Intelligent factory intelligent management system and method

Abstract

This invention discloses an intelligent management system and method for smart factories. The system includes six core functional modules: a unified data platform, a collaborative optimization engine, production scheduling and optimization, and predictive equipment maintenance. The data platform aggregates and processes multi-source heterogeneous data from the IoT sensing layer and enterprise information systems, constructing a real-time updated digital twin of the factory. The core functional modules make local intelligent decisions based on this data, and the collaborative optimization engine coordinates the various modules to achieve global optimization. The system resolves subsystem goal conflicts through deep cross-module linkage, integrates advanced AI algorithms such as DRL and causal discovery to achieve intelligent decision-making, enhances adaptability to internal and external disturbances through online learning, and designs dedicated algorithms for the pain points of high-precision manufacturing industries. The overall benefits far exceed the independent optimization and superposition of subsystems, effectively improving production efficiency, reducing operating costs, and ensuring product quality.

Description

Technical Field

[0001] This invention belongs to the field of ultra-thin flexible glass technology. Specifically, this invention relates to an intelligent management system and method for a smart factory. Background Technology

[0002] With the advancement of Industry 4.0 and smart manufacturing, traditional factory stand-alone automation systems (such as MES, WMS, and EMS) are no longer sufficient to meet the demands of efficient, flexible, and lean production. Existing technological solutions mostly focus on optimizing single processes, for example:

[0003] 1. A genetic algorithm is used for static production scheduling, but it cannot dynamically respond to disturbances such as order insertions or equipment failures.

[0004] 2. Threshold-based alarm systems are considered reactive maintenance and cannot achieve predictive maintenance.

[0005] 3. Quality inspections relying on manual labor or traditional machine vision have a high rate of missed inspections and cannot trace the root cause of defects.

[0006] 4. Each subsystem (production, logistics, energy) operates independently, forming "information silos," and decision-making conflicts lead to low overall resource utilization.

[0007] Especially in high-end manufacturing industries such as ultra-thin flexible glass and precision electronics, higher requirements are placed on the coordination, stability and refined management of the entire process.

[0008] This invention provides an intelligent management system for smart factories, specifically addressing how to break down barriers, achieve data-driven and global dynamic optimization, and realize optimized production scheduling, predictive equipment maintenance, full product quality inspection, automated logistics and warehousing scheduling, intelligent supply chain collaboration, and refined energy management, ultimately aiming to improve production efficiency, reduce operating costs, ensure product quality, and enhance enterprise competitiveness. Summary of the Invention

[0009] This invention provides an intelligent management system for smart factories, which aims to optimize production scheduling, predictive equipment maintenance, full product quality inspection, automatic logistics and warehousing scheduling, intelligent supply chain collaboration, and refined energy management, thereby improving production efficiency, reducing operating costs, and ensuring product quality.

[0010] To achieve the above objectives, the technical solution adopted by the present invention is: an intelligent factory management system, comprising:

[0011] A unified data platform is used to aggregate and process multi-source heterogeneous data from the IoT sensing layer and enterprise information systems, and to build a real-time updated digital twin of the factory.

[0012] A collaborative optimization engine, based on a multi-agent system and a multi-objective optimization framework, is used to coordinate the decisions of various functional modules to achieve the globally optimal objective; and

[0013] The system comprises six core functional modules: production scheduling and optimization, predictive equipment maintenance, quality inspection and control, logistics and warehousing management, supply chain optimization, and energy management. All six modules are communicatively connected to the unified data platform and the collaborative optimization engine. Each module makes local intelligent decisions based on data from the unified data platform and receives coordination instructions from the collaborative optimization engine to achieve dynamic cross-module linkage and global optimization.

[0014] The smart factory integrated optimization management system also includes an online learning and model management platform, which communicates with the six core functional modules and the collaborative optimization engine to continuously monitor algorithm performance, automatically trigger model retraining and A / B testing deployment, and realize the system's self-evolution.

[0015] The multi-source heterogeneous data includes device sensor data, industrial visual image data, mobile device operation data collected by the IoT sensing layer, as well as production plan data, order data, inventory data, supply chain data, and energy consumption data from enterprise information systems.

[0016] The collaborative optimization engine evaluates the impact of local optimization decision schemes output by each core functional module on the global objective through iterative negotiation or centralized optimization, and generates a global collaborative decision instruction set.

[0017] The production scheduling and optimization module is a two-layer dynamic scheduler, including an upper-layer scheduling unit and a lower-layer scheduling unit;

[0018] The upper-level scheduling unit is based on deep reinforcement learning and is used for order priority allocation and production shift recommendation.

[0019] The lower-level scheduling unit adopts a genetic algorithm with hybrid variable neighborhood search, receives constraint information output by the upper-level scheduling unit to perform process-level scheduling, and integrates real-time energy consumption cost parameters into the fitness function of the lower-level scheduling unit.

