Total element-oriented multi-agent collaborative resource optimization method in industrial internet

Abstract

The invention provides a total element-oriented multi-agent collaborative resource optimization method in an industrial internet, and aims to solve the key problems of difficulty in cross-layer coordination, low scheduling efficiency and the like caused by sensing, decision-making and execution of network work at remarkably different time elements (second level, minute level and hour level) in a multi-networked industrial system. According to the method, an MTPHN (full time element prediction hierarchical network) framework is constructed for a multi-networked industrial system environment with time isomerism, three core modules of feature pyramid prediction, cross-time element synchronization and cooperative scheduling are provided, and effective modeling and vertical-horizontal two-dimensional resource coordination of heterogeneous time streams are realized. According to the method, the production efficiency, the equipment utilization rate and the production value in different operation scenes can be effectively improved, and the method is suitable for dynamic resource scheduling and cross-level collaborative optimization application in a multi-networked industrial system.

Description

TECHNICAL FIELD

[0001] The present application relates to the field of industrial internet, and particularly relates to a multi-agent collaborative resource optimization method for all factors in industrial internet. BACKGROUND

[0002] Under the driving of the rapid evolution of industrial internet, modern manufacturing systems deeply integrate artificial intelligence and automation technology, and multiple network architectures have become a prominent feature of modern industrial systems. These systems realize intelligent manufacturing management of the whole chain from data acquisition to production control through the deep collaboration of perception networks, decision networks and execution networks. However, this multi-layer collaborative architecture introduces a fundamental challenge: time heterogeneity, that is, different functional network layers run on completely different time elements. Modern industrial systems exhibit unprecedented architectural complexity through multiple network connections. Its core architecture usually integrates three functional layers: a perception network for real-time monitoring and data acquisition (running at a second-level interval, 1-60 seconds), an intelligent decision network for optimization processing (running at a minute-level cycle, 60-1800 seconds), and an automation execution network for resource coordination (running at a hour-level period, 1800-7200 seconds), which leads to extreme time heterogeneity.

[0003] Firstly, this huge time element difference triggers a cascade of operational inefficiencies. Real-time data generated by the fast perception layer is outdated and ineffective before it can be effectively processed by the slower decision layer, and the execution layer often relies on static resource allocation schemes and cannot dynamically adapt to real-time changing production conditions, resulting in rigid resource utilization and response lag. Secondly, traditional resource scheduling models are mainly based on single-layer or single-time element optimization paradigm, which makes them fundamentally insufficient in dealing with the complex heterogeneous time dynamics between the perception-decision-execution layers. Existing methods have significant limitations in dealing with the challenge of time heterogeneity: multi-agent systems are widely used in industrial environments, but they usually assume uniform time granularity, making it difficult to effectively coordinate the heterogeneous time needs of different network layers; traditional hierarchical control methods can decompose manufacturing processes into different control layers, but lack a time element decomposition mechanism for multiple network characteristics; machine learning methods have shown potential in industrial scheduling, but standard sequential modeling architectures assume uniform time granularity, making it difficult to effectively capture and process all-time element characteristics; edge computing and service optimization technologies provide new architectural possibilities for industrial systems, but still lack in time heterogeneity coordination.

[0004] Therefore, developing a resource optimization model for multiple networked industrial systems within the Industrial Internet that can effectively coordinate the extreme temporal heterogeneity of perception, decision-making, and execution has become a core technical bottleneck that urgently needs to be overcome. This model must possess dynamic coordination capabilities across time elements, resolving the contradiction between rapid perception of data timeliness, decision-making processing delays, and static allocation at the execution layer. It must also overcome the fundamental limitations of existing single-time element optimization paradigms and various coordination technologies in handling the dynamics of all time elements. Constructing a new scheduling architecture that integrates a time element decomposition mechanism and possesses dynamic resource adaptation and real-time response capabilities is crucial for achieving efficient and adaptive operation of multiple networked industrial systems and improving the overall effectiveness of intelligent manufacturing. Summary of the Invention

[0005] Technical issues: Bottlenecks in heterogeneous time stream modeling: Traditional time series processing methods are based on the assumption of unified time elements and lack the ability to differentiate modeling of real-time perception flows (seconds), near-real-time decision information (minutes), and batch-level execution status (hours). It is difficult to design feature extraction and representation mechanisms for different time elements, resulting in incomplete capture of multi-dimensional time features and distorted representation of dynamic evolution laws.

[0006] Cross-layer temporal consistency maintenance bottleneck: The existing data stream transmission mechanism cannot solve the timeliness gap between the perception-decision-execution layers: the value of fast perception data is attenuated due to delays when it is transmitted to the slow decision-making layer, while the long-term static plans of the execution layer are difficult to absorb real-time adjustments, causing a disconnect between information timeliness and production strategies, resulting in resource mismatch and response delays.

