Cross-network interaction method and system based on voice instructions

Abstract

The application provides a cross-network interaction method and system based on voice instructions, relates to the technical field of network communication, and comprises the following steps: receiving a voice instruction, extracting a keyword, performing semantic understanding and intent matching, performing context association analysis, determining an interaction target range and dividing a priority subtask, establishing a network channel to transmit the instruction, deploying execution state tracking and realizing fault recovery, and feeding back an instruction execution result, so that intelligent cross-network interaction based on semantic understanding is realized, the robustness of instruction execution is improved, resource utilization efficiency is optimized, and fault recovery capability is enhanced.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of network communication, and in particular to a cross-network interaction method and system based on voice instructions. BACKGROUND

[0002] With the rapid development of artificial intelligence and voice recognition technology, the human-computer interaction mode based on voice instructions has been widely applied in intelligent home, intelligent office, vehicle-mounted system and other fields. As a natural and convenient interaction mode, voice instructions enable users to complete the control and information query of devices through oral expression.

[0003] The traditional voice interaction system mainly works in a single network environment. With the popularity of Internet of Things technology and the development of heterogeneous network integration, cross-network interaction demand is increasingly prominent. Cross-network interaction technology allows users to complete complex tasks across different network domains through voice instructions, such as controlling devices in an office network through a home network or securely accessing private cloud resources in a public network environment. However, the existing cross-network voice interaction system still has the problems of lacking deep context association analysis capability, being difficult to accurately identify the current interaction scene of the user, not being able to prioritize according to the task characteristics and importance, being prone to low execution efficiency or task blocking when multiple sub-tasks exist in resource competition or dependency, and being unable to achieve fine-grained state recovery, which not only wastes computing resources but also prolongs the user waiting time, seriously affecting the reliability and user experience of the system.

[0004] Therefore, there is an urgent need for a solution to solve the problems in the prior art. SUMMARY

[0005] The present application provides a cross-network interaction method and system based on voice instructions, which can at least solve some of the problems in the prior art.

[0006] In a first aspect, the present application provides a cross-network interaction method based on voice instructions, comprising:

[0007] Receiving a voice instruction sent by a user terminal, extracting keyword information in the voice instruction to generate semantic understanding data, and performing intent matching in a cross-network interaction database to determine target network node information;

[0008] Performing context association analysis on the semantic understanding data, extracting scene feature information in the historical interaction session, identifying the current interaction scene of the user according to the scene feature information, determining the target range of cross-network interaction based on the interaction scene, dividing the target range into multiple sub-tasks, setting an execution priority for each sub-task, and generating a target instruction set with a priority identifier;

[0009] establish a network channel between the user terminal and a target server based on the target network node information, and transmit the target instruction set to the target server;

[0010] An instruction execution state tracker is deployed at the edge node and the core node corresponding to the target network node respectively to record data dependency and intermediate calculation results in the instruction execution process. When an execution exception is detected, a dependency analysis is performed based on a topological sorting algorithm to determine a minimum recovery range. Distributed snapshot technology is used to save the state information of each execution node, and only the necessary execution state is recovered through incremental synchronization during fault recovery to obtain the target instruction execution result.

[0011] The target instruction execution result is received and fed back to the user terminal.

[0012] In an optional implementation,

[0013] The voice instruction sent by the user terminal is received, key word information in the voice instruction is extracted to generate semantic understanding data, and intent matching is performed in the cross-network interaction database. The target network node information includes:

[0014] The voice instruction sent by the user terminal is received, phoneme segmentation and feature vector extraction are performed on the voice instruction, the extracted feature vector is matched with a preset voice feature library, and key word information in the voice instruction is identified;

[0015] The key word information is structured according to a preset semantic rule, the dependency and hierarchical relationship between the key words are extracted, a semantic tree structure is constructed, and semantic understanding data containing semantic association information are generated;

[0016] The semantic understanding data and the multi-dimensional intent feature vector pre-stored in the cross-network interaction database are subjected to similarity calculation, the intent feature vector with the highest matching degree is selected based on the maximum similarity principle, and the target network node information associated with the intent feature vector is obtained.

[0017] In an optional implementation,

[0018] The semantic understanding data is subjected to context association analysis, scene feature information in a historical interaction session is extracted, the current interaction scene of the user is identified according to the scene feature information, the target range of cross-network interaction is determined based on the interaction scene, the target range is divided into multiple sub-tasks, an execution priority is set for each sub-task, and a target instruction set with a priority identifier is generated, including:

[0019] The semantic understanding data is obtained, context association analysis is performed on the semantic understanding data to obtain an association degree calculation parameter, and a time sequence association matrix is constructed based on the association degree calculation parameter.

[0020] extracting feature data in the historical interaction session based on the time sequence correlation matrix by using a sliding time window method, calculating a time decay weight according to the feature data, and performing feature combination on the time decay weight and context information corresponding to the historical interaction session to obtain a scene feature vector;

[0021] inputting the scene feature vector into a multilayer perception machine for classification calculation to obtain a scene classification result, identifying an interaction scene currently experienced by the user based on the scene classification result, and matching the interaction scene with a predefined scene task mapping rule to obtain a target range of cross-network interaction;

[0022] constructing a task dependency graph according to the target range, dividing the target range into a plurality of subtasks based on the task dependency graph, and calculating data dependency values and resource dependency values between the subtasks to obtain a task dependency strength;

[0023] inputting the task dependency strength and the task dependency graph into a preset priority calculation model to obtain a task attribute value of each subtask, setting an execution priority of each subtask based on the task attribute value, and generating a target instruction set, the target instruction set including subtask identification information, priority identification information, and task dependency relationship information.

[0024] In an optional implementation,

[0025] constructing a task dependency graph according to the target range, dividing the target range into a plurality of subtasks based on the task dependency graph, and calculating data dependency values and resource dependency values between the subtasks to obtain a task dependency strength includes:

[0026] performing dependency relationship analysis on the target range to obtain an initial dependency structure, dividing the target range into a plurality of subtasks based on the initial dependency structure, analyzing direct dependency relationships for the subtasks and establishing dependency relationship edges, and combining the subtasks and the dependency relationship edges to construct a complete task dependency graph;

[0027] extracting task type information, execution time information, and resource demand information from the subtasks and combining the information to construct a feature vector, inputting the feature vector into a graph neural network to obtain node representation results, performing message passing operations on the node representation results to obtain subtask state values, and inputting the subtask state values into a multi-head attention mechanism to calculate attention weights;

[0028] performing structural optimization on the complete task dependency graph according to the attention weights to obtain an optimized dependency graph, calculating a data set intersection over union value of a subtask in the optimized dependency graph, and multiplying the data set intersection over union value by a preset data dependency weight to obtain a data dependency value;

[0029] Obtain the resource requirements of subtasks in the optimized dependency graph, calculate the resource requirement ratio between subtasks, sum the resource requirement ratio with the preset resource type weights to obtain the resource dependency value, and combine the data dependency value with the resource dependency value to obtain the task dependency strength.

[0030] In one alternative implementation,

[0031] Establishing a network channel between the user terminal and the target server based on the target network node information, and transmitting the target instruction set to the target server includes:

[0032] Obtain target network node information, calculate the communication quality parameters of each network node based on the target network node information, and determine the network nodes whose communication quality parameters are greater than a preset threshold as network channel nodes to be established;

[0033] For the network channel node to be established, a set of communication paths is obtained by performing routing calculations. Based on the set of communication paths, a multi-path network channel between the user terminal and the target server is constructed, and a primary and backup communication link is established in the multi-path network channel.

[0034] The target instruction set is divided into multiple instruction subsets. Based on the real-time load of the primary and backup communication links, the instruction subsets are distributed to the multi-path network channels for parallel transmission to the target server according to the load balancing strategy.

[0035] In one alternative implementation,

[0036] Instruction execution status tracking is deployed on the edge nodes and core nodes corresponding to the target network node to record data dependencies and intermediate calculation results during instruction execution. When a node execution anomaly is detected, dependency analysis is performed based on a topology sorting algorithm to determine the minimum recovery range, including:

[0037] Deploy instruction execution status tracking on the edge nodes and core nodes corresponding to the target network node, collect instruction execution status to generate edge node status vectors, and collect node association information to generate core node status matrices.