[0020] The equipment predictive maintenance module includes a fault prediction model, which is constructed based on multi-scale spatiotemporal learning and transfer learning, integrates a spatiotemporal feature extraction unit and an equipment topology influence capture unit, and adopts a meta-learning framework; the spatiotemporal feature extraction unit is used to extract spatiotemporal features from multi-source sensor data, and the equipment topology influence capture unit is used to capture the topological correlation influence between devices.

[0021] The quality inspection and control module includes a visual inspection subsystem and a process parameter causal analysis subsystem. The visual inspection subsystem is based on the Transformer model and metric learning and is used for product defect identification. The process parameter causal analysis subsystem is based on a causal discovery algorithm and is used to construct a process parameter causal graph and trace the root cause of defects.

[0022] The logistics and warehousing management module adopts a multi-agent deep reinforcement learning algorithm based on a centralized training and decentralized execution framework for dynamically scheduling logistics and warehousing equipment.

[0023] The reward function of the multi-agent deep reinforcement learning algorithm includes parameters related to task completion, path efficiency, collision avoidance, driving stability, and cleanliness maintenance, and the algorithm is pre-trained in the factory digital twin.

[0024] The supply chain optimization module includes a demand forecasting model and a flexible strategy generation unit;

[0025] The demand forecasting model is built based on federated learning and is used for joint training with supply chain partners;

[0026] The elastic strategy generation unit is based on a robust optimization method and is used to generate procurement and delivery strategies to cope with uncertainties in demand, supply and logistics.

[0027] This invention also provides a smart factory management method, applied to the aforementioned smart factory integrated optimization management system, the method comprising the following steps:

[0028] S1: Collect production site data through the IoT sensing layer and synchronize it with enterprise information system data, aggregate it into a unified data platform, and build and update the factory digital twin;

[0029] S2: Each core functional module runs a core intelligent algorithm based on the real-time and historical data of the factory digital twin to generate a local optimization decision scheme;

[0030] S3: The collaborative optimization engine evaluates the impact of each local optimization decision scheme on the global objective, and generates a global collaborative decision instruction set through iterative negotiation or centralized optimization.

[0031] S4: Send the global collaborative decision instruction set to the execution layer to drive the physical factory to run;

[0032] S5: Collect the execution results and new production data of the global collaborative decision-making instruction set, feed them back to the unified data platform and the algorithm models of each core functional module to form a closed loop, and drive the model to achieve self-evolution through online learning and model management platform.

[0033] The intelligent factory management system of the present invention has the following beneficial effects:

[0034] 1. Global Dynamic Optimization: Through the collaborative optimization engine and the deep linkage mechanism between modules, global dynamic optimization is achieved across production, equipment, quality, logistics, supply chain and energy fields, and the problem of subsystem goal conflict is solved.

[0035] 2. Intelligent and forward-looking decision-making: It innovatively applies a series of advanced AI algorithms such as DRL, causal discovery, federated learning, and physical information neural networks, enabling the system to have high-level intelligence from prediction and diagnosis to optimization, and realizing the transformation from passive response to proactive intervention.

[0036] 3. Strong adaptability and resilience: The system can effectively cope with internal disturbances (such as equipment failure) and external uncertainties (such as market demand fluctuations and supply chain disruptions), and continuously evolve through online learning mechanisms to meet the needs of flexible manufacturing.

[0037] 4. Focusing on industry pain points: For the stringent requirements of high-precision manufacturing industries (such as ultra-thin glass) on quality, cleanliness, and material integrity, specialized algorithms (such as AGV scheduling with stability rewards and causal root cause analysis of quality) have been designed to solve industry-specific problems.

[0038] 5. Generates synergistic added value: The overall benefits generated by system integration far exceed the simple sum of the benefits of independent optimization of each subsystem, reflecting the holistic advantages of systems engineering. Attached Figure Description

[0039] This manual includes the following figures, which illustrate the following:

[0040] Figure 1 This is a schematic diagram of the overall architecture of the intelligent factory integrated optimization management system of the present invention. Detailed Implementation

[0041] The specific embodiments of the present invention will be further described in detail below with reference to the accompanying drawings, in order to help those skilled in the art to have a more complete, accurate and in-depth understanding of the concept and technical solutions of the present invention, and to facilitate its implementation.

[0042] Firstly, such as Figure 1 As shown, this embodiment of the invention provides an intelligent management system for a smart factory, including:

[0043] A unified data platform is used to aggregate and process multi-source heterogeneous data from the IoT sensing layer and enterprise information systems, and to build a real-time updated digital twin of the factory.

