correlation Analysis (CCA) is used to model the correlation between image-text vector pairs. Positive examples are screened using a cosine similarity threshold of 0.85 to generate multimodal alignment data. Based on the alignment results, the image-text blocks are combined and fed into a rule-based semantic parsing engine to perform operations such as entity recognition, attribute extraction, and action detection. The final output is document parsed data in JSON format. Knowledge injection is performed on the document parsed data to construct a unified representation structure. Semantic units consist of triples (entity, attribute, action) and are stored in a structured knowledge table. A library of expert-defined rules is used to semantically enhance these triples. For example, when the "job type" is "high-temperature welding" and the "time" field is empty, a safety time stamp field is automatically inserted. The result of knowledge injection is structured JSON data with 22 fields, including entity categories, semantic role labels, time attributes, and object location indexes. This data is identified as knowledge injection data.

[0011] Step S2: Perform implicit semantic detection based on the knowledge injection data to obtain implicit semantic data; construct an intelligent agent blueprint for user requirements based on the implicit semantic data; construct an intelligent agent workflow based on the intelligent agent blueprint for user requirements; perform task execution scheduling according to the intelligent agent workflow to obtain task execution data; In this embodiment, a context feature representation is constructed based on the knowledge injection data. This process extracts the co-occurrence frequency and sequence order between adjacent triples through a sliding window mechanism, and sets the window length to 5 semantic units. During feature encoding, the co-occurrence frequency matrix (statistical joint probability between triples, threshold 0.05) and the sequence dependency graph (calculating sequence weights through the PageRank algorithm) are calculated respectively, and a context graph structure is constructed, denoted as G=(V,E), where V is the triple node and E is the weighted edge. In the semantic association analysis process, GAT (Graph Attention Mechanism) is used to encode the context graph. Each triple representation is used as the input node feature, and the attention mechanism weights each edge and calculates the context semantic influence factor. Edges with an association strength greater than 0.7 are identified as the core semantic paths, generating semantic association data. LSTM encoding is performed on this semantic path to extract the deep semantic expression, with an output dimension of 256, called the deep semantic data. In the implicit semantic detection stage, a pattern matching mechanism based on logistic regression is used to perform discrimination on the deep semantic data, identifying regions with features of "inference logic chain", such as patterns like "if... then...", "due to... resulting in...", etc. This rule library includes a total of 21 types of patterns, which are matched using a boolean vector, with a matching threshold of 0.9. The extracted pattern sequences form the implicit semantic data, marking the deep implicit relationships. Based on the implicit semantic data, an intelligent agent blueprint for user requirements is constructed using the template matching method. Nine types of blueprint templates are preset (such as data collection type, resource scheduling type, analysis and evaluation type), and mapping and matching are performed according to the semantic label fields and template conditions. For example, if "resource bottleneck + reallocation + optimization path" appears in the semantics, it is matched to the resource scheduling type blueprint. The generated blueprint records information such as task input, output, and resource dependency relationships in a structured form. According to the module dependency relationships and task call sequences defined in the blueprint, a DAG (Directed Acyclic Graph) form of the intelligent agent workflow graph is constructed, where the nodes represent specific functional modules (such as information reading, resource allocation), and the edges represent data transmission dependency relationships. Topological sorting is used to organize the task execution order, and the result is used as the task scheduling input. Task scheduling is performed according to the intelligent agent workflow. The scheduling engine traverses the workflow graph, assigns specific execution locations to each task node, and the scheduling rules perform node optimization based on indicators such as processing capacity (unit: tasks / sec), current load (unit: %), and response latency (unit: ms). Each task execution records the status, duration, and input / output, and the output is task execution data in JSON format.

[0012] Step S3: Perform explicit node topology construction based on the task execution data to obtain explicit node topology data; perform communication anomaly analysis based on the explicit node topology data to obtain node communication anomaly data; identify abnormal communication links based on the node communication anomaly data to obtain abnormal communication link data; perform node resource scheduling based on the abnormal communication link data to obtain node resource scheduling data; In this embodiment, an explicit topology structure is constructed according to the execution node identifiers and communication behaviors recorded in the task execution data. The NetX library is used to construct a directed graph, where the nodes are device identifiers (such as ID001~ID128), and the edges are the actual communication records during task execution. Each edge is marked with the amount of transmitted data (unit: KB), communication timestamp, and communication direction. The edges with more than 3 transmission failure records are set as edges to be detected. Communication logs are extracted for all edges in the topology, and the packet loss rate (expressed as the number of lost frames / total number of frames, with the threshold set to 10%) and out-of-order rate (number of out-of-order packets / total number of packets, with the threshold set to 15%) are calculated respectively. The timestamp sequence is extracted, and the time periods when packet loss or out-of-order occurs each time are recorded. The overlapping time periods are marked as communication anomaly windows. The communication behaviors within the anomaly windows are encapsulated as communication anomaly data. Repeated request analysis is performed on the communication anomaly data. Data channels that continuously send repeated requests to the same target node inward 5 times are identified, and it is determined whether they are abnormal access channels; if the number of packets in the abnormal access channel suddenly increases by more than 2 times the standard deviation within a unit time, they are included in the access surge list. The access surge records are matched. If they match the DoS behavior characteristics (such as the continuous request frequency exceeding 50 times per second) in the attack feature library, they are recorded as denial-of-service attack data. It is further determined whether there are behaviors such as IP spoofing and port scanning, and intrusion behavior data is formed. According to the intrusion behavior data, the current communication protocol type (such as MQTT, HTTP, gRPC) and protocol version information are identified. Compared with the historical version standard library, if there are problems such as version incompatibility and abnormal encryption algorithms, they are recorded as protocol version anomaly data. All abnormal edges are scored and sorted (the comprehensive weight is communication error frequency 0.4 + protocol anomaly 0.3 + intrusion probability 0.3). Those with a score higher than the set threshold (0.7) are identified as abnormal communication links, and the result forms abnormal communication link data. Based on the above link data, the current resource usage of the connected nodes is statistically analyzed. The CPU utilization rate (obtained by sampling the node system information), memory usage rate (unit: MB), and bandwidth occupancy rate (unit: Mbps) are calculated respectively. Nodes with any one of the indicators higher than the set threshold (such as CPU>90%, memory>85%) are identified as bottleneck nodes. Further, a task migration evaluation is performed on the bottleneck nodes, and the communication overhead required to transfer the existing tasks to neighboring nodes is calculated, including the amount of data (MB) and transfer time (s), and it is required that the overhead is not higher than 80% of the time for the original node to maintain the task. After determining the migration path, according to the principle of the shortest transfer path length and the lowest node load first, the node task scheduling scheme is reconstructed, and finally the node resource scheduling data is output.

[0013] Step S4: Construct a reinforcement learning decision engine according to the node resource scheduling data; perform multi-agent decision path optimization according to the reinforcement learning decision engine to obtain decision path optimization data.

[0014] In this embodiment, the task distribution and node status metrics (including CPU usage, average latency, packet loss rate, memory occupancy, etc.) in the node resource scheduling data are extracted. These status quantities are normalized into a unified vector to form the scheduling node status data with a dimension of 128. This status data is input into the reinforcement learning system for hierarchical decision-making modeling. At the macro level, a model-based reinforcement learning (MBRL) method is used to construct a world model. This model uses the historical resource scheduling trajectories (with a recording time span of 90 days) to train the prediction module. The input is the sequence of task node status vectors, and the output is the trend of business path changes in the next 30 days. An LSTM is used to construct a dynamic path prediction module, and the prediction accuracy is controlled within 10% by the mean square error. Quarterly-level policy data is generated according to the prediction output, including the priority task path, resource redundancy adjustment frequency, and other contents. At the tactical level, the proximal policy optimization algorithm (PPO) is used to optimize the task configuration in the short and medium terms. The input is the task scheduling log within a week (granularity in minutes), and the output is the adjustment plan for the workflow task nodes. The clipratio = 0.2 is used to limit the policy update amplitude, and the maximum gradient is set to 0.01 to avoid policy oscillation. At the execution level, a deep Q-network (DQN) is used to fine-tune and optimize the API execution order and degradation plan (such as switching to a standby node) for each task. The ε-greedy strategy is adopted to control exploration (ε is initially set to 0.5 and decreases by 0.01 every 100 steps), and the action set includes operations such as API reordering, task delay, and discarding low-priority tasks. Finally, the policy data at the macro, tactical, and execution levels are integrated to construct a reinforcement learning decision engine, and a policy graph for reasoning is generated. This graph records all policy paths, conditions, and feedback parameters. Based on this decision engine, multi-agent path optimization operations are performed to sequentially extract the task execution priority, node execution path, and optimal resource allocation. The final output is structured JSON-format decision path optimization data, and the fields include information such as the path node sequence, execution time window, and resource configuration.

[0015] Preferably, step S1 includes the following steps: Step S11: Obtain multi-modal data and perform data preprocessing to obtain the multi-modal data to be processed; In this embodiment, an integrated multimodal acquisition system is used to acquire source data. The text data source is structured and unstructured industrial equipment operation log files, and the file formats include three types: CSV, TXT, and XML, with the encoding uniformly being UTF-8. The image data source is high-definition camera devices installed at each key node of the production line. The camera model is Sony IMX577, the resolution is 3840×2160, and the sampling frame rate is 30fps. The preprocessing of text data includes: 1) filtering out abnormal symbols (such as all control characters with ASCII codes less than 32), 2) uniformly processing redundant spaces into one space, and 3) uniformly retaining two decimal places for numbers. The preprocessing of image data includes: 1) performing denoising processing using bilateral filtering, with the filter kernel size being 7×7, the color standard deviation σColor = 75, and the spatial standard deviation σSpace = 75; 2) performing histogram equalization processing to uniformly map the image gray range. After all preprocessing operations are completed, the image and text data are combined in the form of structured key-value pairs to form the multi-modal data to be processed, with the unified naming format being {task_id,image_array,text_content}.

