Multidimensional Behavioral Modeling and Risk Quantification Methods and Systems for Enterprise AI Agents

Abstract

This application provides a method and system for multi-dimensional behavioral modeling and risk quantification of enterprise AI agents, relating to the field of artificial intelligence technology. First, it collects behavioral data streams generated by the enterprise AI agent, including records of behavioral events and their temporal relationships. Then, it inputs this data into a multi-dimensional behavioral decoupling network to extract intent-driven and environment-responsive behavioral features. Next, it constructs a dynamic interaction mapping relationship between the two, determines driving weights and feedback correction parameters, and then iteratively fuses these parameters to generate a behavioral evolution path feature sequence. Finally, it performs behavioral pattern clustering on the behavioral evolution path feature sequence and matches it with a pre-set risk behavioral feature library to generate a behavioral risk quantification assessment result. This invention can scientifically and accurately assess the behavioral risks of enterprise AI agents, ensuring their safe and stable operation.

Description

Technical Field

[0001] This application relates to the field of artificial intelligence technology, and more specifically, to a method and system for multidimensional behavioral modeling and risk quantification of enterprise AI agents. Background Technology

[0002] In the process of enterprise intelligent development, enterprise AI agents are widely used in various business scenarios, and the complexity and diversity of their behaviors are becoming increasingly prominent. The behavior of AI agents is not only driven by their own internal logic and algorithms, but also involves frequent and complex interactions with the external environment. Accurately understanding and evaluating the behavior of enterprise AI agents is crucial for ensuring operational security, optimizing business processes, and improving decision-making quality.

[0003] However, current research on the behavior of enterprise AI agents has many shortcomings. On the one hand, traditional methods often focus only on the single behavioral performance of AI agents, making it difficult to comprehensively capture the multi-dimensional characteristics of their behavior. For example, they may only analyze the efficiency of task execution while ignoring characteristics such as adaptability when interacting with the external environment. On the other hand, risk assessment of AI agent behavior lacks scientific and systematic quantitative methods, relying heavily on subjective judgment or simple rule matching. This fails to accurately reflect the intrinsic relationship between behavior and risk, and is insufficient to meet the needs of enterprises for precise control of AI agent behavior in complex and ever-changing environments. Summary of the Invention

[0004] In view of this, the purpose of this application is to provide a method and system for multi-dimensional behavioral modeling and risk quantification of enterprise AI agents.

[0005] According to a first aspect of this application, a multi-dimensional behavioral modeling and risk quantification method for enterprise AI agents is provided, the method comprising:

[0006] Collect the agent behavior data stream generated by the enterprise AI agent during operation. The agent behavior data stream includes records of behavioral events generated when the agent interacts with the external environment and the temporal correlation between the behavioral event records.

[0007] The agent behavior data stream is input into a pre-constructed multidimensional behavior decoupling network for behavior feature separation processing. The intention-oriented behavior features of the agent are extracted from the behavior event records by the intention decoder in the multidimensional behavior decoupling network, and the environment-responsive behavior features of the agent are extracted from the behavior event records by the environment response decoder in the multidimensional behavior decoupling network.

[0008] A dynamic interaction mapping relationship is constructed between the intent-oriented behavioral features and the environment-responsive behavioral features. The dynamic interaction mapping relationship includes the driving weight parameters of the intent-oriented behavioral features on the environment-responsive behavioral features and the feedback correction parameters of the environment-responsive behavioral features on the intent-oriented behavioral features.

[0009] The intention-oriented behavioral features and the environment-responsive behavioral features are iteratively fused based on the driving weight parameters and the feedback correction parameters to generate a behavioral evolution path feature sequence that reflects the behavioral evolution trend of the agent.

[0010] The behavioral evolution path feature sequence is subjected to behavioral pattern clustering processing. Based on the clustering results, the behavioral pattern category label of the agent is determined. The behavioral pattern category label is matched with a preset risk behavior feature library to generate a quantitative assessment result of the agent's behavioral risk.

[0011] According to a second aspect of this application, a multi-dimensional behavior modeling and risk quantification system for enterprise AI agents is provided. The system includes a machine-readable storage medium and a processor. The machine-readable storage medium stores machine-executable instructions. When the processor executes the machine-executable instructions, the system implements the aforementioned multi-dimensional behavior modeling and risk quantification method for enterprise AI agents.

[0012] Based on any of the above aspects, the technical effect of this application is as follows:

[0013] By collecting behavioral data streams containing behavioral event records and their temporal relationships generated during the operation of an enterprise's AI agent, and utilizing a pre-built multi-dimensional behavioral decoupling network, intention-oriented behavioral features and environment-responsive behavioral features are extracted from the behavioral event records. This achieves multi-dimensional and precise decoupling of the agent's behavior, deeply exploring the different driving factors behind the behavior, constructing a dynamic interactive mapping relationship between intention-oriented and environment-responsive behavioral features, and determining driving weight parameters and feedback correction parameters. This accurately reflects the mutual influence and dynamic changes between the two behavioral features. Based on these parameters, the behavioral features are iteratively fused, and the generated behavioral evolution path feature sequence can present the development trend of the agent's behavior. Finally, by performing behavioral pattern clustering on the behavioral evolution path feature sequence and matching it with a pre-set risk behavior feature library, a quantitative assessment result of the agent's behavioral risk is generated. This helps enterprises to promptly identify potential risks, take effective measures to ensure the safe and stable operation of the AI ​​agent, and improve the overall operational efficiency of the enterprise. Attached Figure Description

[0014] Figure 1 A flowchart illustrating the multi-dimensional behavioral modeling and risk quantification method for enterprise AI agents provided in this application embodiment is shown.

[0015] Figure 2 The diagram shows a schematic of the component structure of the multi-dimensional behavioral modeling and risk quantification system for enterprise AI agents provided in an embodiment of this application. Detailed Implementation

[0016] Figure 1 This paper illustrates a flowchart of a multi-dimensional behavioral modeling and risk quantification method and system for enterprise AI agents provided in an embodiment of this application. The detailed steps include:

[0017] Step S110: Collect the agent behavior data stream generated by the enterprise AI agent during operation. The agent behavior data stream includes records of behavior events generated when the agent interacts with the external environment and the temporal correlation between the behavior event records.

[0018] In this embodiment, the enterprise resource planning system agent interacts with the external environment during operation. This external environment includes the enterprise resource planning system database, enterprise messaging middleware, external vendor application programming interfaces (APIs), and an internal financial approval workflow engine. Step S110 specifically includes the following sub-steps to collect the agent's behavioral data stream.

[0019] Step S111: Monitor the interaction event nodes generated by the enterprise AI agent during data exchange with external data sources when executing enterprise business process tasks, and record each interaction event node as a basic behavior unit. The basic behavior unit includes the identity information of the interaction initiator, the address information of the interaction receiver, the type identification code of the interaction operation, the timestamp information of the interaction operation, and the data payload feature description carried by the interaction operation.

[0020] For example, when an Enterprise Resource Planning (ERP) system agent executes the task of "automatically generating purchase orders," the agent, as the interaction initiator, sends a Structured Query Language (SCL) query request to the ERP database, which is the interaction receiver. This interaction event node is recorded as a basic behavioral unit. Within this basic behavioral unit, the initiator's identity information is recorded as the unique identifier of the ERP agent, such as "AGENT_PUR_001." The receiver's address information is recorded as "JDBC:MYSQL: / / ERP_DB / ORDERS." The interaction operation type identifier, based on the operation type mapping table, maps the "Structured Query Language query" operation to the code "SQL_Q." The interaction operation's timestamp information precisely records the moment the event occurred, in the format "YYYY-MM-DDHH:MM:SS.SSS." The data payload feature description carried by the interactive operation is an abstract representation of the query statement of the Structured Query Language. For example, "SELECT PO_NUMBER, SUPPLIER_ID, TOTAL_AMOUNTFROMPURCHASE_ORDERSWHERESTATUS='PENDING'ANDCREATED_DATE>'2024-01-01'" describes the target data table, fields and filtering conditions of the query.

[0021] Step S112: Extract the data payload feature description carried in the basic behavior unit, perform semantic parsing processing on the data payload feature description, and obtain the business intent type label and the confidence score of the business intent type label corresponding to the basic behavior unit.

[0022] Building upon the aforementioned basic behavioral unit, its data payload feature description is extracted, namely the query statement in the Structured Query Language (SCL). Semantic parsing is then performed on this statement, such as through lexical and syntactic analysis, to identify that the core objective of this operation is to retrieve order information in the "PENDING" state from the "PURCHASE_ORDERS" table. Based on a predefined business intent type dictionary, the objective "retrieving pending purchase orders" is matched as the business intent type tag "INTENT_ORDER_INQUIRY". Simultaneously, based on the degree of matching in the semantic parsing, such as the matching degree between the keywords "PURCHASE_ORDERS" and "ORDER", and the certainty of the "PENDING" state, a confidence score for this business intent type tag is calculated. This score is quantified as a floating-point number between 0 and 1, for example, using the ratio of the number of successfully matched words to the total number of words as a basis, and then adjusting it based on the completeness of the syntactic structure, ultimately yielding the confidence score.

[0023] Step S113: Based on the timestamp information of the interaction operations in the basic behavior units, construct the sequential relationship of the basic behavior units in the time dimension, connect the basic behavior units with direct successor-successor relationships through directed edges to form a temporal association relationship between behavior event records. The temporal association relationship includes the time interval parameter between the predecessor behavior unit identifier and the successor behavior unit identifier, as well as the type transition probability between the interaction operation type of the predecessor behavior unit identifier and the interaction operation type of the successor behavior unit identifier.

[0024] For example, in the "Automatic Purchase Order Generation" task, after basic action unit A (structured query language query) occurs, basic action unit B (the enterprise resource planning system agent initiates an order creation request to the supplier's application programming interface) immediately follows. Based on timestamp information, the timestamp of basic action unit A is earlier than that of basic action unit B, and there are no other action units belonging to this agent between them. Therefore, basic action unit A and basic action unit B are determined to have a direct successor-successor relationship. A directed edge is constructed from basic action unit A to basic action unit B, representing a temporal association. In this temporal association, the predecessor action unit identifier is the unique number "BE_A" of basic action unit A, and the successor action unit identifier is the unique number "BE_B" of basic action unit B. The time interval parameter between them is calculated as the difference between timestamp B and timestamp A. Simultaneously, the type transition probability from type "SQL_Q" to type "API_CREATE_ORDER" is calculated. This type transition probability can be obtained by counting the number of times in history that a basic behavioral unit of type "SQL_Q" was followed by a basic behavioral unit of type "API_CREATE_ORDER", and dividing by the total number of times a basic behavioral unit of type "SQL_Q" occurred.

[0025] Step S114: Organize the basic behavioral units and the temporal associations into an agent behavior data stream according to the ascending order of timestamp information. Each basic behavioral unit in the agent behavior data stream carries a forward reference pointer to its predecessor basic behavioral unit and a backward reference pointer to its successor basic behavioral unit.

[0026] All collected basic behavioral units, such as basic behavioral unit A and basic behavioral unit B mentioned above, and the directed edges (temporal relationships) between them, are arranged in ascending order according to their respective timestamp information, forming an ordered queue, which is the agent behavior data stream. For fast traversal and backtracking, each basic behavioral unit in this agent behavior data stream is additionally assigned two pointers. A forward reference pointer points to the previous basic behavioral unit; for example, the forward reference pointer of basic behavioral unit B points to basic behavioral unit A. A backward reference pointer points to the next basic behavioral unit; for example, the backward reference pointer of basic behavioral unit A points to basic behavioral unit B. The forward reference pointer of the first basic behavioral unit in the queue is null; the backward reference pointer of the last basic behavioral unit in the queue is null.

[0027] Step S115: Aggregate the basic behavioral units with the same interaction initiator identity information in the intelligent agent behavior data stream to form an individual behavior trajectory sequence with a single enterprise AI intelligent agent as the dimension. The individual behavior trajectory sequence contains all the basic behavioral units generated by the enterprise AI intelligent agent at different time points and the complete temporal correlation network between the basic behavioral units.

[0028] An Enterprise Resource Planning (ERP) system may run multiple agents, such as the procurement agent "AGENT_PUR_001" and the financial agent "AGENT_FIN_001". From the agent behavior data stream generated in step S114, all basic behavioral units with the interaction initiator's identity information "AGENT_PUR_001" are selected. These selected basic behavioral units are arranged in chronological order, retaining their original temporal relationships connected by directed edges, thus constructing an individual behavior trajectory sequence specific to the procurement agent. This individual behavior trajectory sequence not only includes all the basic behavioral units executed by the procurement agent but also the complete temporal relationship network between these basic behavioral units, reflecting their operational logic. For example, it could be a path network composed of directed edges from basic behavioral units of type "SQL_Q" to basic behavioral units of type "API_CREATE_ORDER", and possibly basic behavioral units of type "DB_UPDATE".

[0029] Step S116: The basic behavioral units with the same business intent type label in the individual behavior trajectory sequence are segmented and aggregated according to the timestamp information to generate a set of continuous behavior segments reflecting the enterprise AI agent driven by a specific business intent. The set of continuous behavior segments includes the start timestamp of the behavior segment, the end timestamp of the behavior segment, and the arrangement order of the basic behavioral units within the behavior segment.

[0030] Within the individual behavioral trajectory sequence of the procurement agent, there are multiple business intent type labels. For example, there might be a segment starting with the label "INTENT_ORDER_INQUIRY," followed by the label "INTENT_ORDER_CREATE," and then returning to the label "INTENT_ORDER_INQUIRY." Basic behavioral units with the same business intent type label within a continuous time range are aggregated together. For example, if multiple basic behavioral units (such as multiple queries or a single analysis) have the same business intent type label "INTENT_ORDER_INQUIRY" from timestamp T1 to timestamp T2, these basic behavioral units are aggregated into a continuous behavioral segment. This continuous behavioral segment is recorded as follows: the start timestamp is T1, the end timestamp is T2, and the basic behavioral units contained within the segment are arranged in chronological order, forming an ordered list.

