Artificial intelligence-based user intent analysis system and method

Abstract

The present application relates to the technical field of user intention analysis, in particular to a user intention analysis system and method based on artificial intelligence, which generates a single-point intention recognition result representing the current instantaneous operation of the user through the multi-modal behavior data at the current time; generates a time-series intention chain representing the continuous operation of the user according to the multi-modal behavior data at the historical time; when detecting an intention conflict, calculates the optimal transfer path from the current conflict intention node in the time-series intention chain to the candidate target intention node according to the transfer probability weight in the pre-constructed intention transfer probability graph; determines the resource demand corresponding to the candidate target intention node of the optimal transfer path, and dynamically allocates resources to the candidate target intention node and performs intention analysis according to the current available resource state and the resource demand. After effectively detecting and solving the intention conflict, the resource allocation is dynamically optimized, and the accurate understanding and efficient execution of the user intention are realized.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of user intention analysis, in particular to a user intention analysis system and method based on artificial intelligence. BACKGROUND

[0002] In the current booming development of artificial intelligence technology, user intention analysis has become the core function of many intelligent interaction systems. This function aims to accurately understand the current needs and intentions of users through deep mining and analysis of user behavior data, and to provide personalized and intelligent service responses for users.

[0003] However, in practical applications, the current user intention analysis technology faces many challenges. User behavior is highly dynamic and complex, so user intentions will produce intention conflicts. The user intention analysis process requires a large amount of computing resources, such as processor power and memory. Therefore, when intention conflicts occur, how to reasonably and dynamically allocate computing resources becomes a key problem. Existing technologies mostly use static or simple allocation strategies in resource allocation, which cannot accurately adjust according to the actual needs of the current intention analysis task and the real-time state of the resources, resulting in low resource utilization and seriously affecting the overall operation efficiency and performance of the system.

[0004] To solve the above problems, the present application provides a user intention analysis system and method based on artificial intelligence, which can realize more accurate, efficient and reasonable resource utilization of user intention analysis, and significantly improve the performance and user experience of intelligent interaction systems. SUMMARY

[0005] (1) Technical problem to be solved

[0006] The purpose of the present application is to provide a user intention analysis system and method based on artificial intelligence to solve the problem of not being able to effectively detect and solve intention conflicts and dynamically optimize resource allocation after the conflicts, and to realize accurate understanding and efficient execution of user intentions in complex interaction scenarios.

[0007] (2) Technical solution

[0008] To achieve the above purpose, on the one hand, the present application provides a user intention analysis system based on artificial intelligence, which comprises: a multi-modal data acquisition and intention generation module, an intention conflict detection and optimal transfer path calculation module, a resource allocation and intention analysis execution module, and the modules are sequentially connected in communication:

[0009] The multimodal data acquisition and intent generation module is used to acquire multimodal behavior data generated by the user at the current moment and in historical moments, the multimodal behavior data including voice data, image data and text data; generate a single-point intent recognition result representing the user's current instantaneous operation based on the multimodal behavior data at the current moment; and generate a temporal intent chain representing the user's continuous operation based on the multimodal behavior data at historical moments.

[0010] The intent conflict detection and optimal transfer path calculation module is used to detect whether there is an intent conflict, which is the conflict between the single-point intent recognition result and the currently active intent node in the temporal intent chain; when an intent conflict is detected, the optimal transfer path from the currently conflicting intent node in the temporal intent chain to the candidate target intent node is calculated according to the transfer probability weight in the pre-constructed intent transfer probability graph.

[0011] The resource allocation and intent analysis execution module is used to determine the resource requirements corresponding to the candidate target intent nodes of the optimal transfer path, dynamically allocate resources to the candidate target intent nodes according to the current available resource status and the resource requirements, and perform intent analysis.

[0012] Furthermore, the intent conflict detection and optimal transfer path calculation module includes the following steps:

[0013] Extract the semantic feature vector of the single-point intent recognition result and the semantic feature vector of the currently active intent node, and calculate the semantic similarity value between the semantic feature vectors.

[0014] Obtain the duration weight coefficient of the currently active intent node in the temporal intent chain; when the semantic similarity value is less than a preset conflict determination threshold, an intent conflict is determined to exist; when an intent conflict is determined to exist, a conflict confidence value is calculated based on the semantic similarity value and the duration weight coefficient; if the conflict confidence value exceeds the preset conflict confidence threshold, the intent conflict is confirmed.

[0015] Furthermore, the intent conflict detection and optimal transfer path calculation module includes:

[0016] The candidate target intent node set determination module is used to obtain the node set, the directed edges connecting the nodes, and the transition probability weights corresponding to each directed edge in the pre-constructed intent transition probability graph; and to determine the candidate target intent node set according to the predefined filtering rules, taking the current conflicting intent node in the temporal intent chain as the path starting point.

[0017] The maximum cumulative probability and predecessor node analysis module is used to calculate the maximum cumulative probability value and the corresponding predecessor node of each node from the starting point of the path to the candidate target intent node set based on the transition probability weight value using a dynamic programming algorithm.

[0018] The optimal transfer path calculation module is used to traverse the set of candidate target intent nodes, select the node with the largest cumulative probability value as the optimal target intent node, and perform reverse backtracking based on the predecessor node corresponding to the optimal target intent node to obtain the complete transfer path from the starting point of the path to the optimal target intent node as the optimal transfer path.

[0019] Furthermore, the maximum cumulative probability and predecessor node analysis module includes the following steps:

[0020] The maximum cumulative probability value corresponding to the starting point of the path is initialized to a baseline value, and it is marked as having no predecessor node; the maximum cumulative probability values ​​of other nodes are initialized to invalid values; all nodes are traversed sequentially according to the topological sorting order of the pre-constructed intention transition probability graph.

[0021] For the current traversed node, traverse all the predecessor nodes corresponding to the incoming edges of the current traversed node, and calculate the product of the maximum cumulative probability value of the predecessor node and the corresponding incoming edge transition probability weight value as the candidate cumulative probability value; the incoming edge is the directed edge pointing to the current traversed node in the intention transition probability graph.

[0022] Extract the maximum value of the candidate cumulative probability values ​​as the maximum cumulative probability value of the current traversed node, and record the predecessor node corresponding to the incoming edge that generates the maximum cumulative probability value as the predecessor node of the current traversed node.

[0023] Repeat the above traversal process to obtain the maximum cumulative probability value of each node in the candidate target intent node set and its corresponding predecessor node.

[0024] Furthermore, the resource allocation and intent analysis execution module includes:

[0025] The basic resource demand vector calculation module is used to retrieve the basic resource demand vector from the intent resource mapping table based on the type identifier of the candidate target intent node.

[0026] The resource allocation module is used to scan the real-time load status of each computing unit in the resource pool to obtain a multi-dimensional available resource vector; to obtain a resource allocation scheme by combining the basic resource demand vector and the multi-dimensional available resource vector through a resource scheduling optimization method; to update the resource pool after preemptively reclaiming the resources occupied by inactive intent nodes according to the resource allocation scheme; to allocate resources to the candidate target intent nodes according to the updated resource pool; and to perform intent analysis on the analysis task after the resource allocation is completed.

