A supply chain resilience analysis and optimization method, system and computer device

Abstract

The application relates to a supply chain resilience analysis and optimization method, system and computer device. The method comprises the following steps: constructing an initial tensor according to a supply chain field situation and objective constraints, covering information such as a logistics subgraph, task demand, vehicle capacity and main and backup road networks; generating a context vector by network coding fusion of graph features, numerical features and constraint features; performing sequence decoding on the context vector and combining the objective constraints to generate lower-layer scheduling decisions, and then mapping the decision results to obtain a comprehensive level related to supply chain resilience. Therefore, a closed-loop process from situation awareness, optimization decision to resilience level evaluation is realized, and decision support and quantitative basis are provided for supply chain resilience analysis.

Description

Technical Field

[0001] This application relates to the field of data processing technology, and in particular to a method, system and computer device for supply chain resilience analysis and optimization. Background Technology

[0002] Supply chain resilience refers to the ability of a supply chain to maintain continuous operation, recover quickly, and consistently fulfill its obligations under risks and disturbances. Currently, resilience-oriented supply chain scheduling commonly employs the following methods: relying on expert manual scheduling or rule engines, which suffers from long cycles, difficulty in reproducibility, and inconsistent standards across different entities. Furthermore, the rules struggle to simultaneously consider multiple objectives such as dynamic disturbances, redundancy, and alternative paths. End-to-end deep learning also relies on simply regressing scalar merit scores or directly generating natural language solutions, lacking executable decision structures bound to hard constraints such as vehicle capacity and road network structure, making it difficult to guarantee accurate analysis of feasible domain outputs and supply chain resilience. Summary of the Invention

[0003] Based on this, the purpose of this application is to solve the technical problem that the lack of comprehensive integration of multi-source field situation and constraints leads to poor accuracy in supply chain scheduling, which in turn makes it impossible to accurately support quantitative analysis of supply chain resilience.

[0004] Firstly, this application provides a method for supply chain resilience analysis and optimization. This includes: The initial tensor is determined based on the on-site situation and objective constraints of the supply chain; the on-site situation includes at least the logistics sub-map, task requirements, vehicle capacity, and primary and backup road networks; The initial tensor is encoded by a network to obtain a context vector; the context vector includes graph features, numerical features, and constraint features. Sequence decoding of the context vector is performed, and the lower-level scheduling decision for dispatching tasks is determined based on objective constraints; The overall level of supply chain resilience is determined based on lower-level scheduling decisions.

[0005] In one embodiment, the initial tensor includes at least the initial node features of the logistics subgraph, the adjacency of the primary and backup road networks, and a numerical tensor. Network encoding is performed on the initial tensor to obtain a context vector, including: performing two-layer graph convolution based on the initial node features and the adjacency of the primary and backup road networks to obtain target node features; performing linear mapping and mean pooling on the target node features to obtain graph features; performing multi-layer linear transformation and activation on the numerical tensor to obtain numerical features, and extracting features of a preset dimension from the numerical tensor as constraint features; and fusing the graph features, numerical features, and constraint features to obtain the context vector.

[0006] In one embodiment, the objective constraints include that the actual load of the vehicle does not exceed the vehicle capacity; the context vector is sequence decoded, and the lower-level scheduling decision for the vehicle dispatch task is determined based on the objective constraints, including: obtaining historical scheduling decisions, performing sequence decoding based on the context vector and historical scheduling decisions to obtain the original decision score; performing capacity masking on the original decision score based on the objective constraints to mask the overloaded vehicles corresponding to the original decision score; and determining the lower-level scheduling decision for the vehicle dispatch task based on the masked original decision score.

[0007] In one embodiment, the primary and backup road network includes the primary road network; determining the comprehensive level of supply chain resilience based on lower-level scheduling decisions includes: determining the total mileage of the lower-level scheduling decisions under the primary road network; obtaining simulation filing coefficients, and obtaining multi-dimensional upper-level performance decisions based on the total mileage and simulation filing coefficients; weighting the multi-dimensional upper-level performance decisions through multiple performance weights to determine the comprehensive score; and mapping the comprehensive score to a comprehensive level based on the business-side requirements of supply chain resilience.

[0008] In one embodiment, the lower-level scheduling decision is determined by a deep neural network. The training process of the deep neural network includes: acquiring sample tensors and running a teacher algorithm on the sample tensors to obtain a first scheduling decision; performing network encoding and sequence decoding on the sample tensors to obtain a second scheduling decision; using the first scheduling decision as a standard sequence and determining the difference between the standard sequence and the second scheduling decision through a decision loss function; updating the model parameters based on the difference and the backpropagation algorithm until a trained deep neural network is obtained.

[0009] In one embodiment, running the teacher algorithm on the sample tensor to obtain a first scheduling decision includes: obtaining a set of capacity-feasible vehicles, selecting the vehicle with the smallest marginal mileage increment from the set of capacity-feasible vehicles, and obtaining the corresponding first scheduling decision; or, determining the linear objective of vehicle allocation, and solving the linear objective of vehicle allocation through mixed integer programming to obtain the corresponding first scheduling decision.

[0010] In one embodiment, the method further includes: simulating and mapping the first scheduling decision and the second scheduling decision to obtain the first performance decision and the second performance decision; determining the performance loss function based on the mean square error between the first performance decision and the second performance decision; and obtaining the total loss function by combining the decision loss function and the performance loss function based on the preset target weight.

[0011] Secondly, this application also provides a supply chain resilience analysis and optimization system. The system includes a tensor module, an encoder, a decoder, and a result determination module, wherein: The tensor module is used to determine the initial tensor based on the on-site situation and objective constraints of the supply chain; the on-site situation includes the logistics sub-map, task requirements, vehicle capacity, and primary and backup road networks; The encoder is used to perform network encoding on the initial tensor to obtain a context vector; the context vector includes graph features, numerical features, and constraint features. The decoder is used to perform sequence decoding on the context vector and determine the lower-level scheduling decision for dispatching tasks based on objective constraints. The results determination module is used to determine the overall level of supply chain resilience based on lower-level scheduling decisions.