[0038] Based on the edge node state vector and the core node state matrix, the data dependencies and intermediate calculation results during the instruction execution process are recorded to generate an instruction state descriptor. The update time of the instruction state descriptor is adjusted according to the current load to obtain the real-time instruction state.

[0039] According to the real-time instruction state, an instruction dependency graph is constructed, an instruction dependency strength between each pair of instructions in the instruction dependency graph is calculated, and the instruction dependency strength is taken as a weight of a dependency edge to obtain a weighted instruction dependency graph;

[0040] When a node execution exception is detected, a topological sorting algorithm is executed on the weighted instruction dependency graph for dependency analysis, a propagation path of an instruction is obtained, a fault influence range is determined based on the propagation path, and instructions in the fault influence range are determined as a minimum recovery range.

[0041] In an optional implementation,

[0042] In combination with a distributed snapshot technology, state information of each execution node is saved, and when a fault is recovered, only necessary execution states are recovered by an incremental synchronization manner to obtain a target instruction execution result, including:

[0043] In combination with a distributed snapshot technology, state information of each execution node is collected to construct a state snapshot, a consistency score is calculated based on a consistency index and a preset weight coefficient, and effective state information set is obtained by screening state information with a consistency score greater than a threshold value;

[0044] State data of the effective state information set in the minimum recovery range is saved, a state change log is generated according to the state data, a state change cost is calculated to obtain a to-be-synchronized state set;

[0045] Importance and urgency indexes of the to-be-synchronized state set are calculated by an incremental synchronization manner, a state synchronization priority sequence is generated, a state necessity score is obtained by calculating state influence degrees and recoverability in the state synchronization priority sequence, and necessary execution states are marked;

[0046] Batch recovery parameters are calculated according to the number of necessary execution states and system load, a recovery state result is obtained by executing batch recovery, an error rate, a coverage rate and a consistency rate of the recovery state result are calculated to generate a verification score, and a target instruction execution result is determined in combination with a gain factor and a penalty factor.

[0047] A second aspect of the embodiment of the application provides a cross-network interaction system based on a voice instruction, including:

[0048] A first unit is configured to receive a voice instruction sent by a user terminal, extract keyword information in the voice instruction, generate semantic understanding data, and perform intent matching in a cross-network interaction database to determine target network node information.

[0049] The second unit is configured to perform context correlation analysis on the semantic understanding data, extract scene feature information in a historical interaction session, identify an interaction scene in which the user is currently located according to the scene feature information, determine a target range of cross-network interaction based on the interaction scene, divide the target range into a plurality of subtasks, set an execution priority for each subtask, and generate a target instruction set with a priority identifier;

[0050] The third unit is configured to establish a network channel between the user terminal and a target server based on target network node information, and transmit the target instruction set to the target server.

[0051] The fourth unit is configured to deploy instruction execution state tracking in an edge node and a core node corresponding to the target network node, record data dependency and intermediate calculation results in an instruction execution process, perform dependency analysis based on a topological sorting algorithm when detecting node execution abnormity, determine a minimum recovery range, save state information of each execution node in combination with a distributed snapshot technology, and only recover necessary execution states through an incremental synchronization method when recovering from a fault to obtain a target instruction execution result.

[0052] The fifth unit is configured to receive the target instruction execution result and feed back to the user terminal.

[0053] In a third aspect of the embodiments of the present application, an electronic device is provided, comprising:

[0054] A processor and a memory for storing processor-executable instructions, wherein the processor is configured to invoke the instructions stored in the memory to execute the method described above.

[0055] In a fourth aspect of the embodiments of the present application, a computer-readable storage medium is provided, which stores computer program instructions, and the computer program instructions are executed by a processor to implement the method described above.

[0056] In the present application, by receiving voice instructions and extracting keywords to generate semantic understanding data, combining context association analysis to identify user interaction scenarios, the user's intention can be accurately grasped, so as to realize accurate cross-network interaction, improve the convenience and efficiency of user operation, divide the target range into multiple sub-tasks and set the execution priority, generate a target instruction set with priority identification, realize reasonable scheduling of tasks and optimization of resource allocation, effectively improve the system response speed and processing efficiency, at the same time, reduce the network transmission burden, reduce the server resource consumption, deploy the instruction execution state tracking mechanism at the edge node and the core node of the target network node, record the data dependency relationship and the intermediate calculation result, combine the topological sorting algorithm and the distributed snapshot technology, realize accurate positioning and minimum range recovery in fault condition, significantly improve the stability and reliability of the system, and ensure the continuity and integrity of the cross-network interaction process. BRIEF DESCRIPTION OF DRAWINGS

[0057] Figure 1 A flowchart of the cross-network interaction method based on voice instructions of the embodiment of the present application is shown in

[0058] Figure 2 A task dependency analysis and optimization flowchart of the cross-network interaction method based on voice instructions of the embodiment of the present application is shown in

[0059] Figure 3 A system state recovery flowchart of the cross-network interaction method based on voice instructions of the embodiment of the present application is shown in DETAILED DESCRIPTION

[0060] In order to make the purpose, technical scheme and advantages of the embodiments of the present application clearer, the technical scheme in the embodiments of the present application will be described clearly and completely below in combination with the drawings in the embodiments of the present application. Obviously, the described embodiments are only part of the embodiments of the present application, not all the embodiments. Based on the embodiments in the present application, all other embodiments obtained by those skilled in the art without creative labor fall within the scope of protection of the present application.

[0061] The technical scheme of the present application will be described in detail below with specific embodiments. The following specific embodiments can be combined with each other, and the same or similar concepts or processes may not be described in some embodiments.

[0062] Figure 1 A flowchart of the cross-network interaction method based on voice instructions of the embodiment of the present application is shown in Figure 1 The method comprises:

[0063] receive a voice instruction sent by a user terminal, extract keyword information in the voice instruction to generate semantic understanding data, and perform intent matching in a cross-network interaction database to determine target network node information;

[0064] perform context association analysis on the semantic understanding data, extract scene feature information in a historical interaction session, identify an interaction scene in which the user is currently located according to the scene feature information, determine a target range of cross-network interaction based on the interaction scene, divide the target range into a plurality of sub-tasks, set an execution priority for each sub-task, and generate a target instruction set with a priority identifier;

[0065] establish a network channel between the user terminal and a target server based on the target network node information, and transmit the target instruction set to the target server;

[0066] deploy instruction execution state tracking in an edge node and a core node corresponding to the target network node, record data dependency and intermediate calculation results in the instruction execution process, perform dependency analysis based on a topological sorting algorithm when detecting node execution abnormity, determine a minimum recovery range, save state information of each execution node in combination with a distributed snapshot technology, and only recover necessary execution states through incremental synchronization when recovering from a fault to obtain a target instruction execution result;

[0067] receive the target instruction execution result and feed back to the user terminal.

[0068] In an optional implementation manner,

[0069] receiving a voice instruction sent by a user terminal, extracting keyword information in the voice instruction to generate semantic understanding data, and performing intent matching in a cross-network interaction database to determine target network node information comprises:

[0070] receiving a voice instruction sent by a user terminal, performing phoneme segmentation and feature vector extraction on the voice instruction, matching the extracted feature vector with a preset voice feature library, and identifying keyword information in the voice instruction;

[0071] structurally processing the keyword information according to a preset semantic rule, extracting dependency and hierarchical relationships between keywords, constructing a semantic tree structure, and generating semantic understanding data containing semantic association information;

[0072] performing similarity calculation on the semantic understanding data and a plurality of multi-dimensional intent feature vectors pre-stored in a cross-network interaction database, selecting an intent feature vector with the highest matching degree based on a maximum similarity principle, and obtaining target network node information associated with the intent feature vector.

[0073] The voice instruction sent by the user terminal can be a natural language request issued by the user through a smart phone, a smart sound box or other smart terminal devices. The received voice instruction is processed using a phoneme segmentation technology to divide continuous voice signals into basic voice units. For example, the voice signals are processed in frames using a dynamic time warping algorithm, each frame is 25 milliseconds, and the frame shift is 10 milliseconds. The start and end points of the voice are determined through short-time energy and zero-crossing rate analysis. The voice stream is divided into a series of phoneme sequences using a method based on a hidden Markov model. For example, when the user says the voice instruction "turn on the smart lamp in the living room and set it to yellow", the system divides it into the corresponding phoneme sequence.