[0031] Step S117: Extract the type identifier code of the interaction operation of the first basic behavior unit of each continuous behavior segment in the set of continuous behavior segments as the segment entry behavior type identifier, and extract the type identifier code of the interaction operation of the last basic behavior unit of each continuous behavior segment in the set of continuous behavior segments as the segment exit behavior type identifier.

[0032] For the persistent behavior segments aggregated by the "INTENT_ORDER_INQUIRY" tag, extract the type identifier code of the interaction operation of the basic behavior unit with the earliest timestamp in the persistent behavior segment, such as "SQL_Q", as the segment entry behavior type identifier of the persistent behavior segment. Similarly, extract the type identifier code of the interaction operation of the basic behavior unit with the latest timestamp in the persistent behavior segment, such as "CSV_EXPORT" or "REPORT_VIEW", as the segment exit behavior type identifier of the persistent behavior segment.

[0033] Step S118: Construct the boundary behavior type conversion path of the persistent behavior segment based on the segment entry behavior type identifier and the segment exit behavior type identifier. The boundary behavior type conversion path includes the intermediate behavior type sequence from the entry behavior type to the exit behavior type and the statistical value of the number of transitions between adjacent behavior types.

[0034] Based on the segment entry behavior type identifier "SQL_Q" and segment exit behavior type identifier "CSV_EXPORT" extracted in step S117, and combined with the order of the basic behavior units within the persistent behavior segment, a path from the segment entry behavior type identifier to the segment exit behavior type identifier can be constructed. For example, the sequence of basic behavior unit types within the persistent behavior segment might be "SQL_Q", "DATA_TRANSFORM", and "CSV_EXPORT". The boundary behavior type transition path is then recorded as: starting from "SQL_Q", passing through "DATA_TRANSFORM", and finally reaching "CSV_EXPORT". Simultaneously, the number of times each pair of adjacent types (e.g., "SQL_Q" to "DATA_TRANSFORM", "DATA_TRANSFORM" to "CSV_EXPORT") appears in all historical behaviors of the purchasing agent is counted, serving as the statistical value of the number of transitions between adjacent behavior types.

[0035] Step S119: Classify the persistent behavior segments with the same boundary behavior type conversion path in the persistent behavior segment set to generate a cluster of similar behavior segments with the boundary behavior type conversion path as the cluster center. The cluster of similar behavior segments includes all persistent behavior segments belonging to the same boundary behavior type conversion path and the frequency weight of each persistent behavior segment in the cluster of similar behavior segments.

[0036] In the historical behavior of the purchasing agent, there may be multiple persistent behavior segments that all share the same boundary behavior type transition path, for example, all starting from "SQL_Q", passing through "DATA_TRANSFORM", and finally reaching "CSV_EXPORT". These persistent behavior segments with the same boundary behavior type transition path are grouped together to form a cluster of similar behavior segments. Within this cluster, all persistent behavior segments belonging to that path are recorded, and a frequency weight is calculated for each persistent behavior segment. This frequency weight is equal to the number of times the segment appears in the cluster divided by the total number of times all segments in the cluster appear.

[0037] Step S1110: Calculate the duration deviation parameter between the duration of each persistent behavior segment in the cluster of similar behavior segments and the average duration of all persistent behavior segments in the cluster of similar behavior segments. Mark persistent behavior segments whose duration deviation parameter exceeds a preset tolerance threshold as abnormal behavior segments, and mark the boundary behavior type conversion path corresponding to the abnormal behavior segment as an abnormal behavior path identifier.

[0038] For the aforementioned cluster of similar behavior segments centered on the path "SQL_Q-DATA_TRANSFORM-CSV_EXPORT", calculate the average duration of all persistent behavior segments in the cluster. Then, for each persistent behavior segment in the cluster, calculate the difference between its own duration and this average, and divide this difference by the average to obtain the duration deviation parameter. A preset tolerance threshold is set, for example, 0.5. If the duration deviation parameter of a persistent behavior segment is greater than 0.5, it means that the duration of the segment deviates significantly from the average level, and the persistent behavior segment is marked as an abnormal behavior segment. At the same time, the boundary behavior type conversion path "SQL_Q-DATA_TRANSFORM-CSV_EXPORT" to which the abnormal behavior segment belongs is marked as an abnormal behavior path identifier for subsequent risk monitoring and early warning.

[0039] Step S120: Input the agent behavior data stream into a pre-constructed multidimensional behavior decoupling network for behavior feature separation processing. Extract the agent's intent-oriented behavior features from the behavior event records through the intent decoder in the multidimensional behavior decoupling network. Extract the agent's environment-responsive behavior features from the behavior event records through the environment response decoder in the multidimensional behavior decoupling network.

[0040] In this embodiment, in order to separate behavioral features of different dimensions from the complex agent behavior data stream, a multi-dimensional behavior decoupling network is constructed. Its purpose is to decouple the intrinsic driving force (intent) of behavior from its adaptation to the external environment (response). Step S120 specifically includes the following sub-steps to achieve this separation process.

[0041] Step S121: Input the basic behavior unit in the agent behavior data stream into the shared feature encoding module of the multidimensional behavior decoupling network. The shared feature encoding module performs embedding vectorization processing on the identity information of the interaction initiator, the address information of the interaction receiver, the type identification code of the interaction operation, and the data payload feature description of the basic behavior unit to generate the original behavior embedding vector of the basic behavior unit.

[0042] First, each basic behavioral unit in the agent behavior data stream generated in step S110 is passed as input to a shared feature encoding module in the multidimensional behavior decoupling network. This shared feature encoding module contains multiple embedding layers. For the identity information of the interaction initiator in the basic behavioral unit (e.g., "AGENT_PUR_001"), it is mapped to a dense vector of fixed dimension through an identity embedding matrix. Similarly, the address identification information of the interaction receiver (e.g., "JDBC:MYSQL: / / ERP_DB / ORDERS") and the type identification code of the interaction operation (e.g., "SQL_Q") are also mapped to dense vectors of corresponding dimensions through their respective embedding matrices. For the data payload feature description, it is transformed into a semantic vector through a pre-trained language model (e.g., a lightweight Transformer encoder). Finally, the vectors obtained from different information sources are concatenated along the feature dimension to form a comprehensive original behavior embedding vector representing the basic behavioral unit.

[0043] Step S122: Input the original behavior embedding vector into the intention decoder and the environment response decoder in parallel. The intention decoder performs semantic intention parsing processing on the original behavior embedding vector through a self-attention mechanism, and separates the behavior driving factor features directly related to the achievement of business goals from the original behavior embedding vector as intention-oriented behavior features. The intention-oriented behavior features include business goal orientation parameters and behavior purpose persistence parameters.

[0044] The original behavior embedding vector generated in step S121 is simultaneously input into two parallel branches of the multidimensional behavior decoupling network: the intent decoder and the environment response decoder.

[0045] Internally, the intent decoder first processes the input raw behavior embedding vector through a multi-head self-attention network layer. This multi-head self-attention network layer allows the vector at each position to interact with vectors at all other positions, thereby capturing semantic dependencies within the behavior unit. After processing by the self-attention layer, the vector is fed into a feedforward neural network layer for non-linear transformation. The output of the intent decoder is designed as a feature space directly related to the business objective. Its output is a multi-dimensional feature vector, namely intent-oriented behavior features, which is not a single numerical value but consists of multiple dimensions. For example, one dimension can be quantified as a business objective orientation parameter, which is obtained by activating a specific component of the decoder output vector through a sigmoid function, representing the strength of the association between the current behavior and a preset business objective (such as "completing the purchase order"). Another dimension can be quantified as a behavior purpose persistence parameter, which is reflected by calculating the gradient rate of change of the decoder output vector over time, reflecting the stability and persistence tendency of the intent in the time dimension.

[0046] Step S123: The environmental response decoder performs context dependency modeling on the original behavior embedding vector through a conditional random field model, and extracts behavioral adaptation factor features associated with changes in the external environment state from the original behavior embedding vector as environmental response behavioral features. The environmental response behavioral features include environmental state sensitivity parameters and behavioral feedback timeliness parameters.

[0047] The environment response decoder, which processes in parallel with the intent decoder, employs a Conditional Random Field (CRF) model to model the contextual dependencies of the behavior sequence. This CRF model treats the input behavior sequence (i.e., a series of original behavior embedding vectors) as an observation sequence and learns the probability distribution of a state sequence (i.e., an environment response label sequence). The model captures the relationship between the observation sequence and the state sequence, as well as the transition relationships between adjacent states within the state sequence, by defining feature functions. The output of this environment response decoder is a multi-dimensional feature vector, namely, environment-responsive behavior features, which also consists of multiple dimensions. For example, one dimension can be quantified as an environmental state sensitivity parameter, obtained by calculating the derivative of the marginal probability distribution output by the CRF model with respect to changes in the input observation, reflecting the sensitivity of the agent's behavior to changes in external input (such as database results). Another dimension can be quantified as a behavior feedback timeliness parameter, obtained by analyzing the rate of state transitions in the CRF model or calculating the average time delay between changes in environmental state and changes in behavior response, reflecting the speed of the agent's response to external stimuli.

[0048] Step S124: Input the intent-oriented behavior features output by the intent decoder and the environment-responsive behavior features output by the environment response decoder into the feature orthogonalization constraint layer of the multidimensional behavior decoupling network. Calculate the mutual information between the intent-oriented behavior features and the environment-responsive behavior features through the feature orthogonalization constraint layer, and jointly optimize and adjust the network weight parameters of the intent decoder and the environment response decoder based on the goal of minimizing the mutual information.

[0049] To ensure true decoupling of intent-driven and environment-responsive behavioral features—that is, since they each carry different aspects of the behavior—the multidimensional behavior decoupling network introduces a feature orthogonality constraint layer during training. During each forward propagation, it receives intent-driven behavioral features from the intent decoder and environment-responsive behavioral features from the environment-responsive decoder. It then calculates the mutual information between these two feature vectors, which measures the amount of information shared between two random variables. The calculation process first discretizes the feature vectors or estimates their probability distribution using kernel methods, and then uses information theory formulas to calculate the mutual information. The network's loss function includes a mutual information minimization term in addition to the loss from downstream tasks (such as behavior reconstruction). Through backpropagation, the gradient signal simultaneously updates the network weight parameters of both the intent decoder and the environment-responsive decoder, continuously reducing the mutual information between their output feature vectors. This drives the two features towards orthogonality in the representation space, achieving decoupling of the behavioral features.

[0050] Step S125: Perform residual connection processing between the intention-oriented behavior features processed by the feature orthogonalization constraint layer and the original behavior embedding vector to generate enhanced intention-oriented behavior features with compensation for the original behavior information. Perform residual connection processing between the environment-responsive behavior features processed by the feature orthogonalization constraint layer and the original behavior embedding vector to generate enhanced environment-responsive behavior features with compensation for the original behavior information.

[0051] The intent-oriented behavior features processed by the feature orthogonalization constraint layer may lose some detailed information from the original behavior embedding vector. To compensate for this loss, a residual connection technique is used. The intent-oriented behavior features output by the feature orthogonalization constraint layer are added element-wise to the original behavior embedding vector generated in step S121 to obtain a new feature vector, namely the enhanced intent-oriented behavior feature, which contains both the decoupled and optimized intent information and retains the low-level details of the original input. Similarly, the environmental response behavior features processed by the feature orthogonalization constraint layer are added element-wise to the original behavior embedding vector to generate the enhanced environmental response behavior feature.

[0052] Step S126: Input the enhanced intent-oriented behavior features into the first temporal convolution module of the multidimensional behavior decoupling network, and perform sliding window convolution operation on the enhanced intent-oriented behavior features in the time dimension through the first temporal convolution module to extract the change trend features of the enhanced intent-oriented behavior features in the continuous time window as the evolution trajectory of intent-oriented behavior.

[0053] The enhanced intent-driven behavioral features generated in step S125 are arranged into a time series according to their temporal order in the agent's behavioral data stream. This sequence is input into the first temporal convolutional module, which consists of multiple stacked one-dimensional convolutional layers. Each convolutional layer uses a convolutional kernel of a different size to perform a sliding window convolution operation in the time dimension. For example, a convolutional layer with a kernel size of 3 will calculate the weighted sum of the enhanced intent-driven behavioral feature vectors at three consecutive time points, generating a new feature value that reflects the change pattern within that local time window. Through the stacking of multiple convolutions, it can extract the change trends at different time scales. Finally, its output is a new feature sequence, namely the intent-driven behavioral evolution trajectory, which describes how the intent-driven behavioral features evolve and flow along the time axis.

[0054] Step S127: Input the enhanced environmental responsive behavior features into the second temporal convolution module of the multidimensional behavior decoupling network, and perform sliding window convolution operation on the enhanced environmental responsive behavior features in the time dimension through the second temporal convolution module to extract the change trend features of the enhanced environmental responsive behavior features within the continuous time window as the evolution trajectory of environmental responsive behavior.

[0055] Parallel to and structurally identical to step S126, the enhanced environmental responsive behavior features generated in step S125 are input into the second temporal convolutional module, which is also composed of one-dimensional convolutional layers. The module performs sliding window convolution operations on the enhanced environmental responsive behavior features in the time series to extract their change patterns and multi-scale trends within the local time window. Its output is the evolution trajectory of environmental responsive behavior, which describes the dynamic process of environmental responsive behavior features changing over time.

