A distributed artificial intelligence driven intelligent decision making method

Abstract

This invention discloses a distributed AI-driven intelligent decision-making method, relating to the field of AI technology. The method includes: utilizing global feature space information to drive a machine learning decision generation process, generating candidate strategies for each node; performing credibility assessment on the candidate strategies to form a distributed candidate set; performing cross-node collaborative constraint propagation on the distributed candidate set and continuously correcting the strategy consistency relationship between nodes to obtain a collaborative decision evolution sequence; performing stability judgment and strategy optimization reconstruction on the collaborative decision evolution sequence to obtain a globally stable decision strategy, and driving each node to execute the globally stable decision strategy, generating strategy feedback information. This invention, by performing cross-node collaborative constraint propagation on the distributed candidate set and continuously correcting the strategy consistency relationship between nodes, enables candidate strategies to gradually converge to a unified evolutionary trajectory during multiple rounds of interaction.

Description

Technical Field

[0001] This invention relates to the field of artificial intelligence technology, and in particular to a distributed artificial intelligence-driven intelligent decision-making method. Background Technology

[0002] With the development of distributed computing architecture and edge intelligence technology, the collaborative control of distributed nodes in complex scenarios is gradually shifting from rule-driven approaches to adaptive decision-making models centered on machine learning. Deploying local machine learning models on each distributed node to achieve real-time modeling and online updates of environmental and task status data has become a crucial technical path for enhancing the system's autonomous decision-making capabilities. Simultaneously, the continuous improvement of cross-node information interaction, feature alignment, and collaborative optimization mechanisms enables distributed networks to develop a certain degree of global perception and collaborative behavior in dynamic environments, driving distributed artificial intelligence technology towards higher levels of autonomous decision-making.

[0003] However, under conditions of heterogeneous data sources from multiple nodes and temporal drift, existing technologies struggle to maintain the online update capability of machine learning models while achieving consistent alignment of feature spaces across nodes and collaborative decision-making convergence. This leads to deviations between local learning results and global decision objectives, thereby affecting the stability and executability of the overall strategy. Therefore, how to construct a collaborative decision-making mechanism in a distributed network that combines online learning capabilities with global consistency constraints has become a key technical challenge that urgently needs to be addressed in the field of distributed artificial intelligence. Summary of the Invention

[0004] In view of the aforementioned existing problems, the present invention is proposed.

[0005] Therefore, this invention provides a distributed artificial intelligence-driven intelligent decision-making method to solve the problem of difficulty in unifying global consistency and policy stability in the distributed machine learning decision-making process.

[0006] To solve the above-mentioned technical problems, the present invention provides the following technical solution: This invention provides a distributed AI-driven intelligent decision-making method, comprising: collecting environmental state data and task state data from each distributed node, and performing unified time identification, node identification, and semantic normalization to form a node state-aware information set; employing a local machine learning model to perform online representation learning on the state trajectories in the node state-aware information set to form a node feature representation set; performing temporal offset correction on the node feature representation set to obtain corrected feature information, and performing consistent interactive alignment of the corrected feature information within the distributed network to generate global feature space information; using the global feature space information to drive the machine learning decision generation process, generating decision candidate strategies for each node, performing credibility evaluation on the decision candidate strategies for each node to form a distributed decision candidate set; performing cross-node collaborative constraint propagation on the distributed decision candidate set, and continuously correcting the strategy consistency relationship between nodes to obtain a collaborative decision evolution sequence; performing stability discrimination and strategy optimization reconstruction on the collaborative decision evolution sequence to obtain a globally stable decision strategy, and driving each node to execute the globally stable decision strategy to generate strategy feedback information.

[0007] As a preferred embodiment of the distributed artificial intelligence-driven intelligent decision-making method of the present invention, the steps for forming a node state perception information set are as follows: The environmental status data and task status data of each distributed node are concatenated and time offset correction is performed to form unified time series status data. Embedding node identity codes into unified time-series state data forms enhanced state data; The enhanced state data is filtered based on its contribution to state observability to form regularized state data; The data of regular state is compressed to reduce information redundancy and then aggregated to form a set of node state perception information.

[0008] As a preferred embodiment of the distributed artificial intelligence-driven intelligent decision-making method of the present invention, the steps for forming the node feature representation set are as follows: The node state-aware information set is constructed into state trajectory fragments in a unified time order to form an online learning sequence. The online learning sequence is then standardized and temporally smoothed to form a learnable input sequence. The learnable input sequence is input into the local temporal encoder to perform temporal feature extraction, and the extracted temporal features are used to perform state prediction for the next time step, and the state prediction bias is calculated. The state prediction bias is used to drive the local temporal encoder to perform self-supervised online updates and obtain the updated temporal features; The updated temporal features are bound to the corresponding state prediction bias to generate a set of node feature representations.

[0009] As a preferred embodiment of the distributed artificial intelligence-driven intelligent decision-making method of the present invention, the steps for obtaining the corrected feature information are as follows: The node feature representation set is rearranged according to the generation time of each node under a unified time axis, and a continuous time index is established to form a sparse feature sequence. The sparse feature sequence is converted into aligned event time points using a monotonic time mapping mechanism and then reconstructed into a continuous feature trajectory. Perform consistent interactive alignment on continuous feature trajectories within the node neighborhood to generate corrected feature information.

