Intelligent decision method and device for long-chain industrial task, equipment and medium

Abstract

The application relates to the technical field of industrial artificial intelligence, in particular to an intelligent decision-making method and device for long-chain industrial tasks, equipment and a medium, aiming to solve the problems of response lag caused by serial industrial knowledge retrieval and reasoning, easy breaking of long-chain task logic and insufficient reliability of output results in existing industrial complex task processing. The method comprises the following steps: determining a pre-fetching time and a corresponding pre-fetching quantity of next-round industrial knowledge through flow retrieval; determining an industrial knowledge slice queue to be picked up according to the pre-fetching time and the pre-fetching quantity; and generating an industrial scheduling decision sequence through dense vector retrieval and a Monte Carlo tree search decision maker based on a double-loop fusion mechanism of thought chains and decision chains. By using the above scheme, the real-time performance of industrial knowledge perception and reasoning cooperation can be improved, the continuity and consistency between long-chain task steps can be enhanced, and the reliability and executability of industrial decision results can be improved.

Description

Technical Field

[0001] This application relates to the field of industrial artificial intelligence technology, and in particular to an intelligent decision-making method, apparatus, equipment and medium for long-chain industrial tasks. Background Technology

[0002] In handling complex industrial tasks, a large model agent system based on static retrieval enhancement is typically used. This system acquires domain-specific industrial knowledge through vector retrieval and combines it with model reasoning to output results.

[0003] However, these solutions generally employ a sequential processing approach for retrieval and reasoning. As the scale of the industrial knowledge base increases, retrieval latency rises significantly, limiting the overall response speed. Furthermore, single-step reasoning struggles to maintain semantic consistency in long-chain tasks, easily leading to breaks in logical dependencies. Moreover, relying solely on shallow rule validation makes it difficult to guarantee the factual correctness and execution reliability of the output results.

[0004] Therefore, improving the real-time performance of industrial knowledge perception and reasoning response in complex industrial task scenarios, while ensuring the consistency of long-chain logic and the credibility of output, has become a technical problem that needs to be solved. (Industrial knowledge) Summary of the Invention

[0005] This application provides an intelligent decision-making method, apparatus, device, and medium for long-chain industrial tasks to solve the aforementioned technical problems. This solution addresses the need for cross-domain industrial knowledge collaborative computing in complex industrial tasks, coordinating the pace of industrial knowledge acquisition with the task reasoning process. It achieves dynamic coordination between industrial knowledge perception and reasoning decision-making during task execution, while also considering the logical coherence and reliability of results in multi-stage task processing, thereby improving the overall performance of collaborative computing in complex scenarios.

[0006] Firstly, this application provides an intelligent decision-making method for long-chain industrial tasks, the method comprising:

[0007] The timing and quantity of prefetching for the next round of industrial knowledge are determined through streaming retrieval.

[0008] The industrial knowledge slice retrieval queue is determined based on the pre-fetching timing and the pre-fetching quantity;

[0009] Based on the dual-loop fusion mechanism of thinking chain and decision chain, an industrial scheduling decision sequence is generated through dense vector retrieval and Monte Carlo tree search decision engine.

[0010] In one possible implementation, determining the timing and corresponding quantity of prefetching for the next round of industrial knowledge through streaming retrieval includes:

[0011] Determine the expected inference time for this round and the expected retrieval time for the next round;

[0012] The prefetch timing is calculated based on the expected inference time and the expected retrieval time.

[0013] The expected inference time and the expected retrieval time are input into the adaptive prefetch depth algorithm to determine the prefetch quantity.

[0014] In one possible implementation, calculating the prefetch timing based on the expected inference time and the expected retrieval time includes:

[0015] Obtain the actual reasoning time of the previous round and the corresponding historical expected reasoning time;

[0016] Calculate the feedback factor based on the actual reasoning time and the historical expected reasoning time;

[0017] Based on the Markov decision process, the prefetching timing is determined according to the expected inference time, the expected retrieval time, and the feedback factor.

[0018] In one possible implementation, the industrial scheduling decision sequence generated through a dual-loop fusion mechanism based on the thought chain and decision chain, using dense vector retrieval and a Monte Carlo tree search decision-maker, includes:

[0019] The inner loop is started to perform dense vector retrieval on the industrial knowledge slice queue to generate a sequence of inference nodes;

[0020] Start the outer loop and perform a Monte Carlo tree search based on the inference node sequence;

[0021] External tools are invoked to generate decision subsequences, and the execution results are fed back to update the reward values ​​of the search tree nodes;

[0022] Select the path with the highest cumulative reward value and output it as the industrial scheduling decision sequence.

[0023] In one possible implementation, the inner loop specifically includes:

[0024] Convert the current task intent into a query vector;

[0025] Retrieve industrial knowledge slices that semantically match the query vector from the industrial knowledge slice queue, and use the matched industrial knowledge slices as the current inference nodes;

[0026] Traverse all task intentions and combine all reasoning nodes into a reasoning node sequence according to the order of the task intentions.

[0027] In one possible implementation, the outer loop specifically includes:

[0028] Each inference node in the inference node sequence is used as a search tree node;

[0029] For the current search tree node, an external tool is invoked to generate a decision subsequence, and the execution result of the decision subsequence is captured;

[0030] If the execution result indicates an anomaly, the corresponding anomaly signal is injected into the current search tree node and the tool is invoked again;

[0031] If the execution result indicates that it is normal, the Monte Carlo tree search reward value of the path is calculated, and the execution result is fed back to the inference node sequence for correction of subsequent nodes;

[0032] After completing the search of all nodes, select the path with the highest cumulative reward value and output it as the industrial scheduling decision sequence.

[0033] In one possible implementation, after obtaining the industrial scheduling decision sequence, the method further includes:

[0034] Customized prompt words are synthesized using a multi-dimensional architectural feature industrial knowledge base;

[0035] The industrial scheduling decision sequence is constrained and reconstructed using the customized prompt words to generate a target to be verified.

[0036] The semantic equivalence of the target to be verified is evaluated using the CodeBLEU formula, and the industrial scheduling decision sequence is regenerated based on the evaluation results, or the high-level confidence result of the industrial scheduling decision sequence is output.

[0037] Secondly, this application provides an intelligent decision-making device for long-chain industrial tasks, comprising:

[0038] The acquisition module is used to determine the timing and corresponding quantity of prefetching for the next round of industrial knowledge through streaming retrieval.

[0039] The determination module is used to determine the queue for receiving industrial knowledge slices based on the pre-fetching timing and the pre-fetching quantity;

[0040] The control module is used to generate industrial scheduling decision sequences based on a dual-loop fusion mechanism of thought chain and decision chain, through dense vector retrieval and Monte Carlo tree search decision-maker.

[0041] Thirdly, this application provides an electronic device, including: a processor, and a memory communicatively connected to the processor;

[0042] The memory stores computer-executed instructions;

[0043] The processor executes computer execution instructions stored in the memory to implement the intelligent decision-making method for long-chain industrial tasks as described in the first aspect.

[0044] Fourthly, this application provides a computer-readable storage medium storing computer-executable instructions, which, when executed by a computer, are used to implement the intelligent decision-making method for long-chain industrial tasks as described in the first aspect.

[0045] Fifthly, this application also provides a computer program product, including a computer program that, when executed by a processor, can implement the steps of the scheme recommendation method as described in any of the first aspects.

[0046] This application provides an intelligent decision-making method, device, equipment, and medium for long-chain industrial tasks. By determining the pre-fetching timing and corresponding pre-fetching quantity of industrial knowledge for the next round through streaming retrieval, the industrial knowledge acquisition process can be adapted to the task's progress, reducing the impact of retrieval waiting time on response speed under large-scale industrial knowledge base conditions. By determining the queue for receiving industrial knowledge slices based on the pre-fetching timing and quantity, the orderly supply and connection of subsequent industrial knowledge resources can be achieved, thereby improving the real-time and continuous nature of industrial knowledge perception in complex task processing. Through a dual-loop fusion mechanism based on the thought chain and decision chain, combined with dense vector retrieval and Monte Carlo tree search decision generator to generate industrial scheduling decision sequences, the semantic consistency and logical integrity of long-chain tasks can be enhanced during multi-step reasoning, thereby improving the factual correctness and execution reliability of the output results. Attached Figure Description

[0047] The accompanying drawings, which are incorporated in and form part of this specification, illustrate embodiments consistent with this application and, together with the description, serve to explain the principles of this application.