[0007] Technical solution: This invention proposes a full-time factor prediction hierarchical network framework (MTPHN) for the full-time factor collaborative scheduling environment of the industrial Internet, which aims to solve the resource coordination problem of the three-layer perception-decision-execution network in modern industrial systems under the time factor differences. This framework focuses on solving the problems of low efficiency, severe response delay and poor scheduling adaptability caused by huge time differences in the operating cycles of the perception network, decision network and execution network, information transmission lag and static resource allocation in industrial scenarios. The technical solution mainly includes the following core steps and modules:

[0008] First, a feature pyramid prediction module is constructed to convert the original full-factor time series data into prediction information with uncertainty quantification (multi-view prediction results) through adaptive time block segmentation and hierarchical feature pyramid structure. Specifically: the module first divides the data from each layer into manageable time blocks through an adaptive time segment generation and embedding process, while retaining the dependencies between different operation scales. Then, a hierarchical feature pyramid structure is used to capture temporal patterns from multiple resolutions, and the most appropriate pyramid level is dynamically selected through the attention mechanism. Finally, the enhanced features are processed through a temporal convolutional network to generate multi-view prediction results that meet the time requirements of each layer. This process ensures the effective capture of rapidly changing perceptual data and provides accurate input basis for subsequent decision-making and execution.

[0009] Secondly, a cross-time element synchronization module is designed, which adopts a time-aware layered propagation architecture and dynamically adjusts the information transmission mechanism according to the inherent execution cycle of each network layer to ensure the consistency and timeliness of information across time elements (expired data filtering). Specifically: This module establishes a cross-layer communication priority queue and a timestamp matching mechanism to identify whether the perception data has expired or whether it still has decision-making value, thereby avoiding information invalidation caused by the mismatch between fast perception updates and slow decision-making processes. At the same time, a lightweight caching mechanism and an event-triggered transmission strategy are introduced to optimize the efficiency of information flow between networks with different time elements, thereby improving the responsiveness and resource utilization of the overall system. This helps to maintain the consistency and accuracy of cross-layer information, and prevents fast perception updates from becoming invalid too early or delays in slow decision-making processes from affecting the response efficiency of the overall system.

[0010] Finally, a collaborative scheduling module is developed to achieve resource coordination and task scheduling across time elements through an innovative vertical-horizontal dual decoupling mechanism. This module combines the prediction model with the scheduling strategy to realize the task feedback and adjustment mechanism from the execution layer to the decision layer and then to the perception layer in the vertical direction; and supports resource competition and collaboration between multiple subtasks or devices in the same layer in the horizontal direction. Specifically, it includes: predicting resource demand and task load in the future time period based on historical data and current status; automatically adjusting scheduling rules using reinforcement learning strategies to adapt to dynamic production conditions; and by introducing multi-objective optimization functions, taking into account multiple indicators such as task completion rate, resource utilization and system stability, to achieve efficient and flexible multi-cycle scheduling. In addition, in order to deal with the problem of full-time element sequential dependency that is difficult to handle with traditional scheduling models, the present invention further proposes a scheduling representation based on the time element decoupling coding mechanism to ensure that tasks can be accurately parsed and executed between different levels.

[0011] Through the synergistic effect of the above three core modules, the MTPHN framework realizes efficient scheduling and resource coordination of the perception-decision-execution three-layer network in industrial systems under highly time-heterogeneous environments. This framework not only improves the system's processing capacity for full-time factor data, but also significantly enhances the robustness and adaptability of the scheduling system.

[0012] Advantages:

[0013] (1) Improve production efficiency and resource utilization By adopting the hierarchical network (MTPHN) framework based on full-time factor prediction, parallel processing and feature extraction of heterogeneous data streams in the perception, decision, and execution three-layer network can be effectively realized, significantly improving production efficiency in different industrial scenarios, optimizing equipment utilization, and improving overall production value through accurate resource scheduling strategies.

[0014] (2) Support information synchronization and consistency maintenance across time factors The time-aware hierarchical propagation architecture is introduced, which can dynamically adjust the information transmission mechanism according to the inherent execution cycle of each network layer, ensuring the consistency and accuracy of information across time factors. Through the implementation of a lightweight caching mechanism and an event-triggered transmission strategy, not only does it solve the mismatch between fast perception updates and slow decision-making processes, but it also significantly reduces the rate of operational errors caused by information lag, ensuring efficient operation under real-time production conditions.