[0074] Feature vector extraction is performed on the segmented phoneme sequence. Mel frequency cepstral coefficient technology is used to extract 13-dimensional basic feature parameters for each phoneme. First-order difference and second-order difference features are combined to form a 39-dimensional feature vector. The extracted feature vector is matched with a pre-established voice feature library. The feature library contains a set of feature vectors of voice samples of common instruction words, device names, operation actions, etc. The similarity is calculated through a deep neural network model to identify the keyword information in the voice instruction. In the foregoing example, the keywords "turn on", "living room", "smart lamp", "set", "yellow" are identified.

[0075] After identifying the keywords, the keywords are structured according to the preset semantic rules. The dependency and hierarchical relationships between the keywords are analyzed using dependency syntax analysis technology. A dependency syntax analysis model based on a bidirectional long short-term memory network is used. The model is trained on 100,000 domain-related corpora with an accuracy of 94.3%. Through dependency relationship analysis, it is determined that "turn on" is the core predicate verb, "smart lamp" is the patient component, "living room" is the location modifier, "set" is the secondary predicate, and "yellow" is the object component. Based on the dependency relationship, a semantic tree structure is constructed, in which "turn on" is the root node, connected to the "smart lamp" node, the "smart lamp" node is connected to the "living room" modifier node, and the "turn on" node is also connected to the "set" sub-operation node. The "set" node is connected to the "yellow" parameter node.

[0076] After the semantic tree is constructed, semantic understanding data containing semantic association information is generated. The semantic understanding data is represented in JSON format and includes fields such as main operation type, operation object, location information, sub-operation and its parameters. For the above example, the generated semantic understanding data is: {"operation type": "turn on", "operation object": "smart lamp", "location": "living room", "sub-operation": [{"operation type": "set", "parameters": {"color": "yellow"}}]}.

[0077] After generating the semantic understanding data, similarity calculation is performed between the semantic understanding data and the pre-stored multi-dimensional intent feature vectors in the cross-network interaction database. The database contains feature vector representations of various user intents that can be responded to by network nodes, and each intent feature vector has a dimension of 128 and is obtained by training on 100,000 user instruction samples. The cosine similarity algorithm is used to calculate the similarity scores of the semantic understanding data and each intent feature vector in the database. During the calculation, different weights are assigned to the key dimensions in the semantic understanding data, such as operation type, operation object, and location information, with the operation type weight being 0.4, the operation object weight being 0.3, the location information weight being 0.2, and the weight of other information being 0.1, in order to improve the matching accuracy.

[0078] Based on the maximum similarity principle, the intent feature vector with the highest matching degree is selected, and the "smart home-lighting control" intent feature vector with a similarity score of 0.92 may be matched. Subsequently, the target network node information associated with the intent feature vector is obtained, including node ID, network address, communication protocol, authentication information, etc. For example, the target network node information may be: {"node ID": "SL2023", "network address": "192.168.1.105", "protocol type": "MQTT", "authentication key": "dhw7e9r2hdiq8w", "device type": "smart lighting controller"}.

[0079] After obtaining the target network node information, a communication connection with the node is established, the user intent is converted into a format recognizable by the node, and intelligent control across networks is realized. The entire processing flow from voice instruction reception to target node determination is usually completed within 500 milliseconds, meeting the real-time interaction requirements.

[0080] In this embodiment, the most matching intent feature vector is selected based on the maximum similarity principle, thereby accurately obtaining the associated target network node information, constructing a semantic tree structure by analyzing the dependency and hierarchical relationships between keywords, forming semantic understanding data containing complete semantic association information, significantly improving the recognition accuracy of voice instructions, realizing deep semantic understanding of voice content, and ensuring accurate positioning of cross-network interaction targets.

[0081] In an alternative embodiment,

[0082] The semantic understanding data is subjected to context association analysis, scene feature information in historical interaction sessions is extracted, the current interaction scene of the user is identified according to the scene feature information, the target range of cross-network interaction is determined based on the interaction scene, the target range is divided into multiple sub-tasks, and an execution priority is set for each sub-task, and a target instruction set with a priority identifier is generated, including:

[0083] obtaining semantic understanding data, performing context association analysis on the semantic understanding data to obtain an association degree calculation parameter, constructing a time sequence association matrix based on the association degree calculation parameter;

[0084] extracting feature data in a historical interaction session based on the time sequence association matrix using a sliding time window method, calculating a time decay weight according to the feature data, and combining the time decay weight with context information corresponding to the historical interaction session to obtain a scene feature vector;

[0085] inputting the scene feature vector into a multi-layer perception machine for classification calculation to obtain a scene classification result, identifying an interaction scene in which the user is currently located based on the scene classification result, and matching the interaction scene with a predefined scene task mapping rule to obtain a target range of cross-network interaction;

[0086] constructing a task dependency graph according to the target range, dividing the target range into a plurality of sub-tasks based on the task dependency graph, and calculating data dependency values and resource dependency values between the sub-tasks to obtain a task dependency strength;

[0087] inputting the task dependency strength and the task dependency graph into a preset priority calculation model to obtain a task attribute value of each sub-task, setting an execution priority of each sub-task based on the task attribute value, generating a target instruction set, and the target instruction set containing sub-task identification information, priority identification information and task dependency relationship information.

[0088] Obtain semantic understanding data, which may come from the user's voice input, text message or other interactive interface. For example, the user may issue the instruction "help me find the meeting records of the past week and create a summary report". After receiving this semantic understanding data, context association analysis will be performed on it. For example, extract keywords in the semantics (such as "meeting records" and "summary report"), and calculate the similarity with the keywords in the historical session to obtain the association degree calculation parameter. For example, the association degrees of "meeting records" in the current instruction with "video conference" and "meeting minutes" in the historical session are 0.85 and 0.92 respectively. Based on the association degree calculation parameter, a time sequence association matrix is constructed, which records the semantic association degree between user instructions at different time points.

[0089] The sliding time window method is used to process the time series correlation matrix, and the window size is set to the interaction records in the past 7 days. By extracting feature data in the historical interaction session within this time window, the user's recent focus and behavior patterns are identified. For example, it is found that the user has 4 queries related to meetings and 2 report generation operations in the past 5 days. The time decay weight is calculated so that the more recent interactions have higher reference value. For example, the interaction weight 1 day ago is 0.9, and the interaction weight 5 days ago is only 0.5. The time decay weight is combined with the context information corresponding to the historical interaction session to obtain a scene feature vector. This vector may contain user preferences, task frequency, time patterns, and other multi-dimensional features.

[0090] The generated scene feature vector is input into a pre-trained multilayer perceptron for classification calculation. The perceptron includes an input layer, two hidden layers (with 64 and 32 neurons respectively), and an output layer, which processes feature information through a nonlinear activation function and finally outputs the scene classification result. For example, the user's current interaction scene is identified as "work report preparation phase". The system matches this interaction scene with the predefined scene task mapping rule to determine the target range of cross-network interaction. For the "work report preparation phase" scene, the target range includes three main task domains: data query, information integration, and document generation.

[0091] Based on the determined target range, a task dependency graph is constructed. The task dependency graph is a directed acyclic graph, where nodes represent subtasks and edges represent the dependency relationship between tasks. For example, the instruction "help me find the meeting records of the past week and create a summary report" is divided into three subtasks: task one, query the meeting record system to obtain the meeting data of the past week; task two, extract key information from the meeting records; task three, generate a summary report based on the extracted information. The data dependency value and resource dependency value between subtasks are calculated to obtain the task dependency strength. For example, the data dependency value of task two to task one is 0.95 (very high), indicating that without meeting record data, information cannot be extracted; the dependency value of task three to computing resources is 0.75 (high), because generating a report requires more processing power.