[0056] Step S128: Input the intent-oriented behavior evolution trajectory and the environment-responsive behavior evolution trajectory into the trajectory alignment module of the multidimensional behavior decoupling network. Calculate the phase offset between the intent-oriented behavior evolution trajectory and the environment-responsive behavior evolution trajectory in the time dimension through the trajectory alignment module. Perform time axis translation calibration processing on the intent-oriented behavior evolution trajectory based on the phase offset to generate an environment-responsive behavior evolution trajectory that is time-aligned with the intent-oriented behavior evolution trajectory.

[0057] Because there may be a temporal misalignment between intention-driven behavior (driving) and environment-responsive behavior (feedback), for example, a period of time may pass before the environmental response changes are observed after the intention is generated. The trajectory alignment module of the multidimensional behavior decoupling network aims to solve this problem. It receives two trajectory sequences generated in steps S126 and S127. First, the phase shift of the two sequences is determined by calculating their cross-correlation function. The cross-correlation function describes the similarity between the two sequences when one is time-shifted relative to the other. The time shift corresponding to the peak of the cross-correlation function is found to be the phase shift. Then, based on the calculated phase shift, the evolutionary trajectory of the environment-responsive behavior is translated and calibrated on the time axis. After calibration, a new sequence is obtained, which is the evolutionary trajectory of the environment-responsive behavior that is time-aligned with the evolutionary trajectory of the intention-driven behavior, so that the two correspond to causally related behavioral features at the same point in time.

[0058] Step S129: Perform time-point feature concatenation processing on the time-aligned environment-responsive behavior evolution trajectory and the intent-oriented behavior evolution trajectory to generate a multi-dimensional coupled behavior feature sequence with time synchronization correlation. The feature vector corresponding to each time point in the multi-dimensional coupled behavior feature sequence simultaneously contains the intent-oriented behavior feature information and the environment-responsive behavior feature information at that time point.

[0059] After time alignment, the time-aligned environment-responsive behavior evolution trajectory output in step S128 and the intent-oriented behavior evolution trajectory output in step S126 are concatenated along their feature dimensions at each identical time point. For example, at time point t, the intent-oriented behavior evolution trajectory provides a feature vector of dimension 'a', and the time-aligned environment-responsive behavior evolution trajectory provides a feature vector of dimension 'b'. Concatenation results in a new feature vector of dimension (a+b). This process is repeated for all time points to obtain a new feature sequence, namely a multi-dimensional coupled behavior feature sequence. Each feature vector in this sequence simultaneously contains both the intent-oriented behavior feature information and the environment-responsive behavior feature information at that moment, and the two are synchronously correlated in time.

[0060] Furthermore, step S120 may also include:

[0061] Step S1291: Obtain the data payload feature description carried by the interaction operation of the basic behavior unit in the intelligent agent behavior data stream, perform content type identification processing on the data payload feature description, and divide the data payload feature description into a structured data payload part and an unstructured data payload part.

[0062] In this embodiment, the data payload feature descriptions carried by the interactive operations generated by the enterprise resource planning system agent during operation have multiple formats. For example, when the agent initiates a structured query language query request to the enterprise resource planning system database, its data payload feature description is a structured query language statement. This structured query language statement has a strict grammatical structure and field definitions, belonging to the structured data payload part. When the agent receives a response message from an external vendor's application programming interface, the response message may contain a piece of natural language text describing the order status, such as "order confirmed, estimated delivery time is within three working days." This text lacks a predefined structured format and belongs to the unstructured data payload part. Step S1291 performs content type identification on the data payload feature description of each basic behavioral unit. By parsing the format header information and grammatical features of the data payload or by using a pre-trained classification model, the data payload feature description is divided into structured data payload part and unstructured data payload part.

[0063] Step S1292: Input the structured data payload part into the first feature extraction branch of the intent decoder. The first feature extraction branch performs numerical feature mapping processing on the structured data payload part through a multilayer perceptron network to generate structured intent feature components.

[0064] The structured data payload identified in step S1291, such as a structured query language statement "SELECT PO_NUMBER, SUPPLIER_ID, TOTAL_AMOUNTFROMPURCHASE_ORDERSWHERESTATUS='PENDING'", is input into the first feature extraction branch of the intent decoder, which consists of a multilayer perceptron network. First, feature engineering is performed on this structured data payload, converting key fields (such as the table name "PURCHASE_ORDERS", the selected field "PO_NUMBER", etc.), the operation type ("SELECT"), and the condition ("STATUS='PENDING'") into numerical feature vectors. For example, the table name can be mapped to a sparse vector using one-hot encoding, and the field name can be mapped to a dense vector using an embedding matrix. These numerically converted feature vectors are then concatenated into an input vector and fed into the multilayer perceptron network. The multilayer perceptron network consists of multiple fully connected layers, each performing a linear transformation on the input followed by a non-linear activation function, extracting higher-level abstract features layer by layer. The output vector of a multilayer perceptron network is the structured intent feature component, which encodes the intent information contained in the structured data payload.

[0065] Step S1293: Input the unstructured data payload portion into the second feature extraction branch of the intent decoder. The second feature extraction branch performs semantic feature extraction processing on the unstructured data payload portion through a convolutional neural network to generate unstructured intent feature components.

[0066] The unstructured data payload identified in step S1291, such as a natural language text "Order confirmed, estimated delivery time is within three working days," is input into the second feature extraction branch of the intent decoder, which uses a convolutional neural network for semantic feature extraction. First, the text is segmented into words, resulting in a series of word units. Each word unit is mapped to a fixed-dimensional word embedding vector using a pre-trained word embedding matrix (e.g., word vectors pre-trained using a large-scale corpus). The word embedding vectors of all word units are arranged sequentially to form a two-dimensional matrix, where rows correspond to word units and columns correspond to embedding dimensions. This two-dimensional matrix is ​​then input into the convolutional neural network. The convolutional neural network contains multiple convolutional layers, each using a different sized convolutional kernel to perform a sliding window convolution operation along the word unit sequence dimension, extracting semantic features from local word groups. After the convolution operation, the feature map is reduced in dimensionality using pooling layers (e.g., max pooling) to extract the most salient features. Finally, the extracted multi-scale semantic features are integrated into a fixed-dimensional output vector through a fully connected layer, namely the unstructured intent feature component, which encodes the intent information implicit in the unstructured data payload.

[0067] Step S1294: Align the structured intent feature components with the unstructured intent feature components in terms of feature dimensions, and then perform feature splicing on the aligned structured intent feature components and unstructured intent feature components to generate fused intent-oriented behavioral features.

[0068] Since structured and unstructured intent feature components may have different feature dimensions, dimension alignment is required before fusion. Feature dimension alignment is achieved through a linear transformation layer; either a fully connected layer is applied to the feature component with lower dimension to increase its dimension to the same level as the higher-dimensional feature component, or a fully connected layer is applied to the feature component with higher dimension to reduce its dimension. After alignment, the two feature components have the same dimension. Then, the aligned structured and unstructured intent feature components are concatenated along their feature dimensions to form a new feature vector with a dimension equal to the sum of the two. This is the fused intent-oriented behavioral feature, which simultaneously integrates the precise numerical and structured information of the structured data payload, as well as the semantic and contextual information of the unstructured data payload, thus more comprehensively representing the agent's intent.

[0069] Step S1295: Input the structured data payload into the third feature extraction branch of the environmental response decoder. The third feature extraction branch performs time-series modeling of the numerical change pattern of the structured data payload through a recurrent neural network to generate structured environmental response feature components.

[0070] For the structured data payload, such as the continuous numerical data streams received by the agent from sensors (e.g., enterprise server load percentage, database connection count, etc.), it is input into the third feature extraction branch of the environmental response decoder. This branch employs a recurrent neural network, particularly a long short-term memory network or a gated recurrent unit, to capture the temporal variation patterns of these values. The structured data payload is organized into a sequence in chronological order, with the data at each time point (e.g., a vector containing server load values ​​and database connection counts) serving as the input to the recurrent neural network for one time step. The hidden states within the recurrent neural network are updated with each time step, passing historical information to the current moment. Through this temporal modeling, the recurrent neural network can learn the trends, periodicity, and abrupt change patterns of numerical changes. The hidden state of the last time step, or pooling the hidden states of all time steps, yields a vector that represents the structured environmental response feature component, encoding the agent's response characteristics to changes in structured environmental data.

[0071] Step S1296: Input the unstructured data payload into the fourth feature extraction branch of the environment response decoder. The fourth feature extraction branch performs contextual analysis on the semantic content of the unstructured data payload through an attention mechanism to generate unstructured environment response feature components.

[0072] For the unstructured data payload, such as natural language notifications, email content, or log messages received by the agent from a collaborative system, it is input into the fourth feature extraction branch of the environment response decoder. This branch employs an attention mechanism, specifically the multi-head self-attention mechanism in the Transformer architecture, to parse semantic content and perform contextual analysis. First, the unstructured text is segmented and embedded, resulting in a sequence of word embedding vectors. This sequence is then input into a multi-head self-attention layer, which allows each word in the sequence to interact with all other words, calculating the association weights between them. Through this mechanism, the model can focus on the most critical information in the text regarding the current environmental state (e.g., keywords like "urgent," "fault," and "delay" in notifications) and understand the meaning of this information in context. After processing by multiple self-attention layers and feedforward neural network layers, the output sequence is aggregated (e.g., taking the output of the first word or averaging the outputs of all words) to obtain the unstructured environment response feature components, which encode the semantic understanding response features of the agent to unstructured environmental information (such as natural language instructions or notifications).

[0073] Step S1297: Align the structured environmental response feature components with the unstructured environmental response feature components in terms of feature dimensions, and then perform feature splicing on the aligned structured environmental response feature components and the unstructured environmental response feature components to generate fused environmental response behavior features.

[0074] Similar to step S1294, the structured environmental response feature components generated in step S1295 and the unstructured environmental response feature components generated in step S1296 are aligned in feature dimensions, with an alignment layer ensuring that the dimensions are the same. Then, the two aligned feature components are concatenated along their feature dimensions to generate a fused environmental response behavior feature. This fused environmental response behavior feature simultaneously includes the agent's numerical response pattern to changes in structured environmental data and its semantic understanding response pattern to unstructured environmental information, providing a more complete representation of environmental response behavior.

[0075] Step S130: Construct a dynamic interaction mapping relationship between the intent-oriented behavioral features and the environment-responsive behavioral features. The dynamic interaction mapping relationship includes the driving weight parameters of the intent-oriented behavioral features on the environment-responsive behavioral features and the feedback correction parameters of the environment-responsive behavioral features on the intent-oriented behavioral features.

[0076] This step aims to quantify the bidirectional dynamic influence relationship between intention-oriented behavioral characteristics and environment-responsive behavioral characteristics. Step S130 specifically includes the following sub-steps to construct this dynamic interaction mapping relationship.

[0077] Step S131: Use the intention-oriented behavior feature vectors of adjacent time points as forward drive inputs and the environment response behavior feature vectors of adjacent time points as forward drive outputs to construct a forward drive mapping function for intention to environment response.

[0078] From the multidimensional coupled behavioral feature sequence generated in step S129, the intent-oriented behavioral feature vectors and the corresponding environment-responsive behavioral feature vectors at adjacent time points t and t+1 are extracted. The intent-oriented behavioral feature vector at time point t is defined as the forward driving input, and the environment-responsive behavioral feature vector at time point t+1 is defined as the forward driving output. The forward driving mapping function of intent to environment response is defined as the mapping relationship from the input space to the output space.

[0079] Step S132: Calculate the linear transformation matrix required to map the intention-oriented behavioral feature vector to the environment-responsive behavioral feature vector in each adjacent time point pair using the forward driving mapping function. Extract the off-diagonal elements in the linear transformation matrix to generate driving weight parameters of the intention-oriented behavioral feature on the environment-responsive behavioral feature. The driving weight parameters are used to characterize the driving strength of each dimension component of the intention-oriented behavioral feature on each dimension component of the environment-responsive behavioral feature.

[0080] To approximate the aforementioned forward-driving mapping function, it is assumed that it can be approximated as a linear transformation within a local temporal neighborhood. For each pair of adjacent time points (t, t+1), a linear transformation matrix W_f is solved such that the result of multiplying W_f by the intent-driven behavior feature vector (at time point t) is as close as possible to the environment-responsive behavior feature vector (at time point t+1). This linear transformation matrix W_f is a matrix whose number of rows equals the dimension of the environment-responsive behavior feature vector, and whose number of columns equals the dimension of the intent-driven behavior feature vector. The element w_f[i][j] in matrix W_f represents the driving strength of the j-th dimension of the intent-driven behavior feature on the i-th dimension of the environment-responsive behavior feature. All off-diagonal elements in this matrix (i.e., elements where i is not equal to j) are extracted, and the values ​​of these elements are used as driving weight parameters of the intent-driven behavior feature on the environment-responsive behavior feature. Each driving weight parameter corresponds to a specific pair of intent dimension and environment dimension, quantifying how the intent dimension drives the change of the environment dimension.

[0081] Step S133: Use the environmental response-type behavioral feature vectors of adjacent time points as the reverse correction input and the intention-oriented behavioral feature vectors of adjacent time points as the reverse correction output to construct the reverse correction mapping function of environmental response to intention.

[0082] Similarly, environmental response-oriented behavioral feature vectors and intent-oriented behavioral feature vectors at adjacent time points t and t+1 are extracted from the multidimensional coupled behavioral feature sequence. The environmental response-oriented behavioral feature vector at time point t is defined as the back-correction input, and the intent-oriented behavioral feature vector at time point t+1 is defined as the back-correction output. The back-correction mapping function of environmental response to intent is defined as the mapping relationship from the input space to the output space.