[0015] (3) Enhance the robustness and adaptability of the system Through the vertical-horizontal dual decoupling mechanism, resource coordination and task scheduling across time factors are realized, which can automatically adjust the scheduling rules to adapt to different production demands in a dynamically changing industrial environment. In particular, the use of reinforcement learning strategies for prediction and decision optimization enables the system to quickly respond to emergencies or task changes, and through a multi-objective optimization function, it takes into account multiple key indicators such as task completion rate, resource utilization, and system stability, enhancing the stability and emergency response capabilities of the entire system. BRIEF DESCRIPTION OF DRAWINGS

[0016] Figure 1 is a flowchart showing the feature pyramid prediction module of the hierarchical network framework based on full-time factor prediction in the industrial internet.

[0017] Figure 2 is a flowchart showing the cross-time factor synchronization module of the hierarchical network framework based on full-time factor prediction in the industrial internet.

[0018] Figure 3 is a flowchart showing the collaborative scheduling module of the hierarchical network framework based on full-time factor prediction in the industrial internet.

[0019] Figure 4 is the main principle diagram of the method of the present application. DETAILED DESCRIPTION

[0020] The present invention will be further described below with reference to the accompanying drawings and specific embodiments. It should be understood that the following specific embodiments are only used to illustrate the present invention and are not used to limit the scope of the present invention.

[0021] like Figure 4 As shown, the present invention provides a multi-agent collaborative resource optimization method for all factors in the industrial Internet. It is oriented to multiple networked industrial systems with temporal heterogeneity and adopts a full-time factor prediction hierarchical network MTPHN framework consisting of a feature pyramid prediction module, a cross-time factor synchronization module, and a collaborative scheduling module. The method includes the following steps:

[0022] (1) Formalization of the problem based on multi-agent constraints:

[0023] System model definition: Consider a multi-networked industrial system consisting of three types of heterogeneous functional network layers Each layer consists of distributed agents. Each layer l contains an agent network G l , respectively represent the perception, decision and execution networks, layer index Corresponding to three types of functional layers, time elements Different layers have different computational requirements and time intervals. The system operates under three typical time elements, namely τ P =1s(perception), τ D =60s(decision), τ E =3600s (execution), the heterogeneity ratio is Ω=τ E / τ P =3600.

[0024] Task scheduling and optimization objective function model: The goal is to minimize the expected loss of all agents while satisfying multi-agent constraints. The specific optimization objectives are:

[0025]

[0026] in is the state vector of agent i at layer l at time t, π i,l represents the scheduling strategy of agent i at layer l, l l is the loss function of layer l. The objective function minimizes the expected prediction error of all agents and network layers under time heterogeneity while ensuring that resource, safety, and capacity constraints are met.

[0027] Constraint restriction model: The constraints are as follows (total resource limit, response time safety limit, information freshness, cross-layer synchronization error, and time weight normalization):

[0028]

[0029] Among them C l is the resource consumption of layer l, C total is the total system resource capacity, is the actual response time of layer l, is the safety response time limit, is the actual state vector at time t, is (t-τ l ) time,∈ sync is the maximum allowable synchronization error, is the information freshness of layer l, θ min is the minimum freshness threshold, τ l is the characteristic time element of layer l operation, τ l′ / τ l is the inter-layer time scale ratio of layer l, w k is the timing weight, and l→l′ transitions are the cross-layer information transfer process from the source network layer l to the target network layer l′. The constraints ensure the consistency of cross-layer synchronization and the normalization of time weights in all-factor coordination.

[0030] MTPHN is mainly divided into three modules to solve the above-mentioned multi-agent constraint-based problems, with the following functions:

[0031] 1. Feature Pyramid Prediction Module: Generate Multi-temporal Predictions in is the predicted state at time (t+h), is the prediction uncertainty variance of layer l at time t.

[0032] 2. Cross-time element synchronization module: Generate synchronous forecasts

[0033] 3. Collaborative Scheduling Module: Generate Coordinated Actions in Represents the action of agent i at time t in layer l.

[0034] (2) Feature pyramid prediction module: Figure 1 As shown in Figure 2, it transforms raw full-factor time series data into forecast information with uncertainty quantification. This module addresses the challenge of simultaneously modeling temporal patterns with different operating time factors in multiple industrial networks.

[0035] Generation and embedding of time blocks: To achieve effective full-factor time series modeling, a time block-based approach is used to segment multi-agent time series data into manageable time blocks while preserving the dependencies between different operation time elements. are segmented into adaptive-sized time chunks whose size is determined by

[0036]

[0037] where L denotes the history window length, τ l is the characteristic temporal element of layer l. The logarithmic relationship ensures that the time chunk size grows sub-linearly with the temporal element, capturing key temporal patterns while preserving computational efficiency. The minimum value of 3 ensures sufficient local context, while the upper bound of L / 4 prevents over-segmentation of short sequences.