[0012] Thirdly, this application also provides a computer device. The computer device includes a memory and a processor, the memory storing a computer program, and the processor executing the computer program to implement the above-mentioned supply chain resilience analysis and optimization methods.

[0013] Fourthly, this application also provides a computer-readable storage medium. The computer-readable storage medium stores a computer program, which, when executed by a processor, implements the above-described supply chain resilience analysis and optimization methods.

[0014] The aforementioned supply chain resilience analysis and optimization methods, systems, computer equipment, and readable storage media construct an initial tensor based on the supply chain's on-site situation and objective constraints, encompassing information such as logistics sub-graphs, task requirements, vehicle capacity, and primary / backup road networks. A context vector is generated by fusing graph features, numerical features, and constraint features through network coding. The context vector is then sequence-decoded and combined with objective constraints to generate lower-level dispatching decisions for vehicle allocation. Finally, the decision results are mapped to obtain a comprehensive level associated with supply chain resilience. Therefore, a closed-loop process from situational awareness and optimization decision-making to resilience level evaluation is achieved, providing decision support and quantitative basis for supply chain resilience analysis. Attached Figure Description

[0015] Figure 1 This is a diagram illustrating the application environment of a supply chain resilience analysis and optimization method in one embodiment. Figure 2 This is a flowchart illustrating a supply chain resilience analysis and optimization method in one embodiment; Figure 3 This is a flowchart illustrating the deep neural network training process in one embodiment; Figure 4 This is an internal structural diagram of a computer device in one embodiment. Detailed Implementation

[0016] To make the objectives, technical solutions, and advantages of this application clearer, the following detailed description is provided in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the scope of this application.

[0017] The supply chain resilience analysis and optimization method provided in this application can be applied to, for example... Figure 1 In the application environment shown, terminal 102 communicates with server 104 via a network. Server 104 obtains the supply chain situation and objective constraints provided by the user from terminal 102 to determine the initial tensor and performs network encoding on the initial tensor to obtain a context vector. Server 104 also performs sequence decoding on the context vector and determines the lower-level scheduling decision for the dispatch task based on the objective constraints. Based on the lower-level scheduling decision, it determines the comprehensive level of supply chain resilience, so that terminal 102 presents the lower-level scheduling decision and comprehensive level to the user. Terminal 102 can be, but is not limited to, various personal computers, laptops, smartphones, tablets, drones, smart vehicle devices, and portable wearable devices. Server 104 can be implemented using a standalone server or a server cluster composed of multiple servers. It can be a cloud server providing basic cloud computing services such as cloud services, cloud databases, cloud computing, cloud functions, cloud storage, network services, cloud communication, middleware services, domain name services, security services, CDN (Content Delivery Network), and big data and artificial intelligence platforms.

[0018] In one embodiment, such as Figure 2 As shown, a supply chain resilience analysis and optimization method is provided. Application scenarios include, but are not limited to, joint warehouse and distribution scheduling with primary and backup road networks, vehicle capacity, and task queues, as well as situations requiring a consistent "situation-constraint-decision structure" in which the leading enterprise verifies and compares transportation plans and leaves traceable records of executable plans in cooperation renewal evaluation materials. The method includes the following steps: Step 202: Determine the initial tensor based on the on-site situation and objective constraints of the supply chain.

[0019] The on-site situation includes at least the logistics sub-map, task requirements, vehicle capacity, and primary and backup road networks. The on-site situation represents the real-time, actual state of the site, such as inventory quantity, goods in transit, equipment capacity, and production line status; objective constraints refer to hard rules that cannot be violated, such as vehicle load limits, warehouse capacity limits, transportation distances, and cost thresholds.

[0020] Specifically, for the relevant statistics of this batch of supply chain transportation obtained from the user side, the on-site situation includes the primary and backup road networks at the depot and customer nodes, the task demand vector, and the vehicle capacity vector; the objective constraints include gradually increasing vehicle load to not exceed capacity, and the supply chain on-site situation and computable objective constraints are encoded into batch-processable initial tensors. The initial tensor includes: the initial node feature matrix of the logistics subgraph. (like (Including depot, F=8); row normalization adjacency under primary and backup road networks The characterization process involves performing row normalization on the adjacency matrices corresponding to the primary and backup road networks, ensuring that the sum of each row's elements is 1. This is used by the graph encoder to stably extract the topological features of the dual road networks. Numerical tensors... This includes task requirements, vehicle capacity, demand-capacity ratio, depot departure primary and backup side statistics, and redundant agents, among which some dimensions can be placeholders to support time series or text expansion.

[0021] The shape or dimension of the initial tensor is described in Table 1; the main symbols and their corresponding meanings in this application are shown in Table 2.

[0022] Table 1. Description of the shape or dimension of the initial tensor:

[0023] Table 2. Key Symbols and Their Corresponding Meanings in This Application:

[0024] Step 204: Perform network encoding on the initial tensor to obtain the context vector.

[0025] The context vector includes graph features, numerical features, and constraint features. Graph features are determined by encoding using a dual relation graph encoder, numerical features are determined by multi-level linear transformation and activation, and constraint features are determined by truncation of features of a preset dimension from the numerical tensor.

[0026] Specifically, the row-normalized adjacency and initial node features of both the primary and backup road networks are input into a dual-relation graph encoder to encode the road network topology differences, resulting in graph features e_g. In other words, the primary and backup road network topologies are encoded using the dual-relation graph encoder to obtain graph features. The numerical tensor is then processed through two layers of MLP feature extraction to obtain numerical features e_v. Since the numerical tensor contains statistical data representing objective constraints on the primary and backup roads, as well as redundant correlation features, features of a portion of the numerical tensor can be extracted as explicit constraint features p, thus decoupling the extended modalities from the core statistical features under a fixed dimension. Finally, the graph features, numerical features, and constraint features are concatenated and linearly fused to obtain the context vector z.