[0083] Step S134: Calculate the linear transformation matrix required to map the environmental responsive behavior feature vector to the intention-oriented behavior feature vector in each adjacent time point pair using the reverse correction mapping function. Extract the off-diagonal elements in the linear transformation matrix to generate feedback correction parameters of the environmental responsive behavior feature to the intention-oriented behavior feature. The feedback correction parameters are used to characterize the correction strength of each dimension component of the environmental responsive behavior feature to each dimension component of the intention-oriented behavior feature.

[0084] For each pair of adjacent time points (t, t+1), solve for another linear transformation matrix W_b such that the result of multiplying W_b by the environmental responsive behavior feature vector (at time point t) is as close as possible to the intention-oriented behavior feature vector (at time point t+1). The number of rows in this matrix W_b equals the dimension of the intention-oriented behavior feature vector, and the number of columns equals the dimension of the environmental responsive behavior feature vector. The element w_b[p][q] in matrix W_b represents the correction strength of the q-th dimension of the environmental responsive behavior feature to the p-th dimension of the intention-oriented behavior feature. Extract all off-diagonal elements (p ≠ q) from this matrix, and use the values ​​of these elements as feedback correction parameters of the environmental responsive behavior feature to the intention-oriented behavior feature. Each feedback correction parameter quantifies how a specific environmental dimension corrects a specific intention dimension.

[0085] Step S135: Organize the driving weight parameters into a time series sequence according to the time dimension, perform differential operation on the time series sequence of driving weight parameters, and extract the change amplitude value of the driving weight parameters between adjacent time points as the dynamic fluctuation feature of driving weight.

[0086] The driving weight parameters calculated for each pair of adjacent time points in step S132 (usually an expansion of a multidimensional vector or matrix) are arranged in chronological order to form a time-series sequence of driving weight parameters. Then, a first-order difference operation is performed on each parameter in this sequence, that is, the parameter value at the later time point is subtracted from the parameter value at the previous time point, resulting in a new sequence. Each value in this new sequence represents the magnitude of change of the driving weight parameter between adjacent time points. These magnitude values ​​are organized to form the dynamic fluctuation characteristics of the driving weights, reflecting the drastic and unstable nature of the intention's driving effect on the environment over time.

[0087] Step S136: Organize the feedback correction parameters into a time series sequence according to the time dimension, perform differential operation on the time series sequence, and extract the change amplitude value of the feedback correction parameters between adjacent time points as the dynamic fluctuation feature of feedback correction.

[0088] Similarly, the feedback correction parameters calculated for each pair of adjacent time points in step S134 are arranged in chronological order to form a time series sequence of feedback correction parameters. A first-order difference operation is performed on this sequence to obtain the variation amplitude of the feedback correction parameters between adjacent time points. These values ​​are then organized as the dynamic fluctuation characteristics of the feedback correction, reflecting the fluctuation of the environment's effect on the intended correction over time.

[0089] Step S137: Input the dynamic fluctuation feature of the driving weight and the dynamic fluctuation feature of the feedback correction into the interactive coupling analysis module. Calculate the covariance value of the dynamic fluctuation feature of the driving weight and the dynamic fluctuation feature of the feedback correction at the same time point through the interactive coupling analysis module. Use the covariance value as an indicator of the interactive coupling strength between intention-oriented behavioral features and environment-responsive behavioral features.

[0090] The dynamic fluctuation feature sequence of the driving weight generated in step S135 and the dynamic fluctuation feature sequence of the feedback correction generated in step S136 are input into the interaction coupling analysis module, which calculates the covariance of these two sequences at corresponding time points. The covariance calculation process is as follows: For each time point, the difference between the dynamic fluctuation feature value of the driving weight at that time point and the mean of the dynamic fluctuation feature sequence of the driving weight is calculated, multiplied by the difference between the dynamic fluctuation feature value of the feedback correction at that time point and the mean of the dynamic fluctuation feature sequence of the feedback correction, and then the products of all time points are accumulated and averaged. The calculated covariance value is the interaction coupling strength index. The larger the value of the interaction coupling strength index, the more positive it is, indicating that the change in intention-driven behavior and the change in environmental feedback have a strong correlation in the same direction, that is, the tighter the dynamic interaction coupling between the two.

[0091] Step S138: Based on the interaction coupling strength index, the driving weight parameters and feedback correction parameters are filtered and processed. The driving weight parameters at the corresponding time points where the interaction coupling strength index exceeds the preset coupling threshold are extracted as valid driving weight parameters, and the feedback correction parameters at the corresponding time points where the interaction coupling strength index exceeds the preset coupling threshold are extracted as valid feedback correction parameters.

[0092] A preset coupling threshold is set, for example, the average value of the interaction coupling strength index at all time points is taken as the threshold. Time points whose interaction coupling strength index is greater than the threshold are selected. For these selected time points, the corresponding driving weight parameters calculated in step S132 are extracted as effective driving weight parameters; at the same time, the corresponding feedback correction parameters calculated in step S134 are extracted as effective feedback correction parameters. This process aims to select moments where the intention to interact with the environment is significant and tightly coupled, and to eliminate moments with weak interaction or high noise.

[0093] Step S139: Organize the effective driving weight parameters into an effective driving weight sequence according to the chronological order of their corresponding time points, and organize the effective feedback correction parameters into an effective feedback correction sequence according to the chronological order of their corresponding time points. Construct a mapping table of correspondence between the effective driving weight sequence and the effective feedback correction sequence. Each entry in the mapping table contains the effective driving weight parameter and the effective feedback correction parameter at the same time point.

[0094] All effective driving weight parameters selected in step S138 are arranged in chronological order according to their corresponding time points to form an effective driving weight sequence. Similarly, the corresponding effective feedback correction parameters are arranged in chronological order to form an effective feedback correction sequence. Then, a correspondence mapping table is constructed. This mapping table is a data structure in which each entry corresponds to a specific time point and contains the effective driving weight parameters and effective feedback correction parameters at that time point. This mapping table records the quantified values ​​of driving and feedback effects at moments when the intention and environment are closely coupled.

[0095] Furthermore, step S130 may also include: step S1310: obtaining the type identifier code of the interaction operation of the basic behavior unit in the agent behavior data stream, organizing the type identifier code of the interaction operation into a behavior type sequence according to the chronological order of time, performing frequent pattern mining processing on the behavior type sequence, and extracting the frequently occurring continuous behavior type subsequences in the behavior type sequence as behavior type transition patterns.

[0096] In this embodiment, type identifier codes for the interaction operations of each basic behavioral unit generated in step S111 are extracted, such as "SQL_Q", "API_CREATE_ORDER", "DB_UPDATE", and "CSV_EXPORT". These type identifier codes are arranged in chronological order according to their timestamps to form a behavioral type sequence. Then, a frequent pattern mining algorithm, such as the PrefixSpan algorithm or the FP-Growth algorithm, is applied to this behavioral type sequence. These algorithms discover frequently occurring subsequence patterns in the sequence by recursively constructing a projection database or a frequent pattern tree. For example, it might be found that the subsequence "SQL_Q->API_CREATE_ORDER->DB_UPDATE" has appeared very many times historically, and this subsequence is extracted as a behavioral type transition pattern. This behavioral type transition pattern represents a common operational flow of the agent when performing a task.

[0097] Step S1311: Construct a behavior type transfer graph structure based on the behavior type transfer pattern. The behavior type transfer graph structure includes a graph structure representation with the type identifier code of the interaction operation as the graph node and the behavior type transfer pattern as the directed edge. Each directed edge carries the frequency weight of the occurrence of the behavior type transfer pattern.

[0098] Using all the behavior type transition patterns extracted in step S1310 as a foundation, a behavior type transition graph structure is constructed. In this graph structure, each unique interaction operation type identifier code (such as "SQL_Q" or "API_CREATE_ORDER") is defined as a graph node. For each behavior type transition pattern, such as the pattern "SQL_Q->API_CREATE_ORDER->DB_UPDATE", it is decomposed into directed edges between adjacent node pairs: a directed edge from node "SQL_Q" to node "API_CREATE_ORDER", and a directed edge from node "API_CREATE_ORDER" to node "DB_UPDATE". For each directed edge, the sum of the occurrences of all behavior type transition patterns containing that edge is counted, and this sum is used as the frequency weight of the directed edge, which is then appended to the edge. The final constructed behavior type transition graph structure is a directed weighted graph that intuitively displays the transition relationships between agent behavior types and their frequency.

[0099] Step S1312: Map the intent-oriented behavior features to the graph nodes of the behavior type transfer graph structure to generate an intent-annotated behavior type transfer graph with intent-oriented behavior feature annotations; map the environment-responsive behavior features to the graph nodes of the behavior type transfer graph structure to generate an environment-annotated behavior type transfer graph with environment-responsive behavior feature annotations.

[0100] For each graph node in the behavior type transition graph structure (representing a type identifier code for an interactive operation), it is associated with the intent-oriented behavior features corresponding to all basic behavior units belonging to that type. For example, the node "SQL_Q" corresponds to multiple basic behavior units, and each basic behavior unit has an intent-oriented behavior feature vector calculated from step S120. The above feature vectors are averaged or aggregated to obtain a representative intent-oriented behavior feature vector for that node, and this vector is then labeled on the graph node. After this operation, each node in the behavior type transition graph structure is associated with an intent-oriented behavior feature vector, thus forming an intent-labeled behavior type transition graph. Similarly, the environmental response behavior features of the basic behavior units corresponding to each graph node are aggregated and labeled on the corresponding graph nodes to form an environmental-labeled behavior type transition graph.

[0101] Step S1313: Calculate the difference degree of intent-oriented behavior features between adjacent graph nodes in the intent-annotated behavior type transfer graph, use the difference degree as the change magnitude parameter of intent-oriented behavior features in the behavior type transfer process, and attach the change magnitude parameter to the corresponding directed edge as the intention change edge weight.

[0102] Traverse each directed edge in the intent-annotated behavior type transition graph, connecting two graph nodes, for example, from node A to node B. Obtain the intent-oriented behavior feature vectors labeled at node A and node B. Calculate the dissimilarity between these two vectors, for example, by calculating Euclidean or cosine distance. The calculated dissimilarity value represents the magnitude of change in the agent's intent-oriented behavior features when the agent's behavior type transitions from A to B. Attach this dissimilarity value as the intent change edge weight to the directed edge. In this way, each directed edge carries not only a transition frequency weight but also an intent change edge weight, reflecting the drastic evolution of intent during the transition process.

[0103] Step S1314: Calculate the difference degree of environmental responsive behavior features between adjacent graph nodes in the environmental annotation behavior type transfer graph, use the difference degree as the change magnitude parameter of environmental responsive behavior features in the behavior type transfer process, and attach the change magnitude parameter to the corresponding directed edge as the environmental change edge weight.

[0104] Similarly, traverse each directed edge in the environment-labeled behavior type transition graph. Obtain the environment-responsive behavior feature vectors labeled on the two graph nodes connected by the directed edge, and calculate the dissimilarity (e.g., Euclidean distance) between these two vectors. The calculated dissimilarity value represents the magnitude of change in the agent's environment-responsive behavior features when the behavior type transitions. This dissimilarity is then used as the environment change edge weight and appended to the directed edge. Thus, each directed edge in the behavior type transition graph structure carries three weights: transition frequency weight, intention change edge weight, and environment change edge weight.

[0105] Step S1315: Calculate the ratio of the intention change edge weight to the environment change edge weight for the corresponding directed edge, generate the intention-to-environment driving intensity ratio for each directed edge, and use the intention-to-environment driving intensity ratio as the driving weight parameter of intention-oriented behavioral features to environment-responsive behavioral features.

[0106] For each directed edge in the behavior type transition graph, obtain its intention change edge weight (denoted as W_intent) and environment change edge weight (denoted as W_env). Calculate the ratio of the intention change edge weight to the environment change edge weight, i.e., W_intent divided by W_env, to obtain the ratio of the intensity of the intention's driving force on the environment. This ratio reflects the strength of the change in intention relative to the change in environment during the behavior transition process. If the ratio is much greater than 1, it indicates that the change in intention is the dominant factor, and the intention drives the change in the response of the environment; if the ratio is much less than 1, it indicates that the change in environment is more significant. Use this ratio as the driving weight parameter for the intention-oriented behavior feature corresponding to that directed edge on the environment-responsive behavior feature.

[0107] Step S1316: Calculate the ratio of the weight of the environmental change edge to the weight of the intention change edge for the corresponding directed edge, generate the ratio of the environmental feedback intensity to the intention for each directed edge, and use the ratio of the environmental feedback intensity to the intention as the feedback correction parameter of the environmental response behavior feature to the intention-oriented behavior feature.

[0108] Similarly, for each directed edge, the ratio of the edge weight for environmental change to the edge weight for intent change is calculated, i.e., W_env divided by W_intent, to obtain the ratio of environmental feedback intensity to intent. This ratio reflects the strength of the influence of environmental change on intent change during the behavior transition process. This ratio is used as the feedback correction parameter for the environment-responsive behavior feature corresponding to the directed edge on the intent-oriented behavior feature. Through steps S1315 and S1316, the driving weight parameters and feedback correction parameters calculated based on the entire behavior type transition graph structure are combined with or compared with the parameters calculated based on the time sequence in steps S132 and S134.

[0109] Step S140: Iteratively fuse the intent-oriented behavioral features and the environment-responsive behavioral features according to the driving weight parameters and the feedback correction parameters to generate a behavioral evolution path feature sequence that reflects the behavioral evolution trend of the agent.