[0010] As a preferred embodiment of the distributed artificial intelligence-driven intelligent decision-making method of the present invention, the steps for generating global feature space information are as follows: The corrected feature information is used to construct a semantic prototype set in each node, and a consistent exchange summary is extracted to form an exchangeable feature structure. The exchangeable feature structure is randomly propagated among adjacent nodes through a ring exchange mechanism to form an interactive ledger; Based on the interactive ledger, cross-node matching and alignment are performed on the semantic prototypes in the semantic prototype set, and the cross-node matching relationships between semantic prototypes are continuously updated until convergence, generating global feature space information.

[0011] As a preferred embodiment of the distributed artificial intelligence-driven intelligent decision-making method of the present invention, the steps for generating decision candidate strategies for each node are as follows: The global feature space information is transformed into decision condition constraints within each node, forming the input for decision generation; The decision-making input is propagated through circular random interactions between adjacent nodes, forming interactive constraints; Under interactive constraints, condition-driven sampling is performed on the local machine learning model on the node side to form a candidate policy sequence, and the candidate policy sequence is organized into decision candidate policies for each node according to the node identifier.

[0012] As a preferred embodiment of the distributed artificial intelligence-driven intelligent decision-making method of the present invention, the steps of forming a distributed decision candidate set are as follows: Shadow playback is performed on the decision candidate strategies of each node to form a non-consistent evidence sequence; For inconsistent evidence sequences, the credibility of adjacent nodes is calibrated to form a credibility assessment threshold; Based on the inconsistent evidence sequence, the credibility of the decision candidate strategy is calculated, and a consistency audit and screening are performed on the decision candidate strategy according to the credibility assessment threshold to form a distributed decision candidate set.

[0013] As a preferred embodiment of the distributed artificial intelligence-driven intelligent decision-making method of the present invention, the steps for obtaining the collaborative decision-making evolution sequence are as follows: Based on the cross-node matching relationships in the global feature space information, the distributed decision candidate set is constructed into a candidate policy trajectory set with consistency constraint labels; Constraint propagation is performed on the candidate strategy trajectory set through asynchronous interaction between adjacent nodes, generating consistency deviation relationships between nodes; Based on the consistency deviation relationship, the candidate strategy trajectories in the candidate strategy trajectory set are iteratively corrected, and the process is recursively applied according to the interaction rounds to generate a collaborative decision evolution sequence.

[0014] As a preferred embodiment of the distributed artificial intelligence-driven intelligent decision-making method of the present invention, the steps for obtaining a globally stable decision-making strategy are as follows: Windowing is performed on the collaborative decision-making evolution sequence to form cross-round differential trajectories. Based on the cross-round differential trajectories, the stable intervals in the collaborative decision-making evolution sequence are determined, and the strategy trajectory of the stable phase is formed. Based on the strategy trajectory in the stable phase, a change trend sequence is constructed to determine the convergence direction. Then, the near-end projection reconstruction of the strategy trajectory in the stable phase is performed along the convergence direction to generate a globally stable decision strategy.

[0015] As a preferred embodiment of the distributed artificial intelligence-driven intelligent decision-making method of the present invention, the steps of driving each node to execute a globally stable decision-making strategy and generating strategy feedback information are as follows: The global stable decision-making strategy is encapsulated as a strategy execution transaction and committed consistently across all nodes to form an effective strategy version. Under the effective policy version, drive each node to execute the global stable decision policy, record the state change information and policy response information during the execution process, and generate policy feedback information.

[0016] The beneficial effects of this invention are as follows: by performing cross-node collaborative constraint propagation on a distributed set of decision candidates and continuously correcting the policy consistency relationship between nodes, the candidate policies gradually tend towards a unified evolutionary trajectory during multiple rounds of interaction. Thus, a global consistency constraint structure is embedded in the policy formation stage. A cross-node constraint propagation link is constructed on the basis of the local policies generated by machine learning, so that the deviations between policies can be corrected in an orderly manner during the evolution stage, forming a collaborative decision evolution sequence with convergence characteristics. This provides a structured basis for stability judgment and optimization reconstruction, and effectively improves the executability and continuous adaptability of the globally stable decision policy. Attached Figure Description

[0017] To more clearly illustrate the technical solutions of the embodiments of the present invention, the drawings used in the following description of the embodiments will be briefly introduced. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0018] Figure 1 A flowchart for a distributed artificial intelligence-driven intelligent decision-making method.

[0019] Figure 2 A flowchart for forming a set of node state awareness information.

[0020] Figure 3 A flowchart for generating global feature space information.

[0021] Figure 4 A flowchart for generating a collaborative decision-making evolution sequence. Detailed Implementation

[0022] To make the above-mentioned objects, features and advantages of the present invention more apparent and understandable, the specific embodiments of the present invention will be described in detail below with reference to the accompanying drawings.

[0023] Many specific details are set forth in the following description in order to provide a full understanding of the invention. However, the invention may also be practiced in other ways different from those described herein, and those skilled in the art can make similar extensions without departing from the spirit of the invention. Therefore, the invention is not limited to the specific embodiments disclosed below.

[0024] Secondly, the term "one embodiment" or "embodiment" as used herein refers to a specific feature, structure, or characteristic that may be included in at least one implementation of the present invention. The phrase "in one embodiment" appearing in different places in this specification does not necessarily refer to the same embodiment, nor is it a single or selective embodiment that is mutually exclusive with other embodiments.