[0027] Furthermore, the basic resource demand vector calculation module includes the following steps:

[0028] The type identifier of the candidate target intent node is used as the query key to access the intent resource mapping table, which is stored in key-value pair format with the type identifier of the intent node as the key and the corresponding basic resource requirement description as the value.

[0029] The intent resource mapping table is searched for an entry that exactly matches the query key; the basic resource requirement description is extracted from the matched entry; and the extracted basic resource requirement description is parsed to obtain the basic resource requirement vector.

[0030] Furthermore, the resource allocation module includes:

[0031] The resource allocation optimization model construction module is used to establish a resource allocation optimization model with the goal of maximizing the resource demand of candidate target intent nodes and maximizing resource utilization. The basic resource demand vector is used as the demand input constraint of the resource allocation optimization model, and the multidimensional available resource vector is used as the resource supply constraint of the resource allocation optimization model.

[0032] The resource allocation scheme generation module is used to solve the resource allocation optimization model to obtain a set of candidate allocation schemes; calculate the resource utilization efficiency evaluation value corresponding to each candidate allocation scheme in the set of candidate allocation schemes; and select the candidate allocation scheme with the highest resource utilization efficiency evaluation value as the final resource allocation scheme.

[0033] Furthermore, the resource allocation scheme generation module includes the following steps:

[0034] Obtain the resource type and quantity allocated to the candidate target intent node in the candidate allocation scheme to obtain the actual allocated resource vector; obtain the basic resource demand vector corresponding to the candidate target intent node and the current multidimensional available resource vector of the resource pool.

[0035] The resource demand matching degree is calculated based on the actual allocated resource vector and the basic resource demand vector; the resource utilization rate is calculated based on the actual allocated resource vector and the multidimensional available resource vector.

[0036] The resource utilization efficiency evaluation value is obtained by weighting the resource demand matching degree and the resource utilization rate.

[0037] Furthermore, the resource allocation module also includes the following steps:

[0038] When the resource allocation scheme determines that the current multidimensional available resource vector cannot meet the basic resource requirement vector of the candidate target intent node, it scans all intent nodes occupying resources in the resource pool, extracts the operation timestamp sequence of the intent node within a predefined time window, calculates the variance of adjacent operation time intervals as behavior dispersion, and obtains the difference between the current time and the last valid operation timestamp of the intent node as the silence duration.

[0039] The behavior dispersion and silence duration are used to obtain an activity score value through a preset inactivity determination rule. When the activity score value is lower than the preset activity score threshold, the intent node is marked as an inactive intent node to be reclaimed. The inactive intent nodes are sorted in descending order according to their total resource usage, and their resources are reclaimed to the resource pool in turn.

[0040] On the other hand, based on the same inventive concept, the present invention also provides a user intent analysis method based on artificial intelligence, the method comprising:

[0041] The system acquires multimodal behavior data generated by the user at the current moment and in historical moments, including voice data, image data, and text data; generates a single-point intent recognition result representing the user's current instantaneous operation based on the multimodal behavior data at the current moment; and generates a temporal intent chain representing the user's continuous operations based on the multimodal behavior data at historical moments.

[0042] The system detects whether an intent conflict exists, where the intent conflict is a conflict between the single-point intent recognition result and the currently active intent node in the temporal intent chain. When an intent conflict is detected, the system calculates the optimal transfer path from the currently conflicting intent node in the temporal intent chain to the candidate target intent node based on the transfer probability weights in the pre-constructed intent transfer probability graph.

[0043] Determine the resource requirements corresponding to the candidate target intent nodes of the optimal transfer path, dynamically allocate resources to the candidate target intent nodes based on the current available resource status and the resource requirements, and perform intent analysis.

[0044] (3) Beneficial effects

[0045] Compared with the prior art, the beneficial effects of the present invention are:

[0046] 1. By acquiring multimodal behavioral data of users at the current and historical moments, single-point intent recognition results and temporal intent chains are generated respectively. Using methods such as extracting semantic feature vectors, calculating semantic similarity values, and combining duration weighting coefficients to calculate conflict confidence values, the system accurately detects conflicts between single-point intent recognition results and currently active intent nodes in the temporal intent chain. This approach can keenly capture dynamic changes in user intent, avoid analytical biases caused by ignoring intent conflicts, and significantly improve intent recognition accuracy in complex scenarios.

[0047] 2. Upon detecting an intent conflict, the system uses a dynamic programming algorithm to calculate the optimal transition path from the current conflicting intent node to the candidate target intent node, based on a pre-constructed intent transition probability map. This not only ensures a smooth and coherent transition of user intents when conflicts occur, avoiding intent interruption or confusion, but also improves the system's efficiency in handling complex intent changes, enabling the system to quickly respond and adjust to the appropriate intent state, providing users with a smoother interactive experience.

[0048] 3. After determining the optimal transfer path, the basic resource requirement vector is obtained by retrieving the intent resource mapping table based on the type identifier of the candidate target intent node. Simultaneously, a multi-dimensional available resource vector is obtained by scanning the resource pool, and a resource allocation scheme is generated through resource scheduling optimization methods. Furthermore, by analyzing the behavioral dispersion and inactivity duration of intent nodes, resources occupied by inactive intent nodes are identified and reclaimed. Resources are allocated rationally according to actual needs to avoid resource waste and idleness, improve resource utilization, ensure stable system operation under different intent states, and enhance overall system performance. Attached Figure Description

[0049] Figure 1 This is a schematic diagram of the module composition of the user intent analysis system based on artificial intelligence of the present invention.

[0050] Figure 2 This is a flowchart of the user intent analysis method based on artificial intelligence according to the present invention. Detailed Implementation

[0051] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0052] Before providing examples, it is necessary to describe the application scenarios of the present invention. The present invention is a user intent analysis system and method based on artificial intelligence, which is applied in complex interaction scenarios to dynamically optimize resource allocation after analyzing user intent conflicts, so as to achieve accurate understanding and efficient execution of user intent.

[0053] Example 1: As Figure 1 As shown, this embodiment provides an artificial intelligence-based user intent analysis system. The system includes: a multimodal data acquisition and intent generation module, an intent conflict detection and optimal transfer path calculation module, and a resource allocation and intent analysis execution module. These modules are sequentially connected in communication.