[0092] The system inputs the task dependency strength and the task dependency graph into a preset priority calculation model. The model comprehensively considers factors such as the urgency, importance, dependency relationship and resource consumption of the task, and calculates the task attribute value for each subtask. For example, the attribute value of task one is [0.9, 0.8, 0.3], representing the urgency, importance and complexity respectively. Based on these task attribute values, the execution priority of each subtask is set: the priority of task one is "high" (value 9), the priority of task two is "medium" (value 6), and the priority of task three is "medium" (value 5). Finally, the target instruction set is generated, including subtask identification information (such as "T001-Query meeting records"), priority identification information (such as "P9-high priority") and task dependency relationship information (such as "T002 depends on T001").

[0093] The generated target instruction set is distributed to different execution modules, and each subtask is executed in turn according to the set priority. During execution, the task execution state is monitored, and the execution plan is dynamically adjusted according to the actual situation. For example, if it is found that the meeting record system responds slowly, the priority of the subtask may be temporarily reduced, and other independent subtasks are executed first.

[0094] In this embodiment, by performing context association analysis on semantic understanding data and constructing a time sequence association matrix, and combining with the sliding time window method to extract historical interactive session features, and introducing a time decay weight mechanism, the dynamic feature capture of user interactive behavior is realized, the task dependency graph is constructed, the association strength between tasks is quantified by calculating the data dependency value and resource dependency value between subtasks, and these dependency relationships are input into the priority calculation model to generate a target instruction set containing complete task execution information, breaking through the technical bottlenecks of inaccurate scene recognition and extensive task division in traditional methods, improving the intelligent level of cross-network interaction, realizing accurate division and efficient scheduling of tasks, and at the same time ensuring the continuity of the interactive process and the collaboration of task execution in complex scenarios.

[0095] In an alternative embodiment,

[0096] According to the target range, a task dependency graph is constructed, the target range is divided into a plurality of subtasks based on the task dependency graph, and the data dependency value and the resource dependency value between the subtasks are calculated to obtain a task dependency strength, which includes:

[0097] An initial dependency structure is obtained by performing dependency relationship analysis on the target range, the target range is divided into a plurality of subtasks based on the initial dependency structure, direct dependency relationships are analyzed for the subtasks and dependency relationship edges are established, and a complete task dependency graph is constructed by combining the subtasks and the dependency relationship edges;

[0098] Task type information, execution time information and resource requirement information are extracted from the subtasks and combined to build a feature vector, the feature vector is input into a graph neural network to obtain node representation results, message passing operations are performed on the node representation results to obtain subtask state values, and the subtask state values are input into a multi-head attention mechanism to calculate attention weights;

[0099] An optimized dependency graph is obtained by structure optimization of the complete task dependency graph according to the attention weights, a data set intersection over union value of a subtask in the optimized dependency graph is calculated, and a data dependency value is obtained by multiplying the data set intersection over union value by a preset data dependency weight;

[0100] The resource requirement amount of a subtask in the optimized dependency graph is obtained, a resource requirement ratio between subtasks is calculated, a resource dependency value is obtained by weighted summation of the resource requirement ratio and a preset resource type weight, and a task dependency strength is obtained by weighted combination of the data dependency value and the resource dependency value.

[0101] An initial dependency structure is obtained by performing dependency relationship analysis on a target range. Specifically, a static code analysis tool is used to scan the code in the target range to identify function call relationships, variable reference relationships and data flow relationships. For example, in a data processing target range, it is found through analysis that function A calls function B, function B calls function C, and functions A and C both use the same data set D, and these relationships constitute the initial dependency structure.

[0102] Based on the obtained initial dependency structure, the target range is divided into multiple subtasks. The division is based on functional modules and data processing units, and code blocks with similar functions or operating the same data are combined into a subtask. In actual applications, for example, for a data processing system, data acquisition tasks, data cleaning tasks, feature extraction tasks and model training tasks can be divided into subtasks.

[0103] For the divided subtasks, direct dependency relationships are analyzed and dependency relationship edges are established. Direct dependency relationships include data dependency and control dependency. Data dependency refers to the output data of one subtask being the input data of another subtask; control dependency refers to the execution condition of one subtask depending on the execution result of another subtask. By detecting data flow and execution order constraints between subtasks, dependency relationships are determined and corresponding dependency relationship edges are established.

[0104] The sub-tasks and dependency edges are combined to build a complete task dependency graph. The dependency graph adopts a directed graph structure, where nodes represent sub-tasks and edges represent dependency relationships. For example, in a four-node dependency graph, sub-task A points to sub-task B and sub-task C, and sub-task B and sub-task C both point to sub-task D, indicating that sub-task D depends on the completion of sub-task B and sub-task C, while sub-task B and sub-task C depend on the completion of sub-task A.

[0105] Task type information, execution time information, and resource requirement information are extracted from sub-tasks and combined to build a feature vector. Task type information can be categories such as compute-intensive, IO-intensive, or hybrid; execution time information includes historical average execution time, maximum execution time, and minimum execution time; resource requirement information includes CPU usage, memory occupancy, and network bandwidth demand, etc. After normalization processing, the resource requirement information is combined into the feature vector of each sub-task. Specifically, the feature vector of a certain sub-task can be [0.8, 0.5, 0.3, 0.6, 0.4], representing its compute-intensive degree, average execution time, CPU usage, memory occupancy, and network bandwidth demand, respectively.

[0106] The feature vector is input into the graph neural network to obtain the node representation result. The graph neural network adopts a multi-layer structure, each layer containing feature transformation and aggregation operations. Feature transformation maps the feature vector of a node to a new feature space; aggregation operation considers the node's own features and adjacent node features. After multiple layers of processing, a high-dimensional representation result of each sub-task node is obtained, which contains the node's own information and topological structure information.

[0107] Message passing operation is performed on the node representation result to obtain the sub-task state value. Message passing operation updates the node state through iteration, in each iteration, the node receives information from its adjacent nodes and updates its own state. Message passing considers the directionality and weight of the edge, so that the node with stronger dependency has greater influence on the target node. After multiple iterations, the state value of each node tends to be stable, which reflects the importance and dependency of the node in the entire dependency graph.

[0108] The sub-task state value is input into the multi-head attention mechanism to calculate the attention weight. Multi-head attention mechanism processes input data through multiple independent attention calculation units in parallel, each attention head focuses on different aspects of the data, and the results of multiple attention heads are combined. By calculating the similarity between query vector, key vector, and value vector, the attention weight is obtained, which represents the relative importance and correlation strength between nodes.

[0109] An optimized dependency graph is obtained by structurally optimizing the complete task dependency graph according to the attention weights. The structural optimization includes two steps: edge weight adjustment and redundant edge removal. The edge weight adjustment updates the strength of the dependency relationship edges according to the attention weights; the redundant edge removal removes edges with weights below a threshold to simplify the dependency structure. The optimized dependency graph more accurately reflects the actual dependency strength between subtasks.

[0110] The data set intersection-union ratio of the subtasks in the optimized dependency graph is calculated. For each pair of connected subtasks, the number of data set elements of the common operation (intersection) and the total number of data set elements of each operation (union) are counted, and the ratio of the intersection to the union is calculated. For example, if subtask A operates on data set {1, 2, 3, 4} and subtask B operates on data set {3, 4, 5, 6}, their intersection is {3, 4} and their union is {1, 2, 3, 4, 5, 6}, and the intersection-union ratio is 2 / 6=0.33.

[0111] The data dependency value is obtained by multiplying the data set intersection-union ratio by the preset data dependency weight. The data dependency weight is pre-set according to the data type and importance, for example, the weight of structured data is 0.7 and the weight of unstructured data is 0.3. The intersection-union ratio is multiplied by the corresponding weight to obtain the final data dependency value.

[0112] The resource demand of the subtasks in the optimized dependency graph is obtained. The resource demand includes CPU demand, memory demand, and network bandwidth demand, etc. By analyzing the historical execution records and resource occupation of the subtasks, the demand of each resource is estimated. For example, the CPU demand of subtask A is 4 cores, the memory demand is 8 GB, and the network bandwidth demand is 100 Mbps.

[0113] The resource demand ratio between subtasks is calculated. For each pair of connected subtasks, the demand ratio on each type of resource is calculated. For example, if the CPU demand of subtask A is 4 cores and the CPU demand of subtask B is 2 cores, the CPU demand ratio is 4 / 2=2.