[0038] Each time chunk is linearly transformed and positionally encoded PositionalEncoding l to obtain the embedding

[0039]

[0040] where Linear(·) performs the time chunk projection, and the embedding captures the specific temporal and contextual information required for each layer operation. This process converts the full-element raw temporal data into a normalized representation that adapts to the operation requirements.

[0041] • Adaptive feature pyramid construction: A multi-resolution feature pyramid is constructed to capture multi-level temporal patterns, and the optimal pyramid level is selected through a dynamic mechanism. The time chunk embedding is processed through a four-layer pyramid structure, with each layer capturing temporal features of different granularity:

[0042]

[0043] where is the feature of pyramid layer k, Conv k denotes a convolutional layer with gradually increasing receptive fields. Each level targets specific temporal patterns: the first level for fine-grained patterns, the second level for short-term trends, the third level for medium-term patterns, and the fourth level for long-term dependencies.

[0044] A dynamic level selection mechanism according to the attention score Attention l adaptively selects the optimal pyramid level:

[0045]

[0046] where, is the context state vector of layer l at time t, and this attention mechanism can automatically adapt to different industrial conditions,

[0047] without the need for manual parameter adjustment, achieving automatic switching of pyramid levels without human intervention.

[0048] • Full-time element prediction generation: process the extracted pyramid features to generate full-time element predictions that meet the time requirements of each network layer and industry constraints.

[0049] Multi-agent interaction modeling: model the dependency between multiple agents through the CrossAttention mechanism, capturing the coordination patterns between different agents:

[0050]

[0051] where, is the interaction-enhanced feature at time t, is the optimal level feature selected from the feature pyramid at time t through the dynamic level selection mechanism, represents the feature set of all other agents in the same layer at time t; this method ensures that the prediction of a single agent considers the behavior and state of neighboring agents, improving coordination accuracy and preventing conflicting decisions.

[0052] Full-element prediction generation: enhanced features are processed through a TCN network to generate predictions in multiple time ranges:

[0053]

[0054] where, is the predicted state vector at time (t+h), is the interaction-enhanced feature at time t, h represents different prediction perspectives, and adapts to the time elements of each layer; the final output provides uncertainty quantification information simultaneously.

[0055] The feature pyramid prediction module adjusts the block size corresponding to each layer's time element, explicitly handling temporal heterogeneity and ensuring effective capture of rules spanning three orders of magnitude in time difference.

[0056] (3) Cross-time element synchronization module: as shown in Figure 2 , as a key bridge between prediction generation and decision coordination, it ensures that full-element predictions are correctly aligned and synchronized across different time domains, while considering information uncertainty and time sequence misalignment.

[0057] • Full-element information timeliness and quality assessment: to address the inherent time sequence misalignment in full-element industrial systems, an information freshness and quality assessment framework is proposed, adapting to different operational time elements.

[0058] Simplified information freshness model: a hierarchical feature-adaptive freshness decay model is used, and the information freshness is defined as:

[0059]

[0060] where λ l is the attenuation coefficient of each layer, β l is the uncertainty penalty factor, denotes the prediction uncertainty, and Δt is the time lag quantity. This formulation ensures that the information reliability is reduced for all valid inputs. The prediction uncertainty ∈ [0, 1] directly reduces the information reliability. This simplified model, while preserving the basic attenuation behavior, is computationally efficient and empirically validated.

[0061] Quality factor combines multiple industrial context indicators:

[0062]

[0063] where denotes the information relevance, denotes the prediction accuracy, denotes the data integrity, and the weights w rel +w acc +w comp = 1, w rel , w acc , w comp ≥ 0. The time lag quantity Δt = t current -t prediction ensures that the information freshness is appropriately reduced with increasing prediction age, while the context-aware component adjusts the attenuation coefficient according to the current industrial conditions, where t current is the time of the current decision, and t prediction is the original time of the prediction generation.

[0064] The information freshness exponentially decays with the time lag Δt = t current -t prediction . The attenuation coefficient λ l is dynamically adjusted according to the industrial environment:

[0065] λ l = λ base · (1 + β vol · Vol l + β crit · Crit l )

[0066] where λ base is the base attenuation rate, β vol is the volatility weight coefficient, Vol l is the scenario volatility indicator, β crit is the criticality weight coefficient, and Crit lIt is a safety criticality indicator; the attenuation parameter adapts to specific industrial scenarios. It can be seen from the formula that high variability processes will lead to accelerated information attenuation, while safety-critical operations will maintain the validity of information for a longer time.

