A Deep Learning-Based Intelligent Supplier Order Allocation Method and System

Abstract

This invention belongs to the field of supply chain management technology. It discloses a method and system for intelligent allocation of multi-objective supplier orders based on deep learning. By constructing a piecewise softening mapping function for supplier capacity boundaries, the hard capacity ceiling is transformed into a three-segment constraint space containing a safety zone, a transition zone, and a prohibited zone. Urgency and resource consumption intensity are calculated based on order demand characteristics to construct a priority sequence. A capacity margin buffer is established for each supplier. Constraint gradient features are embedded into the loss function of a deep learning model to train and generate a constraint-aware allocation probability model. A transformation from probabilistic to deterministic allocation is achieved through hierarchical constraint projection transformation. A backtracking reallocation is performed when a warning boundary is reached. This invention achieves an organic integration of constraints and optimization, enhancing the resilience and risk resistance of the supply chain under demand fluctuations.

Description

Technical Field

[0001] This invention relates to the field of supply chain management technology, and more specifically, to a method and system for intelligent allocation of multi-objective supplier orders based on deep learning. Background Technology

[0002] Existing supplier order allocation technologies suffer from deficiencies in satisfying hard constraints, particularly when applying deep learning and reinforcement learning methods. Traditional deep learning models typically output probability distributions or continuous value predictions; this "soft output" fundamentally conflicts with insurmountable hard constraints such as supplier capacity limits. In real-world production environments, even minor violations of constraints by the model-predicted allocation scheme (e.g., allocation reaching 101% or 102% of capacity limits) can lead to a chain reaction of production line overload, decreased yield, delivery delays, and even equipment damage. Existing methods often treat constraint handling as a separate step after model output, employing heuristic rules or greedy algorithms for post-processing corrections. This not only undermines the global optimality of the original solution but also introduces additional computational latency. Deep learning-based order allocation systems require mixed-integer programming (MIP) solvers for scheme correction during peak periods, extending inference time from milliseconds to over 30 seconds, impacting decision-making efficiency. Furthermore, existing models typically employ indirect methods such as penalty terms or Lagrange relaxations to handle constraints during training, making it difficult for the model to accurately perceive the "hard" nature of the constraint boundaries. This results in a constraint violation rate as high as 15%-25% for models that perform well in the testing environment but are deployed to the real world. Simultaneously, the integration of deep learning with traditional operations research solvers is usually a loose superposition rather than an organic fusion. The two use different problem representations and optimization paradigms, leading to low collaborative efficiency in large-scale real-time allocation scenarios. This fails to guarantee strict constraint satisfaction and makes it difficult to fully utilize the advantages of deep learning in feature extraction and pattern recognition.

[0003] In view of this, the present invention proposes a multi-objective supplier order intelligent allocation method and system based on deep learning to solve the above problems. Summary of the Invention

[0004] To overcome the aforementioned shortcomings of the prior art and to achieve the above objectives, the present invention provides the following technical solution: a multi-objective supplier order intelligent allocation method based on deep learning, comprising: Obtain supplier capacity constraints and order demand characteristics data; Based on the capacity constraint data, a segmented softening mapping function for the supplier capacity boundary is constructed. The segmented softening mapping function transforms the hard capacity limit into a three-segment constraint space that includes a safe zone, a transition zone, and a prohibited zone. Based on the order demand characteristic data, the urgency and resource consumption intensity of each order are calculated, and an order allocation priority sequence is constructed. In the three-segment constraint space, a capacity margin buffer is established for each supplier, and the boundary value of the capacity margin buffer is determined by the statistical distribution of historical allocation deviations. The gradient features of the segmented softening mapping function are embedded into the loss function of the deep learning model to train and generate a constraint-aware order allocation probability model. The initial allocation probability matrix of each order to each supplier is output through the order allocation probability model. A hierarchical constraint projection transformation is performed on the initial allocation probability matrix, which transforms the soft probability values ​​in the initial allocation probability matrix into a deterministic allocation matrix that satisfies the capacity margin buffer constraint. Detect whether the deterministic allocation matrix touches the warning boundary of the capacity margin buffer; When the warning boundary is detected, a rollback reallocation is performed according to the order allocation priority sequence, migrating the allocation of low-priority orders to suppliers with redundant capacity. Output the final order allocation scheme after the rollback redistribution verification, in which the allocation of all suppliers in the final order allocation scheme strictly does not exceed the hard capacity limit in the capacity constraint data.