[0114] The resource dependency value is obtained by weighted sum of the resource demand ratio and the preset resource type weight. The resource type weight is set according to the importance and scarcity of the resource, for example, the CPU weight is 0.5, the memory weight is 0.3, and the network bandwidth weight is 0.2. The demand ratio of each type of resource is multiplied by the corresponding weight and summed to obtain the final resource dependency value.

[0115] The data dependency value and the resource dependency value are combined by weighting to obtain a task dependency strength. The weighted combination can adopt a linear combination manner, and weights are set according to the relative importance of data dependency and resource dependency. For example, if the data dependency weight is 0.7, the resource dependency weight is 0.3, the data dependency value is 0.8, and the resource dependency value is 0.5, then the final task dependency strength is 0.7*0.8+0.3*0.5=0.71.

[0116] In the embodiment, the target range is divided into multiple sub-tasks through dependency relationship analysis, and a complete task dependency graph is constructed based on direct dependency relationship. A graph neural network and a multi-head attention mechanism are introduced, a feature vector is constructed by extracting information such as task type, execution time and resource demand, deep representation learning of the task node is realized, and a message passing mechanism is used to obtain a sub-task state value, and then the structure of the task dependency graph is optimized based on attention weight;

[0117] In the prior art, when performing task dependency analysis, a simple linear dependency relationship judgment method is often used, which cannot accurately capture the complex dependency relationship between tasks, and lacks comprehensive consideration of task features and resource demand, resulting in inaccurate task division results, which affects the efficiency of cross-network task execution;

[0118] The embodiment calculates the data set intersection and union ratio value and obtains the data dependency value by combining the data dependency weight, simultaneously considers the resource demand ratio and resource type weight between tasks, and finally obtains an accurate task dependency strength by weighted combination, breaks through the technical limitations of rough dependency relationship analysis and incomplete feature extraction in traditional methods, realizes accurate quantification of task dependency relationship and intelligent optimization of task structure, significantly improves the accuracy and comprehensiveness of task dependency analysis, provides a reliable basis for subsequent task scheduling, optimizes the system resource allocation efficiency, and effectively improves the overall performance of cross-network task execution.

[0119] Figure 2 A task dependency analysis and optimization flowchart of the cross-network interaction method based on a voice instruction according to the embodiment of the application;

[0120] In an optional implementation,

[0121] Establishing a network channel between the user terminal and the target server based on the target network node information, and transmitting the target instruction set to the target server comprises:

[0122] Obtaining target network node information, calculating a communication quality parameter of each network node according to the target network node information, and determining a network channel node to be established as a network node whose communication quality parameter is greater than a preset threshold;

[0123] Performing routing calculation on the network channel nodes to be established to obtain a set of communication paths, and constructing a multi-path network channel between the user terminal and the target server based on the set of communication paths, wherein the primary and backup communication links are established in the multi-path network channel;

[0124] Dividing the target instruction set into a plurality of instruction subsets, and distributing the instruction subsets to the multi-path network channel according to a load balancing strategy for parallel transmission to the target server according to the real-time load of the primary and backup communication links.

[0125] Obtain target network node information. The target network node information includes node identification, geographic location, network bandwidth, time delay, packet loss rate and other network state parameters. The network state parameters are obtained by network detection tools, such as ping test using ICMP protocol, or API interface call of network management system. For example, the bandwidth of node A is 100 Mbps, the time delay is 15 ms, and the packet loss rate is 0.2%; the bandwidth of node B is 80 Mbps, the time delay is 20 ms, and the packet loss rate is 0.5%; the bandwidth of node C is 120 Mbps, the time delay is 25 ms, and the packet loss rate is 0.3%, and other detailed information.

[0126] After obtaining the target network node information, the communication quality parameters of each network node are calculated according to the target network node information. The calculation of the communication quality parameters considers the bandwidth, time delay, packet loss rate, node stability and other factors. The weighted average method can be used to assign different weights to different factors. For example, the bandwidth weight can be set to 0.4, the time delay weight can be set to 0.3, the packet loss rate weight can be set to 0.2, and the node stability weight can be set to 0.1. For node A, the communication quality parameter can be calculated as: 0.4x(100 / 120)+0.3x(1-15 / 25)+0.2x(1-0.2 / 0.5)+0.1x0.95=0.81. The system performs similar calculation on all nodes to obtain the communication quality parameter value of each node. Assuming that the preset threshold is 0.75, the nodes with communication quality parameters greater than 0.75 will be determined as the network channel nodes to be established. In this example, nodes A and C that meet the conditions will be selected as the network channel nodes to be established.

[0127] For the to-be-established network channel node, route calculation is performed to obtain a communication path set. The route calculation adopts the Dijkstra algorithm, considers the connectivity, bandwidth, delay, and other factors between nodes, and calculates multiple feasible communication paths. For example, from the user terminal to the target server, the possible paths include: path 1: user terminal→node A→node D→target server, total bandwidth is 80 Mbps, total delay is 35 ms; path 2: user terminal→node C→node E→target server, total bandwidth is 100 Mbps, total delay is 40 ms; path 3: user terminal→node B→node F→target server, total bandwidth is 70 Mbps, total delay is 45 ms, and the multiple paths constitute the communication path set.

[0128] Based on the communication path set, a multi-path network channel between the user terminal and the target server is constructed. The path with the best communication quality is selected as the main communication link, and the remaining paths are used as backup communication links. In the foregoing embodiment, path 1 is selected as the main communication link, and paths 2 and 3 are used as backup communication links. Connections of these links are established through the multi-path TCP protocol (MPTCP) in the TCP / IP protocol family or a custom multi-path communication protocol. The health status of each link is monitored at the same time, and when the main link fails, the backup link is switched immediately to ensure the continuity of communication.

[0129] The target instruction set is divided into multiple instruction subsets, and the instruction set is reasonably divided according to the size and characteristics of the instruction set. For example, for a target instruction set containing 1000 instructions, the instructions can be divided into 10 instruction subsets according to the functional correlation or execution order of the instructions, and each subset contains 100 instructions. The division method can adopt sequential division, functional division, or mixed division strategy. For example, sequential division can simply divide instructions 1-100 into the first subset, instructions 101-200 into the second subset, and so on.

[0130] According to the real-time load of the main and backup communication links, the instruction subsets are distributed to the multi-path network channel for parallel transmission according to a load balancing strategy. The load balancing strategy can be weighted round robin, minimum connection number, or adaptive distribution based on prediction. The bandwidth usage, queue length, response time, and other indicators of each communication link are monitored in real time, and the distribution strategy is dynamically adjusted. For example, when it is detected that the bandwidth usage of the main link (path 1) reaches 85%, part of the instruction subsets are transferred to the lightly loaded backup link (path 2) for transmission. For example, instruction subsets 1-6 are distributed to path 1 for transmission, instruction subsets 7-9 are distributed to path 2 for transmission, and instruction subset 10 is distributed to path 3 for transmission.

[0131] In this embodiment, by calculating the communication quality parameters of the network nodes and selecting suitable network channel nodes, a multi-path network channel between the user terminal and the target server is constructed based on routing calculation, and a primary and backup communication link is established, realizing reliable establishment of the network channel. The target instruction set is divided into multiple instruction subsets, and according to the real-time load of the primary and backup communication links, a load balancing strategy is adopted for instruction transmission distribution, realizing parallel and efficient transmission of instructions, ensuring the communication quality and stability of the network channel, improving the efficiency of instruction transmission, avoiding network congestion through the load balancing mechanism, enhancing the fault tolerance and reliability of the system, and providing stable and efficient data transmission guarantee for cross-network interaction.