[0048] Figure 1 A flowchart illustrating an intelligent decision-making method for long-chain industrial tasks provided in this application embodiment. Figure 1 ;

[0049] Figure 2 A flowchart illustrating an intelligent decision-making method for long-chain industrial tasks provided in this application embodiment. Figure 2 ;

[0050] Figure 3 A flowchart illustrating an intelligent decision-making method for long-chain industrial tasks provided in this application embodiment. Figure 3 ;

[0051] Figure 4 A schematic diagram of the structure of an intelligent decision-making device for long-chain industrial tasks provided in this application embodiment;

[0052] Figure 5 This is a schematic diagram of the structure of an electronic device provided in an embodiment of this application.

[0053] The accompanying drawings illustrate specific embodiments of this application, which will be described in more detail below. These drawings and descriptions are not intended to limit the scope of the concept in any way, but rather to illustrate the concept of this application to those skilled in the art through reference to particular embodiments. Detailed Implementation

[0054] Complex industrial task processing typically occurs in technical fields such as industrial automation control, intelligent manufacturing collaborative scheduling, cross-architecture industrial software migration, and multi-source data-driven real-time decision-making. In these scenarios, the system not only needs to understand multi-dimensional task intents from the equipment, control, and business sides, but also needs to integrate heterogeneous industrial knowledge such as code repositories, process rules, equipment manuals, historical operation records, and online status data within a limited timeframe to form decision results that can directly guide execution. Taking the coordinated control of air compressor groups, the optimization of production process parameters, and the cross-platform adaptation of control programs in an industrial automated production line as examples, tasks often have multiple steps, strong dependencies, rapid feedback, and low fault tolerance. A seemingly simple goal, such as maintaining stable pressure, reducing energy consumption, or completing software migration, actually involves multiple continuous stages such as state perception, rule matching, path deduction, and operation generation. To support such business, existing systems are typically constructed with a processing architecture consisting of a task input module, an industrial knowledge retrieval module, a reasoning generation module, and a result output module. Industrial knowledge content is stored in a retrieval library in the form of document slices, code snippets, or rule entries. After receiving a task, the system retrieves relevant industrial knowledge and then uses models to generate analysis results or control commands. This type of application scenario has industrial-grade requirements for response timeliness, logical integrity, and execution correctness. Any lag in industrial knowledge acquisition, interruption of semantic connection, or decision deviation may lead to scheduling imbalance, software porting failure, or even equipment malfunction. Therefore, it places much higher demands on collaborative computing methods than general question-and-answer or ordinary content generation scenarios.

[0055] In existing technologies, for the aforementioned complex industrial tasks, a large-model agent approach based on static retrieval and enhanced generation is typically employed. The basic idea is to first segment and vectorize domain-specific industrial knowledge offline, establishing a vector index or a similar industrial knowledge base. Online, when the user inputs the industrial task requirements, the system first converts the task description into a query vector, then performs similarity matching in the industrial knowledge base, recalling several candidate text fragments, rule descriptions, or code examples. These contents, along with the original task, are then fed into the model for reasoning or generation, ultimately outputting corresponding analytical conclusions, control suggestions, or program results.

[0056] From a working principle perspective, these solutions rely on a sequential processing model of "retrieval first, then reasoning," assuming that industrial knowledge acquisition precedes task understanding and decision generation. They assume that as long as the retrieved content is sufficiently relevant, it can support the model in completing subsequent tasks. However, in long-chain industrial tasks, this approach reveals significant limitations. First, the strict sequential nature of retrieval and reasoning means that as the industrial knowledge base grows, retrieval time directly adds to the total system latency. This is especially problematic when multiple intermediate steps need to be continuously deduced; the system cannot prepare the next industrial knowledge step in advance based on the context, resulting in response delays and making it difficult to meet the demands of real-time control, rapid scheduling, and online migration. Second, traditional solutions rely more on single or a few inferences to complete tasks, lacking the ability to stably maintain the inherent dependencies between long-chain steps. Even slight deviations in the initial understanding can amplify in subsequent steps, ultimately leading to target deviation, semantic breaks, or incorrect execution order. For example, in cross-architecture software migration, insufficient early understanding of target constraints can cause overall inaccurate code conversion directions; in equipment joint control, an error in parameter judgment at a certain step can cause imbalance in the entire scheduling chain. Furthermore, existing solutions often only verify output results at a formal level, failing to constrain model behavior from the perspective of consistency between task logic and domain-specific industrial knowledge. Therefore, even if the model's output appears grammatically correct and formatted, it may contain inferences inconsistent with industrial rules, suggestions mismatched with equipment status, or operational content that cannot be implemented. Ultimately, existing technologies struggle to simultaneously ensure the real-time nature of industrial knowledge perception, the continuity of long-chain reasoning, and the reliability of output results in complex industrial applications. This is a significant reason limiting their application in demanding industrial settings.

[0057] In view of this, how to improve the real-time performance of industrial knowledge perception and reasoning response in complex industrial task scenarios, while taking into account the consistency of long-chain logic and the credibility of output, has become an urgent technical problem to be solved. To address the above problems, this application provides an intelligent decision-making method for long-chain industrial tasks. When the system receives a request for a complex industrial task, it first determines the timing and quantity of pre-fetching industrial knowledge for the next round through streaming retrieval, so that industrial knowledge acquisition is no longer limited to a fixed serial triggering method, but can be dynamically arranged around the current task progress status. Then, based on the pre-fetching timing and quantity, a queue for receiving industrial knowledge slices is determined, so that subsequent processing can continuously obtain appropriate industrial knowledge support around the task context. On this basis, based on the dual-loop fusion mechanism of the thinking chain and the decision chain, an industrial scheduling decision sequence is generated through dense vector retrieval and Monte Carlo tree search decision-makers, thereby gradually mapping the abstract goals in the complex task into executable decision paths. Combining the common task input, industrial knowledge processing, and decision output architecture in industrial task handling, this method does not simply follow the static sequential process of existing systems. Instead, it improves the methods of industrial knowledge supply and decision generation through collaborative improvements, making industrial knowledge acquisition and task progress more aligned with the dynamic rhythm of real-world scenarios. This also ensures better correlation between the various steps in a long-chain task during continuous evolution. Using this technical approach, applications such as industrial automation control, cross-architecture software processing, and multi-process collaborative decision-making can improve the system's response efficiency to complex tasks, reduce biases caused by industrial knowledge lag and reasoning breaks, and lay the foundation for subsequent stable, coherent, and more suitable processing results.

[0058] Figure 1 A flowchart illustrating an intelligent decision-making method for long-chain industrial tasks provided in this application embodiment. Figure 1 .like Figure 1 As shown, the method includes:

[0059] S101: Determine the timing and corresponding quantity of prefetching for the next round of industrial knowledge through streaming retrieval.

[0060] This step is executed collaboratively by the task orchestration engine, context state management module, industrial knowledge retrieval and scheduling module, and latency prediction module of an intelligent decision-making system for long-chain industrial tasks. Upon receiving a complex industrial task request, the system first integrates the task text, structured process parameters, equipment state variables, control constraints, and historical operation records. Task requests can originate from manufacturing execution systems, distributed control systems, edge gateways, industrial software migration platforms, or manual operation terminals, and can be integrated via API calls, message queue subscriptions, file uploads, or database queries. Subsequently, the task orchestration engine performs context encoding on the current task, mapping the task objective, completed inference nodes, unresolved sub-problems, external constraints, and real-time feedback signals into a context state vector. Based on this context state vector, it predicts the type, source, and granularity of industrial knowledge that may be needed in the next round of inference. It's important to note that this streaming retrieval does not initiate a unified retrieval after all inference is completed. Instead, it starts the retrieval process in advance based on potential industrial knowledge needs for the next round while the current round of inference is still executing, creating an overlapping execution relationship between industrial knowledge acquisition and inference evolution.

