Steel supply chain dynamic optimization and intelligent decision-making system based on deep learning

Abstract

The invention discloses a steel supply chain dynamic optimization and intelligent decision-making system based on deep learning, and belongs to the field of industrial intelligence and supply chain management. Comprising a data sensing layer, a dynamic optimization module and an intelligent decision module. According to the system, a space-time diagram convolutional network is adopted to establish a supply chain state model, multi-objective optimization is realized in combination with an improved NSGA-III algorithm, a risk perception decision is made through deep reinforcement learning, and the simulation precision is improved by using a digital twinborn technology. The dynamic time warping matrix, Hausdorff distance optimization, VAE-GAN adversarial sample generation and other technologies are innovatively introduced, and the defects of a traditional method in the aspects of dynamic adaptability, multi-objective optimization and risk response are overcome. The system further comprises block chain evidence storage and a quantum genetic algorithm, and credible data evidence storage and energy consumption optimization are achieved. The method significantly improves the supply chain scheduling precision and the anti-risk capability, and is suitable for a complex industrial scene.

Description

Technical Field

[0001] This invention relates to the field of industrial intelligence and supply chain management technology, specifically a deep learning-based dynamic optimization and intelligent decision-making system for the steel supply chain. Background Technology

[0002] The steel supply chain is characterized by multiple coupled links (raw material procurement, production, logistics, and inventory), high dynamism (market demand fluctuations, equipment status changes), and multiple conflicting objectives (cost, efficiency, and energy consumption). Traditional optimization methods have the following limitations:

[0003] 1. Static modeling defects: Existing systems mostly rely on linear programming or heuristic algorithms, which make it difficult to capture the dynamic spatiotemporal relationships of the supply chain topology (such as the cascading impact of logistics delays on production scheduling).

[0004] 2. Insufficient integration of heterogeneous data: The lack of unified representation of heterogeneous information from multiple sources such as sensor data and business system data leads to low accuracy of state perception;

[0005] 3. Delayed risk response: The quantitative assessment of sudden disruption events (such as raw material shortages) relies on empirical rules and lacks a real-time decision-making mechanism based on reinforcement learning;

[0006] 4. Bottleneck in energy efficiency optimization: The energy consumption model of steel production has nonlinear process coupling (such as blast furnace-rolling mill coordination), and traditional genetic algorithms are prone to getting trapped in local optima.

[0007] In recent years, deep learning technology has been initially applied in the supply chain field, but the following technological gaps remain:

[0008] 1. Spatiotemporal Graph Convolutional Networks (ST-GCN) have not yet been used to model dynamic supply chain topology; existing methods ignore the time warping matrix (T). k Constraints on state transitions;

[0009] 2. Reference point generation in NSGA-III algorithm for multi-objective optimization The non-uniform distribution characteristics of the Pareto front in the supply chain were not considered;

[0010] 3. When mapping digital twin simulations to real systems, the spectral attenuation constraint of traditional SVD decomposition (Σ) is insufficient, leading to virtual model drift.

[0011] Therefore, based on this, we have studied and improved existing technologies and proposed a deep learning-based dynamic optimization and intelligent decision-making system for the steel supply chain. Summary of the Invention

[0012] This invention aims to address the low optimization efficiency in steel supply chain management caused by insufficient integration of multi-source heterogeneous data, lagging response to dynamic environments, and conflicts among multiple objectives. Specifically, addressing the shortcomings of traditional methods such as difficulty in modeling dynamic supply chain topology (e.g., spatiotemporal coupling of logistics and production), inaccurate quantification of sudden risks, and susceptibility to local optima in energy consumption optimization, this invention integrates the dynamic state transition model of Spatiotemporal Graph Convolutional Network (ST-GCN), an improved multi-objective optimization algorithm of NSGA-III, and a risk perception and decision-making mechanism based on deep reinforcement learning. This enables real-time perception, adaptive optimization, and intelligent decision-making throughout the entire supply chain process, significantly improving resource scheduling accuracy, energy efficiency, and anti-interference capabilities.

[0013] The technical solution adopted in this invention is: a deep learning-based dynamic optimization and intelligent decision-making system for the steel supply chain, comprising:

[0014] The supply chain data perception layer is used to collect multi-source heterogeneous data on raw material procurement, production scheduling, inventory status and logistics transportation in real time.

[0015] The dynamic optimization module uses a spatiotemporal graph convolutional network (ST-GCN) to establish a supply chain state transition model.