[0132] In an alternative embodiment,

[0133] The edge node and the core node corresponding to the target network node are respectively deployed with instruction execution state tracking, and the data dependency relationship and intermediate calculation results in the instruction execution process are recorded. When an abnormal node execution is detected, a dependency analysis is performed based on a topological sorting algorithm to determine the minimum recovery range, including:

[0134] The edge node and the core node corresponding to the target network node are respectively deployed with instruction execution state tracking, and the edge node state vector is generated by collecting the instruction execution state, and the core node state matrix is generated by collecting the node association information;

[0135] Based on the edge node state vector and the core node state matrix, the data dependency relationship and intermediate calculation results in the instruction execution process are recorded to generate an instruction state descriptor, and the real-time instruction state is obtained by adjusting the update time of the instruction state descriptor according to the current load;

[0136] According to the real-time instruction state, an instruction dependency graph is constructed, the instruction dependency strength between each pair of instructions in the instruction dependency graph is calculated, and the instruction dependency strength is taken as the weight of the dependency edge to obtain a weighted instruction dependency graph;

[0137] When an abnormal node execution is detected, a topological sorting algorithm is performed on the weighted instruction dependency graph for dependency analysis to obtain a propagation path of the instructions and determine a fault impact range based on the propagation path. The instructions within the fault impact range are determined as the minimum recovery range.

[0138] The edge node refers to a computing node located at the edge position in the network topology structure, such as an Internet of Things device, a terminal device, etc. The core node refers to a computing node located at the center position in the network topology structure, such as a server, a central processing unit, etc. The instruction execution state tracking is realized by deploying a state collection module on the node.

[0139] The collection process of the edge node state vector is to deploy a state collection probe on the edge node to monitor the instruction execution in real time. The probe collects instruction ID, instruction type, execution timestamp, execution state (not executed, executing, completed, failed), input data identifier, and output data identifier. For example, for a data processing instruction A001, its state vector can be represented as {A001, DATA_PROCESS, 1630145789, COMPLETED, [D001, D002], [D003]}, indicating that the instruction A001 is a data processing type, the execution is completed at timestamp 1630145789, the input data is D001 and D002, and the output data is D003.

[0140] The core node state matrix is collected by deploying a matrix construction module on the core node to collect the association information between nodes, including node pair identifier, interaction type, data flow direction, interaction frequency, latest interaction timestamp, and interaction state. For example, for the association between the edge node E001 and the core node C001, the state matrix element can be represented as {E001-C001, DATA_UPLOAD, E001->C001, 10 / min, 1630145790, NORMAL}, indicating that E001 uploads data to C001 at a frequency of 10 times per minute, the latest interaction time is 1630145790, and the interaction state is normal.

[0141] Based on the edge node state vector and the core node state matrix, the instruction state descriptor is generated by fusing the two parts of information to form the instruction state descriptor, which contains: instruction global ID, instruction source node, instruction target node, instruction dependency set, instruction impact set, instruction state, latest update timestamp, and intermediate calculation result reference. For example, an instruction state descriptor can be represented as {GID001, E001, C001, [GID002, GID003], [GID004, GID005], RUNNING, 1630145800, RESULT_REF001}, indicating that the instruction with global ID GID001 is sent from node E001 to C001, depends on GID002 and GID003, affects GID004 and GID005, the current state is running, the latest update time is 1630145800, and the intermediate result reference is RESULT_REF001.

[0142] Through load awareness, the update time of the instruction state descriptor is adjusted according to the current system load. According to the system CPU usage, memory usage, network bandwidth usage and other indicators, the state update frequency is dynamically adjusted. When the system load is low (such as CPU usage < 30%), the state update frequency can be set to 5 times per second; when the system load is medium (such as 30% ≤ CPU usage < 70%), the state update frequency can be set to 2 times per second; when the system load is high (such as CPU usage ≥ 70%), the state update frequency can be set to 1 time per second.

[0143] Based on the real-time instruction state, an instruction dependency graph is constructed, with each instruction as a node in the graph and the dependency relationship between instructions as a directed edge. For example, if instruction A depends on the output of instruction B, a directed edge is established from B to A in the graph. Exemplarily, all instruction state descriptors are traversed to extract the dependency relationship therein, and an adjacency list or adjacency matrix is constructed to represent the instruction dependency graph.

[0144] Based on the data dependency degree (data sharing amount / total data amount), execution time correlation (common execution time / total execution time), and historical abnormal correlation degree (common abnormal number / total abnormal number), the instruction dependency strength is calculated. For example, for instructions A and B, if the amount of data they share accounts for 80% of the total data amount, the common execution time accounts for 60% of the total execution time, and the number of historical common abnormalities accounts for 40% of the total number of abnormalities, the dependency strength between them can be calculated as (0.8 + 0.6 + 0.4) / 3 = 0.6.

[0145] When a node execution anomaly is detected, a topological sorting algorithm is performed on the weighted instruction dependency graph. Exemplarily, an empty result list is initialized, the in-degree of each node is calculated, all nodes with an in-degree of 0 are added to the queue, a node is taken out of the queue and added to the result list, the in-degree of all neighbor nodes of the node is reduced, and if the in-degree of a neighbor node becomes 0, it is added to the queue. Repeat the above process until the queue is empty.

[0146] By analyzing the order of nodes in the sorting result, the propagation path of the instruction is determined. If node A is located before node B in the sorting result, and there is a dependency edge between A and B, it can be determined that there is a propagation path from A to B. The system will filter out strong dependency relationships according to a dependency strength threshold (such as 0.5) to construct a propagation path graph.

[0147] Based on the propagation path and the abnormal detection result, the fault impact range is determined. Starting from the abnormal node, the downstream is tracked along the propagation path, and all nodes with a dependency strength greater than the threshold (such as 0.7) are included in the impact range. For example, if node X is detected to be abnormal, and the dependency strength from X to Y is 0.8 and the dependency strength from Y to Z is 0.75, then X, Y and Z will all be included in the fault impact range.

[0148] Determine the minimum recovery range based on the fault impact range. Analyze the state and intermediate calculation results of each instruction in the fault impact range to determine the minimum instruction set that needs to be re-executed. For valid intermediate calculation results, keep them; for invalid or possibly contaminated intermediate calculation results, mark the corresponding instructions for re-execution, and generate a recovery plan including the instruction list that needs to be re-executed and its execution order.

[0149] In this embodiment, by deploying instruction execution state tracking mechanism in edge nodes and core nodes, edge node state vector and core node state matrix are generated respectively, realizing comprehensive monitoring of instruction execution state. By recording data dependency relationship and intermediate calculation results in instruction execution process, and dynamically adjusting and updating time according to current load, state descriptor accurately reflecting instruction execution is generated. Weighted instruction dependency graph constructed based on real-time instruction state clearly shows the dependency relationship strength between instructions, providing reliable basis for fault handling. Through topological sorting algorithm, fault propagation path is quickly analyzed, and fault impact range is accurately defined, realizing accurate positioning and minimum recovery of fault, significantly improving fault tolerance capability and recovery efficiency of system, ensuring stability and reliability of cross-network interaction process, and minimizing the impact of fault on overall operation of system.

[0150] In an optional embodiment,

[0151] In combination with distributed snapshot technology, the state information of each execution node is saved, and during fault recovery, only the necessary execution state is recovered through incremental synchronization to obtain the target instruction execution result, including:

[0152] In combination with distributed snapshot technology, the state information of each execution node is collected to construct state snapshot, and based on consistency index and preset weight coefficient, consistency score is calculated to obtain effective state information set by screening state information with consistency score greater than threshold value;

[0153] Save the state data of the effective state information set in the minimum recovery range, generate state change log based on the state data, and calculate state change cost to obtain the state set to be synchronized;

[0154] Calculate the importance index and urgency index of the state set to be synchronized through incremental synchronization method, generate state synchronization priority sequence, calculate the state necessity score of the state influence degree and recoverability in the state synchronization priority sequence, and mark the necessary execution state;

[0155] The batch recovery parameters are calculated according to the necessary execution state quantity and system load, the batched recovery is performed to obtain a recovery state result, the error rate, coverage rate and consistency rate of the recovery state result are calculated to generate a verification score, and the target instruction execution result is determined in combination with the gain factor and the penalty factor.

[0156] The distributed system is composed of multiple execution nodes, each node is responsible for a specific computing task, and a distributed snapshot technology is used to periodically collect state information of each execution node. For example, for each execution node, processor state, memory data, network connection information, task queue and other key data are recorded. For example, for node 1 processing task A, its CPU usage is 75%, memory usage is 2.3GB, the current transaction ID being processed is T10045, and the remaining task queue length is 27.