[0061] Prefetching timing is used to determine the time point for initiating retrieval. Specifically, it can be determined based on the difference between the expected inference time of the current round and the expected retrieval time of the next round. When the system determines that the retrieval time of the next round is close to or exceeds the remaining inference time of the current round, the retrieval task is triggered in advance, ensuring that industrial knowledge is available before it is actually used. Prefetch quantity is used to determine how many industrial knowledge slices need to be recalled in advance. The system can perform joint calculations based on task complexity, contextual uncertainty, candidate industrial knowledge distribution density, and current computing resource load to ensure that the prefetching scale can cover the industrial knowledge needs of subsequent inference without causing excessive cache usage or an excessive accumulation of invalid industrial knowledge.

[0062] In one possible embodiment, the latency prediction module can maintain an inference time model and a retrieval time model. The inference time model outputs the remaining inference time for the current round based on the current thought chain depth, node complexity, number of tools invoked, and historical average processing time. The retrieval time model outputs the next round of industrial knowledge acquisition time based on the industrial knowledge base size, vector index level, candidate set size, network transmission latency, and concurrent load. The system calculates a prefetching trigger threshold accordingly. When the expected retrieval time is greater than or equal to the remaining inference time multiplied by a safety factor, the industrial knowledge retrieval scheduling module immediately starts the next round of streaming retrieval. The safety factor can be set to a value greater than zero and not less than one to absorb communication jitter and computational fluctuations in the industrial environment, thereby ensuring redundancy in the arrival time of industrial knowledge. For the prefetching quantity, the system can employ an adaptive prefetching depth algorithm, taking the task intent change rate, industrial knowledge hit confidence, and acceptable waiting latency as inputs, and outputting the number of industrial knowledge slices to be prefetched in the current round. For example, when the current task is in the syntax mapping stage of software migration, and subsequent tasks may involve instruction substitution rules, platform library function differences, and exception handling strategies, the system will increase the prefetch count to cover multi-branch reasoning. When the current task has converged to a specific device parameter correction path, the system will reduce the prefetch count to reduce redundancy. To improve retrieval speed, the system can also first use approximate nearest neighbor search to perform coarse recall on the vector industrial knowledge base, and then combine it with context consistency scoring for re-ranking, thereby obtaining highly relevant candidate industrial knowledge in a shorter time.

[0063] In this embodiment, streaming retrieval is used to dynamically acquire the industrial knowledge needed for the next round during the inference process. Its core function is to break the traditional static serial processing mode of retrieval first and then inference, allowing the retrieval process and the inference process to overlap on the timeline. Prefetching timing in this embodiment is used to determine when to initiate the advance acquisition of the next round of industrial knowledge. Essentially, it dynamically schedules the retrieval start time to avoid inference blocking caused by industrial knowledge arriving too late. Prefetching quantity in this embodiment is used to determine how many slices of industrial knowledge to acquire in advance, maintaining a balance between the sufficiency of industrial knowledge supply and resource overhead. Based on the above analysis, by predicting the demand for the next round of industrial knowledge before the inference process ends and simultaneously initiating retrieval, the waiting gaps in long-chain industrial tasks can be significantly reduced, avoiding inference pauses caused by delayed industrial knowledge acquisition. Simultaneously, by jointly controlling the prefetching timing and prefetching quantity, the system can achieve a stable balance between the real-time nature of industrial knowledge, cache usage, and retrieval costs, thus better adapting to the requirements of rapid response and continuous deduction in industrial settings. It should be understood that the above examples are merely illustrative and not limiting.

[0064] S102: Determine the queue for receiving industrial knowledge slices based on the pre-fetching timing and pre-fetching quantity.

[0065] In this step, after obtaining the prefetching timing and quantity, the industrial knowledge retrieval scheduling module initiates queries to one or more industrial knowledge sources according to the prefetching trigger plan, and converts the recall results into a queue of continuously consumable industrial knowledge slices. Industrial knowledge sources may include process rule bases, equipment manual bases, control logic bases, code example bases, fault handling case bases, real-time status databases, and vector industrial knowledge bases formed by document parsing. Due to the inconsistent data structures of different industrial knowledge sources, the system first standardizes the original industrial knowledge, uniformly dividing text paragraphs, table entries, rule expressions, code snippets, alarm records, and parameter templates into industrial knowledge slices. Each industrial knowledge slice carries at least the content subject, source identifier, timestamp, applicable scenario tag, associated equipment or software platform tag, and semantic vector representation. Subsequently, the system selects a corresponding number of industrial knowledge slices from the candidate recall set according to the predetermined prefetching quantity, and organizes them into queues for claiming industrial knowledge slices based on the usage order corresponding to the prefetching timing, semantic relevance, and task critical path priority. The industrial knowledge slice queue is used to carry the industrial knowledge units required for subsequent reasoning, so that the reasoning module can retrieve them in sequence and supplement them as needed when generating reasoning nodes or decision nodes, avoiding the need to perform a complete search on the entire industrial knowledge base at each step.

[0066] In one possible embodiment, the system can establish a multi-level queue structure, where the front-end queue stores high-priority industrial knowledge slices that will participate in the current inference round, and the back-end queue stores candidate industrial knowledge slices that may be used in subsequent rounds. When the inference module consumes an industrial knowledge slice from the front-end queue, the queue manager decides, based on the remaining queue length, the current inference progress, and the next round's prefetch plan, whether to add the slice from the back-end queue to the front-end queue, or to trigger incremental retrieval again. To ensure the availability of industrial knowledge slices, the system can also perform deduplication, conflict detection, and granularity adjustment. Deduplication removes duplicate content using slice hash values, source identifiers, and semantic similarity thresholds; conflict detection eliminates expired or contradictory industrial knowledge by judging rule version numbers, device model constraints, and time validity; and granularity adjustment re-segments the industrial knowledge content according to task complexity. For example, in the task of controlling a group of air compressors, the equipment start-up and shutdown rules, pressure threshold ranges, interlocking conditions, and energy consumption curves are divided into independent slices for individual invocation at different inference nodes. For cross-architecture software migration tasks, the source code interface specifications, target platform API (Application Programming Interface) mappings, variable lifecycle rules, and exception handling templates can be divided into corresponding slices for easy retrieval according to the migration steps. To further improve the hit probability of key industrial knowledge, the system can also calculate a priority score for each industrial knowledge slice based on the current task context. This score comprehensively considers semantic matching degree, rule constraint strength, real-time state relevance, and historical successful usage frequency, with higher-priority slices placed at the front of the queue.

[0067] The industrial knowledge slice queue is used to carry industrial knowledge slices determined by prefetching timing and quantity. That is, it organizes scattered, heterogeneous, and dynamically arriving industrial knowledge content into an ordered industrial knowledge stream that can be read continuously. Determining the industrial knowledge slice retrieval queue based on prefetching timing and quantity includes not only simply arranging the recall results, but also screening, rearranging, conflict resolution, and front-end and back-end caching management of industrial knowledge slices, ensuring that the supply of industrial knowledge keeps pace with the subsequent reasoning process.

[0068] Based on the above analysis, it can be seen that by constructing a queue for receiving industrial knowledge slices, the system can stably provide context-adapted industrial knowledge support during the continuous advancement of complex industrial tasks, avoiding inference interruptions and semantic jumps caused by discrete retrieval results, disordered arrival, or mismatched granularity. At the same time, the queued organization method also enables the system to monitor the progress of industrial knowledge consumption and implement incremental replenishment accordingly, thereby enhancing the continuity and controllability of industrial knowledge supply under long-chain tasks.