[0044] A collaborative optimization engine, based on a multi-agent system and a multi-objective optimization framework, is used to coordinate the decisions of various functional modules to achieve the globally optimal objective; and

[0045] The six core functional modules are production scheduling and optimization, predictive maintenance, quality inspection and control, logistics and warehousing management, supply chain optimization, and energy management. All six core functional modules are connected to a unified data platform and a collaborative optimization engine. Each core functional module makes local intelligent decisions based on data from the unified data platform and receives coordination instructions from the collaborative optimization engine to achieve dynamic linkage and global optimization across modules.

[0046] Specifically, embodiments of the present invention provide a highly integrated, data-driven, and autonomously decision-making and optimization-enabled intelligent factory management system. This system aims to achieve optimized production scheduling, predictive equipment maintenance, full product quality inspection, automated logistics and warehousing scheduling, intelligent supply chain collaboration, and refined energy management through artificial intelligence algorithms, ultimately improving production efficiency, reducing operating costs, ensuring product quality, and enhancing enterprise competitiveness.

[0047] The smart factory integrated optimization management system of this invention also includes an online learning and model management platform, which is connected to six core functional modules and a collaborative optimization engine to continuously monitor algorithm performance, automatically trigger model retraining and A / B testing deployment, and realize the self-evolution of the system.

[0048] In this embodiment of the invention, the multi-source heterogeneous data includes device sensor data, industrial visual image data, mobile device operation data collected by the Internet of Things sensing layer, as well as production plan data, order data, inventory data, supply chain data, and energy consumption data from the enterprise information system.

[0049] In this embodiment of the invention, the collaborative optimization engine evaluates the impact of the local optimization decision schemes output by each core functional module on the global objective through iterative negotiation or centralized optimization, and generates a global collaborative decision instruction set.

[0050] In this embodiment of the invention, the production scheduling and optimization module is a two-layer dynamic scheduler that combines deep reinforcement learning (DRL) with a hybrid metaheuristic algorithm, including an upper-layer scheduling unit and a lower-layer scheduling unit. The upper-layer scheduling unit is an upper-layer DRL agent (such as based on DDPG), which learns to prioritize orders and recommend production shifts in a dynamic environment based on deep reinforcement learning. The lower-layer scheduling unit uses a genetic algorithm with hybrid variable neighborhood search, receives constraint information output by the upper-layer scheduling unit to perform process-level scheduling, and integrates real-time energy consumption cost parameters into the fitness function of the lower-layer scheduling unit.

[0051] 1.1 Overall Algorithm Architecture and Design Principles

[0052] In this embodiment of the invention, the production scheduling and optimization module adopts a two-layer collaborative optimization framework, which decomposes the complex production scheduling problem into a strategic decision-making layer and a tactical execution layer, thereby achieving an effective unification of macro scheduling objectives and micro process optimization.

[0053] 1.2 Strategic Decision-Making Layer: Design and Optimization of Deep Deterministic Policy Gradient (DDPG) Agents

[0054] 1.2.1 State Space Construction

[0055] The state space S is defined as a composite vector containing multiple dimensions:

[0056]

[0057] in:

[0058] O (Order Features): Contains feature vectors for n orders. Each order feature includes:

[0059]

[0060] E (Device Status): Contains status information for m devices:

[0061] R (Resource Availability): Key material inventory levels, energy quotas, human resources, etc.

[0062] C (Cost Environment): Time-of-use electricity price curve, transportation cost coefficient, quality loss cost, etc.

[0063] P (Production Environment): Environmental parameters such as workshop temperature, humidity, and cleanliness level.

[0064] 1.2.2 Motion Space Design

[0065] Action space A is a continuous multidimensional space:

[0066]

[0067] in:

[0068] α (Priority Weight): Assign a priority weight α to each order. i ∈ [0,1] indicates its relative importance in scheduling.

[0069] β (Production Shift Recommendation): Recommends a production time window for each order, coded as follows:

[0070]

[0071] γ (Resource allocation tendency): Represents the allocation tendency vector when resources conflict.

[0072] 1.2.3 Reward Function Design

[0073] The compound reward function R contains multiple optimization objectives:

[0074]

[0075] in:

[0076] R_efficiency (efficiency bonus): A bonus based on Overall Equipment Effectiveness (OEE).

[0077] R_efficiency = (Actual OEE - Baseline OEE) / Baseline OEE

[0078] R_timeliness (Timeliness Bonus): A bonus based on the on-time delivery rate of orders.

[0079] R_timeliness = ∑(On-time delivery order weight) - λ•∑(Delay penalty)

[0080] R_cost (cost incentive): A negative incentive based on total production cost.

[0081] R_cost = -(Energy cost + Material cost + Labor cost) / Budgeted cost

[0082] R_quality (quality reward): A reward based on the first pass rate.

[0083] R_stability (stability reward): penalizes large changes in production plans and encourages continuous production.