[0157] After the collection is completed, a consistency score is calculated based on a consistency index and a preset weight coefficient. The consistency index includes data integrity, timestamp consistency, logical sequence consistency, etc. Each state information is given a corresponding weight, such as a data integrity weight of 0.4, a timestamp consistency weight of 0.3, and a logical sequence consistency weight of 0.3. It is calculated that the consistency score of node 1 is 0.85, the consistency score of node 2 is 0.72, and the consistency score of node 3 is 0.91. The preset consistency threshold is 0.75, and the state information with a score greater than the threshold is selected to obtain the state information of node 1 and node 3 as the effective state information set.

[0158] The minimum recovery range is determined according to the dependency relationship analysis. For example, it is found through analysis that task T10045 of node 1 depends on the execution result of task T9032 of node 3, and both tasks are in the effective state information set. Save the state data, including task ID, execution progress, data block reference, etc. For node 1, the saved state data is {"taskId":"T10045","progress":82,"dataRefs":["block_437","block_438"],"dependencies":["T9032"]}。

[0159] Based on the saved state data, a state change log is generated to record the changes of the state from the last snapshot to the current snapshot. For example, the state change log of node 1 shows that the progress of task T10045 increases from 63% to 82%, and a reference to data block "block_438" is added. The state change cost is calculated, considering the changed data volume, change complexity and other factors, and the state change cost of node 1 is 45 units and the state change cost of node 3 is 30 units. These state information constitutes a to-be-synchronized state set.

[0160] The importance index and urgency index of each state in the set of states to be synchronized are calculated. The importance index is based on the degree of influence of the state on the overall system function, and the urgency index is based on the recovery time window requirement. For example, the importance index of task T10045 of node 1 is 0.9 (critical business task), and the urgency index is 0.8 (needs to be recovered within 5 minutes); the importance index of task T9032 of node 3 is 0.7, and the urgency index is 0.5. A state synchronization priority sequence is generated, with the states of node 1 at the front.

[0161] The influence degree and recoverability of each state in the state synchronization priority sequence are calculated. The influence degree represents the influence range of the state recovery failure on the system function, and the recoverability represents the probability of successful recovery under limited resource conditions. The influence degree of the state of node 1 is 0.85 (affects 3 downstream components), and the recoverability is 0.9 (in the case of sufficient resources); the influence degree of the state of node 3 is 0.6, and the recoverability is 0.95. The necessary score of the state is calculated by comprehensive calculation, for example, the necessary score of node 1 is 0.88, and the necessary score of node 3 is 0.72. The states with a necessary score greater than 0.8 are marked as necessary execution states, and in this example, the state of node 1 is marked as a necessary execution state.

[0162] Batch recovery parameters are calculated according to the number of necessary execution states and the current system load. The current system load is 65%, and the available recovery resources are 35% of the total resources, and it is decided to recover 1 necessary execution state per batch. Batch recovery is performed, first recovering the state of node 1 with a success rate of 96%, and then recovering the state of node 3 with a success rate of 98%.

[0163] After recovery is complete, the error rate, coverage rate, and consistency rate of the recovery state result are calculated. The error rate is the proportion of states that have errors in the recovery process, which in this example is 0%; the coverage rate is the proportion of necessary states that are successfully recovered, which in this example is 100%; the consistency rate is the degree of compliance of the logical relationship between the states after recovery, which in this example is 94%, and a verification score of 0.96 is generated.

[0164] The target instruction execution result is determined in combination with the gain factor and the penalty factor. The gain factor considers positive factors such as recovery efficiency and resource saving, and the penalty factor considers negative factors such as time delay and partial function limitation. In this embodiment, the gain factor is 1.05 (efficient recovery and 30% resource saving), and the penalty factor is 0.98 (2 seconds delay). The verification score is multiplied by the gain factor and then multiplied by the penalty factor to obtain a final score of 0.99, which determines that the recovery is successful, and the target instruction execution result is returned as the calculation output of task T10045 {"resultCode": 0, "data": {"itemCount": 1243, "totalValue": 89750.25}, "status": "completed"}.

[0165] In this embodiment, the state information of the execution node is collected by the distributed snapshot technology, and the effective state information is screened based on the consistency scoring mechanism, thereby realizing reliable collection and efficient filtering of the state information. The incremental synchronization mechanism is introduced, the state synchronization priority sequence is constructed in combination with the importance index and the urgency index, the necessary execution state is accurately marked by calculating the state influence degree and the recoverability, the accuracy and efficiency of the system state recovery are significantly improved, the reliability of the instruction execution result is ensured, and the overall performance of the cross-network interaction is improved by the fine state management and verification mechanism.

[0166] Figure 3 A system state recovery flowchart of the cross-network interaction method based on the voice instruction is provided in the embodiment of the application.

[0167] In a second aspect of the embodiment of the application, a cross-network interaction system based on a voice instruction is provided, which comprises:

[0168] A first unit is configured to receive a voice instruction sent by a user terminal, extract keyword information in the voice instruction to generate semantic understanding data, perform intent matching in a cross-network interaction database, and determine target network node information.

[0169] A second unit is configured to perform context association analysis on the semantic understanding data, extract scene feature information in a historical interaction session, identify an interaction scene currently experienced by the user according to the scene feature information, determine a target range of the cross-network interaction based on the interaction scene, divide the target range into a plurality of subtasks, set an execution priority for each subtask, and generate a target instruction set with a priority identifier.

[0170] A third unit is configured to establish a network channel between the user terminal and a target server based on the target network node information, and transmit the target instruction set to the target server.

[0171] A fourth unit is configured to deploy instruction execution state tracking in an edge node and a core node corresponding to the target network node, respectively, record data dependency relationships and intermediate calculation results in an instruction execution process, perform dependency analysis based on a topological sorting algorithm when detecting node execution abnormity, determine a minimum recovery range, save state information of each execution node in combination with a distributed snapshot technology, and obtain a target instruction execution result by restoring only necessary execution states in an incremental synchronization manner during fault recovery.

[0172] A fifth unit is configured to receive the target instruction execution result and feed back to the user terminal.

[0173] In a third aspect of the embodiment of the application, an electronic device is provided, which comprises:

[0174] a processor and a memory for storing processor-executable instructions, wherein the processor is configured to invoke the instructions stored by the memory to perform the method described above.

[0175] In a fourth aspect, the present application provides a computer readable storage medium, having stored thereon computer program instructions, which when executed by a processor implement the method described above.

[0176] The present application can be a method, an apparatus, a system, and / or a computer program product. The computer program product can include a computer readable storage medium having computer readable program instructions stored therein, which, when executed by a computer, perform various aspects of the present application.

[0177] It should be noted that the above-mentioned embodiments are merely used to illustrate the technical solutions of the present application, rather than limiting the present application; although the present application has been described in detail with reference to the above-mentioned embodiments, those skilled in the art should understand that the technical solutions recorded in the above-mentioned embodiments can be modified or equivalent replacements can be made to some or all of the technical features; and the modifications or replacements do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of the present application.

Claims

1. A method for cross-network interaction based on voice instructions, characterized in that, The method comprises the following steps: receiving a voice instruction sent by a user terminal, extracting keyword information in the voice instruction to generate semantic understanding data, and performing intent matching in a cross-network interaction database to determine target network node information; performing context association analysis on the semantic understanding data, extracting scene feature information in a historical interaction session, identifying an interaction scene in which the user is currently located according to the scene feature information, determining a target range of cross-network interaction based on the interaction scene, dividing the target range into multiple sub-tasks, setting an execution priority for each sub-task, and generating a target instruction set with a priority identifier, comprising: obtaining semantic understanding data, performing context association analysis on the semantic understanding data to obtain correlation degree calculation parameters, and constructing a time sequence correlation matrix based on the correlation degree calculation parameters; extracting feature data in a historical interaction session based on the time sequence correlation matrix using a sliding time window method, calculating a time decay weight based on the feature data, and combining the time decay weight with context information corresponding to the historical interaction session to obtain a scene feature vector; inputting the scene feature vector into a multi-layer perception machine for classification calculation to obtain a scene classification result, identifying an interaction scene in which the user is currently located based on the scene classification result, and matching the interaction scene with a predefined scene task mapping rule to obtain a target range of cross-network interaction; constructing a task dependency graph according to the target range, dividing the target range into multiple sub-tasks based on the task dependency graph, and calculating data dependency values and resource dependency values between the sub-tasks to obtain task dependency strengths; inputting the task dependency strengths and the task dependency graph into a preset priority calculation model to obtain task attribute values of each sub-task, setting an execution priority for each sub-task based on the task attribute values, and generating a target instruction set, which includes sub-task identifier information, priority identifier information, and task dependency relationship information; establishing a network channel between the user terminal and a target server based on the target network node information, and transmitting the target instruction set to the target server; deploying instruction execution state tracking on an edge node and a core node corresponding to the target network node, respectively, recording data dependency relationships and intermediate calculation results in the instruction execution process, performing dependency analysis based on a topological sorting algorithm when detecting node execution abnormalities, determining a minimum recovery range, saving state information of each execution node in combination with a distributed snapshot technology, and restoring only necessary execution states through incremental synchronization when recovering from a fault to obtain a target instruction execution result; receiving the target instruction execution result and feeding it back to the user terminal.