[0016] Step S12: Identify the text area based on the multi-modal data to be processed to obtain text area data; perform semantic chunking according to the text area data to obtain semantic chunk data; perform document slicing according to the semantic chunk data to obtain document slice data; In this embodiment, a text area positioning operation is performed on the text content in the multi-modal data to be processed. The sliding window method is used to divide the text content into regions. The sliding window size is 128 characters, the step size is 64 characters, and word frequency statistics and key trigger word matching are performed within the window. The trigger words include 72 keywords such as "step", "parameter", "node", "abnormality", etc. The regions designated as containing core content are extracted as text area data. Semantic chunking processing is performed on the extracted text area data. Using the rule-driven syntactic chunking method, each sentence is divided into clauses according to punctuation marks such as ";", ",", ".", and conjunctions such as "and", "or", "and then", etc. Each semantic chunk is required to have a length of at least 5 words and at least 20 characters, and finally semantic chunk data is obtained. Based on the semantic chunk data, each semantic chunk is divided into separate document fragments. Each document slice needs to retain the original semantic number, the page number and line number of the document where it is located for subsequent semantic mapping. The document slice format is in JSON structure, and the fields include "slice_id", "text_span", "origin_doc_id", "position_index", etc.

[0017] Step S13: Perform image segmentation on the multi-modal data to be processed to obtain image segmentation region data; perform object detection based on the image segmentation region data to obtain image object data; perform object attribute recognition based on the image object data to obtain object attribute data; In this embodiment, image segmentation processing is performed on the image in the multi-modal data to be processed. The K-means segmentation algorithm based on color clustering is used to perform foreground-background distinction operations on the image. The number of cluster centers is set to k = 3, and the RGB value distance of the image pixels is used as the clustering basis. During the segmentation process, the segmentation mask is output as a binary image, the foreground region is labeled and its area is calculated. Regions with an area less than 5% of the total image area will be ignored, and finally the image segmentation region data is output. Based on the image segmentation, the object detection is performed on each foreground region using the feature extraction method based on Histogram of Oriented Gradients (HOG). The sliding window (64×64) is used to scan each pixel in the foreground region, and at the same time, the HOG features of each window are calculated and Euclidean distance matching is performed with the template object features (including categories such as industrial control equipment, tools, and operators). The matching degree threshold is set to 0.85. Regions that meet the matching conditions are marked as image object data, including the bounding box coordinates (x_min, y_min, x_max, y_max) and category labels. For the identified image object regions, attribute recognition is performed. The geometric attribute extraction method based on the template is called to perform rectangular fitting on the object contour, and geometric attributes such as height, width, aspect ratio, and centroid coordinates are extracted; at the same time, the gray distribution characteristics are statistically analyzed by combining the image gray histogram, and the gray threshold is set to 85 to identify the metal texture objects. All the above parameters are unified to form the object attribute data, and the storage format is XML, with 15 attribute fields for each object.

[0018] Step S14: Perform vector quantization encoding on the document slice data to obtain semantic vector data; perform feature encoding based on the object attribute data to obtain object attribute encoding data; In this embodiment, vectorized encoding processing is performed on the document slice data. This processing is implemented through the TF-IDF vectorization method. A bag-of-words model is established, and the size of the bag of words is set to 5000. All document slices are uniformly input in the UTF-8 encoding format. Before encoding, the text content is segmented by a standard word segmentation tool, and stop words are filtered using a predefined stop word list. For each term, its frequency of occurrence in the current slice is counted, and combined with the number of documents in which it appears in all slices, a word vector of length 5000 is generated, where each dimension corresponds to the TF-IDF weight of a specific term. The generated semantic vector data is stored in the form of a sparse matrix and structured and saved in the COO format of the SciPy library to improve the efficiency of subsequent processing. In terms of image feature processing, for the image target data obtained in step S13, a unified feature encoding operation is performed. First, the bounding box of the image target area is extracted to obtain the height and width values, accurate to the pixel level, and the gray mean value range of 0 to 255 is calculated through the gray scale statistics of pixel points. Then, one-hot encoding operation is performed based on the identified target category label. The number of categories is set to 7, corresponding to "equipment", "tool", "personnel", "sign", "instrument", "panel", and "control unit" respectively, and the one-hot encoding dimension is 7 dimensions. After obtaining the numerical features and category encoding of the image, all features are concatenated to form unified 25-dimensional target attribute encoding data, which is stored in the form of a NumPy array. Each array data is mapped one-to-one with the corresponding image target, and the data organization form is a two-dimensional array, where each row represents the complete feature expression of an image target object. The semantic vector data and the target attribute encoding data serve as the basis for multi-modal alignment and fusion, and an index field needs to be reserved in the structure for subsequent cross-processing use.

[0019] Step S15: Perform multi-modal alignment processing based on the semantic vector data and the target attribute encoding data to obtain multi-modal alignment data; perform semantic parsing based on the multi-modal alignment data to obtain document parsing data; In this embodiment, semantic vector data and target attribute encoding data are input into the multi-modal alignment processing module. The mutual information score is calculated for each pair of image-text vectors using a correlation detection method based on mutual information maximization. The mutual information threshold is set to 0.8, and pairs exceeding the threshold are identified as highly correlated image-text pairs and subjected to label binding processing. Semantic parsing is performed on the bound image-text pairs. First, the verb and noun part-of-speech annotation results in the semantic vector are extracted, and a regular rule-based entity-relationship-object structure parser is called to extract the semantic structure in the form of triples, such as: {entity: "Sensor A", relation: "appears", object: "error"}. At the same time, the entity and object categories are further corrected according to the image target label information. For example, the image recognition result of "cable" will overwrite "line" in the text. Finally, each image-text pair outputs 1 to 3 triples, which are integrated to form document parsing data, and the fields include: entity_type, relation_type, object_type, text_span, image_id.

[0020] Step S16: Perform knowledge injection based on the document parsing data to obtain knowledge injection data.

[0021] In this embodiment, a unified knowledge representation structure is constructed based on the document parsing data. The semantic units extracted from the text and images are organized in the form of a structured five-tuple, including an entity identifier, an attribute identifier, a relationship identifier, a confidence score, and source location information. The entity identifier uses a unified numbering method to represent the key objects identified in the text or image. The attribute identifier is used to express the quantitative or qualitative descriptions related to the entity. The relationship identifier is used to describe the logical relationship between the entity and other entities or attributes. The confidence score is calculated in the previous semantic parsing stage. The source location field records the text line number and image coordinate index from which the five-tuple is derived. In the constructed knowledge representation vector, the range of the entity space number is controlled between 0000 and 9999, and the relationship number is restricted within the range of 000 to 199. All numbers are generated by replacing them through a pre-established mapping dictionary. Subsequently, the above-constructed knowledge representation vector is compared with the expert knowledge base deployed locally in sequence. The expert knowledge base contains more than twenty thousand structured rule data and uses a hierarchical retrieval method for matching processing. The matching process is carried out step by step in the order of entity, then attribute, and then relationship. When the system retrieves any rule data that is consistent with the current knowledge representation vector in terms of entity category, attribute range, and relationship context, it is considered a successful match, and semantic enhancement processing is triggered. The semantic enhancement operation appends the enhancement information defined in the rule to the current knowledge unit. For example, after identifying the description information of the motor equipment temperature increase, the semantic label "risk increase" is appended, and system suggestion entries such as "execute cooling instructions" are introduced. All enhanced structured data is re-encoded into a standard RDF triple set, and the output format includes the subject, predicate, object, confidence score, and source index identifier. All data is recorded and saved in the TTL language format for use in subsequent implicit semantic detection and agent reasoning processes.

[0022] Preferably, step S16 includes the following steps: Step S161: Extract semantic unit features based on the document parsing data to obtain semantic unit data; In this embodiment, for the document parsing data, semantic unit feature extraction operations are carried out. The document parsing data includes segmented text segments, image annotation regions, and their corresponding entity category labels. Taking each text segment as the basic processing unit, a rule-based part-of-speech analysis method is used to decompose the syntactic structure, and key components such as noun phrases, verb phrases, quantity phrases, and prepositional phrases are identified. Each phrase is labeled with its semantic category, such as device name, operation behavior, physical state, or measurement parameter. For each image annotation region, by parsing the target attribute coding data within the region, features of the target in terms of spatial dimension, shape boundary, gray-scale distribution, and recognition label are extracted. These text semantic components and image target information are combined into semantic units in a one-to-one correspondence relationship, forming records with the structural form of {text segment ID, image region ID, entity category, semantic category, syntactic label, spatial position}. In this step, a fixed syntactic rule set (a total of 120 rules) is used to define the boundary determination conditions for various semantic structures, and all document segments are scanned item by item, and the matching results are extracted according to the rules as semantic unit data.

[0023] Step S162: Construct a knowledge representation vector based on the semantic unit data to obtain knowledge representation vector data; In this embodiment, after obtaining the structured semantic unit data, a knowledge representation vector for subsequent enhancement and comparison needs to be constructed. For each semantic unit, the "subject-predicate-object" ternary semantic combination is extracted and uniformly encoded into a triple structure. Entity encoding uses a preset five-level hierarchical entity table for number assignment. Among them, the top-level entity numbers are from 1 to 99, and the numbers are extended down to the five-level subclasses level by level. Finally, each entity has a unique integer number. Attribute and relationship encoding use a key-value mapping table, corresponding to descriptive adjectives, numerical parameters, operational verbs, or prepositions, etc., a total of 210 common semantic relationships, each corresponding to a number from 1 to 210. Each knowledge representation vector finally forms a structure of {entity ID, attribute ID, relationship ID, confidence score, text position, image coordinates}. Among them, the confidence score is calculated according to the semantic structure integrity, entity semantic matching degree, and syntactic stability, and the score range is from 0 to 1, with a precision of 0.001.