[0110] This step aims to utilize the dynamic interaction mapping relationship (i.e., the effective driving weight sequence and the effective feedback correction sequence) constructed in step S130 to deeply and multi-roundly fuse intentions and environmental features, in order to simulate the long-term evolution trend of behavior under interaction. Step S140 specifically includes the following sub-steps to achieve this iterative fusion process.

[0111] Step S141: Organize the intent-oriented behavioral features into an intent feature time series matrix according to the chronological order of time, organize the environmental response behavioral features into an environmental feature time series matrix according to the chronological order of time, organize the effective driving weight sequence into a driving weight time series matrix according to the chronological order of time, and organize the effective feedback correction sequence into a feedback correction time series matrix according to the chronological order of time.

[0112] From the multidimensional coupled behavioral feature sequence, the intent-oriented behavioral feature component and the environment-responsive behavioral feature component are separated, and then arranged in chronological order to form an intent feature time series matrix (each row is an intent feature vector at a given time point) and an environment feature time series matrix (each row is an environment feature vector at a given time point). Simultaneously, the effective driving weight sequence and the effective feedback correction sequence generated in step S139 are also arranged in chronological order to form a driving weight time series matrix and a feedback correction time series matrix. These four matrices are aligned in the time dimension.

[0113] Step S142: Perform a feature concatenation operation on the driving weight time series matrix and the intent feature time series matrix to generate a combined feature matrix of intent features and driving weights; perform a feature concatenation operation on the feedback correction time series matrix and the environment feature time series matrix to generate a combined feature matrix of environment features and feedback corrections.

[0114] For the same time point, concatenate one row of the driving weight time series matrix (i.e., the effective driving weight parameter vector at that time point) with one row of the intent feature time series matrix (i.e., the intent feature vector at that time point) along the feature dimension to obtain a new vector. Perform this operation on all time points, and the resulting matrix is ​​the combined feature matrix of intent features and driving weights. Similarly, concatenate one row of the feedback correction time series matrix with one row of the environmental feature time series matrix to obtain the combined feature matrix of environmental features and feedback corrections.

[0115] Step S143: Perform feature concatenation operation on the combined feature matrix of intent features and driving weights and the combined feature matrix of environmental features and feedback correction to generate a first-round combined feature matrix. The feature vector of each time point in the first-round combined feature matrix simultaneously contains the intent feature component weighted by the driving weights and the environmental feature component weighted by the feedback correction.

[0116] The two combined feature matrices generated in step S142 are concatenated again at the same time point. That is, for time point t, the combined feature vector of its intent feature and driving weights and the combined feature vector of its environmental features and feedback corrections are concatenated to form a longer vector. After concatenation at all time points, the first round of combined feature matrices is obtained, in which each feature vector simultaneously contains the original intent feature, the feedback correction parameters affecting the intent, the original environmental features, and the driving weight parameters affecting the environment.

[0117] Step S144: Input the first round combined feature matrix into the feature evolution prediction module, perform time-series prediction processing on the first round combined feature matrix through the feature evolution prediction module, generate the predicted feature vector of the first round combined feature matrix at the next time point, and concatenate the predicted feature vector with the first round combined feature matrix to generate the expanded first round combined feature matrix.

[0118] The feature evolution prediction module is a pre-trained temporal prediction model, such as a model based on a Long Short-Term Memory (LSTM) network. The first-round combined feature matrix is ​​fed into this module as the input sequence. Based on the historical information of the input sequence, it predicts the combined feature vector for the next time point (i.e., time point N+1). Then, this predicted new vector is added to the end of the first-round combined feature matrix as a new time point, thereby extending the sequence length by one bit to form an expanded first-round combined feature matrix. This operation simulates behavior evolving one step into the future through interaction.

[0119] Step S145: Concatenate the expanded first-round combined feature matrix, the driving weight time series matrix, and the feedback correction time series matrix to generate the second-round combined feature matrix.

[0120] The expanded first-round combined feature matrix generated in step S144 is concatenated with the original driving weight time series matrix and feedback correction time series matrix (note that the driving weight and feedback correction sequences themselves are not expanded for prediction and retain their original length) after being aligned in the time dimension. Here, "concatenation" refers to concatenating the feature vector of each time point in the expanded first-round combined feature matrix with the corresponding driving weight vector and feedback correction vector in the feature dimension to generate the second-round combined feature matrix. This operation also incorporates the predicted future state into the interaction framework.

[0121] Step S146: Repeat the iterative process of inputting the previous round combined feature matrix into the feature evolution prediction module to generate extended features, and then concatenating the extended features with the driving weight time series matrix and the feedback correction time series matrix, until the preset number of iteration rounds is reached to obtain the final round combined feature matrix.

[0122] The second-round combined feature matrix obtained in step S145 is input again into the feature evolution prediction module to predict the feature vector for the next new time point. This vector is then appended to the end of the sequence to obtain the expanded second-round combined feature matrix. This expanded matrix is ​​then concatenated with the driving weights and feedback correction matrix to obtain the third-round combined feature matrix. This process iterates, with each round using the combined features from the previous round and the static interaction weight matrix to predict a future point, which is then included in the input for the next round. A preset number of iteration rounds is used, for example, K rounds. When the number of iterations reaches K rounds, the iteration stops, and the resulting combined feature matrix is ​​the final round combined feature matrix. The sequence length of this final round combined feature matrix is ​​K points longer than the original intent / environment feature sequence length, representing behavioral evolution prediction over K future time steps.

[0123] Step S147: Extract the spatial coordinates of the feature vectors of each time point in the final round combination feature matrix in the feature space, connect the spatial coordinates of each time point in the order of time dimension to form a spatial coordinate trajectory line, and use the spatial coordinate trajectory line as the behavior evolution path feature sequence reflecting the behavioral evolution trend of the agent.

[0124] Each eigenvector in the final round-combined feature matrix is ​​a point in a high-dimensional space. It can be viewed as a coordinate in that high-dimensional space. For example, a eigenvector of dimension D can be directly considered a point in D-dimensional space. The eigenvector at the first time point is considered the starting point of the trajectory, the eigenvector at the second time point is considered the second point of the trajectory, and so on. Connecting the points corresponding to all time points in chronological order forms a trajectory line in D-dimensional space. This trajectory line is the behavioral evolution path feature sequence, which intuitively describes the continuous movement path of the agent's behavioral state in the high-dimensional representation space driven by intention and environment interaction.

[0125] Step S148: Calculate the Euclidean distance between the spatial coordinates of adjacent time points in the behavior evolution path feature sequence, organize the Euclidean distance values ​​into a behavior evolution step length sequence according to the chronological order of time, calculate the ratio between adjacent step length values ​​in the behavior evolution step length sequence, and append the ratio as a behavior evolution rate change sequence to the behavior evolution path feature sequence.

[0126] To further enrich the dynamic information of the behavioral evolution path feature sequence, the dynamic properties of the sequence are calculated. First, for each pair of adjacent points (point t and point t+1) in the sequence, their Euclidean distance in the feature space is calculated, which is the behavioral evolution step size, representing the magnitude of behavioral change between two time points. All step sizes are arranged in chronological order to form a behavioral evolution step size sequence. Then, for each pair of adjacent step sizes (step size t and step size t+1) in the behavioral evolution step size sequence, the ratio of step size t+1 to step size t is calculated to obtain the rate of change of behavioral evolution velocity, which reflects the acceleration of behavioral change. All calculated rate of change of velocity are arranged in chronological order as a new sequence and appended to the original behavioral evolution path feature sequence. The final behavioral evolution path feature sequence not only contains the position coordinates of each time point in high-dimensional space, but also the Euclidean distance between adjacent points and the rate of change of distance, thus more comprehensively describing the trajectory, velocity, and acceleration information of behavioral evolution.

[0127] Further, step S140 may also include: step S1410: obtaining the intention change edge weight and environment change edge weight of each directed edge in the behavior type transition graph structure, and combining the intention change edge weight and environment change edge weight of the directed edges into a tuple form to form a set of interaction feature tuples for each directed edge.

[0128] From the behavior type transition graph structure constructed in steps S1313 and S1314, extract the intention change edge weight and environment change edge weight for each directed edge. Combine these two weight values ​​into a tuple, for example, (W_intent_edge_i, W_env_edge_i). All the tuples of directed edges together constitute the interaction feature tuple set, which, from the perspective of graph structure, summarizes the strength characteristics of the interaction between intention and environment on different behavior type transition paths.

[0129] Step S1411: The pairs in the set of interaction feature pairs are weighted and averaged according to the frequency of occurrence of their corresponding directed edges to generate an average interaction feature pair for each directed edge. The intention change edge weight component in the average interaction feature pair is used as the benchmark driving weight of the directed edge, and the environment change edge weight component in the average interaction feature pair is used as the benchmark feedback weight of the directed edge.

[0130] For each directed edge in the behavior type transition graph, in addition to the intention change edge weight and the environment change edge weight, it also carries the occurrence frequency weight calculated in step S1311. Now, the occurrence frequency weight is used as a weighting coefficient to perform a weighted average on the set of interaction feature pairs. Specifically, for all directed edges, the intention change edge weight of each directed edge is multiplied by its occurrence frequency weight, the sum is divided by the sum of all occurrence frequency weights, and the weighted average intention change edge weight is obtained. Similarly, the weighted average environment change edge weight is calculated. These two weighted averages constitute the average interaction feature pair for the directed edge. Then, the intention change edge weight component in the average interaction feature pair is used as the baseline driving weight of the directed edge, and the environment change edge weight component is used as the baseline feedback weight of the directed edge. These two baseline values ​​represent the typical strength of the interaction between intention and environment on the behavior type transition path after considering the transition frequency.

[0131] Step S1412: Map the basic behavioral units in the agent behavior data stream to graph nodes of the behavior type transition graph structure according to the type identifier code of the interaction operation, and use the original behavior embedding vector corresponding to the basic behavioral unit as the node feature vector of the graph node.

[0132] Returning to the agent behavior data stream, for each basic behavior unit, the corresponding graph node is found in the behavior type transition graph structure based on its interaction operation type identifier code. Then, the original behavior embedding vector generated in step S121 for that basic behavior unit is used as the initial node feature vector for that graph node. Since the same graph node may correspond to multiple basic behavior units, the node will obtain an initial feature set consisting of multiple original behavior embedding vectors.

[0133] Step S1413: Use the benchmark driving weight as the propagation attenuation coefficient on the directed edge, and the benchmark feedback weight as the back propagation attenuation coefficient on the directed edge. Perform the node feature iterative propagation operation of the graph neural network on the behavior type transition graph structure. Through multiple iterative propagations, the node feature vectors are diffused and fused along the directed edge according to the propagation attenuation coefficient.

[0134] Now, we run a graph neural network, such as a graph convolutional network or a graph attention network, on a behavior type transition graph structure. The propagation rules of the graph neural network are as follows: For each directed edge from the source node to the target node, the baseline driving weights on it are used as decay coefficients for forward propagation. In each iteration, the feature vector of the target node receives feature vectors from all its predecessor source nodes, and these feature vectors are multiplied by the baseline driving weights on the corresponding directed edges and then aggregated. Simultaneously, to simulate the feedback correction effect, the graph neural network also supports backpropagation: From the target node to the source node, the baseline feedback weights on it are used as decay coefficients for backpropagation. In each iteration, the feature vector of the source node also receives feature vectors from its successor target node, multiplied by the corresponding baseline feedback weights and then aggregated. Through multiple iterations (e.g., T iterations), the node feature vectors diffuse and fuse along the directed edges according to these decay coefficients, allowing feature information to propagate not only between behaviors of the same type but also between different types of behaviors with causal or correlational relationships. The above propagation mechanism simulates the mutual influence of intention-driven (forward) and environmental feedback (backward) in the behavior network.

[0135] Step S1414: Extract the updated node feature vector of each graph node after the iterative propagation operation of the graph neural network, sort the updated node feature vectors according to the timestamp information of the corresponding basic behavioral units, and generate the time sequence of the updated node feature vectors.

[0136] After the iterative propagation of the graph neural network, each graph node obtains an updated node feature vector, which integrates its own initial features and the features of its neighboring nodes propagated through directed edges. Since each basic behavioral unit is mapped to a graph node, each basic behavioral unit can obtain the updated node feature vector of its corresponding graph node. Then, the above updated node feature vectors are sorted according to the timestamp information of the corresponding basic behavioral unit to form a sequence that is temporally aligned with the original behavioral data stream, i.e., the temporal sequence of updated node feature vectors.

[0137] Step S1415: Input the updated node feature vector time series into the time series feature extraction network, and perform long-short-term dependency modeling on the updated node feature vector time series through the time series feature extraction network to extract the time series evolution trend component of the updated node feature vector time series as the behavior evolution path feature sequence.

[0138] The updated node feature vector time sequence generated in step S1414 is input into a temporal feature extraction network, which can be a Long Short-Term Memory (LSTM) network, a gated recurrent unit (GRU), or a temporal convolutional network. Taking an LSM network as an example, the updated node feature vector time sequence is input sequentially step by step. The forget gate, input gate, and output gate within the LSM network control the long-term retention and short-term updates of information through a gating mechanism, thereby effectively capturing long-distance dependencies in the sequence. After processing the entire sequence, the hidden state output at each time step is taken. These hidden state vectors constitute a new sequence, which is the temporal evolution trend component. The above hidden state vector sequence is output as the final behavioral evolution path feature sequence, which not only includes node features that have been propagated through a graph neural network and incorporated with structured interaction relationships, but also models the dynamic evolution trend of these features on the time axis through a temporal network.

[0139] Step S150: Perform behavior pattern clustering processing on the behavior evolution path feature sequence, determine the behavior pattern category label of the agent based on the clustering result, match the behavior pattern category label with the preset risk behavior feature library, and generate the agent behavior risk quantitative assessment result.