[0069] S103: Based on the dual-loop fusion mechanism of thinking chain and decision chain, an industrial scheduling decision sequence is generated through dense vector retrieval and Monte Carlo tree search decision engine.

[0070] In this step, after the industrial knowledge slice queue is established, the reasoning generation module initiates a dual-loop fusion processing of the thought chain and decision chain. The thought chain represents the abstract reasoning process oriented towards the task objective, including task decomposition, causal relationship judgment, constraint identification, state transition analysis, and sub-objective arrangement. The decision chain represents the specific execution actions corresponding to the thought chain, including parameter setting, control instruction generation, code conversion steps, tool call order, verification actions, and rollback actions. The inner loop first generates thought chain nodes around the current task context and the established industrial knowledge slice queue. The system encodes the current task intent into a query vector and performs dense vector retrieval on the slice vectors in the industrial knowledge slice queue. It selects one or more industrial knowledge slices that best match the semantics of the current sub-objective as the industrial knowledge support for the current reasoning node, and then forms a sequence of reasoning nodes, i.e., thought chain nodes, based on these support contents. Here, dense vector retrieval represents the task semantics and industrial knowledge semantics through low-dimensional or medium-dimensional continuous vectors, enabling the system to identify industrial knowledge content with different literal meanings but similar connotations. Therefore, it is highly adaptable to the numerous synonyms, abbreviations, equipment aliases, and cross-platform interface differences that exist in industrial tasks. After the thought chain is generated, the outer loop maps each inference node to the corresponding decision space. The inference node is used as the tree node for Monte Carlo tree search. Selection, expansion, simulation and backhaul processing are carried out around possible execution actions to generate candidate decision subsequences. The reward evaluation mechanism is used to select the path with higher cumulative benefits, and finally form an industrial scheduling decision sequence for the current industrial task.

[0071] Specifically, in the selection phase, the system selects nodes to be expanded from the existing decision tree based on historical cumulative reward values ​​and the exploration benefit balance criterion. In the expansion phase, multiple candidate actions are generated based on the industrial knowledge slices associated with the current inference node and the available toolset. For example, in the equipment control scenario, actions such as starting and stopping a compressor, adjusting pressure thresholds, delaying switching time, and performing interlock verification are generated; in the software migration scenario, actions such as interface replacement, syntax rewriting, dependency library mapping, compilation parameter correction, and compatibility checks are generated. In the simulation phase, the system calls external tools to quickly verify the candidate actions. These external tools can be control rule validators, industrial process simulators, compilers, static analyzers, digital twin models, or historical case matchers. The tools return results such as execution feasibility, constraint satisfaction, risk level, and expected benefits. In the feedback phase, the system converts the above results into reward values ​​and updates the statistics of each node on the search path. The reward value can be composed of multiple weighted indicators, including goal achievement, rule consistency, real-time cost, energy consumption level, fault risk, and operational complexity, with the weights set according to the specific industrial scenario. For example, in the scenario of air compressor group joint control, pressure stability and interlock safety have higher weights, followed by energy consumption optimization; in the scenario of software migration, semantic equivalence and compilation pass rate have higher weights, while execution efficiency and maintenance cost are considered later. After multiple rounds of search, the system selects the path with the highest cumulative reward value or the comprehensive score that meets the threshold as the output industrial scheduling decision sequence. The execution result does not remain in the current round, but flows back into the inner loop of the thought chain to correct the priority of subsequent sub-objectives, update the task context state, and trigger new streaming retrieval and industrial knowledge slice queue adjustments when necessary, thus forming a two-way closed loop between the thought chain and the decision chain.

[0072] In this embodiment, the dual-loop fusion mechanism of the thought chain and decision chain is used to establish a dynamic mapping relationship between abstract reasoning steps and specific execution actions. The inner loop ensures the continuity of reasoning, while the outer loop ensures the executability of actions. Both achieve continuous correction through execution feedback. Dense vector retrieval is used to find industrial knowledge slices in the industrial knowledge slice queue that semantically match the task intent, forming a sequence of reasoning nodes. This ensures that the thought chain is built on a foundation of real-time, relevant, and continuously supplied industrial knowledge. MCTS, or Monte Carlo Tree Search Decision Maker, is used to search and select a better execution path in a multi-branch decision space. By repeatedly trying different action combinations and updating path values ​​based on feedback, it can output a more feasible and consistent industrial scheduling decision sequence under complex constraints.

[0073] Based on the above analysis, this step integrates industrial knowledge retrieval, abstract reasoning, tool verification, and decision optimization into a closed-loop process. This prevents the system from relying on a single inference to arrive at the final answer. Instead, it continuously corrects understanding and action direction as the long-chain task progresses, thereby reducing the risk of prior biases accumulating and spreading to subsequent steps. Simultaneously, utilizing execution feedback from external tools in reward evaluation constrains model generation behavior from two dimensions: domain rule consistency and task feasibility, thus improving the credibility and industrial executability of the output results. It should be understood that the above example is merely illustrative and not limiting.

[0074] In one possible embodiment, during cross-architecture code porting, the system faces the problem of a lack of effective verification mechanism after the industrial scheduling decision sequence is generated, leading to risks such as syntax errors, semantic deviations, or logical illusions in the generated results. Therefore, to ensure the high reliability and cross-architecture semantic equivalence of the industrial scheduling decision sequence, this embodiment introduces a constraint reconstruction and iterative verification mechanism based on a multi-dimensional architectural feature industrial knowledge base. The corresponding specific implementation method is as follows:

[0075] The industrial scheduling decision sequence is constrained and reconstructed using customized prompt words to generate the target to be verified; the semantic equivalence of the target to be verified is evaluated using the CodeBLEU formula, and the industrial scheduling decision sequence is regenerated based on the evaluation results, or the high-level confidence result of the industrial scheduling decision sequence is output.

[0076] The multi-dimensional architecture feature industrial knowledge base includes structured industrial knowledge such as instruction set differences, register mappings, and ABI specifications. This industrial knowledge is stored in the form of graphs and rule tables, supporting fine-grained feature alignment and constraint injection. The ABI specification refers to the Application Programming Interface specification, which defines key constraints in cross-architecture portability, such as function calling conventions, register usage rules, and stack frame layout. Customized prompt words are used to organize the above-mentioned constraint industrial knowledge into a structured context to apply targeted refactoring guidance to the industrial scheduling decision sequence. The target to be verified represents the text or code result to be evaluated after being constrained by the prompt words. The CodeBLEU (Code Bilingual Evaluation Understudy) formula is used to measure the semantic equivalence of the target to be verified from dimensions such as code matching, keyword matching, data flow consistency, and structural consistency, thereby determining its degree of conformity with the target architecture requirements.

[0077] Specifically, the system first synthesizes customized prompt words using a multi-dimensional architectural feature industrial knowledge base. Then, the system uses these customized prompt words to reconstruct constraints on the industrial scheduling decision sequence, generating targets to be verified. The customized prompt word synthesis assembles the retrieved architectural industrial knowledge and source code fragments according to a specific template. This template forcibly distinguishes between immutable rule areas and code areas to be transformed, guiding the large model to prioritize satisfying hard architectural constraints before performing semantic equivalence transformation. After generating the targets to be verified, the system uses the CodeBLEU formula to evaluate their semantic equivalence. CodeBLEU is a formula for calculating code semantic similarity, combining lexical matching and syntax tree structure matching, specifically expressed as:

[0078] .

[0079] in, For standard lexical matching degree, For weighted lexical matching degree, For the abstract syntax tree matching degree, For data stream matching, the hyperparameter settings satisfy:

[0080] .

[0081] The recommended configuration in this embodiment is as follows: =0.1、 =0.1、 =0.4、 =0.4, the physical meaning of this configuration is to significantly strengthen the weight of syntax tree matching and data flow matching, and weaken surface lexical matching. In cross-architecture code porting, variable names and keywords may change completely, but control flow and data flow must be strictly equivalent. This can suppress logical illusions.