[0024] Step S163: Perform semantic enhancement based on the knowledge representation vector data to obtain semantic enhancement data; In this embodiment, semantic enhancement processing is performed based on the knowledge representation vector data. The built-in semantic extension rule set of the system is loaded. This rule set contains approximately 3,200 enhancement templates defined based on semantic mappings such as hyponymy, coordination, causality, and time-dependency relationships. Taking each knowledge representation vector as input, the enhancement conditions in the rule set are matched item by item. For example, when the entity in the recognition vector is "temperature sensor", the attribute is "rising", and the relationship is "monitoring", a template in the rule set is triggered: if the entity is a sensor and the relationship is in the rising direction, then the additional label "device anomaly" is attached. After the enhancement is executed, the original knowledge representation vector adds an enhancement field to record the additional label, enhancement rule ID, and enhancement source evidence number. After all semantic enhancement results are uniformly encoded, extended knowledge vector data is formed.

[0025] Step S164: Perform expert knowledge base matching based on the semantic enhancement data to obtain expert knowledge base matching data; In this embodiment, a three-layer indexing mechanism is used in the matching process: the first layer quickly locates rules based on entity IDs, the second layer matches the attribute ID values, and the third layer uses a semantic distance measurement method to calculate the cosine similarity between the enhanced relationship field and the relationship defined in the rule base. The similarity threshold is fixed at 0.85 or above to be judged as a successful match. The knowledge base ontology contains a five-element rule structure in a fixed format: {entity ID, attribute range, relationship category, semantic context vector, recommended label}, with a total of 22,479 entries. All knowledge vector data sequentially enters the matching engine and is compared with the knowledge base item by item. After a match is hit, the rule ID and recommended label of the knowledge base entry will be returned, and a final matching record will be formed to constitute the expert knowledge base matching data.

[0026] Step S165: Perform knowledge injection based on the expert knowledge base matching data to obtain knowledge injection data.

[0027] In this embodiment, based on the obtained expert knowledge base matching data, the knowledge injection process is carried out. This process relies on a unified data encoding standard to convert each piece of matching data into a triple in the standard RDF format. Each piece of data includes a subject (i.e., entity ID), a predicate (i.e., relationship ID), an object (i.e., attribute value or enhanced label), and an additional metadata field for recording the matching rule ID, source location, and confidence score. The knowledge injection operation adopts a batch processing method to write all RDF triples into the TTL file (Terse RDF Triple Language) of the knowledge graph management module and register the written content through the SPARQL query interface. Finally, complete-structured, uniformly numbered, and cross-queryable knowledge injection data is formed for subsequent semantic detection and path reasoning operations.

[0028] Preferably, step S2 includes the following steps: Step S21: Perform context feature perception based on knowledge injection data to obtain context feature data; In this embodiment, when performing the context feature perception operation, first load the knowledge injection data, which is in the format of a set of RDF-structured triples. Taking each knowledge triple as a processing unit, extract its original context position from the subject (entity ID), predicate (relationship ID), and object (attribute value) in the triple, including the original text position information (such as paragraph number, in-sentence position) and image coordinate information (such as the center point position of the bounding box, region number). Based on the above position data, trace back the previous 3 text fragments and the previous 2 image regions, and extract their corresponding semantic unit information, including part-of-speech tags, semantic tags, time expressions, entity genera, action verbs, and quantitative values. Encode the extracted adjacent context semantic elements into fixed-length vectors to form a context semantic feature set, with fields including upstream and downstream entity names, action verbs, time status, spatial position offset, semantic tag ID, numerical attribute interval, etc. Among them, the time status is discretized into 5 levels (past, current, planned, long-term future, indefinite), and the position offset is quantitatively represented by the Euclidean distance (in pixels), with a control range of 0–500 pixels. All context semantic feature data is merged into a structured table to form a context feature data set.

[0029] Step S22: Perform semantic association analysis based on the context feature data to obtain semantic association data; In this embodiment, group the records in all context feature data labeled with the same main entity ID. In each group, construct a timeline by sorting according to the time status from "past" to "future". Secondly, in each group's timeline, find the verb-like operations that appear more than 2 times in the relationship ID field, and extract their corresponding entity-relationship-attribute chains as potential semantic connection candidates. Furthermore, calculate the coincidence degree of the semantic tag IDs between different context records. If the coincidence degree exceeds 60%, it is determined that there is semantic overlap. Combine the semantic overlap records to perform a relative difference operation on the numerical parameters in the attribute field. If the relative change is greater than 25% and in the same direction, add an attribute trend connection mark. Finally, perform a logical connection through the three conditions of entity coincidence, relationship consistency, and attribute trend consistency. Each group of context records that meet two or more conditions is merged into a semantic association cluster and encoded as a semantic association data set.

[0030] Step S23: Perform deep semantic recognition based on the semantic association data to obtain deep semantic data; In this embodiment, deep semantic recognition operations are carried out based on semantic association data. The specific process includes three stages: First, extract the co-occurrence relationships of entity genera in each semantic association cluster, count the high-frequency entity combinations in the association cluster, such as "motor - temperature - alarm" or "pump - pressure - drop", etc., and construct an entity collaboration matrix with a dimension of 500×500, corresponding to 500 basic entity categories. Second, analyze the combination patterns of entity combinations and relationship types. For each pair of entity relationship chains, count their occurrence frequencies in all semantic association clusters, and combine the time status fields in the context to model the causal chain for the operation sequence, such as the chain operation path of "temperature rises" → "fan turns on" → "current increases", etc. Finally, construct a semantic structure diagram based on the entity combination frequency, causal relationship path length, and attribute change trend direction. Each node represents an entity, the edge represents a relationship, and the direction indicates the order of operations. Obtain all logical chains with three or more layers through graph traversal and convert them into deep semantic data.

[0031] Particularly importantly, step S23 includes the following steps: Step S231: Extract the context sentence window according to the semantic association data to obtain context fragment data; In this embodiment, call the generated semantic association data, which records the semantic connection information between entity pairs and their position numbers in the original text. Taking the entity pairs involved in each group of semantic associations as anchors, perform context window extraction operations. The context window size is set to 2 sentences before and after the sentence where the current entity pair is located, and the window boundary does not exceed the same paragraph. Using the sentence number and paragraph number as the index range, extract the corresponding text content from the original corpus data. All the extracted texts are uniformly encoded in UTF-8, and special characters and line breaks are removed, and standardized into a continuous text segment. Each context fragment data structure contains five fields: entity pair ID, start sentence number, end sentence number, text content, original paragraph number, and finally form a context fragment data table as the input for subsequent part-of-speech analysis and syntactic structure processing.

[0032] Step S232: Perform part-of-speech tagging processing based on the context fragment data to obtain lexical structure data; In this embodiment, for the context fragment data, perform word segmentation and part-of-speech tagging operations one by one. Use the CRF (Conditional Random Field) sequence tagging algorithm to construct a part-of-speech recognition process, and use the bidirectional maximum matching method for word processing. For each context fragment, perform word segmentation and then predict the part of speech for each word to generate a sequence of word and part-of-speech pairs. The processing results are stored in a JSON structured format, and each record includes fragment ID, word order number, word item, part-of-speech category, and position offset in the original text. All lexical structure data is stored in a temporary cache database for subsequent dependency syntactic analysis.

[0033] Step S233: Perform dependency syntactic analysis based on the lexical structure data to obtain syntactic dependency chain data; In this embodiment, a dependency syntactic tree generation algorithm based on a rule set is adopted. For the part-of-speech tagging sequence of each context fragment, word nodes are constructed sentence by sentence, and directional edges are established according to the grammatical collocation rules between adjacent words to form a tree-like dependency graph structure. The dependency relationship is stored in the form of a triple: {modifier number, modified word number, relationship type}. Each syntactic chain is attached with the minimum path distance between words (calculated according to the difference in word order numbers). All syntactic dependency chain data are uniformly encoded and sorted according to the original text fragment number to form a complete syntactic dependency chain data table.

[0034] Step S234: Identify semantic role boundaries based on the syntactic dependency chain data to obtain semantic role unit data; perform semantic paradigm matching according to the semantic role unit data to obtain semantic intention number data; In this embodiment, the action center word (verb class) is identified as the semantic focus according to the syntactic dependency chain data, and the agent, patient, and modifying components of the verb are analyzed one by one. An action-driven role recognition mechanism is adopted. According to the basic structure model of "agent - verb - patient", it is determined that the subject word connected to the verb in the dependency chain is the agent, the object word is the patient, and the adverbial is the modifying component. Each semantic unit includes fields such as action word, role type, term value, position of the word in the sentence, and dependency path. For the extracted semantic role units, the semantic paradigm library is called for matching. This paradigm library is based on the semantic paradigm framework of GB / T 18391-2021 and contains about 200 types of typical semantic intention combinations. The matching method is role sequence matching and keyword coincidence degree evaluation (the coincidence degree threshold is set to 0.65). After successful matching, the corresponding semantic intention number is assigned. The generated data is recorded in a structured table, and the fields are intention ID, action word, semantic paradigm type, role combination structure, etc., and the output is semantic intention number data.

[0035] Step S235: Aggregate semantic units according to the semantic intention number data to obtain structured semantic aggregation data; perform encoding and archiving according to the structured semantic aggregation data and output it as deep semantic data.