[0005] A deep learning-based intelligent multi-objective supplier order allocation system, used to implement the deep learning-based intelligent multi-objective supplier order allocation method, includes: The data acquisition module is used to acquire supplier capacity constraint data and order demand characteristic data; The constraint softening mapping module is used to construct a piecewise softening mapping function for the supplier's capacity boundary based on the capacity constraint data. The piecewise softening mapping function transforms the hard capacity limit into a three-segment constraint space that includes a safety zone, a transition zone, and a prohibited zone. The priority sorting module is used to calculate the urgency and resource consumption intensity of each order based on the order demand characteristic data, and to construct an order allocation priority sequence. The buffer design module is used to establish a capacity margin buffer for each supplier in the three-segment constraint space. The constraint-aware modeling module is used to embed the gradient features of the piecewise softening mapping function into the loss function of the deep learning model, train and generate a constraint-aware order allocation probability model; and output the initial allocation probability matrix of each order to each supplier through the order allocation probability model. The hierarchical projection transformation module is used to perform hierarchical constraint projection transformation on the initial allocation probability matrix. The hierarchical constraint projection transformation transforms the soft probability values ​​in the initial allocation probability matrix into a deterministic allocation matrix that satisfies the capacity margin buffer constraint. The risk detection module is used to detect whether the deterministic allocation matrix touches the warning boundary of the capacity margin buffer. When the warning boundary is detected, the rollback adjustment module is used to perform rollback-style reallocation according to the order allocation priority sequence, and the final order allocation scheme after verification by the rollback-style reallocation is output.

[0006] The technical effects and advantages of the multi-objective supplier order intelligent allocation method and system based on deep learning in this invention are as follows: This invention can output executable allocation schemes that strictly adhere to supplier capacity limits, eliminating the risk of production overload. The organic integration of strict constraints and optimization objectives allows enterprises to maximize order fulfillment rates and resource utilization efficiency while ensuring production safety boundaries. Through an intelligent buffer mechanism, this invention effectively addresses uncertainties in real-world production environments, improving the stability of allocation scheme execution. In high-pressure scenarios of order competition for resources, this invention, through multi-dimensional priority assessment and tiered protection mechanisms, ensures that high-value orders and strategic customer needs are prioritized, effectively enhancing customer satisfaction and the enterprise's market competitiveness. In the face of extreme situations such as sudden demand surges and capacity constraints, the invention's intelligent rollback and rebalancing capabilities provide enterprises with a solid defense against market fluctuations, enhancing the resilience and risk resistance of the supply chain. Attached Figure Description

[0007] Figure 1 This is a schematic diagram of the deep learning-based multi-objective supplier order intelligent allocation method of the present invention; Figure 2 This is a schematic diagram of the multi-objective supplier order intelligent allocation system based on deep learning according to the present invention. Detailed Implementation

[0008] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0009] This application provides a method and system for intelligent allocation of multi-objective supplier orders based on deep learning. The system's execution entities include, but are not limited to, supply chain management systems, production scheduling platforms, order allocation centers, and intelligent manufacturing systems, which can be considered general computing nodes in this application. The intelligent allocation method includes, but is not limited to, at least one of the following: a deep learning-based allocation model, an adaptive constraint processing engine, and a hierarchical projection transformation system.