[0067] Uncertainty-weighted cross-layer fusion: Predictions from different network layers are combined through a sophisticated fusion mechanism, while explicitly accounting for prediction uncertainty and temporal misalignment.

[0068] Time series alignment across time elements: To process the prediction alignment under different time elements, weighted aggregation is used:

[0069]

[0070] in, is the aggregated state information synchronized from the perception layer (P) to the decision layer (D) at time t, is the predicted state sequence of the perception layer agent i in the time window [tk, t], G P is the set of intelligent agents in the perception layer, and W max Maximum aggregation window length;

[0071] Uncertainty-aware information fusion: Exponential decay weighting is used to ensure that the weight of recent predictions is higher. The time characteristic formula of the adaptation layer is:

[0072]

[0073] in, w2=1-w1, is the fusion state information synchronized from layer l1 to layer l2, is the predicted state of the source layer l1 at time t, is the true state of the target layer l2 at time t.

[0074] Adaptive synchronization strategy: Implements a flexible synchronization mechanism to adapt to timing heterogeneity, balancing industrial safety and efficiency.

[0075] Priority information routing: Information priority (t) Defined as:

[0076]

[0077] Among them, α, β, and γ are balance parameters emphasizing freshness, certainty, and criticality respectively. is the freshness indicator, As an indicator, Crit l Indicates key indicators.

[0078] (4) Collaborative scheduling module: Figure 3As shown in Figure 1, the module, as the final coordination component of a multi-layer time series prediction-scheduling network, transforms the synchronization predictions output by the cross-time element synchronization module into coordinated scheduling decisions for the entire network layer. This module addresses the challenge of time series heterogeneity in cross-layer and intra-layer coordination within multi-layer industrial systems.

[0079] Dual decoupling coordination framework: The collaborative scheduling module adopts a dual decoupling coordination mechanism, including:

[0080] Vertical cross-layer coordination: managing asynchronous information flow between different time element layers through timing decision buffers.

[0081] Defining a decision buffer

[0082]

[0083] in To ensure complete temporal coverage, heterogeneous temporal elements require a decision buffering mechanism, where faster layers accumulate decisions for slower layers; is the scheduling decision of agent i in layer l at time tk, w k is the time decay weight of the historical decision at time t, is the freshness of decision information at time tk, is the set of agents in layer l. Specifically, the perception layer aggregates 1-30 decisions per decision cycle, while the decision layer accumulates 1-120 results per execution cycle.

[0084] To maintain temporal relevance, each buffered decision is assigned a temporal weight w k :

[0085]

[0086] Where k is the decision time offset, τ l The inherent temporal element of layer l; this formula reveals that the most recent decision has greater influence in cross-layer coordination.

[0087] The cross-temporal feature aggregation strategy combines temporality and freshness weighted decisions:

[0088]

[0089] in, is the aggregated strategy.

[0090] Horizontal coordination: Ensure consensus among agents in the same time layer through reputation-weighted coordination:

[0091]

[0092] where α l ∈[0,1] controls the coordination strength, represents the neighbor set of agent i in layer l, Q i,l is the local decision function of agent i in layer l, is the local state vector of agent i at layer l at time t, is the synchronized global state of layer l, is the scheduling decision of agent i in layer l at time t, is the scheduling decision of agent j in layer l at time t.

[0093] The reputation weighting coefficient is:

[0094]

[0095] Among them rep j,l (t) represents the reputation score of agent j in layer l at time t, is the neighbor set of agent i The score is the sum of the reputation values ​​of all neighbors in the task; this score is updated based on task completion success and resource efficiency.

[0096] Information propagation across temporal elements: A dual decoupling mechanism enables information flow between layers operating on different temporal elements through adaptive aggregation and interpolation.

[0097]

[0098] Among them, FastToSlow P→D (t) is the information propagation function from the perception layer P to the decision layer D, For the perception layer in the time window The decision buffer set within, w temporal is the time decay weight function, SlowToFast E→D (t) is the information propagation function from the slow layer E to the fast layer D, is the decision buffer set of the execution layer at time t;

[0099] In fast-to-slow aggregation, multiple fast layer decisions are combined using time-decaying weights, while slow-to-fast interpolation distributes slow layer guidance to multiple fast layer cycles.

[0100] The dual decoupling strategy maintains layer autonomy while ensuring cross-layer consistency through bounded synchronization:

[0101]

[0102] where ∈ sync represents the bounded synchronization error tolerance, α lControls the strength of horizontal coordination within each layer, is the cross-layer decision buffer set,

[0103] is the decision set of all agents in the layer, MinimizeSync is to minimize the cross-layer synchronization error, and MaximizeConsensus is to maximize the decision consensus within the layer.