[0036] In this embodiment, the semantic intent number data is grouped by fragment number, and the intent combination of the same paradigm type is extracted. The semantic unit merging operation is performed on the intent numbers in the same group, and the repeated role structure is eliminated and aggregated into a unified semantic intent expression template. The template fields include action paradigm, participating entity code, role coverage, and frequency of occurrence. The merged aggregation results are encoded and archived. The encoding method uses SHA-256 digest encryption to generate a unique identifier for the semantic template content, and the generated encoding field length is 64-bit hexadecimal string. Each deep semantic data record contains a semantic ID, a coding identifier, a semantic paradigm type, a fragment range, and an archiving timestamp. The final deep semantic data is stored in the form of a relational database. Each table can accommodate up to 10,000 aggregated semantic template records. Each record has a fixed structure and index field for subsequent intelligent agent workflow configuration and scheduling calls.

[0037] Step S24: performing implicit semantic detection based on the deep semantic data to obtain implicit semantic data; In this embodiment, all potential paths between nodes in the statistical semantic structure graph that are not directly marked in the original knowledge triples are counted. For each entity-attribute combination in the path, the expert knowledge base is queried to see whether there is a supplementary relationship with a similar structure but not explicitly marked in the current data. If it exists, it is regarded as a potential implicit semantic chain. At the same time, the empty connection nodes in the semantic structure graph (isolated entities without predecessors or successors) are checked to see whether there are adjacent relationship boundary values in their context (such as numerical changes exceeding the set threshold of 15%). Nodes that meet the above conditions are marked as "implicit causal nodes" or "abnormal trigger nodes". All cases of path structure abnormalities, interrupted entity reasoning chains, and unexplained numerical mutations are classified as implicit semantic expressions, and the original knowledge triple number, supplementary semantic information and implicit logical conditions are recorded to generate an implicit semantic data set.

[0038] It is particularly important that step S24 includes the following steps: Step S241: extracting a semantic pointing chain based on the deep semantic data to obtain semantic pointing chain data; In this embodiment, deep semantic data is first loaded, which contains information such as semantic unit identifiers, semantic role boundaries, and syntactic dependency relationships. For each semantic unit, the text is scanned word by word based on a rule matching algorithm to extract semantic pointing relationships. Specifically, the pointing chain between a verb and its corresponding object or modifier is identified. By constructing a directed graph structure, where nodes represent semantic units and edges represent pointing relationships, and the edge weights are calculated comprehensively from the dependency syntactic distance and word frequency statistics, a weighted threshold of 0.7 is used to filter out valid pointing chains. This step uses a custom parsing program to parse the deep semantic data, relying on a fixed syntactic rule library and dictionary to ensure the accuracy of the pointing chain construction, and finally outputs a semantic pointing chain data structure containing the start and end nodes of the link and the link weight.

[0039] Step S242: Extract context logical clues based on the semantic pointing chain data to obtain context logical clue data; In this embodiment, for each pointing chain, using the sliding window technique, the window size is set to 3 sentences, and the sentences within the window are used as the context scope. Sentence-level logical clues are identified through regular expressions and key logical word matching (including more than 20 logical connectives such as "if", "then", "because", "so", etc.). Combining the weight information in the semantic pointing chain, a logical connection matrix is established, and the similarity of matrix elements is calculated. A cosine similarity threshold of 0.65 is used to determine valid clues, generating a context logical clue data table with fields including logical connectives, start and end sentence numbers, weight values, etc., for use in subsequent semantic reasoning.

[0040] Step S243: Perform cross-sentence semantic trigger word matching based on the context logical clue data to obtain potential semantic trigger information; In this embodiment, a semantic trigger word library is pre-constructed, which includes about 1,500 trigger words and phrases related to implicit semantics, such as "cause", "lead to", "due to", "indicate", etc. Matching is performed by scanning the text sentence by sentence in combination with regular expressions, and the sentence numbers and context of the trigger word occurrences are recorded. For consecutive sentences, the relevance of the trigger words is verified based on word frequency statistics and word vector similarity (calculated by the cosine similarity in the word vector space, with a threshold set to 0.7), and potential semantic trigger information is filtered out. The matching results form a data set containing trigger word text, sentence positions, and semantic relevance scores, providing input for constructing the inference path.

[0041] Step S244: Construct an inference path based on the potential semantic trigger information to obtain implicit semantic relationship chain data; In this embodiment, the inference path adopts a chain inference structure and is recursively extended through dependency relationships. Each path consists of multiple semantic nodes and trigger words linked together. During the construction process, the maximum path length is set to 5, and paths exceeding this length are truncated. Using a graph traversal algorithm, starting from the trigger word starting node, it extends along the semantic pointing chain. The node weight comprehensively considers the syntactic dependency distance (threshold ≤ 3) and the context logic weight (threshold ≥ 0.6). After the path construction is completed, the path information is stored in a linked list structure, and the fields include the node serial number, the node semantic unit identifier, the trigger word and its weight, and the total path weight. This implicit semantic relationship chain data provides a structured basis for subsequent induction and aggregation.

[0042] Step S245: Perform semantic induction and aggregation based on the implicit semantic relationship chain data to obtain implicit semantic data.

[0043] In this embodiment, all path linked lists are iteratively compared, and a similarity clustering algorithm (based on the Jaccard similarity coefficient, threshold 0.75) is used to merge the semantic relationship chains to reduce redundant links. During the aggregation process, common semantic nodes and repeated trigger words are extracted, and the weights are merged and weighted averaged to generate an inductive semantic unit. Finally, the aggregation result is converted into a standardized implicit semantic data format, and the fields include the inductive semantic unit identifier, the associated path set, the weighted weight, and the text segment mapping. The data is stored in a JSON structure for easy parsing by subsequent agents and task scheduling.

[0044] Step S25: Construct a blueprint for the user requirement agent based on the implicit semantic data; construct an agent workflow based on the blueprint for the user requirement agent; In this embodiment, in each implicit semantic chain, there is a main entity (such as a control device, an electrical device, etc.) and multiple action behaviors it drives. This main entity is regarded as an agent unit in the blueprint. According to the action behavior type and the corresponding response attributes, each agent unit is mapped to a standard module component. The components include five categories: a sensing module, a decision-making module, an execution module, an exception response module, and a communication module. Each category is set with a fixed parameter template. For example, the sensing module is bound to input types such as temperature and voltage, and the decision-making module executes actions such as frequency increase and load reduction. Multiple agent units are generated into a network structure according to the connection order of the semantic links. All blueprint structures are organized in XML format, and the fields include Agent_ID, Function_ID, Trigger_Type, Precondition, Action, Output, etc. According to the logical connections between the modules in the blueprint, an agent workflow is automatically constructed. Each workflow node is bound to an Agent_ID, and its pre- and post-dependency relationships are clarified, and a standard DAG graph structure (directed acyclic graph) is output and stored as a JSON structure file to complete the construction of the agent workflow.

[0045] Step S26: Perform task execution scheduling according to the agent workflow to obtain task execution data.

[0046] In this embodiment, each workflow node is parsed into a task scheduling unit, and the fields include: task ID, input data type, required resource type (such as computing resources, sensor data), operation duration, priority level (1–5), and upstream and downstream dependent node IDs. Then, the task nodes are sorted in topological order, and a scheduling algorithm based on the greedy strategy is used to traverse the entire task chain sequentially starting from the source node. On the premise of ensuring no resource conflicts, high-priority nodes are preferentially scheduled and independent tasks are processed in parallel as much as possible. The minimum load first strategy is adopted in resource allocation, and among the currently idle resource nodes, the one with the minimum load is selected for task binding. The scheduling result is recorded in a scheduling log table, and the fields include Task_ID, Start_Time, End_Time, Assigned_Node, Upstream_Task, etc. Finally, complete task execution data is generated for subsequent topology construction operations.

[0047] Preferably, step S25 includes the following steps: Step S251: Extract user statement intention features according to the implicit semantic data to obtain user statement intention data; In this embodiment, the implicit semantic data includes text fragments, corresponding entity labels, potential causal relationships, and logical conditions. A text scanning method based on regular expressions is used to match the predefined user statement intention keyword library in the text fragment. The keyword library contains about 3000 specific keywords and phrases, such as "request", "query", "control", "monitor", etc., and each keyword corresponds to a unique number. Subsequently, a rule-based syntactic word segmentation technique is used to break down the text fragment into basic semantic units, ensuring that the word segmentation follows the rules of the linguistic subject-predicate-object structure. The dependency syntactic tree parsing algorithm is applied to each semantic unit to calculate the dependency relationship and word distance between each vocabulary, and clarify the core verb, execution object, and its modifiers. According to the parsing result, the action category number (ranging from 0 to 99), the object category number (ranging from 0 to 199), and the modifier intensity score are extracted. The modifier intensity is represented by a floating point number between 0 and 1, reflecting the emphasis degree of the modifier. At the same time, in combination with the context information, the timestamp and the start and end positions of the sentence in the original text are recorded. All the extracted intention features are summarized in a structured manner to form a user statement intention data table containing the fields intention ID, action code, object code, modifier intensity, and sentence start and end positions for subsequent semantic analysis and task execution process construction.

[0048] Step S252: Identify syntactic dependency relationships based on the user statement intention data to obtain syntactic dependency relationship data; In this embodiment, a statistical syntax analysis tool is used to divide the input text into syntactic components, obtaining the part-of-speech tags of the sentence components and the phrase structure tree. Then, through a dependency syntax analysis algorithm (such as a graph-based shortest path algorithm), the dependency relationships between words are calculated, including types such as subject-predicate, object-predicate, modifier-modified, etc. The dependency relationship data includes the relationship type number (e.g., 1 represents subject-predicate, 2 represents verb-object), the lexical IDs of the dependent head word and the dependent word, the dependency relationship weight (range 0.1 to 1.0, floating point number), and the path length of the dependency chain. The result data is stored in a directed graph structure, with nodes being lexical IDs and edges being dependency relationships, forming a syntactic dependency relationship data set.