[0010] Please see Figure 1 In this embodiment of the invention, the specific implementation process of the deep learning-based multi-objective supplier order intelligent allocation method includes: Acquire supplier capacity constraint data and order demand characteristic data. Supplier capacity constraint data includes key information such as each supplier's maximum production capacity, equipment utilization rate, working time limits, and historical capacity fluctuations. Order demand characteristic data includes features such as order quantity, delivery deadline, complexity, and priority, acquired in real time through a multi-dimensional data interface. Supplier capacity constraint data records the suppliers' hard capacity limits and production efficiency, while order demand characteristic data details the specific requirements and time constraints of each order. This data provides comprehensive raw material for subsequent analysis, ensuring the rationality and feasibility of the allocation plan.

[0011] Based on supplier capacity constraint data, a piecewise softening mapping function is constructed for the supplier capacity boundary. This function transforms the hard capacity ceiling into a three-part constraint space comprising a safe zone, a transition zone, and a prohibited zone, converting the original binary hard constraint into a continuous penalty function. The safe zone corresponds to the interval with low capacity utilization, generating a small linear penalty; the transition zone corresponds to the interval close to the capacity ceiling, generating a smooth transition penalty value through the sigmoid function; and the prohibited zone corresponds to the interval exceeding the hard capacity ceiling, generating an exponentially increasing high penalty value. This three-part constraint space provides a continuously differentiable optimization objective for the deep learning model while maintaining respect for the hard constraints.

[0012] Based on order demand characteristic data, the urgency and resource consumption intensity of each order are calculated to construct an order allocation priority sequence. Urgency is calculated by the difference between the order delivery deadline and the current time, reflecting the time pressing nature of the order; resource consumption intensity is assessed based on the number and complexity of orders, indicating the extent to which orders occupy supplier capacity. By comprehensively considering these two indicators, a priority score is generated for each order, and an allocation priority sequence is constructed accordingly, providing an important basis for subsequent allocation decisions.

[0013] Within the three-segment constraint space, a capacity buffer is established for each supplier. The boundary values ​​of the capacity buffer are determined through the statistical distribution of historical allocation deviations; the upper boundary is close to but strictly below the hard capacity limit, and the lower boundary is located at the midpoint of the transition zone. This buffer serves as a safety mechanism to prevent capacity overruns due to prediction errors or actual execution deviations, thereby enhancing the robustness of the solution.

[0014] The gradient features of a piecewise softening mapping function are embedded into the loss function of a deep learning model to train and generate a constraint-aware order allocation probability model. This model comprises an encoding layer, a constraint-aware layer, and an output layer. The encoding layer extracts order and supplier features, the constraint-aware layer fuses constraint gradient information, and the output layer generates allocation probabilities. Through optimization of a multi-objective loss function, the model simultaneously considers allocation efficiency, load balancing, and constraint satisfaction, achieving intelligent solutions to complex allocation problems.

[0015] The order allocation probability model outputs an initial allocation probability matrix for each order to each supplier. This matrix represents the probability distribution of each order being assigned to each supplier, reflecting the model's evaluation of the order-supplier matching degree and providing a basis for subsequent deterministic allocation.

[0016] A hierarchical constraint projection transformation is performed on the initial allocation probability matrix to transform the soft probability values ​​in the initial allocation probability matrix into a deterministic allocation matrix that satisfies the capacity margin buffer constraint. This transformation process is executed sequentially according to order priority, ensuring that high-priority orders receive resource allocation first, while strictly controlling the cumulative allocation of each supplier to not exceed the upper boundary of its capacity margin buffer.

[0017] The system detects whether the deterministic allocation matrix has reached the warning boundary of the capacity margin buffer. By calculating the capacity safety margin of each supplier, the risk level of the allocation plan is assessed. When multiple suppliers are simultaneously approaching the warning boundary and have important orders, a fallback reallocation mechanism is triggered to further optimize the allocation plan.