[0104] Multi-cycle coordination strategy: Multi-cycle coordination operations have overlapping time periods corresponding to different time elements. Each period maintains independent resource allocation while being coordinated through dual decoupled interfaces.

[0105] Planning cycle aware resources To manage multiple overlapping periods k on different time elements:

[0106]

[0107] Each layer l has its characteristic time element τ l Operate in a way that enables independent cycle management while maintaining coordination through a dual decoupling mechanism.

[0108] Buffer management across cycles To handle time mismatches between cycles:

[0109]

[0110] where α l Balance current use with future planning, and adjust according to the time characteristics of each layer. is the actual resource occupancy rate of layer l at time t, is the layer l in the next cycle (t+τ l )’s pre-planned resource requirements; ultimately achieving collaborative processing of time element differences from seconds to hours.

[0111] The collaborative scheduling module directly addresses temporal heterogeneity through three coordination stages that handle temporal element differences ranging from seconds to hours: full-element decomposition processing adapts to the temporal characteristics of each layer, cross-temporal element synchronization maintains information relevance in temporal element conversion, and double-decoupled scheduling enables autonomous coordination decisions within the time range.

Claims

1. A multi-agent collaborative resource optimization method for all factors in the industrial Internet, characterized by: Aiming at multiple networked industrial systems with temporal heterogeneity, a full-time factor prediction hierarchical network (MTPHN) framework consisting of a feature pyramid prediction module, a cross-time factor synchronization module, and a collaborative scheduling module is adopted. The specific steps are as follows: first, a feature pyramid prediction module is constructed to process the original full-factor time series data through adaptive time block segmentation and hierarchical feature pyramid structure to generate multi-view prediction results with uncertainty quantification; second, a cross-time factor synchronization module is designed to adopt a time-aware hierarchical propagation architecture and a lightweight caching mechanism to achieve information consistency maintenance and expired data filtering across time factors; finally, a collaborative scheduling module is developed to complete multi-objective optimized resource coordination and task scheduling through a vertical-horizontal dual decoupling mechanism combined with a reinforcement learning strategy.

2. The multi-agent collaborative resource optimization method for all factors in the Industrial Internet according to claim 1 is characterized by: The multi-networked industrial system includes three types of heterogeneous functional network layers Each layer consists of distributed agents; each layer l contains an agent network G l , respectively represent the perception, decision and execution networks, layer index Corresponding to three types of functional layers, the time element T={τ P ,τ D ,τ E }, different layers have different computing requirements and time intervals; the system operates under three typical time elements, namely τ P =1s, τ D =60s, τ E =3600s, the isomerism ratio is Ω=τ E / τ P =3600; The goal of task scheduling and optimization is to minimize the expected loss of all agents while satisfying the multi-agent constraints. The specific formula is: in is the state vector of agent i at layer l at time t, π i,l represents the scheduling strategy of agent i at layer l, l l is the loss function of layer l; The objective function is to minimize the expected prediction error of all agents and network layers under time-heterogeneous conditions while ensuring that resource, safety, and capacity constraints are met; Constraints include: Total resource limit: Response time safety margin: Information freshness: Cross-layer synchronization error: Time weight normalization: Among them C l is the resource consumption of layer l, C total is the total system resource capacity, is the actual response time of layer l, is the safety response time limit, is the actual state vector at time t, is (t-τ l ) time,∈ sync is the maximum allowable synchronization error, is the information freshness of layer l, θ min is the minimum freshness threshold, τ l is the characteristic time element of layer l operation, τ l′ / τ l is the inter-layer time scale ratio of layer l, w k is the timing weight, l→l′ transitions is the cross-layer information transmission process from the source network layer l to the target network layer l′; the constraints ensure the cross-layer synchronization consistency and the normalization of the time weight in the full-factor coordination.

3. The method for multi-agent collaborative resource optimization in the industrial Internet for all factors according to claim 2 is characterized by: feature pyramid Generation and embedding of time blocks in the prediction module: To achieve effective full-factor time series modeling, a time block-based approach is used to segment multi-agent time series data into manageable time blocks while preserving the dependencies between different operation time elements; for each layer l, the time series is divided into adaptively sized time blocks of size p l Determined by the following formula: Where L represents the length of the history window, τ l is the characteristic temporal element of layer l; the logarithmic relationship ensures that the temporal block size grows sublinearly with the temporal element, capturing key temporal patterns while maintaining computational efficiency; the minimum value of 3 ensures sufficient local context, while the upper limit of L / 4 prevents over-segmentation of short sequences; Each time block After linear transformation and positional encoding l After embedding Among them, Linear(·) performs time block projection, embedding the timing and context information specific to each layer's operational requirements, and transforming all-factor raw time series data into a standardized representation that adapts to operational requirements.