[0025] Reference Figures 1-4 As one embodiment of the present invention, this embodiment provides a distributed artificial intelligence-driven intelligent decision-making method, comprising the following steps: S1. Collect environmental status data and task status data of each distributed node, and perform unified time identification, node identification and semantic normalization to form a set of node status perception information. S1.1: Combine the environmental status data and task status data of each distributed node, and perform time offset correction to form unified time series status data; Specifically, when collecting environmental and task status data from each distributed node, a time stamp corresponding to the collection time and a distributed node identifier from which the data originated are synchronously added to each piece of environmental and task status data. The environmental and task status data are then mapped one-to-one according to the time stamps and concatenated to form an initial state data sequence. Based on the time stamps, time offset correction is performed on the correspondence between environmental and task status data in the initial state data sequence to correct any time differences, so that the environmental and task status data from each distributed node are arranged continuously under the time stamps, resulting in a unified time series state data.

[0026] It should be noted that environmental status data refers to monitoring information obtained during the operation of each distributed node, reflecting changes in the external environment and its own operating environment. This includes information on changes in the environmental conditions of the node and information on the operating status of the node as a result of environmental influences. Task status data refers to the recorded information obtained during the execution of tasks by various distributed nodes, reflecting the progress and status of the tasks, including information on the task execution stages and status changes during the task execution process.

[0027] S1.2: Embed the unified time series state data into the node identity code to form enhanced state data; Specifically, based on the distributed node identifiers corresponding to the environmental status data and the task status data, the unified time series status data is organized and assigned, and a structured binding relationship is established between the distributed node identifiers and the corresponding unified time series status data, so that the unified time series status data forms enhanced status data with node assignment relationships while maintaining the original time identifiers.

[0028] S1.3: Perform filtering on the enhanced state data based on the contribution of state observability to form regularized state data; Specifically, in the enhanced state data, the monitoring information in the environmental state data is compared with the task execution record information in the task state data one by one according to the same time identifier. The monitoring information that can reflect the changes in task execution under the same time identifier is identified as the operation measurement item, and the task execution record information that records the progress of task execution is identified as the execution state item. The operation measurement items are matched with the execution state items under the same time identifier one by one to confirm the observable correlation between the operation measurement items and the changes in execution state. The operation measurement items that cannot maintain a correspondence under two or more adjacent time identifiers are marked as low contribution candidates and removed. The operation measurement items that can maintain a correspondence with the execution state items under two or more adjacent time identifiers are marked as high contribution candidates and retained, thus forming the regularized state data.

[0029] S1.4: Perform information redundancy compression on the regular state data and aggregate it to form a set of node state perception information.

[0030] Specifically, under the same distributed node identifier, the environmental state data and task state data in the regularized state data are compared one by one according to the order of the time identifiers. When the same environmental state data and the same task state data exist under adjacent time identifiers and correspond to the same task execution stage, they are considered to appear continuously and are retained. When a certain environmental state data and a certain task state data appear only under a single time identifier and do not correspond to the same task execution stage again in adjacent time identifiers, they are considered to appear briefly and are removed. This ensures that the retained environmental state data and task state data maintain a correspondence between adjacent time identifiers and are collected and organized according to the distributed node identifiers to form a set of node state perception information.

[0031] S2. Use a local machine learning model to perform online representation learning on the state trajectories in the node state-aware information set to form a node feature representation set.

[0032] S2.1: Construct the node state-aware information set into state trajectory segments according to a unified time order to form an online learning sequence, and perform standardization and temporal smoothing on the online learning sequence to form a learnable input sequence; Specifically, the node state awareness information set is arranged according to the chronological order of time markers, and environmental state data and task state data corresponding to adjacent time markers are continuously associated under the same distributed node identifier to form state trajectory segments. The state trajectory segments are then sequentially connected according to the chronological order of time markers to form an online learning sequence. In the online learning sequence, the arrangement order of environmental state data and task state data at each time marker position is uniformly aligned using the same distributed node identifier as a unit, so that the environmental state data and task state data maintain a fixed correspondence at each time marker position. At the same time, for positions in the online learning sequence where there are missing or abrupt records under adjacent time markers, existing environmental state data and task state data under adjacent time markers are used to continue and fill in the gaps, so that the online learning sequence maintains continuous connection in the time marker dimension, forming a learnable input sequence.

[0033] S2.2: Input the learnable input sequence into the local temporal encoder to perform temporal feature extraction, and perform state prediction for the next time step on the extracted temporal features, and calculate the state prediction bias; Specifically, the learnable input sequence is fed into the local temporal encoder in chronological order according to the time markers, and the temporal features at the corresponding time marker positions are extracted. At each time marker position, the corresponding environmental state data and task state data are combined in a fixed order to form a unified state vector with a fixed structure, and the length of the unified state vector formed by concatenating the environmental state data and task state data is recorded. Under the same time sequence, the temporal features are matched one by one with the environmental state data and task state data corresponding to the next time marker position, and the state prediction deviation is calculated.

[0034] The expression for calculating the state prediction deviation is: ; ; in, The time identifier is The predicted value of the unified state vector at time t; The time identifier is The unified state vector; This represents the continuity coefficient of the change in the unified state vector under adjacent time markers; The time identifier is The unified state vector; The time identifier is The deviation in state prediction at any given moment; Indicates the length of the uniform state vector; The time identifier is The unified state vector at time t; Indicates the current time and location; It should be noted that the local temporal encoder takes minimizing the prediction bias of the unified state vector at adjacent time marker positions as its self-supervised objective. It uses the actual unified state vector at the next time marker position as the supervision signal. By constructing the sum of squared differences between the predicted unified state vector and the actual unified state vector as the loss function, the state prediction bias gradually converges during the time progression. Thus, it achieves self-supervised online learning of the temporal evolution relationship between environmental state data and task state data without the need for external annotation. A fixed arrangement order refers to arranging the monitoring values ​​in the environmental status data according to priority at each time marker, and then arranging the execution values ​​in the task status data, and maintaining the same arrangement position at all time markers without changing. The continuity coefficient is determined based on whether the uniform state vectors corresponding to adjacent time markers in the unified time series state data maintain the same direction of change as time progresses. The specific steps include selecting the uniform state vectors corresponding to three consecutive time marker positions in the unified time series state data, and comparing the consistency of the change directions of the two segments. When the change directions of the two segments are consistent, the continuity coefficient is determined to be in the higher range; when the change directions of the two segments are inconsistent, the continuity coefficient is determined to be in the lower range. The exemplary value range is 0 to 1. When the change of the uniform state vector is continuous and stable under adjacent time markers, λ can be in the higher range of 0.6 to 0.9; when the state change is unstable, λ can be in the lower range of 0 to 0.6.