[0027] In one embodiment, the context vector also includes placeholder dimension processing: the placeholder dimension can be written to zero or statistics by the data pipeline, so that the network interface dimension is fixed while services can progressively access time-series or text-encoded data.

[0028] Step 206: Sequence decoding of the context vector and determination of the lower-level scheduling decision for the vehicle dispatch task based on objective constraints.

[0029] The vehicle dispatch task represents the process of selecting vehicle resource slots step by step according to the task, such as assigning multiple tasks within the same decision window in a fixed order. The vehicles are assigned to R vehicle slots in sequence.

[0030] Specifically, the context vector is input to a sequence decoder with a causal mask, and the step-by-step task outputs the original decision scores (logits) for each vehicle. The causal mask represents that decisions can only be made sequentially. Under hard constraints such as capacity, a decoding mask is applied to step t, i.e., capacity feasibility is determined based on the initial vehicle load, thus masking the original decision scores corresponding to overloaded vehicles. The next-level scheduling decision for step t is determined based on the masked original decision scores. The task requirements of step t are accumulated into the selected initial vehicle load, and the capacity feasibility for step t+1 is recalculated using the updated load until all tasks are assigned, resulting in the next-level decision sequence for all tasks, i.e., the next-level scheduling decision. The mask is only used to constrain logits and is not used as a learnable parameter of the network.

[0031] In one embodiment, the process of determining capacity feasibility is determined based on vehicle load and task requirements. Specifically, it includes: determining whether the vehicle load of each vehicle r satisfies load_r(t-1)+dt≤Qr. If it does, the original decision score logit remains unchanged, and the vehicle is feasible at step t. If it does not, the original decision score logit is set to a minimum negative number (large negative bias) to preserve the probability quality on the feasible subset. Here, dt is the task requirement of the t-th task; Qr is the capacity of the r-th vehicle.

[0032] In one embodiment, the variable representation of the lower-level scheduling decision for vehicle dispatching tasks includes two forms. Discrete sequence form: , , indicating that the t-th task is assigned to the t-th task. A vehicle. 0-1 format: , , ,in If and only if The relationship between the two is as follows: .

[0033] In one embodiment, the objective constraints include: task uniqueness. Each task t is assigned to one vehicle. The vehicle capacity constraint is as follows: the sum of the task requirements carried by each vehicle shall not exceed the vehicle's capacity.

[0034] ;

[0035] Step 208: Determine the overall level of supply chain resilience based on the lower-level scheduling decisions.

[0036] Specifically, an evaluation index system and a chain of fixed rule formulas are provided, comprising a two-layer structure. The second layer is the lower-level scheduling decision corresponding to the business model; that is, the scheduling decisions of the second layer are the lower-level scheduling decisions. The first layer is the multi-dimensional upper-level performance decision; that is, the performance decisions of the first layer are the upper-level scheduling decisions. This is achieved by tensorizing the initial tensor containing on-site situational data such as graphs, demand, and capacity. The lower-level scheduling decision y is determined by the deterministic simulator. We obtain the raw scalar quantities such as delay, cost, default, and redundancy, and then process them separately. Normalization is mapped to [0, 1] to obtain multi-dimensional upper-level performance decisions. Finally, weighting and grading are performed based on these multi-dimensional upper-level performance decisions to obtain a comprehensive level of supply chain resilience. The chain relationship between the first and second layers is as follows:

[0037] The aforementioned supply chain resilience analysis and optimization method constructs an initial tensor based on the on-site situation and objective constraints of the supply chain, encompassing information such as logistics subgraphs, task requirements, vehicle capacity, and primary / backup road networks. A context vector is generated by fusing graph features, numerical features, and constraint features through network coding. The context vector is then sequence-decoded and combined with objective constraints to generate lower-level scheduling decisions. Finally, the comprehensive level of supply chain resilience is mapped based on the decision results. Therefore, a closed-loop process from situational awareness and optimization decision-making to resilience level evaluation is achieved, providing decision support and quantitative basis for supply chain resilience analysis.

[0038] In one embodiment, network encoding is performed on the initial tensor to obtain a context vector, including: performing two-layer graph convolution processing based on the initial node features and the adjacency of the primary and backup road networks to obtain target node features; performing linear mapping and mean pooling on the target node features to obtain graph features; performing multi-layer linear transformation and activation on the numerical tensor to obtain numerical features, and extracting features of a preset dimension from the numerical tensor as constraint features; and fusing the graph features, numerical features, and constraint features to obtain the context vector.

[0039] The initial tensor includes at least the initial node features of the logistics subgraph, the adjacency of the primary and backup road networks, and the numerical tensor. Graph features and numerical features are responsible for situational and demand-side encoding, while sequence decoding and progressive network encoding masks are responsible for vehicle capacity-side constraints, ensuring that the specificity of the network structure problem and the source of business constraints are consistent.

[0040] Specifically, the dual-relation graph encoder uses the primary relation adjacency of the primary and backup road networks in the two-layer stacked message passing. Adjacent to the preparation relationship and the loop branch road, the Layer node representation Updated by the following formula, For stepwise nonlinear activation, , Learnable parameters: ;

[0041] The above formula represents the stacked form of dual relation graph convolutional encoding, where X is the initial node feature, and the readout... Before can be Then, a linear mapping is performed. For example, taking the hidden dimension. At that time, the target node features of the two-layer message passing are mapped to through a linear layer. Re-pooling yields the encoded graph features. .