[0018] When a warning boundary is detected, a rollback reallocation is performed based on the order allocation priority sequence, migrating the allocation of low-priority orders to suppliers with redundant capacity. The rollback reallocation process starts by adjusting low-priority orders, determining the migration amount by calculating the urgency of suppliers' capacity overruns, and reallocating them to suppliers with sufficient capacity based on the matching score, thus effectively mitigating risks.

[0019] The output is a final order allocation scheme validated through a rollback-based reallocation process, ensuring that the allocation to all suppliers strictly does not exceed the hard capacity limit specified in the capacity constraint data. This final scheme achieves multi-objective optimization of order allocation while guaranteeing constraint satisfaction, providing efficient and feasible decision support for supply chain management.

[0020] In this embodiment of the invention, the detailed implementation steps for constructing a piecewise softening mapping function for the supplier's capacity boundary based on capacity constraint data include: Extract the hard capacity cap for each supplier from the capacity constraint data. The hard capacity cap is the core indicator of capacity constraints, representing the maximum number of orders a supplier can handle within a given time. The extraction process is based on the supplier's historical data and current status, considering factors such as equipment quantity, working hours, and staffing to comprehensively calculate an accurate capacity cap. Specific capacity assessment models are used for different types of suppliers; for example, a time-based assessment method is used for discrete manufacturing suppliers, while a throughput-based assessment method is used for process manufacturing suppliers, ensuring the accuracy and applicability of the extraction results. The extracted hard capacity cap serves as the benchmark for subsequent softening mapping, providing a quantitative basis for determining the constraint boundaries.

[0021] Based on a rigid capacity ceiling, a first percentage threshold is set downwards, and the range where capacity utilization falls below this threshold is marked as the safety zone. This first percentage threshold is typically set at 70%-85% of the rigid capacity ceiling, with the specific value dynamically adjusted based on supplier type and industry characteristics. When a supplier's capacity utilization is within the safety zone, it indicates that the supplier has sufficient production capacity, low allocation risk, and can freely allocate orders without excessive consideration of capacity constraints. The safety zone provides a flexible range for allocation decisions and represents an ideal operating range for capacity utilization, helping suppliers maintain stable and efficient production.

[0022] A transition zone is established between the first percentage threshold and the hard capacity ceiling. The width of the transition zone is determined by a multiple of the supplier's historical capacity fluctuation standard deviation. The transition zone is a key area for constraint softening, representing a state where capacity utilization is close to but has not yet reached the ceiling. The width is typically calculated using 2-3 times the historical capacity fluctuation standard deviation, reflecting the range of uncertainty in capacity forecasting. Setting a transition zone allows for a gradual penalty, rather than a hard cutoff, for allocations approaching the capacity ceiling, providing a buffer between constraint satisfaction and optimization objectives. The specific range of the transition zone is dynamically adjusted based on the supplier's stability; a wider transition zone is set for suppliers with greater volatility, and a narrower transition zone is set for suppliers with less volatility, achieving personalized and refined constraint handling.

[0023] A prohibited zone is established above the hard capacity limit, and the corresponding mapping function value for the prohibited zone grows exponentially. The prohibited zone represents an allocation state exceeding the supplier's actual capacity, and orders should be avoided in this zone as much as possible. The exponentially increasing penalty value ensures a high degree of avoidance of the prohibited zone during optimization, while maintaining the continuity and differentiability of the function. The penalty growth rate is set based on the degree of exceedance, typically using an exponential function form such as... ,in The proportion exceeding the hard capacity limit, The parameter used to control the rate of penalty growth is typically set to 5-10, ensuring that even a small overshoot can produce a significant penalty effect.