[0140] This step aims to identify specific behavioral patterns from behavioral evolution paths and conduct risk assessments based on pre-defined risk knowledge. Step S150 specifically includes the following sub-steps to achieve this risk assessment.

[0141] Step S151: Input the behavior evolution path feature sequence into the behavior pattern clustering network. The behavior pattern clustering network performs dimensionality reduction processing on the behavior evolution path feature sequence through an autoencoder structure to generate a hidden layer representation vector sequence of the behavior evolution path feature sequence in a low-dimensional latent space.

[0142] Behavioral evolution path feature sequences are typically high-dimensional, which is unfavorable for cluster analysis. Therefore, they are input into a behavioral pattern clustering network. The core of this network is an autoencoder structure. The autoencoder first compresses each feature vector in the high-dimensional behavioral evolution path feature sequence into a low-dimensional hidden representation vector through an encoder network (composed of multiple fully connected layers). This encoding process aims to retain the most important essential features in the input data that distinguish different patterns. The decoder network is responsible for reconstructing the hidden representation vector back to the original input dimension. The encoder and decoder are jointly trained by minimizing the reconstruction error. After training, the sequence output by the encoder is the hidden representation vector sequence, which represents the original behavioral trajectory in a low-dimensional, compact space.

[0143] Step S152: Input the hidden layer representation vector sequence into the clustering assignment layer of the behavior pattern clustering network. The clustering assignment layer calculates the membership probability distribution of each hidden layer representation vector to each preset cluster center through a soft assignment mechanism, and uses the cluster center number corresponding to the maximum probability value in the membership probability distribution as the behavior pattern category label corresponding to the hidden layer representation vector.

[0144] The sequence of hidden layer representation vectors is input into the clustering assignment layer. This layer pre-defines multiple cluster center vectors, each representing a different potential behavioral pattern. The clustering assignment layer employs a soft assignment mechanism, such as using a t-distributed random neighborhood embedding similarity metric. For each input hidden layer representation vector, its similarity to each pre-definement cluster center is calculated. These similarities are then normalized to form a probability distribution, representing the probability distribution of the hidden layer representation vector's membership to each cluster center. The cluster center with the highest probability value is selected, and its corresponding number becomes the behavioral pattern category label for that hidden layer representation vector (and thus, the behavior at the original time point).

[0145] Step S153: Perform segmentation and labeling processing on the behavior evolution path feature sequence according to the behavior pattern category label, and aggregate the behavior evolution path feature sequence segments corresponding to consecutive time points with the same behavior pattern category label into a single behavior pattern segment to generate a behavior pattern segment set. The behavior pattern segment set includes the start time point position, end time point position and behavior pattern category label of each behavior pattern segment.

[0146] Based on the behavioral pattern category labels assigned to each time point in step S152, the behavioral evolution path feature sequence is segmented on the time axis. For multiple consecutive time points where the behavioral pattern category labels are identical, the corresponding behavioral evolution path feature sequence segments are aggregated to form a single behavioral pattern segment. This single behavioral pattern segment is recorded as: the start time point position (i.e., the index of the first time point within the segment), the end time point position (i.e., the index of the last time point within the segment), and the behavioral pattern category label for that segment. All of these segments together constitute a set of behavioral pattern segments.

[0147] Step S154: Extract the spatial geometric features of the behavioral evolution path feature sequence fragments of each behavioral pattern fragment in the behavioral pattern fragment set. The spatial geometric features include the number of curvature change extreme points of the behavioral evolution path feature sequence fragments in the feature space and the cumulative amount of torsion change of the behavioral evolution path feature sequence fragments in the feature space.

[0148] For each segment in the set of behavioral pattern segments, the geometric morphology of its corresponding behavioral evolution path feature sequence segment is further analyzed. The curvature of the segment in the feature space is calculated; curvature describes the degree of path bending. The number of extreme points (i.e., local maxima and minima) on the curvature change curve is counted as the number of curvature change extreme points. Simultaneously, for higher-dimensional spaces (three dimensions or more), the torsion of the path is calculated; torsion describes the degree to which the path deviates from the plane. The area under the torsion change curve is calculated, or the absolute value of the torsion is integrated to obtain the cumulative torsion change. These two geometric morphological features describe the intrinsic structural complexity of the behavioral pattern segments from different perspectives.

[0149] Step S155: Associate and store the spatial geometric features with the behavior pattern category labels to generate a set of behavior pattern labeled segments with spatial geometric features. Arrange the behavior pattern segments with the same behavior pattern category labels in the set of behavior pattern labeled segments in chronological order to form a temporal evolution chain of behavior patterns under a single behavior pattern category.

[0150] The two spatial geometric features calculated in step S154—the number of curvature change extrema and the cumulative amount of torsion change—are associated with the behavior pattern category label of the segment in step S153, forming a richer labeled segment. All the above labeled segments constitute a behavior pattern labeled segment set. Then, all segments with the same behavior pattern category label are selected from this set and arranged according to their order of appearance on the original timeline, thereby constructing a behavior pattern temporal evolution chain for each behavior pattern category. This behavior pattern temporal evolution chain reflects how the same type of behavior pattern repeats, alternates, or evolves over time.

[0151] Step S156: Input the temporal evolution chain of the behavior pattern into the risk behavior feature library for matching processing. The risk behavior feature library contains a spatial geometric feature template of a preset risk behavior pattern and a temporal evolution template of a preset risk behavior pattern. Calculate the morphological similarity between the spatial geometric feature of the temporal evolution chain of the behavior pattern and the spatial geometric feature template of the preset risk behavior pattern.

[0152] The risk behavior feature library is pre-built and stores features of known risk behavior patterns. For the temporal evolution chain of the behavior pattern to be evaluated, its corresponding spatial geometric features (i.e., the number of extreme points of curvature change and the cumulative amount of torsion change) are first extracted. Then, this set of features is compared with the spatial geometric feature template of each preset risk behavior pattern in the risk behavior feature library. Morphological similarity can be obtained by calculating the cosine similarity or the reciprocal of the Euclidean distance between two feature vectors; the larger the value, the greater the morphological similarity.

[0153] Step S157: Calculate the sequence edit distance between the transition sequence of the behavior pattern category label in the behavior pattern temporal evolution chain and the transition sequence of the behavior pattern category label in the temporal evolution template of the preset risk behavior pattern, and normalize the sequence edit distance to obtain the temporal evolution matching degree.

[0154] Besides being a single fragment, risky behavior often manifests as an ordered transition of a series of patterns. Therefore, the behavioral pattern category labels in the temporal evolution chain are extracted sequentially to form a transition sequence (e.g., "Pattern A, Pattern B, Pattern C"). Simultaneously, the temporal evolution templates in the risky behavior feature library also contain similar label transition sequences. The edit distance (Levenshtein distance) between these two sequences is calculated. The edit distance measures the minimum number of single-character editing operations (insertion, deletion, replacement) required to transform one sequence into another. The smaller the edit distance, the more similar the two sequences are. The edit distance is divided by the length of the longer of the two sequences and normalized to a value between 0 and 1. Subtracting this normalized edit distance from 1 yields the temporal evolution matching degree; a higher value indicates a more matching temporal evolution patterns.

[0155] Step S158: Calculate the comprehensive risk matching score based on the morphological similarity and the temporal evolution matching degree, mark the corresponding behavioral pattern temporal evolution chain with the comprehensive risk matching score exceeding the preset risk threshold as the risk behavior evolution chain, and output the behavioral pattern category label sequence corresponding to the risk behavior evolution chain as the intelligent agent behavior risk quantification assessment result.

[0156] For each candidate risky behavior pattern, the morphological similarity calculated in step S156 and the temporal evolution matching degree calculated in step S157 are weighted and summed to obtain a comprehensive risk matching score. For example, the weight of morphological similarity can be preset to 0.6, and the weight of temporal evolution matching degree can be preset to 0.4. A preset risk threshold is set, such as 0.7. If the comprehensive risk matching score of a certain behavior pattern's temporal evolution chain with a certain preset risky behavior pattern exceeds 0.7, then the behavior pattern's temporal evolution chain is marked as a risky behavior evolution chain. Finally, the behavior pattern category label sequence (e.g., "pattern X, pattern Y, pattern Z") corresponding to the risky behavior evolution chain is output as the final quantitative assessment result of the agent's behavior risk, used for alarms or subsequent analysis.

[0157] For example, further, step S150 may also include: step S1510: inputting the behavior evolution path feature sequence into the temporal pattern discovery network, wherein the temporal pattern discovery network performs multi-scale temporal feature extraction processing on the behavior evolution path feature sequence through a temporal convolutional encoder to generate a multi-scale temporal feature pyramid of the behavior evolution path feature sequence at different time scales.

[0158] In this embodiment, to discover behavioral patterns of varying durations and complexities from the feature sequences of behavioral evolution paths, these patterns are input into a temporal pattern discovery network. This network first comprises a temporal convolutional encoder. This encoder consists of multiple parallel temporal convolutional branches, each using convolutional kernels with different dilation rates. For example, one branch uses a kernel with a dilation rate of 1 to extract local, short-term patterns; another uses a kernel with a dilation rate of 2 to extract medium-length patterns; and yet another uses a kernel with a dilation rate of 4 to extract global, long-term patterns. The output of each branch is a feature sequence representing the temporal features at that time scale. These feature sequences from different branches are combined (e.g., concatenated along the feature dimension or weighted fusion) to form a multi-scale temporal feature pyramid. Each layer of this multi-scale temporal feature pyramid structure corresponds to a specific time scale, providing a multi-resolution perspective on behavioral evolution paths.

[0159] Step S1511: Input the multi-scale temporal feature pyramid into the pattern candidate generation module of the temporal pattern discovery network. The pattern candidate generation module scans the multi-scale temporal feature pyramid in the time dimension through a sliding window and extracts the local feature vector at each sliding window position as a behavioral pattern candidate fragment.

[0160] The multi-scale temporal feature pyramid generated in step S1510 is input into the pattern candidate generation module, which uses a sliding window of fixed length (e.g., L time steps) to scan each layer of the pyramid in the time dimension. For each sliding window position, feature vectors of all time steps within that window (possibly from different layers of the pyramid) are extracted and aggregated (e.g., through average pooling or max pooling) to form a fixed-dimensional local feature vector. This local feature vector represents a candidate behavioral pattern segment. The sliding window moves with a certain step size, thereby generating a series of overlapping or non-overlapping behavioral pattern candidate segments in the time series.

[0161] Step S1512: Calculate the cosine similarity matrix between the candidate behavior pattern segments, perform spectral clustering on the cosine similarity matrix, and classify the candidate behavior pattern segments with cosine similarity higher than a preset similarity threshold into the same behavior pattern category cluster.

[0162] For all candidate behavioral patterns generated in step S1511, the cosine similarity between each pair of them is calculated to construct a cosine similarity matrix. This cosine similarity matrix is ​​a symmetric matrix, where each element represents the degree of similarity between two corresponding candidate fragments. Then, a spectral clustering algorithm is applied to this cosine similarity matrix. Spectral clustering first converts the similarity matrix into a Laplace matrix, then calculates the eigenvectors of the Laplace matrix, using these eigenvectors to map the original data points (i.e., candidate fragments) to a low-dimensional space, and finally uses traditional clustering algorithms such as K-means to perform clustering in the low-dimensional space. Through spectral clustering, candidate fragments with a cosine similarity higher than a preset similarity threshold (e.g., 0.7) are grouped into the same behavioral pattern category cluster. In this way, the candidate fragments within each cluster are highly similar in the feature space, representing the same behavioral pattern.

[0163] Step S1513: Extract the central feature vector of all candidate fragments of behavior patterns in each behavior pattern category cluster as the behavior pattern prototype vector of that behavior pattern category cluster, associate the behavior pattern prototype vector with the cluster identifier of the behavior pattern category cluster and store it to generate a behavior pattern prototype library.

[0164] For each behavior pattern category cluster formed after spectral clustering, the center of the local feature vector corresponding to all candidate fragments of behavior patterns within the cluster is calculated. The calculation of the central feature vector can be achieved by calculating the averaging of all vectors along their respective dimensions. This central feature vector is the behavior pattern prototype vector for that cluster, representing the most typical features of that pattern. Each behavior pattern prototype vector is associated with the cluster identifier of its corresponding cluster (e.g., "pattern_001", "pattern_002") and stored in a library, which is the behavior pattern prototype library. This behavior pattern prototype library provides a compact representation of various behavior patterns discovered from historical behavior.

[0165] Step S1514: Perform sliding window matching processing on the behavior evolution path feature sequence and the behavior pattern prototype vector in the behavior pattern prototype library, calculate the matching score between the behavior evolution path feature sequence fragment at each sliding window position and each behavior pattern prototype vector, and take the cluster identifier corresponding to the behavior pattern prototype vector with the highest matching score as the behavior pattern category label at that sliding window position.

[0166] In the application phase, the behavior evolution path feature sequence is reacquired. A sliding window of the same length as in step S1511 is slid across the behavior evolution path feature sequence. For each sliding window position, a feature sequence fragment within the window is extracted, and the matching score between this fragment and each behavior pattern prototype vector in the behavior pattern prototype library is calculated. The matching score can be calculated using the reciprocal of the dynamic time-normalized distance or cosine similarity. The behavior pattern prototype vector with the highest matching score is selected, and its corresponding cluster identifier (i.e., behavior pattern category label) is assigned to that sliding window position. By performing this operation on all sliding window positions, each time interval on the entire behavior evolution path feature sequence is labeled with a behavior pattern category label.