[0042] Meanwhile, for the numerical tensor of the numerical branch After two layers of affine mapping and ReLU mapping, it becomes , written as numerical features ,in The final layer is an identity mapping. We take numerical tensors. For example, the last 16 dimensions constitute ,and , After vector concatenation and linear fusion, we get The context vector expression is as follows

[0043] In this embodiment, a two-layer graph convolution is constructed by connecting initial node features with the primary and backup road networks to extract deep topological features of the target nodes. Graph features are generated through linear mapping and mean pooling, and numerical features are obtained by numerical tensor transformation, with constraint features extracted. These three features are fused to obtain a context vector, comprehensively integrating road network structure, numerical status, and constraint prior information, providing global awareness for subsequent sequence decoding and improving the accuracy and robustness of scheduling decisions.

[0044] In one embodiment, performing sequence decoding on the context vector and determining the lower-level scheduling decision for the vehicle dispatch task based on objective constraints includes: obtaining historical scheduling decisions; performing sequence decoding on the context vector and historical scheduling decisions to obtain an original decision score; performing capacity masking on the original decision score based on objective constraints to mask overloaded vehicles corresponding to the original decision score; and determining the lower-level scheduling decision for the vehicle dispatch task based on the masked original decision score.

[0045] The objective constraints include ensuring that the actual vehicle load does not exceed the vehicle's capacity. The process of sequence decoding the context vector is implemented through a sequence decoder, which includes two coding sub-layers, four multi-head self-attention paths per sub-layer, and feedforward hidden dimension extraction. Dropout is set to 0.1, and the self-attention layer uses a causal mask to adjust the position. Only depend Hidden state in each step through have to .

[0046] Specifically, the first The step maps the one-hot or index of the previous historical scheduling decision to the embedding layer. After being concatenated with the context vector, it is projected onto the decoding dimension through a linear layer. , Stacking The algorithm takes a sequence as input and uses a Transformer sequence decoder to output the original decision scores of vehicles task by task, instead of directly outputting the comprehensive level with a separate classification head to replace the link obtained from the simulation of the decision. A capacity feasible mask is constructed based on the real-time load and capacity constraints of the vehicles to constrain the original decision scores, thus masking overloaded vehicles corresponding to the original decision scores. The algorithm then takes the maximum index (argmax) of the masked original decision scores until the lower-level scheduling decision for the current task is obtained. Finally, the vehicle load status is updated based on the decision results to dynamically generate masks for subsequent steps, realizing a stepwise autoregressive lower-level scheduling decision for vehicle dispatching tasks with hard constraints.

[0047] In this embodiment, sequence decoding is performed based on context vectors and historical scheduling decisions to output the original vehicle dispatch decision score. The score is then subjected to capacity masking through objective constraints to directly mask the scores corresponding to overloaded vehicles, ensuring that the decision meets the hard capacity constraint. The next-level scheduling decision is then generated based on the masked decision score. This ensures that the scheduling scheme is compliant and feasible while realizing intelligent vehicle dispatch under constraint awareness, thereby improving the stability of scheduling execution and the practicality of business operations.

[0048] In one embodiment, determining the overall level of supply chain resilience based on lower-level scheduling decisions includes: determining the total mileage of the lower-level scheduling decisions under the main road network; obtaining simulation filing coefficients, and obtaining multi-dimensional upper-level performance decisions based on the total mileage and simulation filing coefficients; weighting the multi-dimensional upper-level performance decisions through multiple performance weights to determine the overall score; and mapping the overall score to an overall level based on the business-side requirements of supply chain resilience.

[0049] The primary and backup road networks include both the primary and backup road networks. The simulation registration coefficient must include at least the registered equivalent vehicle speed. Delayed truncation Path redundancy .

[0050] Specifically, in the deterministic simulator processing, the total mileage of each vehicle's route is first calculated on the main road network according to the lower-level scheduling decision and then summed to obtain the total mileage. (meters); then calculate the capacity violation based on the cumulative demand per vehicle and the vehicle capacity. ; Delay minutes are used to calculate mileage and time when the customer time window is not enabled. This is to distinguish between "time pressure" and "distance cost" represented by total mileage; path redundancy is statistically mapped from the ratio of primary to backup side lengths of each path segment to... Normalized, multi-dimensional upper-level performance decisions include: ; ; ; for Cut to ;

[0051] Next, by weighting multiple performance weights on multi-dimensional upper-level performance decisions, a comprehensive score representing the overall performance level is obtained. For example, the performance weight corresponding to the four-dimensional performance is 1 / 4, so as to maintain interpretable linear aggregation even when the semantics of each dimension are not completely of the same dimension, and the performance weight can be adjusted according to the business needs in actual deployment.

[0052] ;

[0053] Finally, based on the business-side requirements for supply chain resilience, the overall score can be mapped to a comprehensive level, for example, a five-tier ordinal system, in the following way: , This indicates rounding down. The simulation filing coefficients and other parameters mentioned above are all fixed before training and are consistent with the actual data generation side.

[0054] In this embodiment, the total mileage of the main road network is calculated based on the lower-level scheduling decision, and multi-dimensional upper-level performance decisions are generated by combining simulation filing coefficients. The comprehensive score is obtained by weighting multiple performance weights and mapped to a comprehensive level according to the business needs of supply chain resilience. This realizes the quantitative transformation from scheduling execution results to business value evaluation, clearly reflects the economy, compliance and comprehensive effectiveness of the scheduling plan, and provides a standardized and quantifiable business guidance basis for scheduling effect evaluation and optimization.

[0055] Table 3 shows the specific meaning of multi-dimensional upper-level performance decisions; Table 4 is a summary of the levels involved in this application and the symbol conventions, and the parameters between different levels have been semantically aligned.

[0056] Table 3. Specific meanings of multi-dimensional upper-level performance decision-making:

[0057] Table 4. Summary of Hierarchical Levels and Symbol Conventions in this Application:

[0058] In one embodiment, the lower-level scheduling decision is determined by a deep neural network. This involves encoding and decoding the initial tensor until the lower-level scheduling decision is obtained, using a deep neural network containing an encoder and decoder. The reference decision generated by the teacher algorithm in path one serves as the gold standard for training the deep neural network; the student network in path two serves as the inference path for the deep neural network. A decision loss function is used for primary supervision to align the scheduling decisions output by the student network with the standard sequence given by the teacher algorithm; a performance loss function incorporates deviations in performance decisions into validation or testing. Figure 3 As shown, Figure 3 A training process for a deep neural network is provided, including: Step 302: Obtain the sample tensor and run the teacher algorithm on the sample tensor to obtain the first scheduling decision.