[0054] The multimodal data acquisition and intent generation module is used to acquire multimodal behavioral data generated by the user at the current moment and in historical moments. This multimodal behavioral data includes voice data, image data, and text data. Based on the multimodal behavioral data at the current moment, it generates a single-point intent recognition result representing the user's current instantaneous operation. Based on the multimodal behavioral data at historical moments, it generates a temporal intent chain representing the user's continuous operations. For example, the user intent analysis in this embodiment is a smart home control system. When the user interacts with the system through various means such as voice assistants, mobile applications, and touch panels, the system needs to accurately understand the user's true intent and provide corresponding service responses. The system first establishes a comprehensive multimodal data acquisition mechanism. Voice data acquisition is achieved through microphone arrays distributed throughout the room, which can capture the user's voice commands in real time, such as "turn up the living room lights" or "play some light music." Image data acquisition is completed through camera devices installed in key locations. These cameras can not only recognize the user's facial expressions and body movements but also analyze the user's location movement trajectory and dwell time. Text data is primarily acquired through user input on mobile applications or wall-mounted touch panels, including text input, button clicks, and swipe selections. The system labels each piece of collected data with a precise timestamp, data type, acquisition device identifier, and preliminary confidence score to ensure accurate data traceability during subsequent processing. For multimodal behavioral data collected at the current moment, the system runs a single-point intent recognition algorithm to generate instantaneous intent recognition results, i.e., single-point intent recognition results. For voice data, the system first converts audio signals into text content using automatic speech recognition technology, then uses natural language processing algorithms to extract semantic features, identifying keywords, grammatical structures, and emotional tone. For image data, the system uses computer vision algorithms to analyze user movement patterns; for example, a user walking towards a room may indicate an intention to use equipment in that room, and gestures may indicate a need to adjust environmental parameters. For text data, the system directly performs semantic analysis and intent classification. These different modal feature information are comprehensively processed through a multimodal fusion algorithm. The multimodal fusion algorithm assigns different weights based on the reliability and relevance of each modality's data, ultimately generating a unified single-point intent recognition result. For example, it identifies the specific intent category of "ambient light adjustment" and includes corresponding parameter information such as the target brightness value. Parallel to single-point intent recognition is the construction and maintenance of a temporal intent chain. The system continuously tracks user behavior data at historical moments and constructs an intent sequence reflecting the user's continuous operation patterns in chronological order. Each node in the temporal intent chain represents a user's intent operation within a specific time period, and the node contains detailed information such as intent type, start time, end time, related parameters, and execution status.For example, a user's behavioral sequence after returning home at night might include consecutive intent operations such as "turning on the entryway light," "adjusting the living room temperature," "starting the sound system," and "checking security monitoring." The system determines whether these intents constitute a coherent behavioral chain based on the time intervals and semantic relationships between the operations, and identifies which intents are currently still actively executed. The construction of the temporal intent chain uses a sliding time window mechanism, typically retaining user behavior data from the past two hours. Historical data exceeding the time window is archived and stored, and does not participate in the current intent analysis process.

[0055] The intent conflict detection and optimal transfer path calculation module is used to detect whether there is an intent conflict, which is the conflict between the single-point intent recognition result and the currently active intent node in the temporal intent chain; when an intent conflict is detected, the optimal transfer path from the currently conflicting intent node in the temporal intent chain to the candidate target intent node is calculated according to the transfer probability weight in the pre-constructed intent transfer probability graph.

[0056] The resource allocation and intent analysis execution module is used to determine the resource requirements corresponding to the candidate target intent nodes of the optimal transfer path, dynamically allocate resources to the candidate target intent nodes according to the current available resource status and the resource requirements, and perform intent analysis.

[0057] The intent conflict detection and optimal transfer path calculation module includes the following steps:

[0058] The semantic feature vectors of the single-point intent recognition result and the current active intent node are extracted, and the semantic similarity value between the semantic feature vectors is calculated. In the intent conflict detection process of the smart home control system, the system needs to determine whether a genuine intent conflict exists through precise numerical calculations and logical judgments. When the system obtains the single-point intent recognition result "adjust living room light brightness" and the current active intent node "watch movie mode," it first needs to perform deep semantic feature extraction and similarity analysis on these two intents. The semantic feature vector extraction process uses a pre-trained deep learning semantic encoding model, which can convert the intent described in natural language into a 512-dimensional dense vector representation. For the single-point intent "adjust living room light brightness," the system first performs text preprocessing, including word segmentation, part-of-speech tagging, and named entity recognition, and then inputs the processed text into the semantic encoder to obtain the corresponding semantic feature vector. Similarly, the system performs the same feature extraction process for the current active intent node "watch movie mode" to generate the corresponding semantic feature vector. These two semantic feature vectors can accurately reflect the semantic content and contextual information of their respective intents in a high-dimensional semantic space. The system then calculates the cosine similarity between the two semantic feature vectors, achieved by dividing the vector dot product by the product of their magnitudes. Assuming the calculated semantic similarity is 0.72, the system compares this value to a preset conflict determination threshold of 0.75. This threshold is set based on statistical analysis of extensive historical data and system performance optimization, aiming to effectively identify genuine intent conflicts while avoiding oversensitivity that could lead to misjudgments.

[0059] The system obtains the duration weight coefficient of the currently active intent node in the temporal intent chain. If the semantic similarity value is less than a preset conflict determination threshold, an intent conflict is determined to exist. If an intent conflict is determined to exist, a conflict confidence value is calculated based on the semantic similarity value and the duration weight coefficient. If the conflict confidence value exceeds the preset conflict confidence threshold, the intent conflict is confirmed. When the similarity value (0.72) is less than the preset conflict determination threshold (0.75), the system initially determines that a potential intent conflict exists. To further confirm the reliability of the conflict, the system needs to obtain the duration weight coefficient of the currently active intent node in the temporal intent chain. The duration weight coefficient is calculated using an exponential decay function, considering the temporal stability characteristics of the intent; that is, the longer an intent lasts, the higher its importance in the user behavior pattern, and the larger the corresponding weight coefficient. For example, the intent "watch movie mode" has lasted for 25 minutes. According to the system's preset weight calculation formula, its corresponding duration weight coefficient is 0.85, reflecting the stability and importance of this intent in the current user behavior sequence. The system then calculates the conflict confidence score based on the semantic similarity value and the duration weighting coefficient. The conflict confidence score is calculated using a weighted product method, multiplying the semantic similarity value (0.72) by the duration weighting coefficient (0.85) to obtain a conflict confidence score of 0.61. This conflict confidence score comprehensively reflects the severity and credibility of the intent conflict, considering both the semantic level of conflict and the stability factor at the temporal level. Since the conflict confidence score of 0.61 is greater than the preset conflict confidence threshold of 0.5, the system ultimately confirms that an intent conflict does indeed exist and requires processing.