4. The method for multi-agent collaborative resource optimization in the industrial Internet for all factors according to claim 3 is characterized by: feature pyramid Adaptive feature pyramid construction in the prediction module: A multi-resolution feature pyramid is constructed to capture multi-level temporal patterns, and the optimal pyramid level is selected through a dynamic mechanism. Temporal block embedding is processed through a four-layer pyramid structure, with each layer capturing temporal features of different granularity. in is the feature of pyramid layer k at time t, Conv k Represents convolutional layers with progressively larger receptive fields; each level targets a specific temporal pattern: level 1 for fine-grained patterns, level 2 for short-term trends, level 3 for medium-term patterns, and level 4 for long-term dependencies; Dynamic level selection mechanism based on attention score Attention l Adaptively select the optimal pyramid level: in, is the context state vector of layer l at time t. This attention mechanism can automatically adapt to different industrial conditions without manual parameter adjustment, and realizes adaptive switching of pyramid levels without human intervention.

5. The method for multi-agent collaborative resource optimization in the industrial Internet for all factors according to claim 4 is characterized by: feature pyramid Full-time element forecast generation of the forecast module: Processes the extracted pyramid features to generate full-time element forecasts that meet the time requirements of each network layer and industrial constraints; Multi-agent interaction modeling: CrossAttention is used to model the dependencies between multiple agents and capture the collaborative patterns between different agents. in, is the feature after interactive enhancement at time t, is the optimal level feature selected from the feature pyramid by the dynamic level selection mechanism at time t, represents the feature set of all other agents in the same layer at time t; this method ensures that the prediction of a single agent takes into account the behavior and state of neighboring agents, improves coordination accuracy and prevents conflicting decisions; Full-factor prediction generation: The enhanced features are processed by a temporal convolutional network (TCN) to generate predictions at multiple time scales: in, is the predicted state vector at time (t+h), is the feature after interactive enhancement at time t, h represents different prediction perspectives, adapts to the time elements of each layer, and the final output result synchronously provides uncertainty quantification information.

6. The method for multi-agent collaborative resource optimization in the industrial internet for all factors according to claim 5 is characterized by: Full-factor information timeliness and quality assessment in the cross-time factor synchronization module: To address the inherent temporal misalignment in full-factor industrial systems, a framework for information freshness and quality assessment is proposed to adapt to different operational time factors; Simplified information freshness model: Adopting hierarchical feature adaptive freshness decay model, information freshness Defined as: where λ l is the attenuation coefficient of each layer, β l is the uncertainty penalty factor, represents the forecast uncertainty, Δt is the time lag; this formula ensures holds for all valid inputs; prediction uncertainty This directly reduces the reliability of information; this simplified model is computationally efficient and empirically validated while maintaining the basic decay behavior; Quality Factor Combining multiple industry context indicators: in Indicates information relevance, represents the prediction accuracy, Indicates data integrity, and weight w rel +w acc +w comp =1,w rel , w acc , w comp ≥0; Time lag Δt = t current -t prediction Ensure that the information freshness decreases appropriately with the increase of the forecast age, and the context-aware component adjusts the attenuation coefficient according to the current industrial conditions, where t current is the time of the current decision moment, t prediction The original time when the prediction was generated; the freshness of the information lags behind with time Δt=t current -t prediction exponential decay; Attenuation coefficient λ l Dynamic adjustment according to industrial environment: l l =λ base ·(1+β vol ·Volume l +b crit ·Crit l ) where λ base is the basic decay rate, β vol is the volatility weight coefficient, Vol l is the scenario volatility indicator, β crit is the critical weight coefficient, Crit l It is a safety-critical indicator; the attenuation parameter adapts to specific industrial scenarios. The formula shows that high variability processes will lead to accelerated information attenuation, while safety-critical operations will maintain the validity of information for a longer time.