[0059] Among them, the teacher algorithm is used for offline data construction and gold standard reference generation.

[0060] Specifically, along the path, for the acquired sample tensors, the teacher algorithm, under the conditions of satisfying the objective constraints and objective function mentioned above, and according to the greedy algorithm or mixed integer programming (MIP) mentioned above, obtains the first scheduling decision. The objective constraints and associated scheduling decision variables of the teacher algorithm are the same as those in the deep neural network operation process described above, and will not be repeated here.

[0061] Furthermore, let the main network distance matrix be denoted as... deposit is a node Will be assigned to the car The task set is denoted as The order in which the vehicle visits customer nodes is determined by ascending task numbers (consistent with path construction in deterministic simulation), and the closed-loop mileage from the depot back to the depot is calculated to obtain the total path length of the vehicle. .in To meet the aforementioned objective constraints and The lower-level scheduling decisions are uniquely assigned to all vehicles. Therefore, the global ideal objective can be written as minimizing the sum of the path lengths of all vehicles, which is also the objective function of the teacher algorithm: ;

[0062] In this context, the objective function optimizes only the scheduling decisions of the second layer, with the objective quantity being the sum of the closed-loop mileage of the main network. This represents the teacher's operational objective regarding transportation path costs within the feasible region. The first layer's four-dimensional performance decision k has a comprehensive score. The overall level L is not explicitly written into the objective function, but is determined after the first scheduling decision is obtained, through the deterministic simulator Sim and fixed normalization. This is derived from the fact that upper-level performance decisions are multi-indicator mappings of lower-level scheduling decisions through simulation and normalization chains, rather than being derived directly from the objective function mentioned above. Parallel multi-objective weighted average.

[0063] In one embodiment, running the teacher algorithm on the sample tensor to obtain a first scheduling decision includes: obtaining a set of capacity-feasible vehicles, selecting the vehicle with the smallest marginal mileage increment from the set of capacity-feasible vehicles, and obtaining the corresponding first scheduling decision; or, determining the linear objective of vehicle allocation, and solving the linear objective of vehicle allocation through mixed integer programming to obtain the corresponding first scheduling decision.

[0064] Specifically, for the greedy algorithm: determine the set of feasible vehicles. The set of feasible vehicles is the set of all vehicles that satisfy the loading constraint, where the sum of the current load and the task demand does not exceed the vehicle capacity limit when allocating tasks. In step t, the vehicle r with the smallest marginal mileage increment is selected from the set of feasible vehicles. The marginal mileage increment represents the increase in the total mileage of the vehicle compared to before insertion when the current task t is incorporated into each feasible vehicle according to the agreed insertion rules. The vehicle with the smallest marginal mileage increment is selected; if multiple vehicles have the same increment, the vehicle with the smaller number is selected, thus obtaining the first scheduling decision. This algorithm does not guarantee global optimality, but it is computationally deterministic and reproducible, making it suitable as a gold standard for batch processing. For mixed integer programming: deterministic... Vehicle allocation linear targets, under the aforementioned objective constraints and The feasible or optimal solution is then obtained, where a gap can be set within the time window to obtain the first scheduling decision.

[0065] In this embodiment, a first scheduling decision is formed by obtaining a set of vehicles with feasible capacity and selecting the vehicle with the smallest marginal mileage increment; or by constructing a linear objective for vehicle allocation and solving it using mixed integer programming. Under the premise of satisfying the hard constraint of vehicle capacity, feasible solutions can be generated quickly using a greedy strategy, or optimized solutions can be obtained through planning methods, providing high-quality solutions for subsequent scheduling model training.

[0066] Step 304: Perform network encoding and sequence decoding on the sample tensor to obtain the second scheduling decision.

[0067] Specifically, under path two, referring to steps 204 and 206 above, the initial tensor is subjected to network encoding and sequence decoding to obtain the specific process of the lower-level scheduling decision. It can be determined that the second scheduling decision is obtained in the following steps. The process will not be elaborated here. That is, by using network coding, sequence decoding with causal masking and capacity masking, the task allocation is structurally aligned with the road network and capacity constraints, which is beneficial for learning within the feasible region.

[0068] The training phase employs a teacher-mandated approach along the first path, with the first scheduling decision being made. As the decision scores for each step, which are derived from the previous step's labels and unfolded in parallel, the inference stage under the second-path student network unfolds step by step according to autoregression, applying a mask at each step, and updating the vehicle load state for the next step with the original decision scores after masking and the task requirements, until the second scheduling decision is obtained. .

[0069] Step 306: The first scheduling decision is used as the standard sequence, and the difference between the standard sequence and the second scheduling decision is determined by the decision loss function.

[0070] In this case, Path 1 and Path 2 share the same decision variable semantics in the second layer.

[0071] Specifically, the first scheduling decision provided by Path 1 is used as the standard sequence, i.e., the gold standard reference; the parameters of Path 2 are aligned with the scheduling decisions of the lower layers as the main supervision. Unlike the objective function above where the teacher only optimizes the main network mileage, the decision loss function directly targets the learning network. The following formula corresponds to the scheduling decision of the second layer, ensuring that the stepwise decision is consistent with the standard sequence given by the teacher's algorithm. The decision error between the second scheduling decision and the standard sequence is measured by the stepwise mask cross-entropy: ;

[0072] in For standard sequences, For the first Step, First The decision score of the slot after the capacity mask. Therefore, by using the teacher algorithm to generate reference decisions and labels in batches, the reliance on manual line-by-line labeling scheduling schemes is significantly reduced, which facilitates large-scale training and reproduction. Compared with expert line-by-line scheduling, automated inference can continuously forward process the entire batch of samples on fixed hardware, and the decision delay per sample is significantly lower than that of manual scheduling and repeated communication cycles.