[0060] The intent conflict detection and optimal transfer path calculation module includes:

[0061] The candidate target intent node set determination module is used to obtain the node set, directed edges connecting the nodes, and the corresponding transition probability weights of each directed edge from a pre-constructed intent transition probability graph. Taking the currently conflicting intent node in the temporal intent chain as the path starting point, it determines the candidate target intent node set according to predefined filtering rules. The pre-constructed intent transition probability graph is the core knowledge base of the system, built by analyzing the historical behavior data of tens of thousands of users, and includes all common intent types in the smart home environment and their transition relationships. The node set in the graph covers hundreds of specific intent types such as "lighting control," "temperature adjustment," "entertainment playback," "security monitoring," and "environmental perception," with each node having a unique identifier and detailed attribute description. The directed edges connecting the nodes represent possible transition relationships between intents, the direction of the edge indicates the directionality of the transition, and the weight value marked on the edge indicates the probability of the transition occurring. The weight value ranges from 0 to 1, with a larger value indicating a higher probability of transition. The system first uses the currently conflicting intent node "watching movies mode" in the temporal intent chain as the starting point for path search, and then determines the candidate target intent node set according to predefined filtering rules. The filtering rules include multiple dimensions of judgment criteria. First, semantic relevance filtering: the candidate target intent must be semantically related to the currently detected single-point intent "adjust living room light brightness". Second, user preference filtering: the system will exclude intent options that do not conform to user habits based on the user's historical behavior patterns and personal preference settings. Finally, contextual rationality filtering: the candidate intent must be consistent with the current environmental state and device availability. After filtering, the system identified three candidate target intent nodes: "movie viewing light mode", "pause movie and adjust light", and "end movie and restore normal lighting".

[0062] The maximum cumulative probability and predecessor node analysis module is used to calculate the maximum cumulative probability value and the corresponding predecessor node of each node from the starting point of the path to the candidate target intent node set based on the transition probability weight value using a dynamic programming algorithm.

[0063] The optimal transition path calculation module iterates through the set of candidate target intent nodes, selecting the node with the highest cumulative probability value as the optimal target intent node. It then performs a reverse backtracking operation based on the predecessor node corresponding to the optimal target intent node to obtain the complete transition path from the path start point to the optimal target intent node, which is then considered the optimal transition path. During the calculation, the system finds that the maximum cumulative probability value for reaching the "Movie Viewing Lighting Mode" node is 0.72, and the optimal path is "Movie Viewing Mode → Adjust Environmental Parameters → Movie Viewing Lighting Mode". The maximum cumulative probability value for reaching the "Pause Movie, Adjust Lighting" node is 0.65, and the optimal path is "Movie Viewing Mode → Pause Movie, Adjust Lighting". The maximum cumulative probability value for reaching the "End Movie, Restore Normal Lighting" node is 0.48, and the optimal path is "Movie Viewing Mode → User Activity Switch → End Movie, Restore Normal Lighting". The system iterates through all candidate target intent nodes, compares their maximum cumulative probability values, and selects the "Movie Viewing Lighting Mode" node with the highest probability as the optimal target intent node. After determining the optimal target intent node, the system performs a reverse backtracking operation based on the predecessor node information recorded during the dynamic programming process. Starting with the "Movie Viewing Lighting Mode" node, the system searches for its predecessor node "Adjust Environmental Parameters," and then continues searching for the predecessor node of "Adjust Environmental Parameters" "Movie Viewing Mode," until it backtracks to the starting node. Through this reverse backtracking process, the system reconstructs the complete optimal transition path "Movie Viewing Mode → Adjust Environmental Parameters → Movie Viewing Lighting Mode," representing the most likely direction of the user's intent evolution.

[0064] The maximum cumulative probability and predecessor node analysis module includes the following steps:

[0065] The system initializes the maximum cumulative probability value corresponding to the path's starting point as a baseline value and marks it as having no predecessor node. The maximum cumulative probability values ​​of other nodes are initialized to invalid values. All nodes are then traversed sequentially according to the topological sorting order of the pre-constructed intent transition probability graph. The system first initializes all nodes in the intent transition probability graph. For the path's starting point node, "Watch Movie Mode," the system sets its maximum cumulative probability value to the baseline value of 1.0, indicating that the probability of reaching itself from this node is definite. Simultaneously, it marks this node as having no predecessor node, indicating it as null in the predecessor node record. For all other nodes in the intent transition probability graph, the system initializes their maximum cumulative probability values ​​to negative infinity, indicating that no valid path has yet been found, and the predecessor node information is also initialized to an undefined state. The system then traverses all nodes sequentially according to the topological sorting order of the pre-constructed intent transition probability graph. Topological sorting is a linear sorting method for directed acyclic graphs, ensuring that for any directed edge in the intent transition probability graph, the starting node appears before the ending node in the sorting. This sorting method ensures that when processing a node in the dynamic programming process, all predecessor nodes that might affect the optimal solution of that node have been correctly processed, thus guaranteeing the correctness and efficiency of the algorithm. The traversal process processes each node one by one according to the order of the topological sort.

[0066] For the currently traversed node, the system iterates through all the predecessor nodes corresponding to the incoming edges of the current traversed node, and calculates the product of the maximum cumulative probability value of the predecessor node and the corresponding transition probability weight value of the incoming edge as a candidate cumulative probability value. The incoming edge is a directed edge in the intent transition probability graph pointing to the currently traversed node. When the system traverses to a specific node, such as the "Adjust Environmental Parameters" node, it first needs to identify all the incoming edges and corresponding predecessor nodes of that node. An incoming edge is a directed edge in the intent transition probability graph pointing to the currently traversed node; each incoming edge connects a predecessor node and the current node and is labeled with a corresponding transition probability weight value. For the "Adjust Environmental Parameters" node, the system finds two incoming edges, one from the "Watch Movie Mode" node (transition probability weight 0.8) and the other from the "Music Playback Mode" node (transition probability weight 0.6). The system then calculates the candidate cumulative probability value corresponding to each incoming edge. The candidate cumulative probability value is calculated by multiplying the current maximum cumulative probability value of the predecessor node by the transition probability weight value of the corresponding incoming edge. For incoming edges from "Watch Movie Mode", the candidate cumulative probability value is 1.0 multiplied by 0.8, which equals 0.8. For incoming edges from "Music Playback Mode", since the maximum cumulative probability value of the "Music Playback Mode" node is still the initial negative infinity, indicating that no valid path has been found to reach the node, the corresponding candidate cumulative probability value is also negative infinity and is excluded in the actual calculation.

[0067] The system extracts the maximum value of the candidate cumulative probability values ​​as the maximum cumulative probability value of the currently traversed node, and records the predecessor node corresponding to the incoming edge that generates the maximum cumulative probability value as the predecessor node of the currently traversed node. The system extracts the maximum value from all valid candidate cumulative probability values ​​as the maximum cumulative probability value of the currently traversed node. In this example, the "Adjust Environment Parameters" node has only one valid candidate value, 0.8, so this value is set as the maximum cumulative probability value for this node. Simultaneously, the system records the predecessor node "Watch Movie Mode" corresponding to the incoming edge that generates this maximum cumulative probability value as the predecessor node of the "Adjust Environment Parameters" node. This predecessor node information will play an important role in subsequent path backtracking.

[0068] Repeat the above traversal process to obtain the maximum cumulative probability value and corresponding predecessor node for each node in the candidate target intent node set. The traversal and calculation process continues until all nodes in the topological sort have been processed. Through this systematic processing, each node can obtain the maximum cumulative probability value corresponding to the optimal path from the starting node to that node, as well as the information of its direct predecessor node along the optimal path. Finally, the system obtains complete optimal path information for each node in the candidate target intent node set.