[0035] S2.3: Utilize state prediction bias to drive the local temporal encoder to perform self-supervised online updates and obtain the updated temporal features; Specifically, the state prediction deviation is correlated with the temporal features under the corresponding time marker, and the temporal feature extraction process of the local time encoder at the subsequent time marker position is synchronously adjusted according to the change of the state prediction deviation in adjacent time markers, so that the local time encoder gradually approaches the actual unified state vector change as time progresses, and the updated temporal features are obtained.

[0036] It should be noted that the local time sequence encoder is a machine learning model deployed on each distributed node side to represent the time sequence of the learnable input sequence formed by the set of node state awareness information. The local time sequence encoder uses environmental state data and task state data arranged in a uniform time order as learning objects. Through continuous learning of the relationship between state changes during the time process, it forms stable time sequence features. The local time sequence encoder needs to continuously receive state prediction deviations and make online adjustments to maintain a close fit to the actual state changes.

[0037] S2.4: Bind the updated temporal features to the corresponding state prediction bias to generate a set of node feature representations.

[0038] Specifically, the updated time series features are associated and organized according to the corresponding time identifier and the state prediction deviation formed under the same time identifier, so that the updated time series features and the state prediction deviation form a corresponding relationship under the same time identifier, forming a node feature representation; the node feature representations are aggregated according to the distributed node identifier to generate a node feature representation set.

[0039] S3. Perform temporal offset correction on the node feature representation set to obtain the corrected feature information, and perform consistent interactive alignment of the corrected feature information in the distributed network to generate global feature space information.

[0040] S3.1: Rearrange the node feature representation set according to the generation time of each node under a unified time axis, and establish a continuous time index to form a sparse feature sequence; Specifically, the set of node feature representations is arranged on a unified time axis according to time markers, and the node feature representations corresponding to different distributed nodes at each time marker position are rearranged so that the updated time-series features with corresponding relationships and the state prediction deviations form a sequential correspondence on the unified time axis. At the same time, the time marker positions with node feature representations are sequentially labeled on the unified time axis to establish a continuous time index and form a sparse feature sequence.

[0041] S3.2: Use the monotonic time mapping mechanism to convert sparse feature sequences into aligned event time points and reconstruct them into continuous feature trajectories; Specifically, according to the chronological order of each time marker in the unified time axis, the time marker positions of node feature representations in the sparse feature sequence are sequentially mapped while maintaining the time progression relationship. This allows the updated temporal features with corresponding relationships and the state prediction deviation to be mapped to continuous time positions in the unified time axis, forming aligned event time points. Based on the time progression order of the event time points, the node feature representations are connected to form a continuous feature trajectory.

[0042] S3.3: Perform consistent interactive alignment on continuous feature trajectories within the node neighborhood to generate corrected feature information.

[0043] Specifically, following the time marker order of a unified time axis, the continuous feature trajectories corresponding to the same time marker position of each distributed node are compared time-by-time within the neighborhood of the distributed nodes. At each time marker position, only continuous feature trajectories with the same time marker position are allowed to participate in the alignment adjustment. Meanwhile, time marker positions of continuous feature trajectories that are outside the preset alignment time window are excluded from the alignment adjustment, forming an alignment correspondence at the unified time marker position. This ensures that the continuous feature trajectories participating in the alignment adjustment within the node neighborhood are always limited to the same time marker position range of the unified time axis, generating corrected feature information.

[0044] It should be noted that the alignment time window range is based on the interval between adjacent time markers on a unified time axis. A fixed number (the number is an integer of not less than two) of consecutive time marker intervals are selected forward and backward around a certain time marker on the unified time axis to form the preset alignment time window range.

[0045] S3.4: Construct a semantic prototype set within each node using the corrected feature information, and extract a consistent exchange summary to form an exchangeable feature structure; Specifically, within each distributed node, the corrected feature information is collected one by one according to the time identifier order of the unified time axis, and the alignment correspondence under the same time identifier position is used as the classification basis, so that the corrected feature information with alignment correspondence under the same time identifier position is grouped into the same semantic prototype set; the corrected feature information that maintains alignment correspondence under the same time identifier position is extracted from the semantic prototype set as a consistency exchange summary, and the consistency exchange summary and the semantic prototype set are encapsulated together to form an exchangeable feature structure.

[0046] S3.5: The exchangeable feature structure is randomly propagated among adjacent nodes through a ring exchange mechanism to form an interactive ledger; Specifically, based on the adjacency relationships between distributed nodes, the exchangeable feature structures of each distributed node are sequentially transmitted along the closed-loop path formed by the adjacent nodes. In each round of transmission, the exchangeable feature structures are randomly selected according to the order of the adjacent nodes. The randomly selected exchangeable feature structures are then alternately transmitted among the distributed nodes, so that each distributed node gradually receives exchangeable feature structures from different adjacent nodes in multiple rounds of transmission. The receiving process is recorded sequentially according to time identifiers to form an interactive ledger.