[0073] Step 308: Update the model parameters based on the difference and backpropagation algorithm until a trained deep neural network is obtained.

[0074] Specifically, the difference between the label and the second scheduling decision is calculated using the decision loss function to obtain the loss value; backpropagation is performed based on this loss value to update the model parameters layer by layer, so as to continuously reduce the deviation between the prediction result and the true label, complete the model optimization, and obtain the trained deep neural network.

[0075] In the above embodiments, path one runs the teacher algorithm to generate a compliant and optimized first scheduling decision as a reference standard; path two performs network encoding and sequence decoding to output a second scheduling decision. The difference between the two scheduling decisions is calculated using a decision loss function, and deep neural network training is completed through backpropagation. Therefore, relying on high-quality teacher decisions to provide strong supervision enables the model to effectively learn the optimized scheduling logic under constraints and output stable and reliable scheduling results.

[0076] In one embodiment, the method further includes: simulating and mapping the first scheduling decision and the second scheduling decision to obtain the first performance decision and the second performance decision; determining the performance loss function based on the mean square error between the first performance decision and the second performance decision; and obtaining the total loss function by combining the decision loss function and the performance loss function based on the preset target weight.

[0077] Specifically, this involves using the reference decisions generated by the teacher's algorithm as a gold standard reference during the training phase, so that the estimation strategy can approximate the teacher's under the capacity mask constraint. In path two of the verification phase, the consistency between the performance decision and the reference given by the teacher's algorithm is measured using the same simulation caliber. This supports the aligned output of the "decision structure-performance indicators" required for chain leader comparison, collaborative review, and internal debriefing.

[0078] The first scheduling decision is the assignment sequence aligned with the teacher's output during training, and the second scheduling decision is the assignment sequence obtained by progressively assigning the student network output of path two.

[0079] Specifically, the scheduling decision is not directly predicted by the network end classifier as the main task. Instead, the first and second scheduling decisions are input into a deterministic simulator consistent with the business filing and a fixed normalization, respectively, to obtain the first performance decision. Second performance decision The deterministic simulator and the fixed normalization formula are implemented and parameterized in the same way during training, validation, and inference; the network's learnable parameters are not updated with gradients according to the simulator's internal rules. During the training phase, in addition to aligning with the standard sequence provided by the teacher's algorithm through the decision loss function, the first performance decision provided by path one is used as the gold standard reference. The deviation between the first and second performance decisions is incorporated into validation or testing monitoring through the performance loss function. That is, the following formula corresponds to the performance decision of the first layer, ensuring consistency between the performance decision and the reference provided by the teacher's algorithm, thereby explicitly distinguishing between "lower-layer progressive alignment" and "upper-layer performance consistency" at the optimization objective level.

[0080] The performance error is measured by the mean square term between the first and second performance decisions. ;

[0081] Therefore, compared to black-box models that only output scalar scores or natural language solutions, this embodiment constrains the network output to lower-level scheduling decisions consistent with a fixed pattern. Upper-level performance decisions are obtained through deterministic simulation and a chain of fixed rule formulas, satisfying the criteria for auditable and recalculated resilience evaluation. This facilitates alignment with the "executable structure" descriptions in partner or regulatory materials. The aligned results are used to explain scheme differences in collaborative reviews or internal debriefings. The necessity lies in the fact that without a unified tensor and unified simulation, cross-batch and cross-entity comparisons lack common semantics. The sufficiency lies in the fact that lower-level scheduling decisions and upper-level performance decisions are closed by the same formula chain, avoiding the management blind spot of only reporting the total score without drilling down to the assignment sequence.

[0082] In one embodiment, in addition to monitoring the performance error between the first performance decision and the second performance decision at the first level of consistency, upper-level consistency monitoring can be performed based on the difference between the two comprehensive scores after determining the comprehensive scores corresponding to path one and path two.

[0083] Furthermore, based on the preset target weights By weighting the decision loss function and the performance loss function, a multi-level total loss function is obtained. Typically, simulators are deterministic and non-differentiable for discrete decisions; during the training phase, the target weights are taken. The network parameters are backpropagated using only the decision loss function, while the performance loss function is reported during the validation or testing phase to measure the consistency between the performance decisions of the first layer and the teacher reference. At this point, the first layer still converges to the standard sequence with the decision loss function, indirectly aligning under the deterministic simulation chain. On the backpropagation path, the gradient is backpropagated to the decision scores and decoder and encoder parameters at each step via softmax and cross-entropy, without passing through the non-differentiable part of discrete assignment in the simulator.

[0084] In this embodiment, simulation and index mapping are performed on the first scheduling decision and the second scheduling decision respectively to obtain the corresponding first performance decision and the second performance decision; a performance loss function is constructed based on the mean square error between the two performance decisions, and the decision loss function and the performance loss function are weighted and fused together with preset target weights to obtain the total loss function; the joint optimization of lower-level scheduling decision alignment and upper-level performance consistency is realized, thereby improving the constraint compliance and comprehensive benefits of the scheduling scheme.

[0085] In one embodiment, if a differentiable approximation of the simulator is introduced, or a differentiable relaxation is applied to the assignment probability to make If the parameters are differentiable, then the target weight can be chosen. The optimization is performed jointly according to the total loss function, which means that the performance error is also added to the loss during training. The gradient learning is not performed on the registration coefficients of the deterministic simulator and the fixed normalization formula in the total loss function.