[0049] Step S253: Based on the syntactic dependency relationship data, perform requirement slot filling to obtain structured requirement slot data; In this embodiment, a standardized requirement slot template is predefined. The template includes multiple slot categories, including action slots, target slots, time slots, and condition slots, etc. Each slot has a strictly defined field format and data type. The action slot is specifically used to store action codes, and the encoding range is limited to integers from 0 to 99. The target slot is used to store object codes, and the encoding range is limited to integers from 0 to 199. The time slot is designed to include two fields, respectively recording the start time and the end time, with the time format being standard UTC time and the precision reaching the second level. The condition slot is used to store logical judgment expressions, and the expressions are stored in string form, with the maximum character length limited to 256 characters. Relying on the path analysis algorithm of the syntactic dependency chain, the dependency relationships between actions and objects in the syntactic dependency chain are analyzed one by one. Based on the dependency path, the characteristics of actions and objects are accurately located and strictly matched according to the corresponding fields of the slot template. When the action code and the object code appear in the intention characteristics, the corresponding values are filled into the action slot and the target slot. If a certain field fails to find an effective match in the dependency relationship, a default null value is assigned to ensure data integrity. The time slot is filled by extracting timestamp information from the context and combining the explicit or implicit time range in the text. If there is no time information, it is set to null. The condition slot extracts the corresponding expression according to the logical condition description in the text, truncates it to within 256 characters, and then fills it into the slot. All filled slot data is uniformly integrated according to the slot name, filled content, filling credibility, and data source number, forming a structured requirement slot data table. The filling credibility is calculated by a rule engine, and the value range is limited between 0 and 1, reflecting the accuracy and reliability of the filled data. The data source number identifies the corresponding text or parsing module. Finally, the structured requirement slot data table is stored in a relational database table structure, supporting standard SQL queries and update operations to ensure the data access efficiency and accuracy in subsequent processing steps.

[0050] Step S254: Calibrate the task type according to the structured requirement slot data to obtain task identification data; perform capability module binding according to the task identification data to obtain capability module data; construct an intelligent agent blueprint for user requirements based on the capability module data; In this embodiment, the task type library contains 500 predefined task types, and each task type corresponds to a call list of a set of functional modules. First, perform a unique mapping match on the encodings of the action slot and the target slot in the requirement slots, and use an exact matching algorithm to retrieve the task type library. The matching threshold is set to 100% exact match. Only when the action encoding and the target encoding simultaneously match the corresponding task type, calibrate the task type. The task identification data fields include task ID, task name, and matching degree. Subsequently, according to the task function requirements corresponding to the task identification, select the required modules from the capability module library, which includes categories such as sensing modules, control modules, communication modules, and decision-making modules. Module binding is based on the function list required by the task type, strictly matching the module number and version number. The binding information fields include module ID, module name, version number, and interface specification. All bound modules are integrated to form a capability module data set, which uses the JSON format to describe the interface and data flow relationship between modules, facilitating subsequent blueprint construction.

[0051] Step S255: Construct an intelligent agent workflow according to the intelligent agent blueprint for user requirements.

[0052] In this embodiment, blueprint construction first generates a directed acyclic graph (DAG) using graph theory methods based on the functional dependency relationship between modules. The nodes represent capability modules, and the edges represent the data and control flow between modules. Each node is attached with module function description, input and output port definitions, interface protocol version, and processing delay parameter (unit: millisecond). During the construction process, based on the input dependency relationship of the modules, ensure that the data transfer path is complete and loop-free, and the topological relationship between nodes is stored using an adjacency matrix. The blueprint data structure uses the XML format, including module ID, function type, input interface, output interface, and dependency relationship. After the blueprint is completed, an intelligent agent workflow is automatically generated, converting the blueprint nodes into workflow task nodes. Each task node sets execution conditions, timeout time (in seconds), and defines the execution order between tasks. Finally, the intelligent agent workflow is stored in the form of a JSON file, including task ID, dependency relationship, execution time budget, and resource requirements, for subsequent scheduling use.

[0053] Preferably, the communication anomaly analysis based on the node explicit topology data in step S3 includes: Extract real-time communication data based on the node explicit topology data; In this embodiment, the explicit topological structure information of the import node is imported, and the topological structure is stored in the form of a node list and a connection relationship matrix between nodes. The real-time communication data acquisition system uses a network packet capture tool to perform packet capture on each communication link between nodes at intervals of 1 second in a timed polling manner. The captured content includes the packet sequence number, timestamp, source node ID, destination node ID, and packet size (in bytes). The captured data is stored in the cache database in the form of a time series, and the data record fields include the packet sequence number (integer), timestamp (accurate to milliseconds), source node identifier (string), destination node identifier (string), and packet length (integer). By using the predefined node connection relationship, eligible node pairs are screened from the topological matrix, and the communication data of each node pair is collected to achieve accurate extraction of real-time communication data.

[0054] Calculate the communication packet loss rate based on the real-time communication data; In this embodiment, the continuity of the collected packet sequence numbers is detected, and whether a packet loss event exists is determined based on the difference between the sequence numbers. The calculation window is set to 1 minute, and the difference between the actual number of received packets and the maximum sequence number that should be received within this time window is statistically calculated. The communication packet loss rate is calculated by dividing the difference between the actual number of received packets and the expected number of packets by the expected number of packets. The value range of the packet loss rate is limited to between 0 and 1, and the calculation result is updated every minute and stored in the time series database. The fields include the node pair identifier, the start time of the time window, and the packet loss rate value. The packet loss rate threshold is set to 0.05, and the time window exceeding this threshold is marked as an abnormal packet loss time period.

[0055] Statistically calculate the communication packet loss time period based on the communication packet loss rate to obtain the communication packet loss rate time data; In this embodiment, when performing the step of statistically calculating the communication packet loss time period based on the communication packet loss rate, the time windows with a packet loss rate exceeding 0.05 for more than one minute in a row are merged to form continuous communication packet loss time periods. Each time period consists of a start time and an end time. All abnormal time periods are statistically calculated to generate a communication packet loss rate time data set. This data set is saved in tabular form, and the fields include the start time (in UTC time format, accurate to seconds), end time, node pair identifier, and peak packet loss rate. This table supports queries according to the time sequence and node pairs to ensure the accurate invocation of abnormal time periods in subsequent analysis steps.

[0056] Calculate the packet out-of-order rate based on the real-time communication data; In this embodiment, in the step of calculating the packet disorder rate according to the real-time communication data, for the communication data of each node pair, the deviation between the packet arrival order and the transmission sequence number is analyzed. A disorder event is defined as the case where the sequence number of the later-arriving packet is less than that of the earlier-arriving packet. The number of disorder events per minute and the total number of received packets are counted. The disorder rate is calculated by dividing the number of disorder events by the total number of received packets, and the value range is from 0 to 1. It is calculated once per minute and stored in the time-series database. The recorded fields include the node pair identifier, the start time of the time window, and the disorder rate. The disorder rate threshold is set to 0.03, and the time window exceeding the threshold is determined as an abnormal disorder time period.

[0057] Based on the packet disorder rate, the packet disorder time period is statistically analyzed to obtain the packet disorder time data; In this embodiment, time windows with a disorder rate exceeding 0.03 for one minute or more are merged to generate continuous disorder time periods. Each period is defined by the start time and the end time, forming a packet disorder time data set. This set is saved in a standard table format, and the fields include the start time, the end time, the node pair identifier, and the peak disorder rate, supporting retrieval based on time and node pair.

[0058] Based on the communication packet loss rate time data and the packet disorder time data, an intersection operation of the transmission abnormal time is performed to obtain the communication transmission abnormal data; In this embodiment, an overlapping analysis of time periods is performed on the two time data sets. Using the time interval intersection algorithm, the intersection interval of the packet loss time period and the disorder time period is calculated. The intersection interval time length needs to be greater than or equal to 30 seconds to be included in the effective abnormal time period. In this way, the communication abnormal time periods with both packet loss and disorder are screened out to generate a communication transmission abnormal data set. This data set is saved in the form of a time period list, and the fields include the start and end times of the abnormal time, the node pair identifier involved, and the abnormal duration, for subsequent abnormal detection.

[0059] Based on the communication transmission abnormal data, node communication jitter abnormal detection is performed to obtain the node communication abnormal data.