[0082] Meanwhile, to measure the reliability of cross-architecture mapping, the system introduces Cohen's Kappa consistency coefficient to evaluate the fit of multi-dimensional feature alignment. The specific formula is as follows:

[0083] .

[0084] in, To observe the consistency rate, To achieve the desired consistency rate, this module uses multi-dimensional feature alignment to achieve a Kappa coefficient of 0.92 for the architecture mapping, ensuring extremely high consistency in industrial knowledge enhancement. After evaluation, the system either regenerates the industrial scheduling decision sequence based on the evaluation results or outputs the industrial scheduling decision sequence as a highly reliable result.

[0085] Specifically, after generating the industrial scheduling decision sequence, it undergoes dual verification through syntax validation and CodeBLEU semantic evaluation. Syntax validation ensures the code's executability, while semantic evaluation ensures that the converted code is logically equivalent to the source code. If the criteria are not met, the error message is recombined with the original prompt words for iterative generation until a highly reliable industrial scheduling decision sequence with zero illusion is output. This process is essentially a gradient-free optimization process based on execution feedback.

[0086] Key parameters and constraints during this process include: a maximum number of syntax validation attempts of 10 is recommended; if the syntax validation fails after this number, it indicates a systematic misunderstanding in the model, requiring a fallback to the decision reasoning layer for replanning; a maximum number of semantic validation attempts of 10 is recommended to limit the iteration depth of CodeBLEU evaluation and regeneration; a minimum CodeBLEU threshold score of 0.5 is recommended to determine the semantic equivalence of the code, as cross-architecture transformations may lead to a large number of lexical changes, and 0.5 is sufficient to ensure the equivalence of the core data flow; and a large model generation temperature of 1 is recommended during the iterative validation phase to maintain appropriate randomness, allowing it to break out of previous error generation patterns and correct logical errors.

[0087] Through the aforementioned technical means, the system forcibly injects hard architectural constraints into the generation process using a multi-dimensional architectural feature industrial knowledge base and a customized prompt word synthesis mechanism, effectively solving the problem of constraint omission caused by architectural differences in cross-architecture code porting. Furthermore, through dual verification of CodeBLEU semantic evaluation and syntax checking, combined with an iterative feedback optimization mechanism, the system significantly improves the executability and semantic equivalence of the generated code, achieving highly reliable output with zero illusions. Simultaneously, by using a gradient-free optimization process and limiting the maximum number of attempts, computational overhead is controlled while ensuring generation quality, avoiding resource waste caused by infinite iteration. Finally, the quantitative evaluation of multi-dimensional feature alignment using Cohen's Kappa consistency coefficient ensures a high degree of consistency between the industrial knowledge enhancement module and the target architecture, providing a quantitative guarantee for the reliability of cross-architecture porting.

[0088] This application provides an intelligent decision-making method for long-chain industrial tasks, including: determining the pre-fetching timing and corresponding pre-fetching quantity of industrial knowledge for the next round through streaming retrieval; determining the queue for receiving industrial knowledge slices based on the pre-fetching timing and pre-fetching quantity; and generating an industrial scheduling decision sequence through dense vector retrieval and Monte Carlo tree search decision-makers based on a dual-loop fusion mechanism of thought chain and decision chain. In this embodiment, by transforming the industrial knowledge acquisition from a fixed serial process to a dynamic pre-fetching process coupled with the inference process, and further organizing the pre-fetched heterogeneous industrial knowledge into a queue of continuously consumable industrial knowledge slices, and then combining the dual-loop fusion mechanism oriented towards abstract reasoning and concrete execution to complete decision generation, complex industrial tasks can simultaneously possess higher real-time industrial knowledge supply, better long-chain logical continuity, and stronger result verifiability during processing. For scenarios such as industrial automation control, multi-source data-driven scheduling, and cross-architecture industrial software migration, the above technical approach can reduce the waiting time caused by industrial knowledge lag, reduce the target deviation caused by the break in the inference chain, and suppress output content inconsistent with industrial rules through execution feedback closed loop, thereby forming a stable, coherent, and suitable-for-implementation processing result. It should be understood that the above examples are for illustrative purposes only and are not intended to be limiting.

[0089] Figure 2 A flowchart illustrating an intelligent decision-making method for long-chain industrial tasks provided in this application embodiment. Figure 2 This application provides a detailed description of the steps for determining the prefetching timing and corresponding prefetching quantity of industrial knowledge for the next round through streaming retrieval. For example... Figure 2 As shown, the method includes:

[0090] S201. Determine the expected inference time for this round and the expected retrieval time for the next round.

[0091] In this step, the expected inference time represents the estimated time required for the current round to complete industrial knowledge fusion, content understanding, and result generation, while the expected retrieval time represents the estimated time required for the next round of industrial knowledge to be retrieved from the retrieval database and returned. Together, they form the basis of temporal collaborative judgment, enabling prefetching actions to be triggered in advance around the inference process and avoiding significant waiting conflicts between the next round of retrieval and the current round of inference.

[0092] In practice, the system can estimate the expected inference time for the current round based on the complexity of the current task, historical response records, model generation length, and retrieval library load. At the same time, it can estimate the expected retrieval time for the next round by combining the size of the industrial knowledge base, index matching cost, network transmission overhead, and the number of candidate fragments.

[0093] S202. Calculate the prefetch timing based on the expected reasoning time and the expected retrieval time.

[0094] In this step, the traditional sequential execution of retrieval and reasoning—retrieving the full amount of industrial knowledge before performing single-step reasoning—incurs significant first-word delay and reasoning starvation. Therefore, to eliminate retrieval waiting gaps and achieve frequency resonance between industrial knowledge supply and cognitive reasoning, this embodiment introduces an adaptive prefetching timing calculation mechanism based on Markov decision processes. The corresponding specific implementation steps are as follows:

[0095] Obtain the actual reasoning time of the previous round and the corresponding historical expected reasoning time; calculate the feedback factor based on the actual reasoning time and the historical expected reasoning time; and determine the prefetching timing based on the Markov decision process, the expected reasoning time, the expected retrieval time, and the feedback factor.

[0096] The actual inference time characterizes the true runtime of the previous inference round, while the historical expected inference time provides a prediction benchmark for the same round; together, they constitute the basis for deviation assessment. The feedback factor reflects the degree of deviation between the actual inference time and the historical expected inference time, mapping this deviation to a correction amount for subsequent prefetching strategies. The Markov decision process is used for state-driven selection among candidates at different prefetching times, ensuring that the current decision simultaneously considers the inference load of this round, the retrieval cost of the next round, and the impact of historical deviations.

[0097] Specifically, the system first obtains the actual inference time of the previous round and the corresponding historical expected inference time. The system calculates a feedback factor based on the actual inference time and the historical expected inference time. The feedback factor is an adjustment amount that dynamically compensates for the actual delay difference of the preceding steps, and is used to eliminate timing mismatch caused by system fluctuations.

[0098] Subsequently, the system determines the prefetching timing based on the Markov decision process, the expected inference time, the expected retrieval time, and the feedback factor. A Markov decision process is a stochastic decision process with Markov properties, where the system's state at the next moment depends only on the current state and is independent of historical states. In this embodiment, the large-scale model inference process is modeled as a decision process with Markov properties, meaning the inference content at step n+1 depends only on the output at step n. The prefetching timing refers to the point in time when the system initiates the next round of industrial knowledge retrieval before the end of the current inference phase; its calculation is based on a dynamic estimate of historical inference and retrieval times. The feedback factor is calculated as follows:

[0099] .

[0100] in, The actual reasoning time for step n-1, The time required for historical reasoning at step n-1 is... This is the damping coefficient. When the actual inference time exceeds the expected time, the feedback factor is positive, delaying subsequent prefetching; conversely, it is prefetched earlier to prevent the industrial knowledge queue from becoming empty. Damping coefficient. A value of 0.2 is recommended to prevent feedback oscillations from causing the timing of the prefetch to fluctuate wildly between early and late.