[0084] The weighting coefficients W1-W5 are dynamically adjusted through multi-objective Pareto optimization to adapt to the main objectives of different production stages.

[0085] 1.2.4 Network Architecture Optimization

[0086] Actor-Critic Network Enhanced with Attention Mechanism:

[0087] Actor Network: It adopts a 3-layer fully connected network, and the last layer uses the Tanh activation function to output a normalized action vector.

[0088] Critic Network: Introduces a multi-head attention mechanism, enabling the network to focus on the most critical order and device features in the current state.

[0089] Experience replay: The Priority Experience Replay (PER) mechanism is adopted to allocate sampling priority according to TD error, thereby improving learning efficiency.

[0090] Target network: Use a soft update strategy to improve training stability.

[0091] 1.3 Tactical Execution Layer: Detailed Description of the Improved Hybrid Metaheuristic Algorithm

[0092] 1.3.1 Problem Modeling and Coding Scheme

[0093] Extended model of the Flexible Job Shop Scheduling Problem (FJSP):

[0094] Process set: J = { J1, J2, ..., J n}

[0095] Machine set: M = {M1, M2, ..., M} m}

[0096] Each process J i Available in machine subsets Upward processing, processing time p ij It varies depending on the machine.

[0097] Extended constraints: equipment maintenance window, material arrival time, energy consumption limits, cleanliness requirements.

[0098] Double-layer chromosome coding scheme:

[0099]

[0100] Gene sequence layer: based on process coding, each gene represents a process, and the processes for the same workpiece are arranged in the process sequence.

[0101] Machine allocation layer: Assign processing machines to each process, taking into account differences in machine capabilities and energy consumption characteristics.

[0102] Time parameter layer: The start time offset of the encoding process, used for time window optimization.

[0103] 1.3.2 Improved Adaptive Genetic Algorithm (IAGA)

[0104] 1. Adaptive parameter adjustment mechanism:

[0105]

[0106] in:

[0107] P_c(t), P_m(t): Crossover and mutation probabilities in generation t.

[0108] g(t): The variance of the fitness of the current generation population, which measures population diversity.

[0109] d(t): The continuous unimproved algebra of the current optimal solution

[0110] k: decay coefficient, G_max: maximum number of generations

[0111] 2. Elite Retention Strategy:

[0112] The top E individuals from each generation are retained and directly enter the next generation to prevent the loss of excellent genes.

[0113] 3. Enhanced selection operator:

[0114] Tournament selection: Randomly select k individuals, and choose the one with the highest fitness.

[0115] Roulette wheel selection: fitness ratio selection, preserving population diversity.

[0116] The two selection methods can be used in combination in a certain proportion.

[0117] 4. Multi-point priority retention of crossover (MPOX):

[0118]

[0119] Innovation: The number of crossover points is dynamically adjusted, and the optimal segmentation is determined based on the correlation analysis of gene blocks.

[0120] 5. Composite mutation operator:

[0121] Exchange mutation: Two gene loci are randomly selected and exchanged.

[0122] Insertion mutation: Randomly select a gene and insert it into another random location.

[0123] Reverse mutation: Randomly select a fragment and reverse the genetic sequence.

[0124] Adaptive machine redistribution variation: Based on machine load balancing, processes are redistributed to more suitable machines.

[0125] 1.3.3 Deep Optimization of Variable Neighborhood Search (VNS)

[0126] 1. Multi-level neighborhood structure design:

[0127]

[0128] 2. Adaptive Neighborhood Selection Strategy:

[0129] Record the historical success rate of improvements to the structure of each neighborhood:

[0130]

[0131] The success rate statistics are updated regularly, and the neighborhood selection probability is dynamically adjusted.

[0132] 3. Tabu Search-Based VNS Enhancement:

[0133] Introducing a short-term memory mechanism avoids cyclical searching around local optima:

[0134] Tag list: Records the characteristics of recently accessed solutions (such as the sequence of critical steps);

[0135] Desire criterion: Accept a candidate solution if it is better than the historical best;

[0136] Taboo length: adaptively adjusted based on the complexity of the search space;

[0137] 1.3.4 Improvement of Simulated Annealing Acceptance Criteria

[0138] 1. Adaptive temperature scheduling:

[0139]

[0140] 2. Dynamic acceptance probability:

[0141] It considers not only the difference in the objective function Δf, but also the degree of constraint violation Δc:

[0142]

[0143] The weights W1 and W2 are dynamically adjusted based on the importance of the constraints.

[0144] 3. Restart mechanism:

[0145] If no new solution is accepted after R consecutive iterations, the search process is restarted with a certain probability:

[0146] Keep the current optimal solution

[0147] Reinitialize other solutions

[0148] Increase the temperature appropriately to expand the search area.