[0060] In this embodiment, the communication transmission abnormal data obtained previously is called. This data includes the start time and end time of multiple abnormal time periods, as well as the corresponding node pair identifiers. For each abnormal time period, the arrival timestamp sequence of all communication data packets passing through this node pair within this time period is extracted, and the timestamp is accurate to milliseconds. The calculation process uses the time interval between adjacent packet arrivals within this time period as the basic unit. Specifically, it is the arrival time of the latter packet minus the arrival time of the former packet to form a time interval sequence. Subsequently, the standard deviation of this time interval sequence is calculated as the communication delay jitter index. The standard deviation calculation uses the unbiased estimation method, and the degrees of freedom in the formula is the sample size minus one to ensure the statistical rationality of the calculation result. The time window strictly corresponds to the start and end times of the abnormal time period to avoid confusion across time periods. The jitter value is stored in milliseconds, and the record is accurate to three decimal places. The set threshold is 50 milliseconds, and the jitter value exceeding this threshold is marked as an abnormal state. The threshold is derived from network communication standards and actual measurement experience and can effectively reflect the delay fluctuation of the communication link. The calculation result for each abnormal time period includes the node pair identifier, the start and end times of the abnormal time period, the calculated jitter value, and the jitter abnormal state flag (boolean type, 1 indicates abnormal, 0 indicates normal). All results are summarized to form a node communication abnormal data set. The data structure uses a relational table and supports retrieval according to time, node pair, and abnormal state. This data set not only records the abnormal time period but also details the specific numerical changes in jitter, providing accurate quantitative indicators and an abnormal alarm basis for subsequent network maintenance, link optimization, and node resource scheduling. <> <>

[0061] Preferably, in step S3, identifying the abnormal communication link based on the node communication abnormal data includes: <> Performing repeated request analysis based on the node communication abnormal data to obtain repeated request abnormal access data; <> In this embodiment, the key fields of each communication record are extracted from the node communication abnormal data, including the source IP address of the request, the destination IP address, the request type identifier, the request timestamp, and the request sequence number. For the extracted data, an index structure based on a hash table is established, and the same request identifier (composed of the source IP, destination IP, and request type) is mapped to the corresponding time series list. The set time window is 60 seconds, and for all requests within this time window, the number of occurrences of the same request identifier is calculated. If the number of occurrences of the same request identifier exceeds 5 times within this time window, it is determined as a repeated request anomaly. The repeated request abnormal access data is stored in a structured table form, and the fields include the request identifier, the start and end times of the time window, the number of repeated requests, and the request summary information. The storage format is a relational database table for easy subsequent query and analysis. <> <>

[0062] Performing access surge statistics based on the repeated request abnormal access data to obtain access surge data; <> In this embodiment, the access traffic data of each node is divided by time period, and the length of the time period is fixed at 10 seconds. For each time period, the total number of all requests received by the node is counted to form time-series access data. The sliding window algorithm is adopted, and the length of the sliding window is 6 time periods (i.e., 60 seconds). Calculate the ratio of the number of accesses in the current time period to the average number of accesses in the past 5 time periods within the sliding window. When the ratio exceeds 3, the current time period is marked as a sudden increase period of access. The access sudden increase data is stored in fields including the start and end times of the time period, the node identifier, the total number of accesses, and the sudden increase ratio. The format adopts the standard format of the time-series database, supporting efficient retrieval and statistics.

[0063] Based on the access sudden increase data, a denial-of-service attack detection is performed to obtain denial-of-service attack data; In this embodiment, for each time period marked as a sudden increase in access, the total number of requests and the proportion of repeated requests within this time period are counted. The denial-of-service attack threshold is set such that the total number of requests per unit time is greater than 1000 times and the proportion of repeated requests exceeds 20%. The time periods that meet this condition are determined as denial-of-service attack events. The denial-of-service attack data includes the start and end times of the attack time period, the affected node identifier, the total number of requests, the proportion of repeated requests, and the event status flag, and is stored in the structured security event log format.

[0064] Based on the denial-of-service attack data, an intrusion behavior detection is performed to obtain intrusion behavior data; In this embodiment, the source IP address information of all requests during the denial-of-service attack event is collected, and the source IP is matched using a pre-set blacklist IP library. The blacklist library contains more than 100,000 malicious IP addresses, and the matching operation is implemented through a hash index for fast retrieval. If a blacklist IP access occurs during the denial-of-service attack time period, an intrusion behavior event is generated. The fields of the intrusion behavior data include the intrusion event time, the node identifier, the attack source IP address, the intrusion type identifier, and the blacklist library entry number, and are stored in a format compatible with the Security Information and Event Management (SIEM) system.

[0065] According to the intrusion behavior data, the communication link protocol identification is statistically analyzed to obtain communication link protocol data; In this embodiment, all network packets within the intrusion event time window are collected, and the protocol type fields such as TCP, UDP, HTTP, HTTPS, etc. are extracted. The packets are classified and statistically analyzed for the protocol, and the percentages of the number of packets and the transmitted byte count occupied by each protocol are calculated. The protocol type with a protocol ratio exceeding 50% is recorded as the main protocol. The fields of the communication link protocol data include the start and end times of the time window, the node pair identifier, the protocol type, and the protocol ratio, and the data is in the structured protocol analysis report format.

[0066] Perform version anomaly analysis based on communication link protocol data to obtain protocol version anomaly data; In this embodiment, first extract protocol version fields, such as TLS version number, HTTP protocol version, etc. Establish a version whitelist that includes all normally allowed protocol versions. Count the occurrence frequencies of each version within a time window. Versions with a frequency lower than 1% or versions not in the whitelist are marked as abnormal versions. If the number of occurrences of an abnormal version exceeds 10 times, it is recorded as a version anomaly event. The protocol version anomaly data includes the time window, node pair identifier, protocol type, abnormal version number, and the number of abnormal occurrences, and the format adopts the security event log standard.

[0067] Identify abnormal communication links based on the protocol version anomaly data to obtain abnormal communication link data.

[0068] In this embodiment, comprehensively analyze all version anomaly events, filter out the cases of abnormal versions that continuously appear in more than 3 time windows, and count the number of abnormal node pairs involved. If the number of abnormal node pairs exceeds 2 and the abnormal version continuously appears for more than 30 seconds, then determine that this link is an abnormal communication link. The abnormal communication link data includes the start and end nodes of the link, the start and end of the abnormal time period, the details of the abnormal version, and the link status flag, and adopts the network topology anomaly report format to provide data support for subsequent network security protection and node resource scheduling.

[0069] Preferably, the node resource scheduling based on the abnormal communication link data in step S3 includes: Statistically analyze the node resource status based on the abnormal communication link data to obtain node resource status data; In this embodiment, extract the identification information of all nodes involved from the abnormal communication link data, and combine it with the resource usage conditions of each node collected by the real-time monitoring system, including indicators such as CPU utilization rate (percentage, value range 0 to 100), memory occupancy (unit: MB), network bandwidth usage rate (percentage), disk IO latency (unit: milliseconds), and node queue length. The collection frequency is set to once per second, and the statistical period is the data within the most recent five minutes. Perform time series analysis on the node resource data to calculate statistical quantities such as the mean, maximum value, and standard deviation. Aggregate the above indicators by node to generate a node resource status data table, the fields of which include node ID, CPU utilization rate mean, memory occupancy mean, bandwidth usage rate maximum value, disk IO average latency, and queue length mean, and store it in a relational database for subsequent analysis and call.

[0070] Determine resource bottleneck nodes based on the node resource status data to obtain bottleneck node data; In this embodiment, the resource status data of all nodes is determined for bottlenecks. Specific thresholds are set as follows: when the CPU utilization rate exceeds 85%, the memory occupancy exceeds 90% of the total capacity, the network bandwidth utilization rate exceeds 80%, the disk I / O latency exceeds 50 milliseconds, and the queue length exceeds 100 items, it is regarded as a resource tension indicator. For each node, it is judged item by item whether it exceeds the threshold, and the number of exceeded indicators is counted. The node with the largest number of exceeded indicators and the most serious exceeding (such as the CPU utilization rate reaching more than 95% or the memory occupancy reaching more than 95%) is identified as the bottleneck node. The bottleneck node data includes the node ID, the type of exceeded resource, the actual values of each resource indicator, and the exceeding level (mild, moderate, severe). The data format is stored as a structured table, supporting filtering and sorting operations.

[0071] Based on the bottleneck node data, task migration analysis is performed to obtain task migration data; In this embodiment, the list of tasks to be migrated is determined. The tasks include the task ID, the task resource requirements (number of CPU cores, memory size, network bandwidth requirements), and the ID of the current node where the task is located. The tasks on the high-load nodes are screened out using the bottleneck node data, and the resource status of the target node after the task migration is analyzed. The target node needs to meet the task resource requirements and the resource utilization rate is lower than the corresponding threshold (CPU lower than 70%, memory lower than 75%). The task migration analysis uses an exhaustive method to traverse the list of target nodes, and calculates the resource matching degree and load balancing indicators for each migration plan. The task migration data is output, including the task ID, the source node ID, the target node ID, the resource matching score, and the load balancing improvement value. The results are stored as a task migration plan list.

[0072] Based on the task migration data, the migration communication overhead is calculated to obtain the migration communication overhead data; In this embodiment, first, the size of the data volume (in MB) and the estimated migration duration (in seconds) involved in the task migration process are obtained. The data volume comes from the task memory snapshot size and the dependent file transfer volume. The communication bandwidth data is obtained from the network bandwidth utilization rate and the maximum bandwidth value of the target node. The communication overhead calculation formula is: migration communication overhead = data volume / available bandwidth, in seconds. Considering the network latency factor, the network latency is based on the average latency in the past five minutes, in milliseconds, and is converted to seconds and added to the migration time. The final migration communication overhead data includes the task ID, the start and end nodes of the migration, the data volume size, the bandwidth rate, the network latency, and the total communication overhead value. The data format is a structured report, which is convenient for sorting and filtering.

[0073] According to the migration communication overhead data, node scheduling path analysis is performed to obtain the node scheduling path data; In this embodiment, using the communication topology diagram and task dependencies, a path search algorithm based on graph traversal is adopted to traverse all scheduling paths from the source node to the target node. The communication overhead of each path is the sum of the communication overheads between all nodes on the path. The maximum allowable communication overhead threshold is set to 30 seconds, and paths exceeding the threshold are filtered. For paths meeting the threshold, the path load balance degree and failure risk level (based on the historical communication anomaly rate) are calculated. The final node scheduling path data records the path ID, node sequence, total communication overhead, load balance degree, and failure risk score, and is stored in the path database.

[0074] Node resource scheduling is performed based on the node scheduling path data to obtain node resource scheduling data.