[0047] S3.6: Perform cross-node matching alignment on the semantic prototypes in the semantic prototype set according to the interaction ledger, and continuously update the cross-node matching relationship between semantic prototypes until convergence, generating global feature space information.

[0048] Specifically, according to the transmission order of exchangeable feature structures in the interactive ledger among the distributed nodes, at the same time marker position, the corrected feature information with alignment correspondence in the semantic prototype set is associated accordingly. This establishes cross-node correspondence between corrected feature information from different distributed nodes that have transmission order records in the interactive ledger. As the exchangeable feature structure continues to be transmitted along adjacent nodes and received by each distributed node, the receiving process is continuously appended to the interactive ledger according to the time marker, forming a transmission order record. Based on the continuously appended transmission order records in the interactive ledger, the cross-node correspondence between the corrected feature information is synchronously adjusted until the cross-node correspondence between the corrected feature information remains consistent during the continuous appending process and no new or replaced correspondences are added, generating global feature space information.

[0049] It should be noted that the transmission order is the sequential transmission order formed by the exchangeable feature structure being transmitted node by node along the closed loop path formed by adjacent nodes under the ring exchange mechanism. That is, the time identifier sequence record corresponding to the transmission of a certain exchangeable feature structure from one distributed node to the next adjacent distributed node, and then from the adjacent distributed node to the next adjacent distributed node.

[0050] S4. Utilize global feature space information to drive the machine learning decision generation process, generate decision candidate strategies for each node, perform credibility evaluation on the decision candidate strategies of each node, and form a distributed decision candidate set. S4.1: Transform the global feature space information into decision condition constraints within each node to form the input for decision generation; Specifically, within each distributed node, the corrected feature information that has established cross-node correspondences in the global feature space information is grouped and organized according to the time identifier order of the unified time axis. This ensures that the corrected feature information with the same time identifier position in the unified time axis is grouped into the same time identifier group. Within each time identifier group, the corrected feature information corresponding to the same time identifier position in each time identifier group is arranged in order according to the distributed node identifiers. The corrected feature information corresponding to the same time identifier position in each time identifier group is then organized into constraint description content. This establishes a fixed correspondence between the constraint description content and the corresponding time identifier position, forming decision condition constraints. Finally, the decision condition constraints are arranged according to the time identifier order of the unified time axis to form the decision generation input.

[0051] S4.2: The decision generation input is propagated through circular random interactions between adjacent nodes to form interactive constraints; Specifically, based on the existing adjacency relationships between distributed nodes, the decision generation inputs held by each distributed node are sequentially transmitted along the closed-loop path formed by the adjacent nodes. In each round of transmission, the decision generation inputs are randomly selected within the range of adjacent nodes and transmitted sequentially among the distributed nodes along the closed-loop path. This allows each distributed node to gradually receive decision generation inputs from different adjacent nodes during multiple rounds of transmission. During the receiving process, the received decision generation inputs are recorded sequentially according to the time identifier of a unified time axis, forming interactive constraints within each distributed node that include constraints from multiple nodes.

[0052] It should be noted that the range of adjacent nodes is the set of distributed nodes that have a direct connection with a certain distributed node in the distributed network topology. That is, distributed nodes that can be reached in one step in the network connection relationship, excluding distributed nodes that need to be reached through intermediate nodes.

[0053] S4.3: Under interactive constraints, perform condition-driven sampling on the local machine learning model on the node side to form a candidate policy sequence, and organize the candidate policy sequence into decision candidate policies for each node according to the node identifier.

[0054] Specifically, within each distributed node, the interaction constraints and the decision generation inputs at corresponding time markers are sent to the Long Short-Term Memory (LSTM) network on the node side in a time marker order along a unified time axis. This allows the LTM network to read the decision generation input corresponding to the current time marker at each time marker. Simultaneously, the interaction constraints are included as additional constraints in the policy generation process. When generating policy content, the LTM network only outputs the policy content at the corresponding time marker within the interaction constraints. The policy content generated sequentially at each time marker is then connected and arranged in a unified time axis order to form a candidate policy sequence. The candidate policy sequences are then distinguished and organized according to the distributed node identifiers, so that the candidate policy sequences corresponding to each distributed node are aggregated to form the decision candidate policies for each node.

[0055] It should be noted that the training process of the Long Short-Term Memory (LSTM) network is as follows: The set of node state-aware information is organized into a continuous training sequence according to the time marker order of a unified time axis. The training sequence is then used to establish a supervised correspondence between the state changes of the current time marker position and the next time marker position, forming training samples. The training samples are fed into the LSM network and the LSM network parameters are initialized, making each LSM network parameter trainable. During the training process, the training samples are iterated repeatedly in time order, allowing the LSM network to gradually establish a mapping relationship from the state at the current time marker position to the state at the next time marker position. After each round of training, the change in the state prediction deviation corresponding to the training samples is recorded. When the change in the state prediction deviation remains stable and no longer decreases during two or more consecutive rounds of training, the training is considered to have reached a stable state, and the training process is stopped, resulting in a trained LSM network.