2. The method of claim 1, wherein, The method comprises the following steps: receiving a voice instruction sent by a user terminal, extracting keyword information in the voice instruction to generate semantic understanding data, and performing intent matching in a cross-network interaction database to determine target network node information, comprising: receiving a voice instruction sent by a user terminal, performing phoneme segmentation and feature vector extraction on the voice instruction, matching the extracted feature vector with a preset voice feature library, and identifying keyword information in the voice instruction; The keyword information is structured according to a preset semantic rule, a dependency relationship and a hierarchical relationship between keywords are extracted, a semantic tree structure is constructed, and semantic understanding data containing semantic association information is generated; The semantic understanding data is subjected to similarity calculation with a plurality of multi-dimensional intention feature vectors pre-stored in a cross-network interaction database, an intention feature vector with the highest matching degree is selected based on a maximum similarity principle, and target network node information associated with the intention feature vector is obtained.

3. The method of claim 1, wherein, According to the target range, a task dependency graph is constructed, the target range is divided into a plurality of sub-tasks based on the task dependency graph, and a data dependency value and a resource dependency value between the sub-tasks are calculated to obtain a task dependency strength, including: Performing dependency relationship analysis on the target range to obtain an initial dependency structure, dividing the target range into a plurality of sub-tasks based on the initial dependency structure, analyzing direct dependency relationships for the sub-tasks and establishing dependency relationship edges, and combining the sub-tasks and the dependency relationship edges to construct a complete task dependency graph; Task type information, execution time information, and resource demand information are extracted from the sub-tasks and combined to construct a feature vector, the feature vector is input into a graph neural network to obtain node representation results, message passing operations are performed on the node representation results to obtain sub-task state values, and the sub-task state values are input into a multi-head attention mechanism to calculate attention weights; According to the attention weights, the complete task dependency graph is structurally optimized to obtain an optimized dependency graph, a data set intersection over union value of the sub-tasks in the optimized dependency graph is calculated, and the data set intersection over union value is multiplied by a preset data dependency weight to obtain a data dependency value; The resource demand amount of the sub-tasks in the optimized dependency graph is obtained, a resource demand ratio between sub-tasks is calculated, the resource demand ratio is weighted and summed with a preset resource type weight to obtain a resource dependency value, and the data dependency value and the resource dependency value are weighted and combined to obtain a task dependency strength.

4. The method of claim 1, wherein, Based on the target network node information, a network channel between the user terminal and the target server is established, and the target instruction set is transmitted to the target server, including: Target network node information is obtained, communication quality parameters of each network node are calculated according to the target network node information, and network channel nodes to be established are determined as network nodes with communication quality parameters greater than a preset threshold; A communication path set is obtained by performing routing calculation on the network channel nodes to be established, a multi-path network channel between the user terminal and the target server is constructed based on the communication path set, and a primary and backup communication link is established in the multi-path network channel; The target instruction set is divided into a plurality of instruction subsets, the instruction subsets are distributed to the multi-path network channel according to a load balancing strategy based on the real-time load of the primary and backup communication link, and are transmitted in parallel to the target server.

5. The method of claim 1, wherein, The edge node and the core node corresponding to the target network node are respectively deployed with instruction execution state tracking, and the data dependency relationship and the intermediate calculation result in the instruction execution process are recorded, when the node execution exception is detected, the dependency analysis is carried out based on the topological sorting algorithm, and the minimum recovery range is determined to include: The edge node and the core node corresponding to the target network node are respectively deployed with instruction execution state tracking, and the edge node state vector is generated by collecting the instruction execution state, and the core node state matrix is generated by collecting the node association information; Based on the edge node state vector and the core node state matrix, the data dependency relationship and the intermediate calculation result in the instruction execution process are recorded to generate the instruction state descriptor, and the update time of the instruction state descriptor is adjusted according to the current load to obtain the real-time instruction state; According to the real-time instruction state, an instruction dependency graph is constructed, the instruction dependency strength between each instruction pair in the instruction dependency graph is calculated, and the weighted instruction dependency graph is obtained by taking the instruction dependency strength as the weight of the dependent edge; When the node execution exception is detected, the topological sorting algorithm is executed on the weighted instruction dependency graph to perform dependency analysis, the propagation path of the instruction is obtained, and the fault influence range is determined based on the propagation path, and the instructions in the fault influence range are determined as the minimum recovery range.

6. The method of claim 1, wherein, The state information of each execution node is saved in combination with the distributed snapshot technology, and only the necessary execution state is recovered through the incremental synchronization method during fault recovery to obtain the target instruction execution result, including: The state information of each execution node is collected in combination with the distributed snapshot technology to construct a state snapshot, a consistency score is calculated based on a consistency index and a preset weight coefficient, and the effective state information set is obtained by screening the state information with a consistency score greater than a threshold value; The state data of the effective state information set in the minimum recovery range is saved, the state change log is generated according to the state data, the state change cost is calculated to obtain the state set to be synchronized; The importance index and the urgency index of the state set to be synchronized are calculated through the incremental synchronization method, the state synchronization priority sequence is generated, the state necessity score is obtained by calculating the influence degree and the recoverability of each state in the state synchronization priority sequence, and the necessary execution state is marked; The batch recovery parameters are calculated according to the number of necessary execution states and the system load, the recovery state result is obtained by executing batch recovery, the error rate, the coverage rate and the consistency rate of the recovery state result are calculated to generate the verification score, and the target instruction execution result is determined in combination with the gain factor and the penalty factor.

7. A voice command based cross-network interaction system for implementing the method of any of the preceding claims 1-6, characterized by, It includes: The first unit is used for receiving the voice instruction sent by the user terminal, extracting the keyword information in the voice instruction to generate semantic understanding data, and performing intent matching in the cross-network interaction database to determine the target network node information; The second unit is used for context association analysis on the semantic understanding data, extracting scene feature information in the historical interaction session, identifying the current interaction scene of the user according to the scene feature information, determining the target range of cross-network interaction based on the interaction scene, dividing the target range into multiple sub-tasks, setting execution priority for each sub-task, and generating a target instruction set with priority identifier; a third unit configured to establish a network channel between the user terminal and a target server based on the target network node information, and transmit the target instruction set to the target server; a fourth unit configured to deploy instruction execution state tracking on an edge node and a core node corresponding to the target network node respectively, record data dependency and intermediate calculation results in the instruction execution process, when detecting node execution exception, perform dependency analysis based on a topological sorting algorithm, determine a minimum recovery range, save state information of each execution node in combination with a distributed snapshot technology, and when recovering from a fault, only recover necessary execution states through an incremental synchronization manner to obtain a target instruction execution result; a fifth unit configured to receive the target instruction execution result and feed back to the user terminal.

8. An electronic device, comprising: comprising: a processor; a memory for storing processor-executable instructions; wherein the processor is configured to invoke the instructions stored in the memory to execute the method of any one of claims 1 to 6.

9. A computer-readable storage medium having stored thereon computer program instructions, wherein, The computer program instructions, when executed by the processor, implement the method of any one of claims 1 to 6.

Images