[0069] The resource allocation and intent analysis execution module includes:

[0070] The basic resource demand vector calculation module is used to retrieve the basic resource demand vector from the intent resource mapping table based on the type identifier of the candidate target intent node.

[0071] The resource allocation module scans the real-time load status of each computing unit in the resource pool to obtain a multi-dimensional available resource vector. It then uses a resource scheduling optimization method to obtain a resource allocation scheme by combining the basic resource demand vector and the multi-dimensional available resource vector. Based on the resource allocation scheme, it preemptively reclaims resources occupied by inactive intent nodes and updates the resource pool. Resources are then allocated to the candidate target intent nodes according to the updated resource pool. After resource allocation, intent analysis is performed on the analysis task. After resource allocation is complete, the system initiates the intent analysis processing flow according to the determined resource allocation scheme. The intent analysis processing flow initializes the computing environment allocated to the task, including creating an independent process space, allocating a specified memory region, establishing a network communication channel, and preparing necessary data storage space. Then, the system loads the corresponding intent analysis algorithm module, calls the corresponding processing logic based on the specific intent type "movie viewing lighting mode," and performs specific operations such as environmental perception, device control, and parameter adjustment to ultimately achieve accurate identification and intelligent response of the user's intent.

[0072] The basic resource demand vector calculation module includes the following steps:

[0073] The system uses the type identifier of the candidate target intent node as the query key to access the intent resource mapping table. This table stores key-value pairs, with the type identifier of the intent node as the key and the corresponding basic resource requirement description as the value. During resource requirement determination, the system needs to obtain the basic resource configuration information required by the candidate target intent node through precise data query and parsing operations. When the system determines that the type identifier of the candidate target intent node is "MOVIE_LIGHTING_MODE", it immediately uses this identifier as the query key to initiate the access process of the intent resource mapping table. The intent resource mapping table is the core data structure of the system resource management subsystem. It organizes data using a key-value pair storage method, featuring fast querying and efficient access. The keys in the intent resource mapping table are the type identifiers of various intent nodes. These identifiers use a unified naming convention to uniquely identify each intent type. The values ​​in the intent resource mapping table are detailed resource requirement descriptions for the corresponding intent type, containing all the computational resource information required to perform intent analysis for that type. The data in the intent resource mapping table comes from performance testing during the system design phase and statistical analysis during operation, ensuring the accuracy and practicality of the resource requirement descriptions.

[0074] The system searches for an entry in the intent resource map table that exactly matches the query key; extracts the stored basic resource requirement description from the matched entry; and parses the extracted basic resource requirement description to obtain a basic resource requirement vector. The system uses the query key "MOVIE_LIGHTING_MODE" to perform an exact match search in the intent resource map table. An exact match requires the query key to be completely identical to the key value stored in the table, including all details such as character case, special characters, and character encoding. This strict matching method avoids query failures or incorrect results due to key-value mismatches. The system's query algorithm first calculates the hash value of the query key, then searches for a matching item at the corresponding position in the hash table. If a hash collision is found, a collision resolution strategy is used to continue the search until an exact match is found or it is confirmed that no match exists. Once the system successfully locates an entry in the map table that exactly matches the query key, it immediately extracts the stored basic resource requirement description information from that entry. The basic resource requirement description information is stored in a structured data format, typically using standard formats such as JSON or XML, and contains specific requirement values ​​for multiple resource types. For example, the extracted resource requirement description might include detailed information such as 2 CPU cores, 4096MB of memory, 0.5 GPU computing units, 10Mbps network bandwidth, 2048MB of temporary storage, and 512MB of persistent storage. Each resource type has specific quantity requirements and performance specifications. The system then calls a dedicated data parsing module to process the extracted basic resource requirement description. The parsing module first verifies the completeness and correctness of the data format, checking whether it contains all the necessary resource type information and whether the numerical format meets the system requirements. Then, the parsing module converts the structured description information into a standard numerical vector format according to the predefined resource type order. The conversion process includes unit standardization, data type conversion, and numerical precision adjustment to ensure that the generated vector is compatible with the data formats of other modules in the system. After parsing, the system obtains a standard format basic resource requirement vector. Each dimension of the basic resource requirement vector corresponds to a specific resource type, and the numerical value represents the required quantity of that resource. For example, the final generated basic resource demand vector is [2,4096,0.5,10,2048,512], which corresponds to the demand for six resource types: CPU cores, memory capacity (MB), GPU units, network bandwidth (Mbps), temporary storage (MB), and persistent storage (MB).

[0075] The resource allocation module includes:

[0076] The resource allocation optimization model construction module is used to establish a resource allocation optimization model with the objectives of maximizing the satisfaction of resource requirements of candidate target intention nodes and maximizing resource utilization. The basic resource requirement vector is used as the requirement input constraint of the resource allocation optimization model, and the multi-dimensional available resource vector is used as the resource supply constraint. After obtaining the basic resource requirement vector and the multi-dimensional available resource vector, the system needs to generate the optimal resource allocation scheme through a complex resource scheduling optimization method. The core of the resource scheduling optimization method is to establish a mathematical model that comprehensively considers multiple optimization objectives, ensuring that the resource requirements of candidate target intention nodes are fully satisfied while ensuring that the resource utilization of the entire system reaches an optimal state. The process of establishing the resource allocation optimization model first requires clarifying the specific form of the optimization objective function. The system designs a bi-objective optimization function: the first objective is to maximize the satisfaction of resource requirements of candidate target intention nodes, and the second objective is to maximize the resource utilization of the entire system. The resource requirement satisfaction is calculated by comparing the actual allocated resource quantity with the basic requirement quantity; the satisfaction reaches its highest value when the actual allocated quantity equals or exceeds the basic requirement quantity. System resource utilization is measured by calculating the proportion of all allocated resources to the total available resources, reflecting the effective use of system resources. Since these two objectives may conflict in some cases—maximizing demand satisfaction requires prioritizing the fulfillment of nodes' basic resource demand vectors, potentially leading to fragmented resource allocation; and maximizing resource utilization requires the most compact resource allocation (e.g., reducing idle resources), which may sacrifice the demand satisfaction of some nodes—the system uses a weighted summation approach to combine them into a single comprehensive objective function, balancing the importance of different objectives by adjusting the weighting coefficients. The basic resource demand vector serves as the model's demand input constraint, stipulating that the quantity of each type of resource allocated to candidate target intent nodes cannot be less than the corresponding basic demand value. This is the minimum requirement to ensure the normal execution of the intent analysis task. The multidimensional available resource vector serves as the model's resource supply constraint, limiting the total allocation of each type of resource to a quantity that cannot exceed the current available resource quantity, reflecting the system's physical resource limitations.