[0149] 1.3.5 Multi-objective optimization processing

[0150] 1. Pareto front maintenance:

[0151] The algorithm maintains an external archive to store non-dominated solutions:

[0152] File update strategy: Add a new solution to the file when it is not dominated by any solution in the file;

[0153] File pruning mechanism: When the file size exceeds the upper limit, the solution in the densest region is deleted based on the crowding distance;

[0154] Elite solution selection: Select representative solutions from the archives to participate in the genetic operation.

[0155] 2. Reference Point Method (NSGA-III Improvement):

[0156] Define a set of reference points in the target space to guide the search toward a uniformly distributed Pareto front:

[0157]

[0158] Ensure that the algorithm obtains a balanced solution set among multiple competing objectives.

[0159] 1.4 Two-layer collaborative mechanism and information interaction

[0160] 1.4.1 Constraint Transmission and Transformation

[0161] The abstract decisions output from the strategic level need to be translated into concrete constraints at the tactical level:

[0162]

[0163] 1.4.2 Feedback Learning Loop

[0164] The execution results from the tactical layer are fed back to the strategic layer to update the DRL agent.

[0165] 1. Execution Effectiveness Evaluation: KPI data after the actual scheduling plan is executed;

[0166] 2. Constraint Feasibility Analysis: Identify situations where overly strict constraints lead to no feasible solution;

[0167] 3. Strategy Value Update: Adjust the state-action value function estimation based on actual results;

[0168] 4. Parameter Adaptation: Adjust algorithm parameters to adapt to changes in the production environment.

[0169] 1.4.3 Real-time rescheduling mechanism

[0170] Real-time rescheduling is triggered when the following disturbance events occur:

[0171]

[0172] Rescheduling strategy:

[0173] Full rescheduling: Re-optimize from scratch (for major disturbances);

[0174] Local repair: only adjust the affected process and its subsequent processes (for minor disturbances);

[0175] Rolling window scheduling: partial rescheduling within a fixed time window.

[0176] 1.5 Algorithm Complexity Analysis and Acceleration Strategies

[0177] 1.5.1 Computational Complexity Control

[0178] 1. Layered dimensionality reduction: The strategic layer handles macro-level decisions, reducing the search space dimensions of the tactical layer.

[0179] 2. Approximate evaluation function: Use a lightweight surrogate model to quickly evaluate the quality of candidate solutions.

[0180] 3. Parallel computing architecture:

[0181]

[0182] 1.5.2 Memory Optimization Strategies

[0183] 1. Decompress and store: Only store chromosome codes, and decode them into a complete scheduling scheme when needed.

[0184] 2. Hash deduplication: Use a Bloom filter to detect duplicate solutions and avoid duplicate calculations.

[0185] 3. Incremental evaluation: Only the objective function values ​​affected by the modification are recalculated.

[0186] 1.6 Special optimizations for the manufacturing of ultra-thin flexible glass;

[0187] 1.6.1 Cleanroom Production Constraint Modeling:

[0188] 1. Cleanliness decay model:

[0189] .

[0190] 2. Clean recovery time: Insert necessary cleaning time windows between two production tasks.

[0191] 3. Cleanroom Path Planning: AGV scheduling takes into account access restrictions in clean areas.

[0192] 1.6.2 Optimization of fragile material handling:

[0193] 1. Handling stability constraints:

[0194] .

[0195] 2. Temperature gradient control: Limiting the rate of temperature change between adjacent processes.

[0196] 3. Minimize stress accumulation: Optimize the processing sequence to reduce residual stress.

[0197] 1.6.3 Special Objective Function Term:

[0198] 1. Surface quality weight: Assign higher priority to processes that affect surface quality.

[0199] 2. Thickness uniformity incentive: Encourages continuous production of products with similar thickness specifications.

[0200] 3. Defect traceability cost: Consider the difficulty and cost of tracing quality issues.

[0201] 1.7 Performance Evaluation and Verification:

[0202] 1.7.1 Benchmark Question Set

[0203] Verify algorithm performance using the standard FJSP test suite and real-world production data:

[0204] Standard test suite: Classic FJSP examples such as Brandimarte and Barnes

[0205] Extended test suite: Real-world examples with added energy and maintenance constraints.

[0206] Proprietary Test Set for Ultra-Thin Glass: Unique Problems Constructed Based on Actual Production Data

[0207] 1.7.2 Performance Indicators:

[0208] 1. Scheduling quality indicators:

[0209] Maximum completion time (Makespan)

[0210] Total Tardiness

[0211] Machine utilization

[0212] Energy efficiency.

[0213] 2. Algorithm efficiency metrics:

[0214] Convergence Speed

[0215] Computation Time

[0216] Solution stability.

[0217] 3. Actual production targets:

[0218] On-time delivery rate

[0219] First-pass yield

[0220] Total cost per unit.