[0075] In this embodiment, the path with the lowest communication overhead and the highest load balance degree is selected from the scheduling path data as the preferred scheduling path. Combining the bottleneck node information, the task assignment is adjusted, and tasks are preferentially assigned to non-bottleneck nodes or nodes with lower loads. During the scheduling process, the node resource status is updated in real time, and the node task load ratio is adjusted to ensure that the node resource utilization rate is maintained within a safe range. The node resource scheduling data includes the node ID, assigned task list, task resource occupancy details, current resource utilization rate, and scheduling timestamp, and is recorded in a structured log format, providing accurate node load information for the intelligent agent to optimize the decision-making path.

[0076] Preferably, step S4 includes the following steps: Step S41: Extract node status features from the node resource scheduling data to obtain scheduling node status data; In this embodiment, node status features are extracted from the node resource scheduling data. The node resource scheduling data includes the CPU usage rate (percentage, 0 - 100%), memory usage (in MB), network bandwidth utilization rate (percentage), disk read and write rate (MB / s), current task queue length (number of tasks), node response time (in milliseconds), and node operation error rate (percentage) of each node during the scheduling process. By collecting time-series data with a sampling period of 1 second and using a 5-minute sliding time window, the statistical features of each index, including the mean, maximum value, minimum value, and standard deviation, are calculated as the input parameters of the node status features. For abnormal indexes, such as the time period when the error rate is greater than 2% or the response time exceeds 500 ms, an additional abnormal status field is marked. All feature fields are integrated according to the node ID to form a structured scheduling node status data set, and the fields are stored in a table form, including node identification, mean value, peak value, fluctuation range of each resource index, and abnormal status mark.

[0077] Step S42: At the macro level, use a reinforcement learning algorithm to predict the path to achieve the long-term business goals for the scheduling node status data, and generate a quarterly-level optimization strategy at the macro level; In this embodiment, a reinforcement learning algorithm is applied based on the scheduling node status data to achieve quarterly-level business goal path prediction. The state space is composed of the node status feature vectors in Step S41, and the action space is defined as the set of node resource scheduling strategies, including resource priority adjustment, load distribution ratio adjustment, etc. The reward function is set with the system throughput (unit: tasks per second), average response time (milliseconds), and resource utilization rate (percentage) within a quarter as evaluation indicators, and the comprehensive weighting coefficients are 0.5, 0.3, and 0.2 respectively. The discount factor γ = 0.95 is used. During the training process, an experience replay mechanism is used to store the most recent 50,000 state-action-reward records, and the target network parameters are synchronously updated every 1000 steps. The reinforcement learning model uses a value function approximation method, processes the input state through a deep neural network, and outputs the optimal action strategy. The model output is a quarterly-level optimization strategy file, and the specific content includes the resource scheduling priority list between nodes and the resource allocation ratio matrix, which is saved in the standard JSON format for subsequent calls.

[0078] Step S43: At the tactical level, use the proximal policy optimization algorithm to configure the workflow nodes for the scheduling node status data, and generate tactical-level configuration data; In this embodiment, for the scheduling node status data, the proximal policy optimization (PPO) algorithm is applied for workflow node configuration. The state input is the node resource status feature and the current workflow task status information, and the action space is defined as node task allocation, task start order adjustment, and load balancing parameter setting. The training parameters include the clipping coefficient ε = 0.2, the learning rate 0.0003, and the batch size 64. A multi-layer perceptron network structure is used, including two hidden layers with 256 and 128 nodes respectively, and the activation function is ReLU. During the training process, a multi-threaded environment is used to accelerate sample collection to ensure the timeliness of policy updates. The output is tactical-level configuration data, and the format includes the task allocation table, the workflow execution order, and the node load threshold, which is saved as a relational database record for easy real-time call and update by the scheduling system.

[0079] Step S44: At the execution level, apply a deep Q-network to configure the API order for the scheduling node status data, and generate execution-level configuration data; In this embodiment, a Deep Q-Network (DQN) is applied to adjust and optimize the API call order in the scheduling node status data. The input state includes the current task execution status of the node, the API call queue, and historical call latency data. The action space is an adjustable sequence of API call orders. The reward function is designed to minimize the total task completion time and maximize the API call success rate, with a weight ratio of 0.7 and 0.3. The capacity of the experience replay buffer is set to 10,000 entries. In the ε-greedy policy, ε linearly decays from 1.0 to 0.1, and the target network is updated every 1,000 steps. The neural network consists of three fully connected layers with 512, 256, and 128 nodes respectively, and the activation function is ReLU. After training, the execution layer configuration data is output, specifically the API call order adjustment plan, accompanied by a timestamp and an estimated response time, and stored in a structured form for easy invocation by the scheduling execution module.

[0080] Step S45: Construct a reinforcement learning decision engine based on the macro-level quarterly optimization strategy, the tactical layer configuration data, and the execution layer configuration data; In this embodiment, a reinforcement learning decision engine is constructed by integrating the macro-level quarterly optimization strategy, the tactical layer configuration data, and the execution layer configuration data. First, the three-layer data is standardized to unify the field format and unit, and the Z-score normalization method is used to process numerical fields. The weighted fusion algorithm is used to merge the policy parameters of each layer, and the weight distribution is 0.5 for the macro layer, 0.3 for the tactical layer, and 0.2 for the execution layer. The weights are adjusted regularly according to the historical scheduling performance feedback results. The fusion result forms a multi-level decision parameter set, which is input to the core module of the decision engine. This module includes three sub-modules: state evaluation, action selection, and policy execution, supporting real-time data input and online policy fine-tuning. The decision engine outputs a multi-agent coordinated scheduling plan in the format of node task paths, resource allocation ratios, and scheduling priorities, which is saved as an executable scheduling instruction set to support the invocation of the dynamic scheduling system.

[0081] Step S46: Optimize the multi-agent decision path according to the reinforcement learning decision engine to obtain decision path optimization data.

[0082] In this embodiment, based on the input scheduling strategy and node status, a path search algorithm combined with a reinforcement learning strategy is used to dynamically plan the task paths of each agent, considering the inter-node communication latency (value range 0 - 200ms), the resource load threshold (not exceeding 80% CPU utilization), and the task priority (numerical value 1 - 10). The path adjustment is based on real-time feedback, with an adjustment period of 5 seconds, and the path data is efficiently transmitted through an incremental update mechanism. The output decision path optimization data includes agent node IDs, task path sequences, resource allocation ratios, and estimated completion times (in seconds), stored in JSON format to support the docking of the scheduling execution system and achieve the coordination and efficient operation of the multi-agent system.

[0083] Preferably, step S46 includes the following steps: Step S461: Extract decision weight parameters according to the reinforcement learning decision engine to obtain the engine decision weight parameters; In this embodiment, by accessing the weight matrix file of the decision engine, the file contains the weight values in the multi-layer network. The weight values are represented as floating-point numbers, ranging from -1.0 to 1.0, and the precision is reserved to six decimal places. The weight parameters include state input weights, action output weights, and hidden layer connection weights. Use a matrix parsing tool to parse the weight matrix layer by layer, and extract the weight parameters corresponding to key indicators such as node resource utilization, task priority, and communication delay. Normalize the weight parameters, using the maximum-minimum normalization method to linearly map the weight values to between 0 and 1. The weight extraction results are saved in the form of a structured array, and the fields include weight name, layer index, and normalized value, forming an engine decision weight parameter data set for subsequent policy network model construction and invocation.

[0084] Step S462: Construct a policy network model based on the engine decision weight parameters, and perform decision inference operations according to the policy network model to obtain decision inference data; In this embodiment, determine the number of layers of the policy network, the number of nodes in each layer, and the type of activation function according to the weight parameters. The network structure uses three fully connected layers. The number of input layer nodes corresponds to the input state feature dimension. The middle two layers are set to 128 and 64 nodes respectively, and the activation function is ReLU. The number of output layer nodes is the same as the action space size, and Softmax activation is used to output the decision probability distribution. When initializing the network parameters, use the normalized weight parameters in step S461 to assign the corresponding connection weights. Subsequently, perform the decision inference operation of the policy network. The input is the current system state feature vector, including node load rate (percentage, 0-100), task urgency (integer from 1 to 10), and communication delay (milliseconds). Calculate the output action probability through forward propagation. The inference result includes the probability values of each action, which are saved as decision inference data. The data format is a two-dimensional floating-point array, containing action IDs and corresponding probabilities, and supports dynamic scheduling system invocation.

[0085] Step S463: Perform intelligent agent time series conflict detection according to the decision inference data to obtain conflict annotation path data; In this embodiment, the task time window information and the action time series are used to detect the time overlap and resource contention conflicts between different agent tasks. Specifically, by constructing a timing graph, the task execution time period of each agent is mapped to a time interval, and the overlapping part of the intervals is detected. A conflict threshold is set, and if the time overlap exceeds 50 milliseconds, it is determined as a conflict. For all conflicting task pairs, conflict annotations are recorded, including agent ID, conflicting task ID, start and end of the conflict time period, and conflict resource type (CPU, memory, communication link). The conflict annotation path data is stored in a table form, and the fields include path ID, task sequence, conflict flag, conflict level (an integer from 1 to 5, with 1 being the lowest), for subsequent conflict resolution scheduling.

[0086] Step S464: Perform decision priority sorting based on the conflict annotation path data to obtain decision priority data; In this embodiment, the sorting rule adopts the multi-factor weighting method, and the weight coefficients include task urgency (0.4), conflict severity (0.35), and historical task execution success rate (0.25). The task urgency is based on the scheduling input parameters, the conflict severity is based on the conflict level value, and the execution success rate is calculated by statistical analysis of historical scheduling records, with the success rate ranging from 0 to 1. The priority scores are calculated for all conflict paths according to the above weighted scores, and the higher the value, the higher the priority. The sorting result is output as a priority list, including task ID, agent ID, priority score, and sorting sequence number, and is saved as a structured data file (CSV format), supporting real-time access and adjustment by the scheduling system.