[0056] S4.4: Perform shadow playback on the decision candidate strategies of each node to form a non-consistent evidence sequence; Specifically, within each distributed node, following the time-marked order of a unified timeline, each decision candidate strategy is mapped one by one to the node state-aware information set and global feature space information at the same time-marked position, without actual execution. The strategy content in the decision candidate strategy is then compared item by item with the operating state reflected in the node state-aware information set at the corresponding time-marked position and the cross-node consistency relationship determined by the global feature space information. When the strategy content contained in the decision candidate strategy exceeds the allowable state range reflected in the node state-aware information set and is inconsistent with the cross-node consistency relationship corresponding to the time-marked position in the global feature space information, the corresponding situation at the time-marked position is recorded as an inconsistency, and the inconsistencies are arranged in the unified timeline order to form a sequence of inconsistent evidence.

[0057] It should be noted that the allowed state range is the range of executable actions corresponding to the current state of the environmental state data and task state data reflected by the node state awareness information set at the corresponding time marker position. For example, when the node state awareness information set shows that a certain time marker position is in a low-load operation state, only the strategy type that matches the low-load operation state is allowed to be executed, while high-load and emergency adjustment strategies are not allowed to be executed.

[0058] S4.5: Perform credibility calibration between adjacent nodes for inconsistent evidence sequences to form a credibility assessment threshold; Specifically, based on the existing adjacency relationships between distributed nodes, the sequence of inconsistent evidence recorded by each distributed node is passed round by round along the closed-loop path formed by the adjacent nodes. This allows each distributed node to compare inconsistent evidence from different adjacent nodes at the same time marker position, and count the number of distributed nodes that record the same inconsistent event at the same time marker position. The proportion of the number of distributed nodes to the total number of distributed nodes participating in the interaction is used as the credibility assessment threshold.

[0059] It should be noted that the credibility assessment threshold is set based on the number of distributed nodes and the coverage of interactions between neighboring nodes. The total number of distributed nodes participating in the same time-identified location interaction is used as the benchmark, and the proportion of distributed nodes recording the same inconsistency is used as the judgment criterion. The critical range is determined by combining the recurrence of similar inconsistencies among most distributed nodes in the historical interaction process. An exemplary value range is between 30% and 70%. When the proportion of distributed nodes recording the same inconsistency is less than 30%, it indicates that the inconsistency only occurs in a few distributed nodes and belongs to local differences. When the proportion is higher than 70%, it indicates that there is a risk of cross-node consistency in the decision candidate strategy.

[0060] S4.6: Calculate the credibility of decision candidate strategies based on the inconsistent evidence sequence, and perform consistency audit and screening on the decision candidate strategies according to the credibility assessment threshold to form a distributed decision candidate set.

[0061] Specifically, within each distributed node, following a unified timeline, the number of times inconsistent evidence appears at each time marker position of the decision candidate strategy is proportionally converted to the total number of distributed nodes participating in the circular random interaction between adjacent nodes. The credibility of the decision candidate strategy is then calculated, and the credibility of each time marker position is summarized according to the unified timeline to form the overall credibility of the decision candidate strategy. The overall credibility is compared with the credibility assessment threshold. If the overall credibility is within the range allowed by the credibility assessment threshold, the decision candidate strategy is retained. If the overall credibility is higher than the upper limit of the credibility assessment threshold, it is determined to have cross-node consistency risk and is eliminated. If the overall credibility is lower than the lower limit of the credibility assessment threshold, it is determined to be a local difference and is retained, thus forming a distributed decision candidate set.

[0062] The expression for calculating the credibility of candidate strategies is: ; in, This represents the credibility of the candidate decision strategy, with a value range of [0, 1]. This indicates the total number of time markers on the unified timeline. Indicates the first The number of distributed nodes that record the inconsistencies corresponding to the decision candidate strategies at each time marker position; This represents the total number of distributed nodes participating in the circular random interaction between adjacent nodes.

[0063] S5. Perform cross-node collaborative constraint propagation on the distributed decision candidate set and continuously correct the policy consistency relationship between nodes to obtain the collaborative decision evolution sequence. S5.1: Based on the cross-node matching relationships in the global feature space information, construct the distributed decision candidate set into a candidate strategy trajectory set with consistency constraint labels; Specifically, following the time marker order of a unified time axis, the decision candidate strategies corresponding to each distributed node in the distributed decision candidate set are arranged side-by-side at the same time marker position. The cross-node correspondence in the global feature space information is used as a consistency reference framework to compare the decision candidate strategies of each distributed node at the time marker position. Decision candidate strategies that satisfy the cross-node correspondence are marked as consistency satisfied, and decision candidate strategies that do not satisfy the cross-node correspondence are marked as consistency deviated. The marked content of each time marker position is arranged according to the unified time axis order to form a set of candidate strategy trajectories with consistency constraint labels.

[0064] S5.2: Constraint propagation is performed on the candidate strategy trajectory set through asynchronous interaction between adjacent nodes to generate consistency deviation relationships between nodes; Specifically, asynchronous exchange is performed according to the adjacency relationship between distributed nodes, so that adjacent nodes compare their respective candidate policy trajectory sets one by one at the same time identifier position. When two distributed nodes have differences in policy content, execution order and consistency constraint label in the candidate policy trajectory corresponding to the same time identifier position, the differences are recorded as the consistency deviation relationship between the nodes.

[0065] S5.3: Perform iterative correction on the candidate strategy trajectories in the candidate strategy trajectory set according to the consistency deviation relationship, and recursively generate a collaborative decision evolution sequence according to the interaction rounds.

[0066] Specifically, following the time marker order of a unified time axis, each candidate strategy trajectory in the candidate strategy trajectory set that has a consistency deviation relationship at the same time marker position is compared. The candidate strategy trajectory with the same strategy content and consistency constraint label at the corresponding time marker position and the most frequent occurrence is used as the adjustment direction to correct the candidate strategy trajectory with consistency deviation relationship, so that the distributed nodes gradually become consistent at the corresponding time marker positions. The corrected candidate strategy trajectory continues to participate in the next round of interaction, so that the candidate strategy trajectory changes progressively with the interaction rounds and is arranged in the order of the unified time axis, generating a collaborative decision evolution sequence.