[0016]

[0017] in Represents the supply chain topology diagram, T k Let Φ be the dynamic time-normalized matrix, and Φ(·) be the spectral regularization term.

[0018] As a further aspect of the present invention: the dynamic optimization module includes a multi-objective programming sub-module, whose objective function is:

[0019]

[0020] Where L is the supply chain process correlation matrix, and α, β, γ are adaptive weight parameters, dynamically adjusted by the following formula:

[0021]

[0022] As a further aspect of the present invention: the multi-objective programming employs an improved NSGA-III algorithm, and its reference point generation method is as follows:

[0023]

[0024] Individual fitness assessment uses an improved index based on Hausdorff distance:

[0025]

[0026] As a further aspect of this invention: the intelligent decision-making module employs a deep reinforcement learning framework, and its Q-function is approximately:

[0027]

[0028] The state value function is updated using the double-delay DDPG algorithm:

[0029]

[0030] As a further aspect of the present invention: the deep reinforcement learning framework includes a supply chain risk perception submodule, and its risk quantification model is as follows:

[0031]

[0032] Where W is the Hessian matrix of supply chain disruption events, and ∈ is the risk sensitivity coefficient.

[0033] As a further aspect of the present invention: the system includes a digital twin simulation module, whose virtual-reality mapping relationship is represented as follows:

[0034]

[0035] The singular value decomposition Σ satisfies Spectral attenuation constraint.

[0036] As a further aspect of the present invention: the digital twin simulation module uses a variational autoencoder to generate adversarial examples, and its loss function is:

[0037]

[0038] Where f(t) = -log(1 + exp(-t)) is the smoothing adversarial loss.

[0039] As a further aspect of the present invention: the system includes a blockchain notarization module, whose consensus algorithm adopts an improved PoS mechanism, and the probability of selecting a verification node is:

[0040]

[0041] Where C i For credit scores, T i The holding period, and meets the following requirements.

[0042] As a further aspect of the present invention: the system includes an energy consumption optimization submodule, and its steel production energy consumption model is as follows:

[0043]

[0044] Where κ is the process coupling coefficient, optimized using a quantum genetic algorithm:

[0045]

[0046] As a further aspect of the present invention: the confidence assessment of the system's output decision recommendations adopts the DS evidence theory.

[0047]

[0048] The basic probability assignment function m(·) is calibrated through the softmax output of the deep neural network:

[0049]

[0050] The beneficial effects of this invention are:

[0051] 1. Enhanced dynamic optimization capabilities

[0052] The supply chain topology is modeled using a spatiotemporal graph convolutional network (ST-GCN), combined with a dynamic time warping matrix (T). k It effectively captures the spatiotemporal correlations of multiple links such as raw material procurement, production scheduling, and logistics transportation, and improves the optimization accuracy in dynamic environments.

[0053] The improved multi-objective optimization algorithm (NSGA-III) combined with Hausdorff distance fitness assessment makes the supply chain scheduling scheme more in line with actual needs and balances the conflicts of multiple objectives such as cost, efficiency, and energy consumption.

[0054] 2. Intelligent decision-making and risk perception

[0055] A decision-making framework based on deep reinforcement learning (DDPG) combined with a supply chain risk quantification model. Real-time assessment of unexpected risks such as logistics disruptions and equipment failures, and generation of optimal response strategies to enhance system robustness.

[0056] The DS evidence theory is used to assess the confidence level of decision-making suggestions, thereby improving the credibility of intelligent recommendations and reducing the risk of misjudgment.

[0057] 3. Digital Twin and Simulation Optimization

[0058] By generating adversarial examples through variational autoencoders (VAE-GAN), the generalization ability of digital twin models is enhanced, making virtual simulations closer to the real supply chain operation and optimizing simulation verification before decision-making.

[0059] Spectral attenuation constraint mapping based on singular value decomposition (SVD) ensures high fidelity between the twin model and the physical system, reducing simulation bias.

[0060] 4. Blockchain-based evidence storage and trusted management

[0061] It adopts an improved Proof-of-Stake (PoS) consensus mechanism, combined with a credit score (C) i ) and holding time (T) i Dynamically adjust the weight of verification nodes to improve the credibility and tamper-proof capability of supply chain data storage.

[0062] 5. Energy consumption optimization and green production

[0063] The energy consumption model for steel production (including nonlinear coupling terms) is optimized based on quantum genetic algorithm, which significantly reduces energy consumption, improves resource utilization, and contributes to green and low-carbon manufacturing. Attached Figure Description

[0064] Figure 1 This is a framework diagram of a deep learning-based dynamic optimization and intelligent decision-making system for the steel supply chain, as proposed in this invention.