[0087] Step S465: Optimize the multi-agent decision path according to the decision priority data to obtain decision path optimization data.

[0088] In this embodiment, according to the priority order, combined with the node resource status and communication link topology information, the task execution path of each agent is planned. The path planning considers the node load limit (CPU utilization not exceeding 85%), the link bandwidth limit (minimum 100 Mbps), and the dependency relationship between tasks. The depth-first search algorithm with constraints is used for path search to ensure that high-priority tasks obtain the optimal node and time period resource allocation, while avoiding conflicts with other agents. The path optimization output includes agent ID, task sequence, node allocation plan, time node mapping, and resource allocation ratio. The final decision path optimization data is stored in JSON format, and the fields cover agent identification, path node list, execution timestamp, and resource usage details, for the scheduling execution module to call to achieve the efficient coordinated operation of the multi-agent system.

[0089] Therefore, from any perspective, the embodiments should be regarded as exemplary and non-limiting. The scope of the present invention is defined by the appended claims rather than the above description. Thus, all changes falling within the meaning and scope of the equivalent elements of the application documents are intended to be encompassed within the present invention.

[0090] The above are only specific embodiments of the present invention, enabling those skilled in the art to understand or implement the present invention. Various modifications to these embodiments will be obvious to those skilled in the art. The general principles defined herein can be implemented in other embodiments without departing from the spirit or scope of the present invention. Therefore, the present invention will not be limited to these embodiments shown herein, but rather will conform to the broadest scope consistent with the principles and novel features invented herein.

Claims

1. An intelligent agent decision path optimization method that integrates explicit topology and implicit semantics, characterized in that It includes the following steps: Step S1: Obtain multimodal data; Perform document parsing based on the multimodal data to obtain document parsing data; Perform knowledge injection based on the document parsing data to obtain knowledge injection data; Step S2: Perform implicit semantic detection based on the knowledge injection data to obtain implicit semantic data; construct an intelligent blueprint for user requirements based on the implicit semantic data; construct an intelligent agent workflow based on the intelligent blueprint for user requirements; perform task execution scheduling based on the intelligent agent workflow to obtain task execution data; Step S3: Construct an explicit node topology based on the task execution data to obtain explicit node topology data; perform communication anomaly analysis based on the explicit node topology data to obtain node communication anomaly data; Identify abnormal communication links based on the node communication anomaly data to obtain abnormal communication link data; perform node resource scheduling based on the abnormal communication link data to obtain node resource scheduling data; Step S4: Construct a reinforcement learning decision engine based on the node resource scheduling data; Perform multi-agent decision path optimization based on the reinforcement learning decision engine to obtain decision path optimization data.

2. The intelligent agent decision path optimization method that fuses explicit topology and implicit semantics according to claim 1, wherein Step S1 includes the following steps: Step S11: Obtain multimodal data and perform data preprocessing to obtain preprocessed multimodal data; Step S12: Identify text regions based on the preprocessed multimodal data to obtain text region data; perform semantic chunking based on the text region data to obtain semantic chunk data; perform document slicing based on the semantic chunk data to obtain document slice data; Step S13: Perform image segmentation based on the preprocessed multimodal data to obtain image segmentation region data; perform object detection based on the image segmentation region data to obtain image object data; perform object attribute recognition based on the image object data to obtain object attribute data; Step S14: Perform vector quantization encoding based on the document slice data to obtain semantic vector data; perform feature encoding based on the object attribute data to obtain object attribute encoding data; Step S15: Perform multimodal alignment processing based on the semantic vector data and the object attribute encoding data to obtain multimodal alignment data; perform semantic parsing based on the multimodal alignment data to obtain document parsing data; Step S16: Perform knowledge injection based on the document parsing data to obtain knowledge injection data.

3. The intelligent agent decision path optimization method that fuses explicit topology and implicit semantics according to claim 2, characterized in that, Step S16 includes the following steps: Step S161: Extract semantic unit features based on the document parsing data to obtain semantic unit data; Step S162: Construct a knowledge representation vector based on the semantic unit data to obtain knowledge representation vector data; Step S163: Perform semantic enhancement based on the knowledge representation vector data to obtain semantic enhancement data; Step S164: Perform expert knowledge base matching based on the semantic enhancement data to obtain expert knowledge base matching data; Step S165: Perform knowledge injection based on the expert knowledge base matching data to obtain knowledge injection data.

4. The intelligent agent decision path optimization method that integrates explicit topology and implicit semantics according to claim 1, wherein Step S2 includes the following steps: Step S21: Perform context feature perception based on the knowledge injection data to obtain context feature data; Step S22: Perform semantic association analysis based on the context feature data to obtain semantic association data; Step S23: Perform deep semantic recognition based on semantic association data to obtain deep semantic data; Step S24: Perform implicit semantic detection based on the deep semantic data to obtain implicit semantic data; Step S25: Construct an intelligent agent blueprint for user requirements based on the implicit semantic data; construct an intelligent agent workflow based on the intelligent agent blueprint for user requirements; Step S26: Perform task execution scheduling according to the intelligent agent workflow to obtain task execution data.

5. The intelligent agent decision path optimization method that fuses explicit topology and implicit semantics according to claim 4, characterized in that, Step S25 includes the following steps: Step S251: Extract user statement intention features from the implicit semantic data to obtain user statement intention data; Step S252: Identify syntactic dependency relationships based on the user statement intention data to obtain syntactic dependency relationship data; Step S253: Perform requirement slot filling based on the syntactic dependency relationship data to obtain structured requirement slot data; Step S254: Calibrate the task type according to the structured requirement slot data to obtain task identification data; perform ability module binding based on the task identification data to obtain ability module data; construct an intelligent agent blueprint for user requirements based on the ability module data; Step S255: Construct an intelligent agent workflow according to the intelligent agent blueprint for user requirements.

6. The intelligent agent decision path optimization method that fuses explicit topology and implicit semantics according to claim 1, wherein The communication anomaly analysis based on the node explicit topology data in Step S3 includes: Extract real-time communication data based on the node explicit topology data; Calculate the communication packet loss rate according to the real-time communication data; Statistically analyze the communication packet loss time period based on the communication packet loss rate to obtain communication packet loss rate time data; Calculate the packet out-of-order rate according to the real-time communication data; Statistically analyze the packet out-of-order time period based on the packet out-of-order rate to obtain packet out-of-order time data; Perform an intersection operation on the transmission anomaly time based on the communication packet loss rate time data and the packet out-of-order time data to obtain communication transmission anomaly data; Perform node communication jitter anomaly detection based on the communication transmission anomaly data to obtain node communication anomaly data.

7. The intelligent agent decision path optimization method that fuses explicit topology and implicit semantics according to claim 1, wherein The identification of abnormal communication links based on the node communication anomaly data in Step S3 includes: Perform repeated request analysis based on the node communication anomaly data to obtain repeated request abnormal access data; Perform access surge statistics according to the repeated request abnormal access data to obtain access surge data; Perform denial-of-service attack detection based on the access surge data to obtain denial-of-service attack data; Perform intrusion behavior detection based on the denial-of-service attack data to obtain intrusion behavior data; Statistically analyze the communication link protocol identification according to the intrusion behavior data to obtain communication link protocol data; Perform version anomaly analysis based on the communication link protocol data to obtain protocol version anomaly data; Identify abnormal communication links according to the protocol version anomaly data to obtain abnormal communication link data.

8. The intelligent agent decision path optimization method that fuses explicit topology and implicit semantics according to claim 1, wherein The node resource scheduling based on the abnormal communication link data in Step S3 includes: Statistically analyze the node resource status based on the abnormal communication link data to obtain node resource status data; Determine the resource bottleneck nodes according to the node resource status data to obtain bottleneck node data; Perform task migration analysis based on the bottleneck node data to obtain task migration data; Calculate the migration communication overhead based on the task migration data to obtain migration communication overhead data; Perform node scheduling path analysis according to the migration communication overhead data to obtain node scheduling path data; Node resource scheduling is performed according to the node scheduling path data to obtain node resource scheduling data.

9. The intelligent agent decision path optimization method that fuses explicit topology and implicit semantics according to claim 1, wherein Step S4 includes the following steps: Step S41: Extract node status features from the node resource scheduling data to obtain scheduled node status data; Step S42: At the macro level, use a reinforcement learning algorithm to predict the path to achieve long-term business goals for the scheduled node status data, and generate a quarterly-level optimization strategy at the macro level; Step S43: At the tactical level, use the proximal policy optimization algorithm to configure the workflow nodes for the scheduled node status data, and generate tactical-level configuration data; Step S44: At the execution level, apply a deep Q network to adjust the API order for the scheduled node status data, and generate execution-level configuration data; Step S45: Construct a reinforcement learning decision engine based on the quarterly-level optimization strategy at the macro level, the tactical-level configuration data, and the execution-level configuration data; Step S46: Optimize the multi-agent decision path according to the reinforcement learning decision engine to obtain decision path optimization data.

10. The intelligent agent decision path optimization method that fuses explicit topology and implicit semantics according to claim 9, wherein Step S46 includes the following steps: Step S461: Extract decision weight parameters according to the reinforcement learning decision engine to obtain engine decision weight parameters; Step S462: Construct a policy network model based on the engine decision weight parameters, and perform decision inference operations according to the policy network model to obtain decision inference data; Step S463: Detect the temporal conflicts of agents according to the decision inference data to obtain conflict annotation path data; Step S464: Sort the decision priorities based on the conflict annotation path data to obtain decision priority data; Step S465: Optimize the multi-agent decision path according to the decision priority data to obtain decision path optimization data.

Images