[0101] After obtaining the feedback factor, the prefetch timing is calculated based on the feedback factor. The calculation formula is as follows:

[0102] .

[0103] in, The expected reasoning time for the nth step. This represents the expected retrieval time for the (n+1)th step. The above formula is used to characterize: before the current inference ends, prefetching is initiated in advance by the retrieval time plus the feedback adjustment amount, so that the inference ends exactly when the retrieval is completed, eliminating the waiting gap.

[0104] Lead time coefficient The recommended value range is 2 to 4, depending on the fluctuation of the underlying storage network. The greater the network latency variance, the higher the value should be set to reserve a buffer.

[0105] Feedback adjustment coefficient The recommended value range is 4 to 6, which is used to amplify the influence of the feedback signal and ensure the sensitivity of dynamic compensation.

[0106] Through this specific implementation, the system can reduce the idle time of inference waiting for retrieval results and reduce latency fluctuations caused by early or late retrieval, thus maintaining a more stable synchronization between industrial knowledge supply and inference progress. Simultaneously, with the dynamic compensation of feedback factors, the system has a stronger adaptability to different task complexities and execution fluctuations in different rounds, thereby improving the overall real-time response and the continuity of long-chain task processing.

[0107] S203. Input the expected inference time and expected retrieval time into the adaptive prefetch depth algorithm to determine the number of prefetches.

[0108] In this step, when determining the prefetch quantity, the system inputs the expected inference time and expected retrieval time into an adaptive prefetch depth algorithm. This algorithm, combined with the task chain length, industrial knowledge slice granularity, cache usage, and real-time resource status, outputs the prefetch depth. This prefetch depth corresponds to the number of industrial knowledge slices that should be preloaded in the next round. It can increase the prefetch reserve when the inference pace is fast and appropriately expand the prefetch range when the retrieval time is long, thereby achieving a dynamic match between the prefetch quantity and the task progress speed.

[0109] The calculation formula is:

[0110] .

[0111] in, This is the pipeline parallelism coefficient. This formula ensures that within a single inference time window, the retrieval module can provide a sufficient number of industry knowledge slices for subsequent consumption.

[0112] In addition, the system introduces a dynamic pipeline scheduler to monitor the consumption rate of inference nodes and the output rate of retrieval nodes. When inference consumption is too fast, leading to starvation in the industrial knowledge queue, the scheduler triggers concurrent retrieval expansion; when excessive retrieval output leads to the risk of memory overflow, the scheduler applies back pressure to limit retrieval concurrency.

[0113] The key parameters and constraints are as follows:

[0114] The Recall@k loss threshold is a precision loss constraint introduced by the prefetching mechanism. Prefetching is based on prediction rather than the actual input, and it must ensure that the Recall@k decreases by no more than 2%; otherwise, a safety mechanism to degrade to synchronous retrieval must be triggered. Approximate nearest neighbor search is a vector retrieval technique that sacrifices a small amount of precision for a significant speed improvement. Recall@k represents the proportion of truly relevant results among the k results returned by the retrieval.

[0115] This application provides an intelligent decision-making method for long-chain industrial tasks. During operation, it first performs collaborative prediction along the time dimension, and then jointly determines the prefetching timing and scale, enabling streaming retrieval to prepare industrial knowledge before it is actually used in the next round of inference. Since both the prefetching timing and quantity are constrained by the current inference state and the cost of the next round of retrieval, it reduces idle time caused by retrieval waiting, and mitigates industrial knowledge gaps due to insufficient prefetching and resource consumption due to excessive prefetching. Using this approach, the system can maintain a stable pace of industrial knowledge supply in continuous inference scenarios of complex industrial tasks, improve the efficiency of the connection between industrial knowledge retrieval and inference generation, reduce response delays caused by retrieval lags, and enhance the real-time performance and continuity of multi-round industrial knowledge invocation, thereby strengthening the coherence and executability of the overall decision output.

[0116] Figure 3 A flowchart illustrating an intelligent decision-making method for long-chain industrial tasks provided in this application embodiment. Figure 3 This application provides a detailed description of the steps involved in generating an industrial scheduling decision sequence. For example... Figure 3 As shown, the method includes:

[0117] S301. Start the inner loop, perform dense vector retrieval on the industrial knowledge slice queue, and generate a sequence of inference nodes.

[0118] In this step, during the processing of long-chain tasks, traditional large-scale model reasoning suffers from a disconnect between thought process and tool execution. Furthermore, it lacks an effective mechanism to map natural language logic to specific operational paths, leading to logical breaks and the inability to dynamically correct tool call failures during task execution. Therefore, to achieve precise alignment between thought process and tool execution, and to improve the logical robustness and decision path optimization of long-chain tasks, this embodiment introduces a dual-loop fusion mechanism of thought chain and decision chain, generating industrial scheduling decision sequences through dense vector retrieval and MCTS.

[0119] The system initiates an inner loop to perform intensive vector retrieval on the industrial knowledge slice queue, generating a sequence of inference nodes. Specifically, the inner loop includes: converting the current task intent into a query vector; retrieving industrial knowledge slices from the industrial knowledge slice queue that semantically match the query vector, and using the matched industrial knowledge slices as the current inference nodes; and traversing all task intents, combining all inference nodes into a sequence of inference nodes according to the task intent order.

[0120] The industrial knowledge slice queue is a set of structured industrial knowledge fragments output by the aforementioned retrieval module. Dense vector retrieval is a semantic retrieval method based on the vector space model, which achieves semantic matching by calculating the similarity between the query vector and the document vector. All task intents are traversed, and all inference nodes are combined into an inference node sequence according to the task intent order. This stage only performs logical deduction and does not produce external side effects. Dense vector retrieval is used to match historical decision patterns to construct the logical framework of the task.

[0121] The precise matching mechanism relies on a cosine similarity calculation model based on dense vector retrieval. Let the query vector be... The vector corresponding to the industrial knowledge slice in the industrial knowledge slice queue is Then its matching degree is:

[0122] .

[0123] This similarity score is used in the inner loop to recall the most relevant industrial knowledge slices from the industrial knowledge slice queue, providing prior guidance for subsequent Monte Carlo tree search. The Top-K retrieval parameter is the number of candidate industrial knowledge slices returned by dense vector retrieval; a value of 5 is recommended, as too high a value introduces noise, while too low a value leads to a loss of diversity. The system traverses all task intents, combining all inference nodes into an inference node sequence according to the task intent order. This stage only performs logical deduction and does not produce external side effects. It constructs the logical framework of the task by matching historical decision patterns through dense vector retrieval.

[0124] S302. Start the outer loop, perform Monte Carlo tree search based on the inference node sequence, call external tools to generate decision subsequences, and feed back the execution results to update the reward values ​​of the search tree nodes. Select the path with the highest cumulative reward value and output it as the industrial scheduling decision sequence.

[0125] In this step, the system initiates an outer loop to perform a Monte Carlo tree search based on the inference node sequence. Specifically, the outer loop includes: using each inference node in the inference node sequence as a search tree node, for the current search tree node, calling an external tool to generate a decision subsequence, and capturing the execution result of the decision subsequence. If the execution result indicates an anomaly, the corresponding anomaly signal is injected into the current search tree node and the tool is called again.

[0126] The maximum number of outer loop retries limits the number of replanning attempts after a failed outer loop call to the tool, preventing infinite loops; a value of 3 is recommended. If the execution result indicates success, the Monte Carlo tree search reward value for that path is calculated, and the execution result is fed back to the inference node sequence for subsequent node corrections.

[0127] The decision chain optimization logic is based on Monte Carlo Tree Search (MCTS), an algorithm that combines stochastic simulation with tree search to optimize decisions. In decision state s, the strategy for choosing action a follows the UCB1 formula:

[0128] .

[0129] in, The action value represents the historical average return of this decision-making path; For the number of state visits, For the number of times the action is accessed; Let c be the exploration constant. This formula represents the balance between utilizing high-value historical paths and exploring unvisited paths. As the number of simulations increases, the weight of the exploration item gradually decreases, and the algorithm converges to the globally optimal decision path. The recommended value for the exploration constant c is 1.414, i.e., It is suitable for most decision space exploration needs.