[0067] S6. Perform stability judgment and policy optimization and reconstruction on the collaborative decision-making evolution sequence to obtain a globally stable decision-making policy, and drive each node to execute the globally stable decision-making policy to generate policy feedback information.

[0068] S6.1: Perform windowing processing on the collaborative decision-making evolution sequence to form cross-round differential trajectories, and determine the stable interval in the collaborative decision-making evolution sequence based on the cross-round differential trajectories to form the strategy trajectory of the stable stage; Specifically, following the time marker order of a unified time axis, the collaborative decision-making evolution sequence is divided into multiple adjacent intervals based on a fixed number of adjacent time markers. Each interval forms a sliding window, and candidate strategy trajectories at the same time marker position are compared one by one between adjacent sliding windows. The changes in the strategy content and consistency constraint labels of candidate strategy trajectories at the same time marker position between the previous and subsequent sliding windows are recorded in time marker order to form cross-round differential trajectories. Continuous time marker intervals in the cross-round differential trajectories that remain unchanged are observed along the unified time axis. When the candidate strategy trajectories at the same time marker position do not change in at least three consecutive sliding windows, the corresponding continuous time marker interval is determined as a stable interval, and the candidate strategy trajectories within the stable interval are aggregated to form the strategy trajectory of the stable stage.

[0069] S6.2: Construct a trend sequence based on the strategy trajectory in the stable phase, determine the convergence direction, and perform near-end projection reconstruction on the strategy trajectory in the stable phase along the convergence direction to generate a globally stable decision strategy.

[0070] Specifically, following the time marker order of a unified time axis, the changes in policy content and consistency constraint labels at each time marker position of the stable phase are arranged sequentially to form a trend sequence. When the same policy content and consistency constraint label appear in at least three consecutive time marker positions at the end of the stable interval, the repeated policy content is determined as the convergence direction. The policy trajectory of the stable phase is uniformly corrected at the corresponding time marker positions along the convergence direction, so that the policy content and consistency constraint label at each time marker position are concentrated in the convergence direction, generating a globally stable decision policy.

[0071] S6.3: Encapsulate the global stable decision-making strategy into a strategy execution transaction and commit it consistently across all nodes to form an effective strategy version; Specifically, the globally stable decision-making strategy is encapsulated according to the time identifier order of a unified timeline and the distributed node identifiers. The strategy content corresponding to each time identifier position, along with the consistency constraint label, is written into the strategy execution transaction, and a version identifier and an effective time identifier are attached to the strategy execution transaction. Based on the existing adjacency relationships between distributed nodes, the strategy execution transaction is sequentially transmitted to each distributed node. When each distributed node receives a strategy execution transaction, it compares the time identifier in the strategy execution transaction with the time identifiers of its current unified timeline, and checks the version identifier in the strategy execution transaction against its own version identifier. When the time identifiers are consecutive and the version identifier is higher than the version identifier of the current node, the strategy execution transaction is recorded as pending effectiveness. When all participating distributed nodes record the same time identifier and version identifier, the strategy execution transaction is uniformly marked as effective, forming an effective strategy version.

[0072] S6.4: Drive each node to execute the global stable decision-making strategy under the effective strategy version, and record the state change information and strategy response information during the execution process to generate strategy feedback information.

[0073] Specifically, each distributed node executes the corresponding policy content of the global stability decision policy one by one according to the time identifier in the effective policy version. This ensures that the global stability decision policy is implemented sequentially at the current time identifier position of each distributed node. During the execution, the environmental state data and task state data corresponding to the current time identifier position are recorded synchronously to form state change information. At the same time, the control command triggering status, task progress stage changes, and resource usage status changes caused by the execution of the global stability decision policy at each time identifier position are recorded sequentially to form policy response information. The state change information and policy response information under the same time identifier are assigned and organized according to the distributed node identifier to generate policy feedback information.

[0074] In summary, this invention achieves this by: performing cross-node collaborative constraint propagation on a distributed set of decision candidates and continuously correcting the policy consistency relationship between nodes, so that candidate policies gradually tend towards a unified evolutionary trajectory during multiple rounds of interaction. This embeds a globally consistent constraint structure at the policy formation stage, and constructs a cross-node constraint propagation link based on the local policies generated by machine learning, enabling deviations between policies to be corrected in an orderly manner during the evolution stage, forming a collaborative decision evolution sequence with convergence characteristics. This provides a structured foundation for stability judgment and optimization reconstruction, effectively improving the executability and continuous adaptability of globally stable decision policies.

[0075] It should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit it. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical solutions of the present invention without departing from the spirit and scope of the technical solutions of the present invention, and all such modifications or substitutions should be covered within the scope of the claims of the present invention.

Claims

1. A distributed artificial intelligence-driven intelligent decision-making method, characterized in that, include: Collect environmental status data and task status data from each distributed node, and perform unified time identification, node identification and semantic normalization to form a set of node status awareness information. A local machine learning model is used to perform online representation learning on the state trajectories in the node state-aware information set, forming a set of node feature representations. Temporal offset correction is performed on the node feature representation set to obtain the corrected feature information. The corrected feature information is then aligned in a consistent manner within the distributed network to generate global feature space information. The machine learning decision generation process is driven by global feature space information, generating decision candidate strategies for each node, and performing credibility evaluation on the decision candidate strategies of each node to form a distributed decision candidate set. Cross-node collaborative constraint propagation is performed on the distributed decision candidate set, and the policy consistency relationship between nodes is continuously corrected to obtain the collaborative decision evolution sequence; The stability judgment and policy optimization reconstruction are performed on the collaborative decision-making evolution sequence to obtain a globally stable decision-making policy, and each node is driven to execute the globally stable decision-making policy to generate policy feedback information.