[0086] In summary, this application, aiming to achieve supply chain resilience analysis and optimization, employs a resilient supply chain scheduling method based on on-site situation and objective constraint sample learning, encompassing the fields of intelligent scheduling and deep learning technologies. By encoding the supply chain's on-site situation and computable objective constraints into an initial tensor, this initial tensor undergoes network encoding and sequence decoding to generate lower-level scheduling decisions aligned with a fixed pattern. Under a deterministic simulator and a fixed normalization formula, the lower-level scheduling decisions are used to calculate the upper-level performance decisions and overall rating. Therefore, this application achieves standardized scheduling decisions and a fixed, reproducible evaluation process, providing a stable and reliable quantitative basis for supply chain resilience analysis.

[0087] This invention has completed experimental simulations of data generation, model training, and test set evaluation. The data acquisition mode is as follows: the gold standard is primarily generated by a combination of teacher algorithms and deterministic simulation verification, rather than manual questionnaire-style annotation; this mode is consistent with the "situation-constraint-reference solution" batch generation pipeline, and parameters and random seeds are registered to ensure reproducibility. Regarding the data: multiple sets of instances are generated according to a fixed pattern, each instance containing primary and backup road networks, demand, capacity, and teacher reference assignments; the total sample size is 320, and 42 random seeds are generated. The partitioning ratio is approximately 80% for the training set, 10% for the validation set, and 10% for the test set; the partitioning is performed by shuffling the sample index once under a fixed random seed before splitting, ensuring no overlap between training, validation, and testing. The evaluation script and training script share the same partitioning logic for reproducibility.

[0088] It should be understood that although the steps in the flowcharts of the above embodiments are shown sequentially according to the arrows, these steps are not necessarily executed in the order indicated by the arrows. Unless explicitly stated herein, there is no strict order restriction on the execution of these steps, and they can be executed in other orders. Moreover, at least some steps in the flowcharts of the above embodiments may include multiple steps or multiple stages. These steps or stages are not necessarily completed at the same time, but can be executed at different times. The execution order of these steps or stages is not necessarily sequential, but can be performed alternately or in turn with other steps or at least some of the steps or stages of other steps.

[0089] Based on the same inventive concept, this application also provides a supply chain resilience analysis and optimization apparatus for implementing the supply chain resilience analysis and optimization method described above. The solution provided by this apparatus is similar to the solution described in the above method; therefore, the specific limitations in one or more embodiments of the supply chain resilience analysis and optimization apparatus provided below can be found in the limitations of the supply chain resilience analysis and optimization method described above, and will not be repeated here.

[0090] In one embodiment, a supply chain resilience analysis and optimization system is provided, comprising a tensor module, an encoder, a decoder, and a result determination module, wherein: The tensor module is used to determine the initial tensor based on the on-site situation and objective constraints of the supply chain; the on-site situation includes the logistics sub-map, task requirements, vehicle capacity, and primary and backup road networks; The encoder is used to perform network encoding on the initial tensor to obtain a context vector; the context vector includes graph features, numerical features, and constraint features. The decoder is used to perform sequence decoding on the context vector and determine the lower-level scheduling decision for dispatching tasks based on objective constraints. The results determination module is used to determine the overall level of supply chain resilience based on lower-level scheduling decisions.

[0091] Furthermore, the tensor quantization module can also be called the situation and constraint tensor quantization module, the encoder can also be called the dual relation graph and numerical encoding module, the decoder can also be called the masked sequence decoding module, and the result determination module can also be called the deterministic simulation and performance normalization module (the implementation carrier for the fixed rule formula chain). The system also includes a training and inference scheduling module for training deep neural networks. The data interfaces of each module are defined by tensor shapes and fixed patterns to facilitate engineering implementation and auditing. These modules can be deployed as software components on standalone machines, server clusters, or edge computing nodes.

[0092] The modules in the aforementioned supply chain resilience analysis and optimization system can be implemented entirely or partially through software, hardware, or a combination thereof. These modules can be embedded in or independent of the processor in a computer device, or stored in the computer device's memory as software, so that the processor can invoke and execute the corresponding operations of each module.

[0093] In one embodiment, a computer device is provided, which may be a server, and its internal structure diagram may be as follows: Figure 4 As shown, this computer device includes a processor, memory, input / output interfaces (I / O), and a communication interface. The processor, memory, and I / O interfaces are connected via a system bus, and the communication interface is also connected to the system bus via the I / O interfaces. The processor provides computational and control capabilities. The memory includes non-volatile storage media and internal memory. The non-volatile storage media stores the operating system, computer programs, and a database. The internal memory provides the environment for the operation of the operating system and computer programs stored in the non-volatile storage media. The database stores lower-level scheduling decisions. The I / O interfaces are used for exchanging information between the processor and external devices. The communication interface is used for communication with external terminals via a network connection. When the computer program is executed by the processor, it implements a supply chain resilience analysis and optimization method.

[0094] Those skilled in the art will understand that Figure 4 The structure shown is merely a block diagram of a portion of the structure related to the present application and does not constitute a limitation on the computer device to which the present application is applied. Specific computer devices may include more or fewer components than those shown in the figure, or combine certain components, or have different component arrangements.

[0095] In one embodiment, a computer device is also provided, including a memory and a processor, wherein the memory stores a computer program, and the processor executes the computer program to implement the steps in the above method embodiments.

[0096] In one embodiment, a computer-readable storage medium is provided storing a computer program that, when executed by a processor, implements the steps in the above method embodiments.

[0097] In one embodiment, a computer program product or computer program is provided, the computer program product or computer program including computer instructions stored in a computer-readable storage medium. A processor of a computer device reads the computer instructions from the computer-readable storage medium, and executes the computer instructions, causing the computer device to perform the steps in the above method embodiments.