[0221] In this embodiment of the invention, the equipment predictive maintenance module includes a fault prediction model, which is constructed based on multi-scale spatiotemporal learning (MSTL) and transfer learning, integrates a spatiotemporal feature extraction unit and an equipment topology influence capture unit, and adopts a meta-learning framework; the spatiotemporal feature extraction unit is used to extract the spatiotemporal features of multi-source sensor data, and the equipment topology influence capture unit is used to capture the topological correlation influence between devices.

[0222] The fault prediction model integrates a one-dimensional convolutional neural network with a bidirectional LSTM attention mechanism to extract spatiotemporal features from multi-source sensor data, and introduces a graph neural network to capture the topological influence between devices. A meta-learning framework is employed to address the problem of predicting small-sample faults in new devices or rare faults.

[0223] In this embodiment of the invention, the quality inspection and control module includes a visual inspection subsystem and a process parameter causal analysis subsystem; the visual inspection subsystem is based on the Transformer model and metric learning and is used for product defect identification; the process parameter causal analysis subsystem is based on a causal discovery algorithm and is used to construct a process parameter causal graph and realize defect root cause tracing.

[0224] The quality inspection and control module consists of a high-precision visual inspection subsystem and a process parameter causal analysis subsystem. The visual inspection uses Vision Transformer and metric learning to improve the ability to identify defects in small samples; the causal analysis subsystem applies causal discovery algorithms to construct a causal graph of process parameters, enabling automatic tracing from defect manifestations to the root cause of production.

[0225] In this embodiment of the invention, the logistics and warehousing management module employs a multi-agent deep reinforcement learning algorithm based on a centralized training and decentralized execution framework to dynamically schedule logistics and warehousing equipment. The reward function of the multi-agent deep reinforcement learning algorithm includes parameters related to task completion, path efficiency, collision avoidance, driving stability, and cleanliness maintenance, and the algorithm is pre-trained in a factory digital twin simulation.

[0226] The logistics and warehousing management module employs a multi-agent deep reinforcement learning algorithm based on a centralized training and decentralized execution framework to dynamically schedule AGVs and stacker cranes. Its reward function integrates multiple objectives such as task completion, path efficiency, collision avoidance, driving stability (protecting fragile materials), and cleanliness maintenance, and is pre-trained in a digital twin environment.

[0227] In this embodiment of the invention, the supply chain optimization module includes a demand forecasting model and a resilient strategy generation unit; the demand forecasting model is built based on federated learning and is used for joint training with supply chain partners; the resilient strategy generation unit is based on a robust optimization method and is used to generate procurement and distribution strategies to cope with uncertainties in demand, supply and logistics.

[0228] A resilient supply chain network based on federated learning and robust optimization is constructed through a supply chain optimization module. Vertical federated learning is used to jointly train a demand forecasting model with partners while protecting data privacy. Two-stage robust optimization or degenerate robust optimization methods are employed to generate resilient procurement and delivery strategies capable of coping with multiple uncertainties in demand, supply, and logistics.

[0229] In this embodiment of the invention, the energy management module includes an energy consumption prediction unit and a real-time optimization unit; the energy consumption prediction unit is constructed based on a physical information neural network, and its loss function embeds physical law constraints; the real-time optimization unit is based on a model predictive control framework and is used to solve energy efficiency optimization problems with safety and process hard constraints.

[0230] The energy management module applies a physical information neural network for energy consumption prediction and real-time optimization under safety constraints. Physical laws such as mass conservation and energy conservation are embedded as soft constraints in the LSTM network loss function to improve the reliability and extrapolation capability of the predictions. Based on a model predictive control framework, it solves the real-time energy efficiency optimization problem with hard constraints related to safety and process requirements.

[0231] In this embodiment of the invention, the key linkage mechanism includes:

[0232] The production scheduling and optimization module calls the energy consumption prediction data and electricity price curve output by the energy management module in real time to achieve a dynamic balance between production scheduling and energy consumption costs.

[0233] When the predictive maintenance module generates a maintenance work order, it triggers the logistics and warehousing management module to query spare parts inventory and triggers the supply chain optimization module to execute the procurement process.

[0234] When the quality inspection and control module detects batch defects, it locates the root cause of the defects through the process parameter cause-effect analysis subsystem and pushes the defect root cause information to the equipment predictive maintenance module and the production scheduling and optimization module to trigger preventive adjustments.

[0235] Secondly, embodiments of the present invention provide a smart factory management method, applied to the aforementioned smart factory integrated optimization management system, the method comprising the following steps:

[0236] S1: Collect production site data through the IoT sensing layer and synchronize it with enterprise information system data, aggregate it into a unified data platform, and build and update the factory digital twin;

[0237] S2: Each core functional module runs core intelligent algorithms based on real-time and historical data from the factory's digital twin to generate local optimization decision-making schemes;

[0238] S3: The collaborative optimization engine evaluates the impact of each local optimization decision scheme on the global objective, and generates a global collaborative decision instruction set through iterative negotiation or centralized optimization.