7. The method for multi-agent collaborative resource optimization in the industrial internet for all factors according to claim 6 is characterized by: Uncertainty-weighted cross-layer fusion and adaptive synchronization strategy for cross-temporal element synchronization modules; Uncertainty-weighted cross-layer fusion: This combines predictions from different network layers through a fusion mechanism, while explicitly accounting for prediction uncertainty and temporal misalignment. Time series alignment across time elements: To process the prediction alignment under different time elements, weighted aggregation is used. The formula is as follows: in, is the aggregated state information synchronized from the perception layer P to the decision layer D at time t, is the predicted state sequence of the perception layer agent i in the time window [tk, t], G P is the set of intelligent agents in the perception layer, W max is the maximum aggregation window length; Uncertainty-aware information fusion: Exponential decay weighting is used to ensure that the weight of recent predictions is higher. The time characteristic formula of the adaptation layer is: in, w2=1-w1, is the fusion state information synchronized from layer l1 to layer l2, is the predicted state of the source layer l1 at time t, is the true state of the target layer l2 at time t; Adaptive synchronization strategy: Implements a flexible synchronization mechanism to adapt to timing heterogeneity, balancing industrial safety and efficiency; Priority information routing: Information priority (t) Defined as: Among them, α, β, and γ are balance parameters emphasizing freshness, certainty, and criticality respectively. is the freshness indicator, As an indicator, Crit l Indicates key indicators.

8. The method for multi-agent collaborative resource optimization for all factors in the industrial Internet according to claim 7 is characterized by: The dual decoupling coordination framework of the collaborative scheduling module specifically includes: Vertical cross-layer coordination: Manage asynchronous information flow between different time element layers through timing decision buffers; define decision buffers in To ensure complete temporal coverage, heterogeneous temporal elements require a decision buffering mechanism, where faster layers accumulate decisions for slower layers; is the scheduling decision of agent i in layer l at time tk, w k is the time decay weight of the historical decision at time t, is the freshness of decision information at time tk, is the set of agents in layer l; To maintain temporal relevance, each buffered decision is assigned a temporal weight w k : Where k is the decision time offset, τ l The inherent temporal element of layer l; this formula reveals that the most recent decision has a greater influence in cross-layer coordination; The cross-temporal feature aggregation strategy combines temporality and freshness weighted decisions: in, is the aggregated strategy; Horizontal intra-layer coordination: Ensure consensus among agents in the same time layer through reputation-weighted coordination: where α l ∈[0,1] controls the coordination strength, represents the neighbor set of agent i in layer l, Q i,l is the local decision function of agent i in layer l, is the local state vector of agent i at layer l at time t, is the synchronized global state of layer l, is the scheduling decision of agent i in layer l at time t, is the scheduling decision of agent j in layer l at time t; Reputation weighting coefficient for: Among them rep j,l (t) represents the reputation score of agent j in layer l at time t, is the neighbor set of agent i The score is the sum of the reputation values ​​of all neighbors in the task; this score is updated based on task completion success and resource efficiency.

9. The method for multi-agent collaborative resource optimization in the industrial internet for all factors according to claim 8 is characterized by: Cross-time element information propagation of collaborative scheduling module: The dual decoupling mechanism enables information flow between layers operating on different time elements through adaptive aggregation and interpolation. The formula is as follows: Among them, FastToSlow P→D (t) is the information propagation function from the perception layer P to the decision layer D, For the perception layer in the time window The decision buffer set within, w temporal is the time decay weight function, SlowToFast E→D (t) is the information propagation function from the slow layer E to the fast layer D, is the decision buffer set of the execution layer at time t; In fast-to-slow aggregation, multiple fast-layer decisions are combined using time-decaying weights, while slow-to-fast interpolation distributes slow-layer guidance to multiple fast-layer cycles; The dual decoupling strategy maintains layer autonomy while ensuring cross-layer consistency through bounded synchronization: where ∈ sync represents the bounded synchronization error tolerance, α l Controls the strength of horizontal coordination within each layer, is the cross-layer decision buffer set, is the decision set of all agents in the layer, MinimizeSync is to minimize the cross-layer synchronization error, and MaximizeConsensus is to maximize the decision consensus within the layer.

10. The method for multi-agent collaborative resource optimization in the industrial internet for all factors according to claim 9 is characterized by: Multi-cycle coordination strategy of the collaborative scheduling module: Multi-cycle coordination operations have overlapping time periods corresponding to different time elements; each cycle maintains independent resource allocation while being coordinated through dual decoupled interfaces; Planning cycle aware resources To manage multiple overlapping periods k on different time elements, the formula is as follows: Each layer l has its characteristic time element τ l Operate to achieve independent cycle management while maintaining coordination through a dual decoupling mechanism; Buffer management across cycles To handle the time mismatch between cycles, the formula is as follows: where α l Balance current use with future planning, and adjust according to the time characteristics of each layer. is the actual resource occupancy rate of layer l at time t, is the layer l in the next cycle (t+τ l )’s pre-planned resource requirements; ultimately achieving collaborative processing of time element differences from seconds to hours.

Images