2. The distributed artificial intelligence-driven intelligent decision-making method as described in claim 1, characterized in that, The steps for forming the node state awareness information set are as follows: The environmental status data and task status data of each distributed node are concatenated and time offset correction is performed to form unified time series status data. Embedding node identity codes into unified time-series state data forms enhanced state data; The enhanced state data is filtered based on its contribution to state observability to form regularized state data; The data of regular state is compressed to reduce information redundancy and then aggregated to form a set of node state perception information.

3. The distributed artificial intelligence-driven intelligent decision-making method as described in claim 2, characterized in that, The steps for forming the node feature representation set are as follows: The node state-aware information set is constructed into state trajectory fragments in a unified time order to form an online learning sequence. The online learning sequence is then standardized and temporally smoothed to form a learnable input sequence. The learnable input sequence is input into the local temporal encoder to perform temporal feature extraction, and the extracted temporal features are used to perform state prediction for the next time step, and the state prediction bias is calculated. The state prediction bias is used to drive the local temporal encoder to perform self-supervised online updates and obtain the updated temporal features; The updated temporal features are bound to the corresponding state prediction bias to generate a set of node feature representations.

4. The distributed artificial intelligence-driven intelligent decision-making method as described in claim 1 or 3, characterized in that, The steps to obtain the corrected feature information are as follows: The node feature representation set is rearranged according to the generation time of each node under a unified time axis, and a continuous time index is established to form a sparse feature sequence. The sparse feature sequence is converted into aligned event time points using a monotonic time mapping mechanism and then reconstructed into a continuous feature trajectory. Perform consistent interactive alignment on continuous feature trajectories within the node neighborhood to generate corrected feature information.

5. The distributed AI-driven intelligent decision-making method as described in claim 4, characterized in that, The steps for generating global feature space information are as follows: The corrected feature information is used to construct a semantic prototype set in each node, and a consistent exchange summary is extracted to form an exchangeable feature structure. The exchangeable feature structure is randomly propagated among adjacent nodes through a ring exchange mechanism to form an interactive ledger; Based on the interactive ledger, cross-node matching and alignment are performed on the semantic prototypes in the semantic prototype set, and the cross-node matching relationships between semantic prototypes are continuously updated until convergence, generating global feature space information.

6. The distributed artificial intelligence-driven intelligent decision-making method as described in claim 1 or 5, characterized in that, The steps for generating decision candidate strategies for each node are as follows: The global feature space information is transformed into decision condition constraints within each node, forming the input for decision generation; The decision-making input is propagated through circular random interactions between adjacent nodes, forming interactive constraints; Under interactive constraints, condition-driven sampling is performed on the local machine learning model on the node side to form a candidate policy sequence, and the candidate policy sequence is organized into decision candidate policies for each node according to the node identifier.

7. The distributed artificial intelligence-driven intelligent decision-making method as described in claim 6, characterized in that, The steps for forming a distributed decision candidate set are as follows: Shadow playback is performed on the decision candidate strategies of each node to form a non-consistent evidence sequence; For inconsistent evidence sequences, the credibility of adjacent nodes is calibrated to form a credibility assessment threshold; Based on the inconsistent evidence sequence, the credibility of the decision candidate strategy is calculated, and a consistency audit and screening are performed on the decision candidate strategy according to the credibility assessment threshold to form a distributed decision candidate set.

8. The distributed artificial intelligence-driven intelligent decision-making method as described in claim 5 or 7, characterized in that, The steps to obtain the collaborative decision-making evolution sequence are as follows: Based on the cross-node matching relationships in the global feature space information, the distributed decision candidate set is constructed into a candidate policy trajectory set with consistency constraint labels; Constraint propagation is performed on the candidate strategy trajectory set through asynchronous interaction between adjacent nodes, generating consistency deviation relationships between nodes; Based on the consistency deviation relationship, the candidate strategy trajectories in the candidate strategy trajectory set are iteratively corrected, and the process is recursively applied according to the interaction rounds to generate a collaborative decision evolution sequence.

9. The distributed artificial intelligence-driven intelligent decision-making method as described in claim 8, characterized in that, The steps to obtain a globally stable decision strategy are as follows: Windowing is performed on the collaborative decision-making evolution sequence to form cross-round differential trajectories. Based on the cross-round differential trajectories, the stable intervals in the collaborative decision-making evolution sequence are determined, and the strategy trajectory of the stable phase is formed. Based on the strategy trajectory in the stable phase, a change trend sequence is constructed to determine the convergence direction. Then, the near-end projection reconstruction of the strategy trajectory in the stable phase is performed along the convergence direction to generate a globally stable decision strategy.

10. The distributed artificial intelligence-driven intelligent decision-making method as described in claim 9, characterized in that, The steps for driving each node to execute a globally stable decision-making strategy and generating strategy feedback information are as follows: The global stable decision-making strategy is encapsulated as a strategy execution transaction and committed consistently across all nodes to form an effective strategy version. Under the effective policy version, drive each node to execute the global stable decision policy, record the state change information and policy response information during the execution process, and generate policy feedback information.

Images