[0239] S4: Distribute the global collaborative decision-making instruction set to the execution layer to drive the physical factory to run;

[0240] S5: Collects the execution results of the global collaborative decision-making instruction set and new production data, feeds them back to the unified data platform and the algorithm models of each core functional module to form a closed loop, and drives the model to achieve self-evolution through online learning and model management platform.

[0241] In step S3 above, the overall objectives include total operating costs and comprehensive energy efficiency.

[0242] In step S4 above, the execution layer includes manufacturing execution system, warehouse management system and programmable logic controller, etc.

[0243] In step S5 above, the results of instruction execution and new production data are collected and fed back to the data platform and various algorithm models to form a closed loop of "perception-decision-execution-optimization" and drive the model to learn and evolve online.

[0244] Example

[0245] In this embodiment, an ultra-thin flexible glass manufacturing plant is used as the application scenario for the intelligent management system of a smart factory.

[0246] System Deployment:

[0247] Edge computing nodes and a central cloud platform are deployed within the ultra-thin flexible glass manufacturing plant. Vibration sensors, temperature sensors, and current sensors are installed on production line equipment; high-resolution industrial cameras and polarized light devices are deployed at key quality inspection stations; and communication modules are provided for AGVs, stacker cranes, smart meters, and other equipment. All data is connected to a unified data platform through an industrial IoT gateway.

[0248] Model training and initialization:

[0249] 1. Using data collected from equipment sensors over the past year, train the MSTL-PHM prediction model to initially predict the remaining service life of core equipment (such as laser cutting machines and precision rolling mills).

[0250] 2. Collect tens of thousands of glass images containing various defects (scratches, chipping, foreign objects), train a ViT-based DefectNet model, and use a causal discovery algorithm to construct a causal graph of process parameters and final quality indicators for each stage of melting, forming, and cutting.

[0251] 3. In the warehouse digital twin model, the MADRL algorithm is used to perform millions of steps of simulation training on the AGV cluster to learn multi-objective collaborative scheduling strategies.

[0252] 4. Initialize the supply chain federated learning model and robust optimization parameters based on historical data.

[0253] System running example:

[0254] One day, the system received a batch of urgent orders. The collaborative optimization engine initiated a global optimization process:

[0255] 1. Production Scheduling: The DRL agent in the scheduling module combines the order delivery date, current equipment status, and the time-of-use electricity price for the next 24 hours provided by the energy module (predicting peak hours in the afternoon) to determine whether to schedule some non-critical processes for this batch of orders to be produced at night. The lower-level GA-VNS algorithm quickly generates a detailed process schedule.

[0256] 2. Quality Control: During the production process, the visual inspection system determines the product's qualification in real time. Simultaneously, the cause-effect analysis subsystem continuously monitors process parameters. When a strong correlation is detected between "annealing furnace temperature curve fluctuations" and "edge stress values," it provides an early warning of potential batch-scale hidden crack risks.

[0257] 3. Equipment Maintenance and Interconnection: The predictive maintenance module determines that the RUL (Round-Under-Limit) of the precision grinding machine bearing is only 48 hours away, with a failure probability exceeding 85%. The system automatically generates a maintenance work order and triggers the following interconnections: the logistics module immediately reserves the required bearing spare parts; the supply chain module initiates a rapid procurement process, placing an order with the best backup supplier to ensure that the spare parts arrive before the maintenance window.

[0258] 4. Energy and Logistics Synergy: The MPC controller in the energy management module dynamically adjusts the operating combination of the air compressor station and chiller unit according to the new production plan to achieve peak shaving and valley filling. The AGV in the logistics module transports raw materials from the clean warehouse to the production line at a steady speed according to real-time scheduling instructions, avoiding vibration damage to the glass sheets.

[0259] Implementation results:

[0260] After applying the system of this invention, the ultra-thin flexible glass factory achieved the following improvements: overall production efficiency increased by 25%, unplanned equipment downtime decreased by 40%, overall product yield increased to 99.97%, overall energy consumption per unit output decreased by 18%, and average order delivery cycle shortened by 30%. The synergistic optimization between modules has generated a significant added value effect.

[0261] The present invention has been described above by way of example with reference to the accompanying drawings. Obviously, the specific implementation of the present invention is not limited to the above-described manner. Any non-substantial improvements made using the inventive concept and technical solution; or the direct application of the inventive concept and technical solution to other situations without modification, are all within the protection scope of the present invention.