[0098] Those skilled in the art will understand that all or part of the processes in the methods of the above embodiments can be implemented by a computer program instructing related hardware. The computer program can be stored in a non-volatile computer-readable storage medium. When executed, the computer program can include the processes of the embodiments of the above methods. Any references to memory, databases, or other media used in the embodiments provided in this application can include at least one of non-volatile and volatile memory. Non-volatile memory can include read-only memory (ROM), magnetic tape, floppy disk, flash memory, optical memory, high-density embedded non-volatile memory, resistive random access memory (ReRAM), magnetic random access memory (MRAM), ferroelectric random access memory (FRAM), phase change memory (PCM), graphene memory, etc. Volatile memory can include random access memory (RAM) or external cache memory, etc. By way of illustration and not limitation, RAM can take many forms, such as Static Random Access Memory (SRAM) or Dynamic Random Access Memory (DRAM). The databases involved in the embodiments provided in this application may include at least one type of relational database and non-relational database. Non-relational databases may include, but are not limited to, blockchain-based distributed databases. The processors involved in the embodiments provided in this application may be general-purpose processors, central processing units, graphics processing units, digital signal processors, programmable logic devices, quantum computing-based data processing logic devices, etc., and are not limited to these.

[0099] The technical features of the above embodiments can be combined in any way. For the sake of brevity, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this specification.

[0100] The above embodiments are merely illustrative of several implementation methods of this application, and their descriptions are relatively specific and detailed. However, they should not be construed as limiting the scope of this application. It should be noted that those skilled in the art can make various modifications and improvements without departing from the concept of this application, and these all fall within the protection scope of this application. Therefore, the protection scope of this application should be determined by the appended claims.

Claims

1. A method for supply chain resilience analysis and optimization, characterized in that, The method includes: The initial tensor is determined based on the on-site situation and objective constraints of the supply chain; the on-site situation includes at least one or more of the following: logistics sub-graph, task requirements, vehicle capacity, and primary and backup road networks. The initial tensor is encoded using a network to obtain a context vector; the context vector includes graph features, numerical features, and constraint features. The context vector is sequence decoded, and the lower-level scheduling decision for the vehicle dispatch task is determined based on the objective constraints. The overall level of supply chain resilience is determined based on the lower-level scheduling decisions. The lower-level scheduling decision is determined by a deep neural network. The training process of the deep neural network includes: acquiring a sample tensor and running a teacher algorithm on the sample tensor to obtain a first scheduling decision; performing network encoding and sequence decoding on the sample tensor to obtain a second scheduling decision; using the first scheduling decision as a standard sequence and determining the difference between the standard sequence and the second scheduling decision through a decision loss function; updating the model parameters based on the difference and the backpropagation algorithm until a trained deep neural network is obtained. The method further includes: simulating and mapping the first scheduling decision and the second scheduling decision respectively to obtain a first performance decision and a second performance decision; determining a performance loss function based on the mean square error between the first performance decision and the second performance decision; and obtaining a total loss function by combining the decision loss function and the performance loss function according to a preset target weight, thereby achieving joint optimization of lower-level scheduling decision alignment and upper-level performance consistency.

2. The method according to claim 1, characterized in that, The initial tensor includes at least the initial node features of the logistics subgraph, the adjacency of the primary and backup road networks, and any one or more numerical tensors; network encoding is performed on the initial tensor to obtain a context vector, including: Based on the initial node features and the adjacency of the primary and backup road networks, a two-layer graph convolution process is performed to obtain the target node features; The target node features are linearly mapped and mean pooled to obtain graph features; The numerical tensor is subjected to multi-level linear transformation and activation to obtain numerical features, and features of a preset dimension are extracted from the numerical tensor as constraint features. The graph features, numerical features, and constraint features are fused to obtain a context vector.

3. The method according to claim 1, characterized in that, The objective constraints include that the actual load of the vehicle does not exceed the vehicle capacity. Sequence decoding of the context vector and determination of the lower-level scheduling decision for the vehicle dispatch task based on the objective constraints include: Obtain historical scheduling decisions, and perform sequence decoding based on the context vector and the historical scheduling decisions to obtain the original decision score; Based on the objective constraints, the original decision scores are capacity-masked to mask out the overloaded vehicles corresponding to the original decision scores; Based on the original decision score after masking, the lower-level scheduling decision for the vehicle dispatch task is determined.

4. The method according to claim 1, characterized in that, The primary and backup road network includes the primary road network; the determination of the comprehensive level of supply chain resilience based on the lower-level scheduling decision includes: Determine the total mileage traveled under the main road network for the lower-level scheduling decision; Obtain the simulation filing coefficient, and based on the total driving mileage and the simulation filing coefficient, obtain multi-dimensional upper-level performance decisions; The comprehensive score is determined by weighting multiple performance factors to make multi-dimensional upper-level performance decisions. Based on the business-side requirements for supply chain resilience, the comprehensive score is mapped to a comprehensive grade.

5. The method according to any one of claims 1 to 4, characterized in that, The teacher algorithm is run on the sample tensor to obtain a first scheduling decision, including: Obtain a set of vehicles with available capacity, and select the vehicle with the smallest marginal mileage increment from the set of vehicles with available capacity to obtain the corresponding first scheduling decision; Alternatively, determine the linear objective for vehicle allocation, and solve the linear objective for vehicle allocation using mixed integer programming to obtain the corresponding first scheduling decision.

6. A supply chain resilience analysis and optimization system, characterized in that, For implementing the method according to any one of claims 1 to 5, the system comprises a tensor quantization module, an encoder, a decoder, and a result determination module, wherein: The tensor quantization module is used to determine the initial tensor based on the on-site situation and objective constraints of the supply chain; the on-site situation includes the logistics sub-map, task requirements, vehicle capacity, and primary and backup road networks; An encoder is used to perform network encoding on the initial tensor to obtain a context vector; the context vector includes graph features, numerical features, and constraint features. A decoder is used to perform sequence decoding on the context vector and determine the lower-level scheduling decision for the vehicle dispatch task based on the objective constraints. The result determination module is used to determine the overall level of supply chain resilience based on the lower-level scheduling decisions.

7. A computer device comprising a memory and a processor, wherein the memory stores a computer program, characterized in that, When the processor executes the computer program, it implements the steps of the method according to any one of claims 1 to 5.

8. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by a processor, it implements the steps of the method according to any one of claims 1 to 5.

Images