[0065] Figure 2 This is a module interaction framework diagram of a deep learning-based dynamic optimization and intelligent decision-making system for the steel supply chain according to the present invention. Detailed Implementation

[0066] The present invention will be further described below.

[0067] Please see Figure 1-2

[0068] Example 1: Dynamic Supply Chain State Modeling Based on ST-GCN

[0069] Technical solution

[0070] Data Acquisition: Real-time collection of raw material procurement data, production scheduling data, inventory status and logistics transportation data of the steel supply chain through Internet of Things (IoT) devices to construct a multi-source heterogeneous dataset.

[0071] Dynamic optimization module: The spatiotemporal graph convolutional network (ST-GCN) is used to model the supply chain state transitions, as shown in the following formula:

[0072]

[0073] in, Represents the supply chain topology diagram, T k This is a dynamic time warping matrix used to adjust the weights at different time steps.

[0074] Optimization objective: Minimize supply chain delay costs, inventory backlog, and logistics transportation losses.

[0075] Applicable Scenarios

[0076] It is suitable for multi-factory collaborative production scenarios, such as dynamic scheduling optimization among multiple steelmaking plants and rolling mills under a steel group.

[0077] Comparative analysis

[0078] index Example 1 (ST-GCN) Traditional methods (linear programming) Dynamic adaptability Strong (automatic adjustment of time weights) Weak (fixed time window) computational efficiency High (GPU accelerated) Low (slow to solve large-scale problems) Optimize accuracy High (considering spatiotemporal correlation) General (static optimization only)

[0079] Advantages: Compared to traditional linear programming, ST-GCN can automatically learn dynamic changes in the supply chain, resulting in higher optimization accuracy.

[0080] Example 2: Multi-objective optimization based on improved NSGA-III

[0081] Technical solution

[0082] Objective function:

[0083]

[0084] Where L is the supply chain process correlation matrix, and a, β, γ are adaptive weight parameters.

[0085] Optimization Algorithm: An improved NSGA-III algorithm is adopted, and the reference point generation method is as follows:

[0086]

[0087] Fitness assessment uses Hausdorff distance to ensure uniform distribution of the Pareto front.

[0088] Applicable Scenarios

[0089] It is suitable for scenarios with multiple conflicting objectives, such as simultaneously optimizing production costs, energy consumption, and delivery cycles.

[0090] Comparative analysis

[0091] index Example 2 (Improved NSGA-III) Traditional NSGA-II Pareto front distribution More uniform (Hausdorff distance optimization) Possible local clustering computational complexity Relatively high (reference point needs to be dynamically adjusted). lower applicability More suitable for high-dimensional target optimization Applicable to low-dimensional problems

[0092] Advantages: The improved NSGA-III can optimize multiple objectives more evenly, avoiding the local optima problem of traditional algorithms in complex supply chain problems.

[0093] Example 3: Risk Perception Decision Based on Deep Reinforcement Learning

[0094] Technical solution

[0095] Q-function approximation: The double-delay DDPG (TD3) algorithm is used for update.

[0096]

[0097] Risk quantification model:

[0098]

[0099] Where W is the Hessian matrix of supply chain disruption events, and ∈ is the risk sensitivity coefficient.

[0100] Applicable Scenarios

[0101] It is suitable for high-risk supply chain environments, such as situations where international logistics are affected by policies or raw material prices fluctuate drastically.

[0102] Comparative analysis

[0103] index Example 3 (DDPG + Risk Model) Traditional rule-based decision-making Risk response speed Real-time (millisecond-level decision-making) Lag (manual adjustment) Decision robustness High (considering second-order risk) Low (rule of experience) Adaptability Dynamic adjustment strategy Fixed strategy

[0104] Advantages: Compared to human experience and rules, deep reinforcement learning can perceive risks in real time and adjust strategies, thereby improving supply chain resilience.

[0105] Example 4: Simulation Optimization Based on Digital Twin

[0106] Technical solution

[0107] Virtual-Reality Mapping: Building a Digital Twin Model Using SVD Decomposition

[0108]

[0109] Adversarial example generation: VAE-GAN is used to enhance the diversity of simulation data, with the loss function being:

[0110]

[0111] Applicable Scenarios

[0112] It is suitable for supply chain simulation and forecasting, such as capacity simulation before a new factory goes into operation and logistics route optimization.