Claims

1. A smart factory intelligent management system, characterized in that, include: A unified data platform is used to aggregate and process multi-source heterogeneous data from the IoT sensing layer and enterprise information systems, and to build a real-time updated digital twin of the factory. The collaborative optimization engine, based on a multi-agent system and a multi-objective optimization framework, is used to coordinate the decisions of various functional modules to achieve the global optimal goal. as well as The system comprises six core functional modules: production scheduling and optimization, predictive equipment maintenance, quality inspection and control, logistics and warehousing management, supply chain optimization, and energy management. All six modules are communicatively connected to the unified data platform and the collaborative optimization engine. Each module makes local intelligent decisions based on data from the unified data platform and receives coordination instructions from the collaborative optimization engine to achieve dynamic cross-module linkage and global optimization.

2. The intelligent factory integrated optimization management system according to claim 1, characterized in that, It also includes an online learning and model management platform, which communicates with the six core functional modules and the collaborative optimization engine to continuously monitor algorithm performance, automatically trigger model retraining and A / B testing deployment, and realize the system's self-evolution.

3. The intelligent factory integrated optimization management system according to claim 1, characterized in that, The multi-source heterogeneous data includes device sensor data, industrial visual image data, mobile device operation data collected by the IoT sensing layer, as well as production plan data, order data, inventory data, supply chain data, and energy consumption data from enterprise information systems.

4. The intelligent factory integrated optimization management system according to any one of claims 1 to 3, characterized in that, The collaborative optimization engine evaluates the impact of local optimization decision schemes output by each core functional module on the global objective through iterative negotiation or centralized optimization, and generates a global collaborative decision instruction set.

5. The intelligent factory integrated optimization management system according to any one of claims 1 to 3, characterized in that, The production scheduling and optimization module is a two-layer dynamic scheduler, including an upper-layer scheduling unit and a lower-layer scheduling unit; The upper-level scheduling unit is based on deep reinforcement learning and is used for order priority allocation and production shift recommendation. The lower-level scheduling unit adopts a genetic algorithm with hybrid variable neighborhood search, receives constraint information output by the upper-level scheduling unit to perform process-level scheduling, and integrates real-time energy consumption cost parameters into the fitness function of the lower-level scheduling unit.

6. The intelligent factory integrated optimization management system according to any one of claims 1 to 3, characterized in that, The equipment predictive maintenance module includes a fault prediction model, which is constructed based on multi-scale spatiotemporal learning and transfer learning, integrates a spatiotemporal feature extraction unit and an equipment topology influence capture unit, and adopts a meta-learning framework; the spatiotemporal feature extraction unit is used to extract spatiotemporal features from multi-source sensor data, and the equipment topology influence capture unit is used to capture the topological correlation influence between devices.

7. The intelligent factory integrated optimization management system according to any one of claims 1 to 3, characterized in that, The quality inspection and control module includes a visual inspection subsystem and a process parameter causal analysis subsystem; the visual inspection subsystem is based on the Transformer model and metric learning and is used for product defect identification. The process parameter causal analysis subsystem is based on a causal discovery algorithm and is used to construct a process parameter causal graph and trace the root cause of defects.

8. The intelligent factory integrated optimization management system according to any one of claims 1 to 3, characterized in that, The logistics and warehousing management module adopts a multi-agent deep reinforcement learning algorithm based on a centralized training and decentralized execution framework for dynamically scheduling logistics and warehousing equipment. The reward function of the multi-agent deep reinforcement learning algorithm includes parameters related to task completion, path efficiency, collision avoidance, driving stability, and cleanliness maintenance, and the algorithm is pre-trained in the factory digital twin.

9. The intelligent factory integrated optimization management system according to any one of claims 1 to 3, characterized in that, The supply chain optimization module includes a demand forecasting model and a flexible strategy generation unit; The demand forecasting model is built based on federated learning and is used for joint training with supply chain partners; The elastic strategy generation unit is based on a robust optimization method and is used to generate procurement and delivery strategies to cope with uncertainties in demand, supply and logistics.

10. A smart factory management method, characterized in that, The method, applied to the smart factory integrated optimization management system according to any one of claims 1 to 9, comprises the following steps: S1: Collect production site data through the IoT sensing layer and synchronize it with enterprise information system data, aggregate it into a unified data platform, and build and update the factory digital twin; S2: Each core functional module runs a core intelligent algorithm based on the real-time and historical data of the factory digital twin to generate a local optimization decision scheme; S3: The collaborative optimization engine evaluates the impact of each local optimization decision scheme on the global objective, and generates a global collaborative decision instruction set through iterative negotiation or centralized optimization. S4: Send the global collaborative decision instruction set to the execution layer to drive the physical factory to run; S5: Collect the execution results and new production data of the global collaborative decision-making instruction set, feed them back to the unified data platform and the algorithm models of each core functional module to form a closed loop, and drive the model to achieve self-evolution through online learning and model management platform.

Images