[0167] Step S1515: Perform similarity matching between the preset risk behavior patterns in the risk behavior feature library and the behavior pattern prototype vectors in the behavior pattern prototype library, and associate the risk level label corresponding to the preset risk behavior pattern with a similarity exceeding a preset matching threshold with the behavior pattern category label corresponding to the matched behavior pattern prototype vector.

[0168] The risk behavior feature library stores known risk behavior patterns, each associated with a risk level label (e.g., "low risk," "medium risk," "high risk"). For each behavior pattern prototype vector in the behavior pattern prototype library, its similarity to the feature vector of each preset risk behavior pattern in the risk behavior feature library is calculated. If the calculated similarity exceeds a preset matching threshold (e.g., 0.8), the behavior pattern represented by the behavior pattern prototype vector is considered highly similar to the corresponding risk behavior pattern. In this case, the risk level label of the risk behavior pattern is associated (or mapped) to the behavior pattern category label corresponding to the behavior pattern prototype vector. Thus, each behavior pattern category label (e.g., "pattern_001") may be associated with one or more risk level labels.

[0169] Step S1516: Generate a quantitative assessment result of agent behavior risk based on the behavior pattern category label and its associated risk level label at the position of the sliding window. The quantitative assessment result of agent behavior risk includes the risk level label corresponding to each time interval in the behavior evolution path feature sequence and the confidence score of the risk level label.

[0170] For the behavior pattern category label determined for each sliding window position (i.e., each time interval) in step S1514, the risk level label associated with that behavior pattern category label is obtained through the mapping relationship in step S1515. If the same behavior pattern category label is associated with multiple risk level labels (e.g., due to matching multiple preset risk behavior patterns), the risk level label with the highest similarity is selected, or a weighted average risk level is calculated. Simultaneously, the matching score calculated in step S1514 is used as the confidence score for that risk level label. Finally, a quantitative assessment result of agent behavior risk is generated, which includes the risk level label (e.g., "medium risk") corresponding to each time interval (e.g., "time interval [T_start, T_end]") on the behavior evolution path feature sequence, and the confidence score for that risk assessment (e.g., "0.85").

[0171] Figure 2This application illustrates a multi-dimensional behavior modeling and risk quantification system 100 for enterprise AI agents, comprising a processor 1001 and a memory 1003. The processor 1001 and memory 1003 are connected, for example, via a bus 1002. Optionally, the multi-dimensional behavior modeling and risk quantification system 100 for enterprise AI agents may further include a transceiver 1004. The transceiver 1004 can be used for data interaction between this multi-dimensional behavior modeling and risk quantification system for enterprise AI agents and other multi-dimensional behavior modeling and risk quantification systems for enterprise AI agents, such as sending and / or receiving data. It should be noted that in actual scheduling, the transceiver 1004 is not limited to one, and the structure of this multi-dimensional behavior modeling and risk quantification system 100 for enterprise AI agents does not constitute a limitation on the embodiments of this application.

[0172] The memory 1003 is used to store program code for executing the embodiments of this application, and its execution is controlled by the processor 1001. The processor 1001 is used to execute the program code stored in the memory 1003 to implement the steps shown in the foregoing method embodiments.

[0173] The above description is only an optional implementation method for some implementation scenarios of this application. It should be noted that for those skilled in the art, other similar implementation methods based on the technical concept of this application, without departing from the technical concept of this application, also fall within the protection scope of the embodiments of this application.

Claims

1. A multi-dimensional behavioral modeling and risk quantification method for enterprise AI agents, characterized in that, The method includes: Collect the agent behavior data stream generated by the enterprise AI agent during operation. The agent behavior data stream includes records of behavioral events generated when the agent interacts with the external environment and the temporal correlation between the behavioral event records. The agent behavior data stream is input into a pre-constructed multidimensional behavior decoupling network for behavior feature separation processing. The intention-oriented behavior features of the agent are extracted from the behavior event records by the intention decoder in the multidimensional behavior decoupling network, and the environment-responsive behavior features of the agent are extracted from the behavior event records by the environment response decoder in the multidimensional behavior decoupling network. A dynamic interaction mapping relationship is constructed between the intent-oriented behavioral features and the environment-responsive behavioral features. The dynamic interaction mapping relationship includes the driving weight parameters of the intent-oriented behavioral features on the environment-responsive behavioral features and the feedback correction parameters of the environment-responsive behavioral features on the intent-oriented behavioral features. The intention-oriented behavioral features and the environment-responsive behavioral features are iteratively fused based on the driving weight parameters and the feedback correction parameters to generate a behavioral evolution path feature sequence that reflects the behavioral evolution trend of the agent. The behavioral evolution path feature sequence is subjected to behavioral pattern clustering processing. Based on the clustering results, the behavioral pattern category label of the agent is determined. The behavioral pattern category label is matched with the preset risk behavior feature library to generate a quantitative assessment result of the agent's behavioral risk. The data stream of agent behavior generated by the enterprise's AI agent during operation includes: The system monitors the interaction event nodes generated by the enterprise AI agent during data exchange with external data sources while executing enterprise business process tasks. Each interaction event node is recorded as a basic behavior unit. The basic behavior unit includes the identity information of the interaction initiator, the address information of the interaction receiver, the type identification code of the interaction operation, the timestamp information of the interaction operation, and the data payload feature description carried by the interaction operation. Extract the data payload feature description carried in the basic behavior unit, perform semantic parsing on the data payload feature description, and obtain the business intent type label and the confidence score of the business intent type label corresponding to the basic behavior unit. Based on the timestamp information of the interaction operations in the basic behavior units, the sequential relationship of the basic behavior units in the time dimension is constructed. Basic behavior units with direct successor-successor relationships are connected by directed edges to form a temporal association relationship between behavior event records. The temporal association relationship includes the time interval parameter between the predecessor behavior unit identifier and the successor behavior unit identifier, as well as the type transition probability between the interaction operation type of the predecessor behavior unit identifier and the interaction operation type of the successor behavior unit identifier. The basic behavioral units and the temporal associations are organized into an agent behavior data stream in ascending order of timestamp information. Each basic behavioral unit in the agent behavior data stream carries a forward reference pointer to its predecessor basic behavioral unit and a backward reference pointer to its successor basic behavioral unit. The basic behavioral units with the same interaction initiator identity information in the intelligent agent behavior data stream are aggregated to form an individual behavior trajectory sequence with a single enterprise AI intelligent agent as the dimension. The individual behavior trajectory sequence contains all the basic behavioral units generated by the enterprise AI intelligent agent at different time points and the complete temporal correlation network between the basic behavioral units. The basic behavioral units with the same business intent type label in the individual behavioral trajectory sequence are segmented and aggregated according to the timestamp information to generate a set of continuous behavioral segments that reflect the enterprise AI agent under the drive of a specific business intent. The set of continuous behavioral segments includes the start timestamp of the behavioral segment, the end timestamp of the behavioral segment, and the arrangement order of the basic behavioral units within the behavioral segment. Extract the type identifier code of the interaction operation of the first basic behavior unit of each continuous behavior segment in the set of continuous behavior segments as the segment entry behavior type identifier, and extract the type identifier code of the interaction operation of the last basic behavior unit of each continuous behavior segment in the set of continuous behavior segments as the segment exit behavior type identifier. The boundary behavior type conversion path of the persistent behavior segment is constructed based on the segment entry behavior type identifier and the segment exit behavior type identifier. The boundary behavior type conversion path includes the intermediate behavior type sequence from the entry behavior type to the exit behavior type and the statistical value of the number of transitions between adjacent behavior types. The persistent behavior segments with the same boundary behavior type conversion path in the persistent behavior segment set are classified and processed to generate a cluster of similar behavior segments with the boundary behavior type conversion path as the cluster center. The cluster of similar behavior segments includes all persistent behavior segments belonging to the same boundary behavior type conversion path and the frequency weight of each persistent behavior segment in the cluster of similar behavior segments. Calculate the duration deviation parameter between the duration of each persistent behavior segment in the cluster of similar behavior segments and the average duration of all persistent behavior segments in the cluster of similar behavior segments. Mark persistent behavior segments whose duration deviation parameter exceeds a preset tolerance threshold as abnormal behavior segments, and mark the boundary behavior type conversion path corresponding to the abnormal behavior segment as an abnormal behavior path identifier.

2. The multi-dimensional behavioral modeling and risk quantification method for enterprise AI agents according to claim 1, characterized in that, The step of inputting the agent's behavior data stream into a pre-constructed multidimensional behavior decoupling network for behavior feature separation processing, extracting the agent's intent-oriented behavior features from the behavior event records using the intent decoder in the multidimensional behavior decoupling network, and extracting the agent's environment-responsive behavior features from the behavior event records using the environment response decoder in the multidimensional behavior decoupling network includes: The basic behavioral units in the agent behavior data stream are input into the shared feature encoding module of the multidimensional behavior decoupling network. The shared feature encoding module performs embedding vectorization processing on the identity information of the interaction initiator, the address information of the interaction receiver, the type identification code of the interaction operation, and the data payload feature description of the basic behavioral unit to generate the original behavior embedding vector of the basic behavioral unit. The original behavior embedding vector is input into the intention decoder and the environment response decoder in parallel. The intention decoder performs semantic intention parsing on the original behavior embedding vector through a self-attention mechanism, and extracts the behavior driving factor features that are directly related to the achievement of business goals from the original behavior embedding vector as intention-oriented behavior features. The intention-oriented behavior features include business goal orientation parameters and behavior purpose persistence parameters. The environmental response decoder performs context dependency modeling on the original behavior embedding vector using a conditional random field model. It then extracts behavioral adaptation factor features associated with changes in the external environment state from the original behavior embedding vector as environmental response behavioral features. These environmental response behavioral features include environmental state sensitivity parameters and behavioral feedback timeliness parameters. The intent-oriented behavior features output by the intent decoder and the environment-responsive behavior features output by the environment response decoder are input into the feature orthogonalization constraint layer of the multidimensional behavior decoupling network. The mutual information between the intent-oriented behavior features and the environment-responsive behavior features is calculated through the feature orthogonalization constraint layer. Based on the goal of minimizing the mutual information, the network weight parameters of the intent decoder and the environment response decoder are jointly optimized and adjusted. The intention-oriented behavior features processed by the feature orthogonalization constraint layer are coupled with the original behavior embedding vector through residual connection processing to generate enhanced intention-oriented behavior features with compensation for the original behavior information. The environment-responsive behavior features processed by the feature orthogonalization constraint layer are coupled with the original behavior embedding vector through residual connection processing to generate enhanced environment-responsive behavior features with compensation for the original behavior information. The enhanced intent-oriented behavior features are input into the first temporal convolution module of the multidimensional behavior decoupling network. The enhanced intent-oriented behavior features are subjected to sliding window convolution operation in the time dimension through the first temporal convolution module. The change trend features of the enhanced intent-oriented behavior features within the continuous time window are extracted as the evolution trajectory of intent-oriented behavior. The enhanced environmental responsive behavior features are input into the second temporal convolution module of the multidimensional behavior decoupling network. The enhanced environmental responsive behavior features are then subjected to sliding window convolution operation in the time dimension through the second temporal convolution module. The changing trend features of the enhanced environmental responsive behavior features within a continuous time window are extracted as the evolution trajectory of the environmental responsive behavior. The intent-driven behavior evolution trajectory and the environment-responsive behavior evolution trajectory are input into the trajectory alignment module of the multidimensional behavior decoupling network. The trajectory alignment module calculates the phase offset between the intent-driven behavior evolution trajectory and the environment-responsive behavior evolution trajectory in the time dimension. Based on the phase offset, the intent-driven behavior evolution trajectory is subjected to time axis translation calibration to generate an environment-responsive behavior evolution trajectory that is time-aligned with the intent-driven behavior evolution trajectory. The time-aligned environmental responsive behavior evolution trajectory and the intent-oriented behavior evolution trajectory are subjected to time-point feature concatenation processing to generate a multi-dimensional coupled behavior feature sequence with time synchronization correlation. The feature vector corresponding to each time point in the multi-dimensional coupled behavior feature sequence simultaneously contains the intent-oriented behavior feature information and the environmental responsive behavior feature information at that time point.

3. The multi-dimensional behavioral modeling and risk quantification method for enterprise AI agents according to claim 1, characterized in that, The construction of the dynamic interaction mapping relationship between the intent-oriented behavioral features and the environment-responsive behavioral features includes: Using the intention-oriented behavioral feature vectors at adjacent time points as forward driving inputs and the environmental response behavioral feature vectors at adjacent time points as forward driving outputs, a forward driving mapping function for intention to environmental response is constructed. The forward driving mapping function is used to calculate the linear transformation matrix required to map the intention-oriented behavioral feature vector to the environment-responsive behavioral feature vector in each adjacent time point pair. The off-diagonal elements in the linear transformation matrix are extracted and processed to generate the driving weight parameters of the intention-oriented behavioral feature on the environment-responsive behavioral feature. The driving weight parameters are used to characterize the driving strength of each dimension component of the intention-oriented behavioral feature on each dimension component of the environment-responsive behavioral feature. By using the environmental response-oriented behavioral feature vectors at adjacent time points as the inverse correction input and the intention-oriented behavioral feature vectors at adjacent time points as the inverse correction output, a reverse correction mapping function for environmental response to intention is constructed. The linear transformation matrix required to map the environmental responsive behavior feature vector to the intention-oriented behavior feature vector in each adjacent time point pair is calculated using the reverse correction mapping function. The off-diagonal elements in the linear transformation matrix are extracted and processed to generate feedback correction parameters of the environmental responsive behavior feature to the intention-oriented behavior feature. The feedback correction parameters are used to characterize the correction strength of each dimension component of the environmental responsive behavior feature to each dimension component of the intention-oriented behavior feature. The driving weight parameters are organized into a time series sequence according to the time dimension. The time series sequence of driving weight parameters is processed by difference operation to extract the change amplitude value of the driving weight parameters between adjacent time points as the dynamic fluctuation feature of driving weight. The feedback correction parameters are organized into a time series sequence according to the time dimension. The time series sequence is processed by difference operation to extract the change amplitude value of the feedback correction parameters between adjacent time points as the dynamic fluctuation feature of feedback correction. The driving weight dynamic fluctuation feature and the feedback correction dynamic fluctuation feature are input into the interactive coupling analysis module. The interactive coupling analysis module calculates the covariance value of the driving weight dynamic fluctuation feature and the feedback correction dynamic fluctuation feature at the same time point. The covariance value is used as an indicator of the interactive coupling strength between intention-oriented behavioral features and environment-responsive behavioral features. The driving weight parameters and feedback correction parameters are filtered based on the interaction coupling strength index. The driving weight parameters at the corresponding time points where the interaction coupling strength index exceeds the preset coupling threshold are extracted as effective driving weight parameters, and the feedback correction parameters at the corresponding time points where the interaction coupling strength index exceeds the preset coupling threshold are extracted as effective feedback correction parameters. The effective driving weight parameters are organized into an effective driving weight sequence according to their corresponding time points, and the effective feedback correction parameters are organized into an effective feedback correction sequence according to their corresponding time points. A mapping table of correspondence between the effective driving weight sequence and the effective feedback correction sequence is constructed. Each entry in the mapping table contains the effective driving weight parameter and the effective feedback correction parameter at the same time point.