[0113] Comparative analysis

[0114]

[0115]

[0116] Advantages: Digital twins combined with generative adversarial networks (GANs) can more realistically simulate supply chain operations and improve the reliability of pre-decision verification.

[0117] Comparison and summary of implementation examples

[0118] Example Core technologies Applicable Scenarios Advantages limitation 1 ST-GCN Multi-factory collaborative scheduling Strong dynamic adaptability High computing resource requirements 2 Improved NSGA-III Multi-objective optimization Pareto frontier is better High algorithm complexity 3 DDPG+ Risk Model High-risk environment Real-time risk response Requires a large amount of training data 4 Digital Twin + VAE-GAN Simulation prediction High-fidelity simulation Training costs are high

[0119] Recommended best option:

[0120] 1. Dynamic scheduling optimization → Example 1 (ST-GCN).

[0121] 2. Multi-objective conflict optimization → Example 2 (Improved NSGA-III).

[0122] 3. High-risk supply chain management → Example 3 (DDPG).

[0123] 4. Simulation and Prediction → Example 4 (Digital Twin).

[0124] This invention can combine different embodiments according to actual needs to build a more comprehensive intelligent decision-making system for the steel supply chain.

[0125] The above embodiments are only used to illustrate the technical solutions of the present invention, and are not intended to limit them. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. A deep learning-based dynamic optimization and intelligent decision-making system for the steel supply chain, characterized in that, include: The supply chain data perception layer is used to collect multi-source heterogeneous data on raw material procurement, production scheduling, inventory status and logistics transportation in real time. The dynamic optimization module uses a spatiotemporal graph convolutional network (ST-GCN) to establish a supply chain state transition model. in Represents the supply chain topology diagram, T k Let Φ be the dynamic time-normalized matrix, and Φ(·) be the spectral regularization term.

2. The deep learning-based dynamic optimization and intelligent decision-making system for the steel supply chain according to claim 1, characterized in that, The dynamic optimization module includes a multi-objective programming sub-module, whose objective function is: Where L is the supply chain process correlation matrix, and α, β, γ are adaptive weight parameters, dynamically adjusted by the following formula:

3. The deep learning-based dynamic optimization and intelligent decision-making system for the steel supply chain according to claim 2, characterized in that, The multi-objective programming algorithm employs an improved NSGA-III algorithm, and its reference point generation method is as follows: Individual fitness assessment uses an improved index based on Hausdorff distance:

4. The deep learning-based dynamic optimization and intelligent decision-making system for the steel supply chain according to claim 1, characterized in that, The intelligent decision-making module employs a deep reinforcement learning framework, and its Q-function is approximately: The state value function is updated using the double-delay DDPG algorithm:

5. The deep learning-based dynamic optimization and intelligent decision-making system for the steel supply chain according to claim 4, characterized in that, The deep reinforcement learning framework includes a supply chain risk perception submodule, whose risk quantification model is as follows: Where W is the Hessian matrix of supply chain disruption events, and ∈ is the risk sensitivity coefficient.

6. The deep learning-based dynamic optimization and intelligent decision-making system for the steel supply chain according to claim 1, characterized in that, The system includes a digital twin simulation module, whose virtual-reality mapping relationship is represented as follows: The singular value decomposition Σ satisfies Spectral attenuation constraint.

7. The deep learning-based dynamic optimization and intelligent decision-making system for the steel supply chain according to claim 6, characterized in that, The digital twin simulation module uses a variational autoencoder to generate adversarial examples, and its loss function is: Where f(t) = -log(1 + exp(-t)) is the smoothing adversarial loss.

8. The deep learning-based dynamic optimization and intelligent decision-making system for the steel supply chain according to claim 1, characterized in that, The system includes a blockchain notarization module, and its consensus algorithm adopts an improved PoS mechanism, with the probability of selecting a verification node being: Where C i For credit scores, T i The holding period, and meets the following requirements.

9. The deep learning-based dynamic optimization and intelligent decision-making system for the steel supply chain according to claim 1, characterized in that, The system includes an energy consumption optimization submodule, and its steel production energy consumption model is as follows: Where κ is the process coupling coefficient, optimized using a quantum genetic algorithm:

10. A deep learning-based dynamic optimization and intelligent decision-making system for the steel supply chain according to claim 9, characterized in that, The confidence assessment of the system's output decision recommendations uses the DS evidence theory: The basic probability assignment function m(·) is calibrated through the softmax output of the deep neural network:

Images