[0077] The resource allocation scheme generation module is used to solve the resource allocation optimization model to obtain a set of candidate allocation schemes; calculate the resource utilization efficiency evaluation value corresponding to each candidate allocation scheme in the set; and select the candidate allocation scheme with the highest resource utilization efficiency evaluation value as the final resource allocation scheme. The system uses a mixed-integer linear programming algorithm to solve the established resource allocation optimization model. The mixed-integer linear programming algorithm can handle complex optimization problems containing both continuous and discrete variables, and is suitable for handling resource allocation situations that involve both continuous resources (such as memory capacity and network bandwidth) and discrete resources (such as the number of processor cores and device units). The mixed-integer linear programming algorithm uses a branch-and-bound method to solve the problem, approximating the optimal solution by continuously subdividing the problem space and calculating upper and lower bounds. Simultaneously, it uses pruning strategies to eliminate branches that cannot produce the optimal solution, improving the solution efficiency. The solution process generates multiple candidate allocation schemes that satisfy the constraints, and each candidate allocation scheme specifies in detail the quantity and source of various resources allocated to the candidate target intention node. For example, the first candidate solution provides a minimum configuration to meet basic requirements, allocating 2 CPU cores, 4096MB of memory, and 0.5 GPU units. This solution's resource allocation is conservative but ensures the basic execution needs of the task. The second candidate solution, while meeting the basic requirements, appropriately increases the resource allocation, allocating 3 CPU cores, 6144MB of memory, and 1 GPU unit, providing better execution performance. The third candidate solution adopts a relatively aggressive resource allocation strategy, allocating 4 CPU cores, 8192MB of memory, and 1.5 GPU units, pursuing optimal execution performance but consuming more system resources.

[0078] The resource allocation scheme generation module includes the following steps:

[0079] The system obtains the actual allocated resource vector by acquiring the resource types and quantities allocated to candidate target intent nodes in the candidate allocation schemes; it also acquires the basic resource demand vector corresponding to the candidate target intent nodes and the current multi-dimensional available resource vector of the resource pool. During the calculation of the resource utilization efficiency evaluation value, the system needs to conduct a comprehensive performance analysis and quantitative evaluation of each candidate allocation scheme to ensure that the selected final scheme achieves the best balance between meeting task requirements and optimizing resource utilization. The first step in the calculation process is to acquire the detailed resource information actually allocated to candidate target intent nodes in the candidate allocation schemes. For the selected candidate allocation scheme, the system extracts the recorded allocation quantities of various resources to form the actual allocated resource vector. For example, this scheme allocates 3 CPU cores, 6144MB memory, 1 GPU computing unit, 15Mbps network bandwidth, 3072MB temporary storage, and 768MB persistent storage to the "movie viewing lighting mode" intent node, corresponding to the actual allocated resource vector [3,6144,1,15,3072,768]. The actual allocated resource vector accurately reflects the specific resource configuration of the allocation scheme. The system simultaneously re-acquires the basic resource requirement vector corresponding to the candidate target intent node and the current multi-dimensional available resource vector of the resource pool, ensuring that the calculation process uses the latest and most accurate data. The basic resource requirement vector is [2, 4096, 0.5, 10, 2048, 512], reflecting the minimum resource requirements for executing this intent type. The multi-dimensional available resource vector is [8, 16384, 2, 50, 10240, 5120], representing the available quantity of each type of resource in the current resource pool.

[0080] The resource demand matching degree is calculated based on the actual allocated resource vector and the basic resource demand vector; the resource utilization rate is calculated based on the actual allocated resource vector and the multi-dimensional available resource vector; the system then calculates the resource demand matching degree index, which reflects the degree to which the candidate allocation scheme meets the basic resource demand. The calculation method is to compare the actual allocated resource vector and the basic resource demand vector dimension by dimension. For each resource dimension, the ratio of the actual allocated amount to the basic demand amount is calculated, and then the minimum value of this ratio and 1 is taken as the matching degree of that dimension. The CPU core matching degree is min(3 / 2,1)=1, indicating that the demand is fully met. The memory matching degree is min(6144 / 4096,1)=1, indicating that the demand is fully met. The GPU matching degree is min(1 / 0.5,1)=1, indicating that the demand is over-met. The network bandwidth matching degree is min(15 / 10,1)=1, indicating that the demand is over-met. The temporary storage matching degree is min(3072 / 2048,1)=1, indicating that the demand is over-met. The persistent storage matching degree is min(768 / 512,1)=1, indicating that the demand is over-met. The system calculates a weighted average of the matching degrees of each dimension, resulting in an overall resource demand matching degree of 1.0, indicating that the allocation scheme fully meets the basic resource requirements. The system then calculates the resource utilization rate, which reflects the efficiency of the allocation scheme in using the overall system resources. The calculation method compares the actual allocated resource vector with the multi-dimensional available resource vector. For each resource dimension, the proportion of the actual allocated amount to the total available resources is calculated. The CPU utilization rate is 3 / 8=0.375, indicating that 37.5% of the available CPU resources are used. The memory utilization rate is 6144 / 16384=0.375, indicating that 37.5% of the available memory resources are used. The GPU utilization rate is 1 / 2=0.5, indicating that 50% of the available GPU resources are used. The network bandwidth utilization rate is 15 / 50=0.3, indicating that 30% of the available network bandwidth is used. The temporary storage utilization rate is 3072 / 10240 = 0.3, indicating that 30% of the available temporary storage is being used. The persistent storage utilization rate is 768 / 5120 = 0.15, indicating that 15% of the available persistent storage is being used. The system calculates resource utilization rates based on the importance weights of various resources.

[0081] The resource utilization efficiency evaluation value is obtained by weighting the resource demand matching degree and the resource utilization rate. Finally, the system weights the resource demand matching degree and resource utilization rate according to a preset weight ratio to obtain the comprehensive resource utilization efficiency evaluation value of the candidate allocation scheme.

[0082] The resource allocation module further includes the following steps:

[0083] When the resource allocation scheme determines that the current multidimensional available resource vector cannot meet the basic resource requirements of the candidate target intent node, it scans all intent nodes occupying resources in the resource pool, extracts the operation timestamp sequence of the intent node within a predefined time window, calculates the variance of adjacent operation time intervals as the behavior dispersion, and obtains the difference between the current time and the last valid operation timestamp of the intent node as the silent duration. When the system finds through the evaluation of the resource allocation scheme that the current multidimensional available resource vector cannot fully meet the basic resource requirements of the candidate target intent node, it needs to activate a preemptive resource reclamation mechanism to release the computing resources occupied by low-activity tasks. A comprehensive scan and analysis of all intent nodes currently occupying resources in the resource pool is required. The system scan process will traverse all active intent nodes registered in the resource manager, obtaining basic information of each active intent node, including node identifier, intent type, creation time, resource occupancy, etc. For each intent node currently occupying resources, the system needs to deeply analyze its recent activity patterns and behavioral characteristics to determine whether the intent node is still active or can be safely reclaimed. The core of activity analysis is to extract detailed operation records of each intent node within a predefined time window. The system's time window is typically set to the past 30 minutes, which captures short-term behavioral patterns of nodes while avoiding interference from outdated information due to excessively long windows. For each analyzed intent node, the system extracts all operation timestamps within the time window from the operation log, forming a chronologically ordered timestamp sequence. For example, the timestamp sequence for a "background music playback" intent node might be [14:15:30, 14:18:45, 14:22:10, 14:25:55], recording four operations within the past 30 minutes. The system then calculates the time interval between adjacent operation timestamps to assess the regularity and continuity of node operations. For the above timestamp sequence, the adjacent intervals are 195 seconds, 205 seconds, and 225 seconds, respectively. The system further calculates the statistical variance of these time intervals as an indicator of the behavioral dispersion of the intent node. A smaller variance value indicates more regular operation intervals and more stable node behavior; a larger variance value indicates more irregular operation patterns, potentially signifying decreased node activity. Calculated in parallel with the behavior dispersion is the silence duration metric, which reflects the length of time since the intent node's most recent operation. The system obtains the difference between the current system time and the timestamp of the node's last valid operation to calculate the silence duration. If the current time is 14:45:30 and the node's last operation time is 14:25:55, then the silence duration is 1175 seconds, approximately 19.6 minutes. A longer silence duration indicates a longer period of inactivity for the node, and a greater likelihood that it will be judged as inactive.

[0084] The system calculates an activity score by applying the behavior dispersion and silence duration to a preset inactivity determination rule. When the activity score is lower than a preset activity score threshold, the intent node is marked as an inactive intent node to be reclaimed. Inactive intent nodes are then sorted in descending order of their total resource usage, and their resources are reclaimed to the resource pool sequentially. The system uses a preset inactivity determination rule to comprehensively evaluate intent nodes. This rule is a mathematical function that takes behavior dispersion and silence duration as input parameters and outputs the activity score of the intent node. The rule is designed based on the analysis of a large amount of historical data and the training of a machine learning model, accurately reflecting the activity level of nodes corresponding to different behavior patterns. The formula for calculating the activity score considers both time decay and behavior stability factors. For example, the activity score is calculated as: Activity Score = (1 - Silence Duration / Maximum Allowable Silence Duration) × (1 - Behavior Dispersion / Maximum Dispersion Threshold) × Baseline Score. If an intent node has an activity score of 0.25 and the preset activity score threshold is 0.3, it is marked as an inactive intent node to be reclaimed. The system adds all nodes marked as inactive intent nodes to a reclamation candidate list, then sorts them in descending order based on the total resources they occupy. The calculation of the total resource amount considers the weight and value of different types of resources, prioritizing the reclamation of nodes occupying more high-value resources (such as GPU computing units and large-capacity memory). The resource reclamation process involves several coordinated steps. First, a termination signal is sent to the selected inactive intent nodes to notify the relevant processing processes to prepare to release resources. Then, the system waits for the processes to complete their current operations and save necessary state information to ensure no data loss. Next, the system forcibly terminates all associated processes, releases their occupied memory space, reclaims the processor cores allocated to the node, disconnects network connections, and cleans up temporary files. The system remarks the reclaimed resources as available, updates the resource pool's state information, and makes these resources available for new high-priority tasks. Through this preemptive resource reclamation mechanism, the system can dynamically adjust resource allocation, ensuring that important user intent analysis tasks receive sufficient computational support while preventing resources from being ineffectively occupied by long-term idle tasks, thus improving the overall system's resource utilization efficiency and response performance.

[0085] Example 2: Based on the same inventive concept, such as Figure 2 As shown, this embodiment also provides a user intent analysis method based on artificial intelligence, the method including;

[0086] The system acquires multimodal behavior data generated by the user at the current moment and in historical moments, including voice data, image data, and text data; generates a single-point intent recognition result representing the user's current instantaneous operation based on the multimodal behavior data at the current moment; and generates a temporal intent chain representing the user's continuous operations based on the multimodal behavior data at historical moments.

[0087] The system detects whether an intent conflict exists, where the intent conflict is a conflict between the single-point intent recognition result and the currently active intent node in the temporal intent chain. When an intent conflict is detected, the system calculates the optimal transfer path from the currently conflicting intent node in the temporal intent chain to the candidate target intent node based on the transfer probability weights in the pre-constructed intent transfer probability graph.

[0088] Determine the resource requirements corresponding to the candidate target intent nodes of the optimal transfer path, dynamically allocate resources to the candidate target intent nodes based on the current available resource status and the resource requirements, and perform intent analysis.

[0089] Finally, it should be noted that although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art can still modify the technical solutions described in the foregoing embodiments or make equivalent substitutions for some of the technical features. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A user intent analysis system based on artificial intelligence, characterized in that, The system includes: a multimodal data acquisition and intent generation module, an intent conflict detection and optimal transfer path calculation module, and a resource allocation and intent analysis and execution module. These modules are sequentially connected and communicate with each other. The multimodal data acquisition and intent generation module is used to acquire multimodal behavior data generated by the user at the current moment and in historical moments, the multimodal behavior data including voice data, image data and text data; generate a single-point intent recognition result representing the user's current instantaneous operation based on the multimodal behavior data at the current moment; and generate a temporal intent chain representing the user's continuous operations based on the multimodal behavior data at historical moments. The intent conflict detection and optimal transfer path calculation module is used to detect whether there is an intent conflict, which is the conflict between the single-point intent recognition result and the currently active intent node in the temporal intent chain; when an intent conflict is detected, the optimal transfer path from the currently conflicting intent node in the temporal intent chain to the candidate target intent node is calculated according to the transfer probability weight in the pre-constructed intent transfer probability graph. The resource allocation and intent analysis execution module is used to determine the resource requirements corresponding to the candidate target intent nodes of the optimal transfer path, dynamically allocate resources to the candidate target intent nodes according to the current available resource status and the resource requirements, and perform intent analysis. The resource allocation and intent analysis execution module includes: The basic resource demand vector calculation module is used to retrieve the basic resource demand vector from the intent resource mapping table based on the type identifier of the candidate target intent node. The resource allocation module is used to scan the real-time load status of each computing unit in the resource pool to obtain a multi-dimensional available resource vector; to obtain a resource allocation scheme by combining the basic resource demand vector and the multi-dimensional available resource vector through a resource scheduling optimization method; to update the resource pool after preemptively reclaiming the resources occupied by inactive intent nodes according to the resource allocation scheme; to allocate resources to the candidate target intent nodes according to the updated resource pool; and to perform intent analysis on the analysis task after the resource allocation is completed. The basic resource demand vector calculation module includes the following steps: The type identifier of the candidate target intent node is used as the query key to access the intent resource mapping table, which is stored in key-value pair format with the type identifier of the intent node as the key and the corresponding basic resource requirement description as the value. The intent resource mapping table is searched for an entry that exactly matches the query key; the basic resource requirement description is extracted from the matched entry; and the extracted basic resource requirement description is parsed to obtain the basic resource requirement vector.