[0130] After completing the search of all nodes, the system selects the path with the highest cumulative reward value and outputs it as an industrial scheduling decision sequence.

[0131] This application provides an intelligent decision-making method for long-chain industrial tasks. Through the fusion of an inner and outer loop, the system aligns the logical deduction in the natural language space with the operational mapping in the tool space across modalities, effectively solving the problem of disconnect between thinking and execution in traditional large-scale model reasoning and avoiding semantic shift. Simultaneously, the introduction of dense vector retrieval and historical decision pattern matching through the inner loop provides high-quality prior guidance for the Monte Carlo tree search in the outer loop, significantly improving the search efficiency and matching accuracy of decision paths. Furthermore, the introduction of an execution result feedback mechanism and an anomaly signal injection replanning mechanism through the outer loop enables dynamic correction of tool call failures and adaptive adjustment of subsequent nodes, enhancing the robustness of long-chain task execution. Finally, by balancing the exploration and utilization of the UCB1 formula and employing the path selection strategy with the highest cumulative reward value, the global optimality of the industrial scheduling decision sequence is ensured, improving the completion quality and reliability of complex tasks.

[0132] Building upon the aforementioned embodiments, further, based on the dual-loop fusion mechanism of the thinking chain and decision chain, an industrial scheduling decision sequence is generated through dense vector retrieval and MCTS, including: initiating an inner loop to perform dense vector retrieval on the industrial knowledge slice queue and generate an inference node sequence; initiating an outer loop to perform Monte Carlo tree search based on the inference node sequence; calling external tools to generate decision sub-sequences and feeding back the execution results to update the reward values ​​of the search tree nodes; selecting the path with the highest cumulative reward value and outputting it as the industrial scheduling decision sequence.

[0133] The inference node sequence carries multiple inference nodes generated by the inner loop and serves as the basis for the outer loop to perform searches and call tools. The industrial knowledge slice queue carries pre-fetched and retrieved domain industrial knowledge slices, which can come from equipment manuals, process rules, code snippets, or historical operation records. The external tool generates decision sub-sequences for the current search tree node and obtains the execution results. The tool can be a rule engine, simulator, code executor, or status verification module. In practical applications, the tool can also be selected from other models or implementation forms, which are not limited in this embodiment. The search tree node corresponds to the current node in the inference node sequence and carries search branch information. The reward value is used to characterize the degree to which the current path satisfies the task objective.

[0134] In practical implementation, after receiving a complex industrial task, the system writes the received industrial knowledge slices into a retrieval index in a vectorized representation. The inner loop performs a dense vector search on the industrial knowledge slice queue based on the task intent vector, obtaining multiple candidate fragments that are closest to the current semantics, and organizes them into a sequence of inference nodes according to the task intent order. The outer loop builds a search tree with the nodes in the inference node sequence as the root state, and performs a Monte Carlo tree search around each search tree node. During the search, it calls external tools based on the context constraints corresponding to the current node to generate a decision subsequence that matches the current node. After the external tool outputs the execution result, the system performs state analysis on the result. If the result indicates that the decision is executable and satisfies the constraints, the reward value is calculated based on the execution completion degree, constraint satisfaction degree, and target deviation, and the reward value is fed back to the corresponding search tree node. If the result indicates an anomaly, an anomaly signal is injected and the reward of that branch is reduced. After several rounds of expansion and feedback, the system compares the cumulative reward values ​​of each path, selects the highest one to form the final industrial scheduling decision sequence, and outputs it.

[0135] This mechanism decouples industrial knowledge retrieval from decision search into inner and outer loops, enabling inference nodes to continuously provide semantic support for the search tree, while the execution feedback from external tools can correct the search direction, thus forming a closed-loop optimization. Because the reward value accumulates continuously and is bound to the actual execution result, the system can retain branches that better conform to industrial constraints among multi-branch candidate paths, reducing invalid inferences and erroneous decisions, and improving the continuity, executability, and reliability of results for long-chain tasks.

[0136] Figure 4 This is a schematic diagram of the structure of an intelligent decision-making device for long-chain industrial tasks, provided as an embodiment of this application. Figure 4 As shown, the intelligent decision-making device 40 for long-chain industrial tasks includes:

[0137] The acquisition module 401 is used to determine the timing and corresponding quantity of prefetching of the next round of industrial knowledge through streaming retrieval.

[0138] Module 402 is used to determine the queue for receiving industrial knowledge slices based on the prefetching timing and prefetching quantity;

[0139] The control module 403 is used to generate industrial scheduling decision sequences based on the dual-loop fusion mechanism of the thinking chain and the decision chain, through dense vector retrieval and Monte Carlo tree search decision-maker.

[0140] In one possible implementation, the acquisition module 401 is further configured to:

[0141] Determine the expected inference time for this round and the expected retrieval time for the next round;

[0142] Calculate the prefetch timing based on the expected reasoning time and expected retrieval time;

[0143] The expected inference time and expected retrieval time are input into the adaptive prefetch depth algorithm to determine the number of prefetches.

[0144] In one possible implementation, the acquisition module 401 is further configured to:

[0145] Obtain the actual reasoning time of the previous round and the corresponding historical expected reasoning time;

[0146] The feedback factor is calculated based on the actual reasoning time and the historical expected reasoning time.

[0147] Based on the Markov decision process, the prefetching timing is determined according to the expected inference time, the expected retrieval time, and the feedback factor.

[0148] In one possible implementation, the control module 403 is further configured to:

[0149] Start the inner loop to perform dense vector retrieval on the industrial knowledge slice queue and generate a sequence of inference nodes;

[0150] Start the outer loop and perform a Monte Carlo tree search based on the inference node sequence;

[0151] External tools are invoked to generate decision subsequences, and the execution results are fed back to update the reward values ​​of the search tree nodes;

[0152] Select the path with the highest cumulative reward value and output it as the industrial scheduling decision sequence.

[0153] In one possible implementation, the inner loop within the control module 403 specifically includes:

[0154] Convert the current task intent into a query vector;

[0155] Retrieve industrial knowledge slices that semantically match the query vector from the industrial knowledge slice queue, and use the matched industrial knowledge slices as the current inference nodes;

[0156] Traverse all task intentions and combine all reasoning nodes into a reasoning node sequence according to the order of the task intentions.

[0157] In one possible implementation, the outer loop within the control module 403 specifically includes:

[0158] Each inference node in the inference node sequence is used as a search tree node;

[0159] For the current search tree node, call an external tool to generate a decision subsequence and capture the execution results of the decision subsequence;

[0160] If the execution result indicates an error, the corresponding error signal will be injected into the current search tree node and the tool will be invoked again.

[0161] If the execution result indicates that it is normal, the Monte Carlo tree search reward value of the path is calculated, and the execution result is fed back to the inference node sequence for subsequent node correction.

[0162] After completing the search of all nodes, select the path with the highest cumulative reward value and output it as the industrial scheduling decision sequence.

[0163] In one possible implementation, after obtaining the industrial scheduling decision sequence, the control module 403 is further configured to:

[0164] Customized prompt words are synthesized using a multi-dimensional architectural feature industrial knowledge base;

[0165] The industrial scheduling decision sequence is constrained and reconstructed using customized prompts to generate targets to be verified.

[0166] The CodeBLEU formula is used to evaluate the semantic equivalence of the target to be verified, and the industrial scheduling decision sequence is regenerated based on the evaluation results, or the high-level confidence result of the industrial scheduling decision sequence is output.

[0167] This embodiment provides an intelligent decision-making device for long-chain industrial tasks, which can execute the intelligent decision-making method for long-chain industrial tasks provided in the above method embodiment. Its implementation principle and technical effect are similar, and will not be described in detail here.