4. The multi-dimensional behavioral modeling and risk quantification method for enterprise AI agents according to claim 3, characterized in that, The iterative fusion processing of the intent-oriented behavioral features and the environment-responsive behavioral features based on the driving weight parameters and the feedback correction parameters to generate a behavioral evolution path feature sequence reflecting the behavioral evolution trend of the agent includes: The intent-oriented behavioral features are organized into an intent feature time series matrix in chronological order of time; the environmental response behavioral features are organized into an environmental feature time series matrix in chronological order of time; the effective driving weight sequence is organized into a driving weight time series matrix in chronological order of time; and the effective feedback correction sequence is organized into a feedback correction time series matrix in chronological order of time. The driving weight time series matrix and the intent feature time series matrix are concatenated to generate a combined feature matrix of intent features and driving weights. The feedback correction time series matrix and the environmental feature time series matrix are concatenated to generate a combined feature matrix of environmental features and feedback corrections. The combined feature matrix of the intent feature and the driving weight and the combined feature matrix of the environment feature and the feedback correction are concatenated to generate the first round of combined feature matrix. The feature vector of each time point in the first round of combined feature matrix contains both the intent feature component weighted by the driving weight and the environment feature component weighted by the feedback correction. The first round of combined feature matrix is ​​input into the feature evolution prediction module. The feature evolution prediction module performs time-series prediction processing on the first round of combined feature matrix to generate the predicted feature vector of the first round of combined feature matrix at the next time point. The predicted feature vector is concatenated with the first round of combined feature matrix to generate the expanded first round of combined feature matrix. The expanded first-round combined feature matrix, the driving weight time series matrix, and the feedback correction time series matrix are concatenated to generate the second-round combined feature matrix. The process of repeatedly inputting the previous round combined feature matrix into the feature evolution prediction module to generate extended features, and then concatenating the extended features with the driving weight time series matrix and the feedback correction time series matrix, is repeated until the preset number of iteration rounds is reached to obtain the final round combined feature matrix. Extract the spatial coordinates of the feature vectors of each time point in the final round combination feature matrix in the feature space, connect the spatial coordinates of each time point in the order of time dimension to form a spatial coordinate trajectory line, and use the spatial coordinate trajectory line as the behavior evolution path feature sequence that reflects the behavior evolution trend of the agent. Calculate the Euclidean distance between the spatial coordinates of adjacent time points in the behavioral evolution path feature sequence, organize the Euclidean distance values ​​into a behavioral evolution step length sequence according to the chronological order of time, calculate the ratio between adjacent step length values ​​in the behavioral evolution step length sequence, and append the ratio as a behavioral evolution rate of change sequence to the behavioral evolution path feature sequence.

5. The multi-dimensional behavioral modeling and risk quantification method for enterprise AI agents according to claim 1, characterized in that, The process of performing behavioral pattern clustering on the feature sequences of the behavioral evolution path, determining behavioral pattern category labels for the agent based on the clustering results, and matching the behavioral pattern category labels with a preset risk behavior feature library to generate a quantitative assessment result of the agent's behavioral risk includes: The behavior evolution path feature sequence is input into the behavior pattern clustering network. The behavior pattern clustering network performs dimensionality reduction processing on the behavior evolution path feature sequence through an autoencoder structure to generate a hidden layer representation vector sequence of the behavior evolution path feature sequence in a low-dimensional latent space. The hidden layer representation vector sequence is input into the clustering allocation layer of the behavior pattern clustering network. The clustering allocation layer calculates the membership probability distribution of each hidden layer representation vector to each preset cluster center through a soft allocation mechanism. The cluster center number corresponding to the maximum probability value in the membership probability distribution is used as the behavior pattern category label corresponding to the hidden layer representation vector. The behavior evolution path feature sequence is segmented and labeled according to the behavior pattern category label. The behavior evolution path feature sequence segments corresponding to consecutive time points with the same behavior pattern category label are aggregated into a single behavior pattern segment to generate a behavior pattern segment set. The behavior pattern segment set includes the start time point position, end time point position and behavior pattern category label of each behavior pattern segment. Extract the spatial geometric features of the behavioral evolution path feature sequence fragments of each behavioral pattern fragment in the behavioral pattern fragment set. The spatial geometric features include the number of curvature change extreme points of the behavioral evolution path feature sequence fragments in the feature space and the cumulative amount of torsion change of the behavioral evolution path feature sequence fragments in the feature space. The spatial geometric features are associated with the behavioral pattern category labels and stored to generate a set of behavioral pattern labeled fragments with spatial geometric features. The behavioral pattern fragments with the same behavioral pattern category labels in the set of behavioral pattern labeled fragments are arranged in chronological order to form a temporal evolution chain of behavioral patterns under a single behavioral pattern category. The behavioral pattern temporal evolution chain is input into the risk behavior feature library for matching processing. The risk behavior feature library contains a spatial geometric morphological feature template of a preset risk behavior pattern and a temporal evolution template of a preset risk behavior pattern. The morphological similarity between the spatial geometric morphological features of the behavioral pattern temporal evolution chain and the spatial geometric morphological feature template of the preset risk behavior pattern is calculated. Calculate the sequence edit distance between the transition sequence of behavior pattern category labels in the temporal evolution chain of behavior patterns and the transition sequence of behavior pattern category labels in the temporal evolution template of preset risk behavior patterns, and normalize the sequence edit distance to obtain the temporal evolution matching degree. A comprehensive risk matching score is calculated based on the morphological similarity and the temporal evolution matching degree. The temporal evolution chain of the corresponding behavior pattern that exceeds the preset risk threshold is marked as a risk behavior evolution chain. The behavior pattern category label sequence corresponding to the risk behavior evolution chain is output as the quantitative assessment result of the agent's behavior risk.

6. The multi-dimensional behavioral modeling and risk quantification method for enterprise AI agents according to claim 1, characterized in that, The step of inputting the agent behavior data stream into a pre-constructed multidimensional behavior decoupling network for behavior feature separation processing, extracting the agent's intent-oriented behavior features from the behavior event records through the intent decoder in the multidimensional behavior decoupling network, and extracting the agent's environment-responsive behavior features from the behavior event records through the environment response decoder in the multidimensional behavior decoupling network, further includes: The data payload feature description carried by the interaction operation of the basic behavioral unit in the intelligent agent behavior data stream is obtained, and the data payload feature description is processed by content type identification to divide the data payload feature description into a structured data payload part and an unstructured data payload part. The structured data payload is input into the first feature extraction branch of the intent decoder. The first feature extraction branch performs numerical feature mapping processing on the structured data payload through a multilayer perceptron network to generate structured intent feature components. The unstructured data payload is input into the second feature extraction branch of the intent decoder. The second feature extraction branch performs semantic feature extraction processing on the unstructured data payload through a convolutional neural network to generate unstructured intent feature components. The structured intent feature components and the unstructured intent feature components are aligned in terms of feature dimensions. The aligned structured intent feature components and the unstructured intent feature components are then concatenated to generate a fused intent-oriented behavioral feature. The structured data payload is input into the third feature extraction branch of the environmental response decoder. The third feature extraction branch performs time-series modeling of the numerical change pattern of the structured data payload through a recurrent neural network to generate structured environmental response feature components. The unstructured data payload is input into the fourth feature extraction branch of the environmental response decoder. The fourth feature extraction branch performs contextual association analysis on the semantic content of the unstructured data payload through an attention mechanism to generate unstructured environmental response feature components. The structured environmental response feature components and the unstructured environmental response feature components are aligned in terms of feature dimensions. The aligned structured environmental response feature components and the unstructured environmental response feature components are then spliced ​​together to generate a fused environmental response behavior feature.

7. The multi-dimensional behavioral modeling and risk quantification method for enterprise AI agents according to claim 1, characterized in that, The construction of the dynamic interaction mapping relationship between the intent-oriented behavioral features and the environment-responsive behavioral features also includes: Obtain the type identifier code of the interaction operation of the basic behavior unit in the intelligent agent behavior data stream, organize the type identifier code of the interaction operation into a behavior type sequence according to the chronological order of time, perform frequent pattern mining processing on the behavior type sequence, and extract the frequently occurring continuous behavior type subsequences in the behavior type sequence as behavior type transition patterns. A behavior type transfer graph structure is constructed based on the behavior type transfer pattern. The behavior type transfer graph structure includes a graph structure representation with the type identifier code of the interaction operation as the graph node and the behavior type transfer pattern as the directed edge. Each directed edge carries the frequency weight of the occurrence of the behavior type transfer pattern. The intent-oriented behavioral features are mapped to the graph nodes of the behavior type transfer graph structure to generate an intent-annotated behavior type transfer graph with intent-oriented behavioral feature annotations. The environment-responsive behavioral features are mapped to the graph nodes of the behavior type transfer graph structure to generate an environment-annotated behavior type transfer graph with environment-responsive behavioral feature annotations. Calculate the difference degree of intent-oriented behavior features between adjacent graph nodes in the intent-annotated behavior type transfer graph, use the difference degree as the change magnitude parameter of intent-oriented behavior features in the behavior type transfer process, and attach the change magnitude parameter to the corresponding directed edge as the intention change edge weight. Calculate the difference degree of environmental responsive behavior features between adjacent graph nodes in the environmental annotation behavior type transfer graph, use the difference degree as the change magnitude parameter of environmental responsive behavior features in the behavior type transfer process, and attach the change magnitude parameter to the corresponding directed edge as the environmental change edge weight. The ratio of the intention change edge weight to the environment change edge weight is calculated for the corresponding directed edge to generate the intention-to-environment driving strength ratio for each directed edge. The intention-to-environment driving strength ratio is used as the driving weight parameter of intention-oriented behavioral features to environment-responsive behavioral features. The ratio of the weight of the environmental change edge to the weight of the intention change edge is calculated for the corresponding directed edge to generate the environment-to-intention feedback intensity ratio for each directed edge. The environment-to-intention feedback intensity ratio is used as the feedback correction parameter for the environment-responsive behavior feature to the intention-oriented behavior feature.

8. The multi-dimensional behavioral modeling and risk quantification method for enterprise AI agents according to claim 7, characterized in that, The step of iteratively fusing the intent-oriented behavioral features and the environment-responsive behavioral features according to the driving weight parameters and the feedback correction parameters to generate a behavioral evolution path feature sequence reflecting the behavioral evolution trend of the agent also includes: Obtain the intention change edge weight and environment change edge weight of each directed edge in the behavior type transition graph structure, and combine the intention change edge weight and environment change edge weight of the directed edge into a tuple form to form the interaction feature tuple set of each directed edge. The pairs in the set of interaction feature pairs are weighted and averaged according to the frequency of occurrence of their corresponding directed edges to generate an average interaction feature pair for each directed edge. The intention change edge weight component in the average interaction feature pair is used as the reference driving weight of the directed edge, and the environment change edge weight component in the average interaction feature pair is used as the reference feedback weight of the directed edge. The basic behavioral units in the agent behavior data stream are mapped to graph nodes of the behavior type transition graph structure according to the type identifier code of the interaction operation, and the original behavior embedding vector corresponding to the basic behavioral unit is used as the node feature vector of the graph node. The benchmark driving weights are used as propagation attenuation coefficients on the directed edges, and the benchmark feedback weights are used as back propagation attenuation coefficients on the directed edges. The node feature iterative propagation operation of the graph neural network is performed on the behavior type transition graph structure. Through multiple iterative propagations, the node feature vectors are diffused and fused along the directed edges according to the propagation attenuation coefficients. Extract the updated node feature vector of each graph node after the iterative propagation operation of the graph neural network, sort the updated node feature vectors according to the timestamp information of the corresponding basic behavioral units, and generate the time sequence of updated node feature vectors. The updated node feature vector time series is input into the time series feature extraction network. The time series feature extraction network performs long-short-term dependency modeling on the updated node feature vector time series and extracts the time series evolution trend component of the updated node feature vector time series as the behavior evolution path feature sequence.

9. A multi-dimensional behavioral modeling and risk quantification system for enterprise AI agents, characterized in that, The method includes a processor and a computer-readable storage medium storing machine-executable instructions that, when executed by the processor, implement the multi-dimensional behavior modeling and risk quantification method for enterprise AI agents as described in any one of claims 1-8.

Images