2. The user intent analysis system based on artificial intelligence according to claim 1, characterized in that, The intent conflict detection and optimal transfer path calculation module includes the following steps: Extract the semantic feature vector of the single-point intent recognition result and the semantic feature vector of the currently active intent node, and calculate the semantic similarity value between the semantic feature vectors; Obtain the duration weight coefficient of the currently active intent node in the temporal intent chain; when the semantic similarity value is less than a preset conflict determination threshold, an intent conflict is determined to exist; when an intent conflict is determined to exist, a conflict confidence value is calculated based on the semantic similarity value and the duration weight coefficient; if the conflict confidence value exceeds the preset conflict confidence threshold, the intent conflict is confirmed.

3. The user intent analysis system based on artificial intelligence according to claim 1, characterized in that, The intent conflict detection and optimal transfer path calculation module includes: The candidate target intent node set determination module is used to obtain the node set, the directed edges connecting the nodes, and the transition probability weights corresponding to each directed edge in the pre-constructed intent transition probability graph; and to determine the candidate target intent node set according to predefined filtering rules, taking the current conflicting intent node in the temporal intent chain as the path starting point. The maximum cumulative probability and predecessor node analysis module is used to calculate the maximum cumulative probability value and the corresponding predecessor node of each node from the starting point of the path to the candidate target intent node set according to the transition probability weight using a dynamic programming algorithm. The optimal transfer path calculation module is used to traverse the set of candidate target intent nodes, select the node with the largest cumulative probability value as the optimal target intent node, and perform reverse backtracking based on the predecessor node corresponding to the optimal target intent node to obtain the complete transfer path from the starting point of the path to the optimal target intent node as the optimal transfer path.

4. The user intent analysis system based on artificial intelligence according to claim 3, characterized in that, The maximum cumulative probability and predecessor node analysis module includes the following steps: Initialize the maximum cumulative probability value corresponding to the starting point of the path to a baseline value and mark it as having no predecessor node; initialize the maximum cumulative probability value of other nodes to invalid values; traverse all nodes sequentially according to the topological sorting order of the pre-constructed intention transition probability graph; For the current traversed node, traverse all the predecessor nodes corresponding to the incoming edges of the current traversed node, and calculate the product of the maximum cumulative probability value of the predecessor node and the corresponding incoming edge transition probability weight value as the candidate cumulative probability value; the incoming edge is the directed edge pointing to the current traversed node in the intention transition probability graph. Extract the maximum value of the candidate cumulative probability values ​​as the maximum cumulative probability value of the current traversed node, and record the predecessor node corresponding to the incoming edge that generates the maximum cumulative probability value as the predecessor node of the current traversed node; Repeat the above traversal process to obtain the maximum cumulative probability value of each node in the candidate target intent node set and its corresponding predecessor node.

5. The user intent analysis system based on artificial intelligence according to claim 1, characterized in that, The resource allocation module includes: The resource allocation optimization model construction module is used to establish a resource allocation optimization model with the goal of maximizing the resource demand of candidate target intent nodes and maximizing resource utilization; the basic resource demand vector is used as the demand input constraint of the resource allocation optimization model, and the multidimensional available resource vector is used as the resource supply constraint of the resource allocation optimization model. The resource allocation scheme generation module is used to solve the resource allocation optimization model to obtain a set of candidate allocation schemes; calculate the resource utilization efficiency evaluation value corresponding to each candidate allocation scheme in the set of candidate allocation schemes; and select the candidate allocation scheme with the highest resource utilization efficiency evaluation value as the final resource allocation scheme.

6. The user intent analysis system based on artificial intelligence according to claim 5, characterized in that, The resource allocation scheme generation module includes the following steps: Obtain the resource type and quantity allocated to the candidate target intent node in the candidate allocation scheme to obtain the actual allocated resource vector; obtain the basic resource demand vector corresponding to the candidate target intent node and the current multidimensional available resource vector of the resource pool; The resource demand matching degree is calculated based on the actual allocated resource vector and the basic resource demand vector; the resource utilization rate is calculated based on the actual allocated resource vector and the multidimensional available resource vector. The resource utilization efficiency evaluation value is obtained by weighting the resource demand matching degree and the resource utilization rate.

7. The user intent analysis system based on artificial intelligence according to claim 1, characterized in that, The resource allocation module further includes the following steps: When the resource allocation scheme determines that the current multidimensional available resource vector cannot meet the basic resource demand vector of the candidate target intent node, it scans all intent nodes occupying resources in the resource pool, extracts the operation timestamp sequence of the intent node within a predefined time window, calculates the variance of adjacent operation time intervals as behavior dispersion, and obtains the difference between the current time and the last valid operation timestamp of the intent node as the silence duration. The behavior dispersion and silence duration are used to obtain an activity score value through a preset inactivity determination rule. When the activity score value is lower than the preset activity score threshold, the intent node is marked as an inactive intent node to be reclaimed. The inactive intent nodes are sorted in descending order according to their total resource usage, and their resources are reclaimed to the resource pool in turn.

8. A user intent analysis method based on artificial intelligence, used to execute the user intent analysis system based on artificial intelligence as described in any one of claims 1-7, characterized in that, The method includes; Acquire multimodal behavior data generated by the user at the current moment and in historical moments, the multimodal behavior data including voice data, image data and text data; generate a single-point intent recognition result representing the user's current instantaneous operation based on the multimodal behavior data at the current moment; Generate a temporal intent chain representing the user's continuous operations based on the multimodal behavioral data at the aforementioned historical moments; Detect whether there is an intent conflict, where the intent conflict is the conflict between the single-point intent recognition result and the currently active intent node in the temporal intent chain; When an intent conflict is detected, the optimal transfer path from the current conflicting intent node in the temporal intent chain to the candidate target intent node is calculated based on the transfer probability weights in the pre-constructed intent transfer probability graph. Determine the resource requirements corresponding to the candidate target intent nodes of the optimal transfer path, dynamically allocate resources to the candidate target intent nodes according to the current available resource status and the resource requirements, and perform intent analysis. The basic resource requirement vector is obtained by retrieving the intent resource mapping table based on the type identifier of the candidate target intent node; The real-time load status of each computing unit in the resource pool is scanned to obtain a multi-dimensional available resource vector; the basic resource demand vector and the multi-dimensional available resource vector are combined with a resource scheduling optimization method to obtain a resource allocation scheme; the resource pool is updated after the resources occupied by inactive intent nodes are preemptively reclaimed according to the resource allocation scheme, and resources are allocated to the candidate target intent nodes according to the updated resource pool; intent analysis is performed on the analysis task after the resource allocation is completed. The type identifier of the candidate target intent node is used as the query key to access the intent resource mapping table, which is stored in key-value pair format with the type identifier of the intent node as the key and the corresponding basic resource requirement description as the value. The intent resource mapping table is searched for an entry that exactly matches the query key; the basic resource requirement description is extracted from the matched entry; and the extracted basic resource requirement description is parsed to obtain the basic resource requirement vector.

Images