[0168] Figure 5 This is a schematic diagram of the structure of an electronic device provided in an embodiment of this application. Figure 5 As shown, the electronic device 50 includes a processor and a memory communicatively connected to the processor. Optionally, the electronic device 50 also includes a communication component 503. The processor 501, memory 502, and communication component 503 are connected via a bus 504.

[0169] The memory stores instructions that the computer executes;

[0170] The processor executes computer execution instructions stored in memory to implement intelligent decision-making methods for long-chain industrial tasks, as described above.

[0171] At least one processor 501 may be a central processing unit (CPU), an application-specific integrated circuit (ASIC), or one or more integrated circuits configured to implement the embodiments of this application.

[0172] Optionally, in specific implementations, the processor 501 and memory 502 are implemented independently. In this case, the processor 501 and memory 502 can be interconnected via a bus to complete communication between them. The bus can be an Industry Standard Architecture (ISA) bus, a Peripheral Component Interconnect (PCI) bus, or an Extended Industry Standard Architecture (EISA) bus, etc. Buses can be categorized as address buses, data buses, control buses, etc., but this does not imply that there is only one bus or one type of bus.

[0173] Optionally, in a specific implementation, if the processor 501 and the memory 502 are integrated on a single chip, the processor 501 and the memory 502 can communicate through an internal interface.

[0174] This application also provides a computer storage medium storing computer execution instructions, which, when executed by a processor, implement the aforementioned technical solution for intelligent decision-making methods for long-chain industrial tasks.

[0175] The aforementioned computer-readable storage medium can be implemented by any type of volatile or non-volatile storage device or a combination thereof, such as static random access memory (SRAM), electrically erasable programmable read-only memory (EEPROM), erasable programmable read-only memory (EPROM), programmable read-only memory (PROM), read-only memory (ROM), magnetic storage, flash memory, magnetic disk, or optical disk. The computer-readable storage medium can be any available medium accessible to a general-purpose or special-purpose computer.

[0176] An exemplary readable storage medium is coupled to a processor, enabling the processor to read information from and write information to the readable storage medium. Alternatively, the readable storage medium can be an integral part of the processor. Both the processor and the readable storage medium can reside in an Application Specific Integrated Circuit (ASIC). Alternatively, the processor and the readable storage medium can exist as discrete components in the control device of a garment handling apparatus.

[0177] This unit division is merely a logical functional division; in actual implementation, there may be other division methods. For example, multiple units or components may be combined or integrated into another system, or some features may be ignored or not executed. Furthermore, the coupling or direct coupling or communication connection shown or discussed may be indirect coupling or communication connection through some interfaces, devices, or units, and may be electrical, mechanical, or other forms.

[0178] The units described as separate components may or may not be physically separate. The components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the units can be selected to achieve the purpose of this embodiment according to actual needs.

[0179] In addition, the functional units in the various embodiments of the present invention can be integrated into one processing unit, or each unit can exist physically separately, or two or more units can be integrated into one unit.

[0180] If this function is implemented as a software functional unit and sold or used as an independent product, it can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of this invention, or the part that contributes to the prior art, or a part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods indicated in the various embodiments of this invention. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0181] Those skilled in the art will understand that all or part of the steps of the above-described method embodiments can be implemented by hardware related to program instructions. The aforementioned program can be stored in a computer-readable storage medium. When executed, the program performs the steps of the above-described method embodiments; and the aforementioned storage medium includes various media capable of storing program code, such as ROM, RAM, magnetic disks, or optical disks.

[0182] The technical solutions of this application have been described above with reference to the preferred embodiments shown in the accompanying drawings. However, it is readily understood by those skilled in the art that the scope of protection of this application is obviously not limited to these specific embodiments. The above embodiments are only used to illustrate the technical solutions of this application and are not intended to limit them. Although this application has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some or all of the technical features therein. These modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of this application.

Claims

1. An intelligent decision-making method for long-chain industrial tasks, characterized in that, The method includes: The timing and quantity of prefetching for the next round of industrial knowledge are determined through streaming retrieval. The industrial knowledge slice retrieval queue is determined based on the pre-fetching timing and the pre-fetching quantity; Based on the dual-loop fusion mechanism of thinking chain and decision chain, an industrial scheduling decision sequence is generated through dense vector retrieval and Monte Carlo tree search decision engine.

2. The method according to claim 1, characterized in that, The process of determining the timing and corresponding quantity of prefetching for the next round of industrial knowledge through streaming retrieval includes: Determine the expected inference time for this round and the expected retrieval time for the next round; The prefetch timing is calculated based on the expected inference time and the expected retrieval time. The expected inference time and the expected retrieval time are input into the adaptive prefetch depth algorithm to determine the prefetch quantity.

3. The method according to claim 2, characterized in that, The step of calculating the prefetch timing based on the expected inference time and the expected retrieval time includes: Obtain the actual reasoning time of the previous round and the corresponding historical expected reasoning time; Calculate the feedback factor based on the actual reasoning time and the historical expected reasoning time; Based on the Markov decision process, the prefetching timing is determined according to the expected inference time, the expected retrieval time, and the feedback factor.

4. The method according to claim 1, characterized in that, The aforementioned mechanism, based on the dual-loop fusion of thought chain and decision chain, generates industrial scheduling decision sequences through dense vector retrieval and Monte Carlo tree search decision-makers, including: The inner loop is started to perform dense vector retrieval on the industrial knowledge slice queue to generate a sequence of inference nodes; Start the outer loop and perform a Monte Carlo tree search based on the inference node sequence; External tools are invoked to generate decision subsequences, and the execution results are fed back to update the reward values ​​of the search tree nodes; Select the path with the highest cumulative reward value and output it as the industrial scheduling decision sequence.

5. The method according to claim 4, characterized in that, The inner loop specifically includes: Convert the current task intent into a query vector; Retrieve industrial knowledge slices that semantically match the query vector from the industrial knowledge slice queue, and use the matched industrial knowledge slices as the current inference nodes; Traverse all task intentions and combine all reasoning nodes into a reasoning node sequence according to the order of the task intentions.

6. The method according to claim 5, characterized in that, The outer loop specifically includes: Each inference node in the inference node sequence is used as a search tree node; For the current search tree node, an external tool is invoked to generate a decision subsequence, and the execution result of the decision subsequence is captured; If the execution result indicates an anomaly, the corresponding anomaly signal is injected into the current search tree node and the tool is invoked again; If the execution result indicates that it is normal, the Monte Carlo tree search reward value of the path is calculated, and the execution result is fed back to the inference node sequence for correction of subsequent nodes; After completing the search of all nodes, select the path with the highest cumulative reward value and output it as the industrial scheduling decision sequence.

7. The method according to claim 1, characterized in that, After obtaining the industrial scheduling decision sequence, the method further includes: Customized prompt words are synthesized using a multi-dimensional architectural feature industrial knowledge base; The industrial scheduling decision sequence is constrained and reconstructed using the customized prompt words to generate a target to be verified. The semantic equivalence of the target to be verified is evaluated using the CodeBLEU formula, and the industrial scheduling decision sequence is regenerated based on the evaluation results, or the high-level confidence result of the industrial scheduling decision sequence is output.

8. An intelligent decision-making device for long-chain industrial tasks, characterized in that, include: The acquisition module is used to determine the timing and corresponding quantity of prefetching for the next round of industrial knowledge through streaming retrieval. The determination module is used to determine the queue for receiving industrial knowledge slices based on the pre-fetching timing and the pre-fetching quantity; The control module is used to generate industrial scheduling decision sequences based on a dual-loop fusion mechanism of thought chain and decision chain, through dense vector retrieval and Monte Carlo tree search decision-maker.

9. An electronic device, characterized in that, include: A processor, and a memory communicatively connected to the processor; The memory stores computer-executed instructions; The processor executes computer execution instructions stored in the memory to implement the method as described in any one of claims 1 to 7.

10. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores computer-executable instructions, which, when executed by a processor, are used to implement the intelligent decision-making method for long-chain industrial tasks as described in any one of claims 1 to 7.

Images