[0024] A piecewise softening mapping function is constructed. This function outputs a linear penalty value in the safe region, a smooth transition penalty value based on the Sigmoid function in the transition region, and an exponential penalty value in the forbidden region. The mathematical expression of the piecewise softening mapping function is: ; in, For the value of the mapping function, For capacity utilization, This is the first percentage threshold (upper bound of the safe zone). This is a hard capacity ceiling. The linear penalty coefficient within the safe zone. The penalty value for the transition zone and the prohibited zone. The base penalty coefficient for the prohibited area. The growth rate of penalties for prohibited areas, The center point of the Sigmoid function (usually set to ) ), This is the scaling parameter for the Sigmoid function, controlling the steepness of the transition.

[0025] The output of the piecewise softening mapping function is used as a constraint penalty term during the training process of the deep learning model. During model optimization, the cumulative allocation amount for each supplier is converted into a corresponding penalty value through the mapping function and added to the overall loss function, guiding the model to learn a constraint-aware allocation strategy. The introduction of the penalty term enables the model to learn the boundary characteristics of capacity constraints during the training phase, achieving constraint softening while maintaining the practicality and feasibility of the solution.

[0026] In this embodiment of the invention, the detailed implementation steps for calculating the urgency and resource consumption intensity of each order based on order demand characteristic data and constructing an order allocation priority sequence include: Order delivery deadlines, order amounts, order quantities, and order complexity are extracted from order demand feature data. These features are key dimensions for evaluating order priority, comprehensively reflecting the time urgency, economic importance, and resource requirements of an order. The extraction process employs structured data processing techniques to transform the raw order data into standardized feature vectors, facilitating subsequent calculations and comparisons. For missing or outlier feature values, specific data repair strategies, such as mean imputation, regression prediction, or category inference, are used to ensure feature integrity and consistency. These extracted features constitute the foundational dataset for order priority calculation, providing input for subsequent urgency and consumption intensity calculations.

[0027] A time urgency factor is calculated based on the difference between the order's delivery deadline and the current time. This factor maps the remaining time using an exponentially decaying function. The time urgency factor reflects the degree of time pressure on the order and increases as the delivery deadline approaches. The calculation formula is: ; in, As for the time urgency factor, For order delivery deadline, For the current time, This is a time decay parameter that controls the growth rate of the urgency coefficient. The value of the time urgency coefficient is typically set based on industry characteristics. Fast-response industries use larger values ​​(e.g., 0.1-0.2), while planned production industries use smaller values ​​(e.g., 0.05-0.08). The exponential form of the time urgency coefficient ensures that the urgency increases rapidly and non-linearly as the deadline approaches, which aligns with the changing patterns of time pressure in actual business operations.

[0028] The urgency level is calculated by multiplying the order amount by the time urgency factor. Urgency level comprehensively considers both the economic value and time urgency of the order, reflecting its overall importance. Orders with high value and close to their delivery deadline receive the highest urgency score and are given priority in resource allocation. The calculation formula is as follows: ; in, For orders The urgency of the demand For the corresponding order amount, The urgency factor represents the time constraint. Calculating the urgency of demand enables a comprehensive assessment across time and value dimensions, providing a primary basis for order prioritization. In multi-objective optimization, urgency serves as a key indicator, guiding the prioritization of high-value, time-sensitive orders, thereby improving overall service quality and economic efficiency.

[0029] Based on order quantity and complexity, the expected consumption of supplier capacity by a single order is calculated, denoted as resource consumption intensity. Resource consumption intensity quantifies the degree of demand for supplier resources from an order and is a crucial factor in allocation decisions. The calculation formula is: ; in, For orders resource consumption intensity For the corresponding order quantity, This represents the corresponding order complexity (normalized to a value between 0 and 1). This is a complexity weighting coefficient, typically ranging from 0.5 to 2, reflecting the degree to which complexity affects resource consumption. Calculating resource consumption intensity accurately assesses the resource requirements of each order, providing a quantitative basis for meeting capacity constraints and offering crucial decision variables for order allocation optimization.

[0030] The ratio of demand urgency to resource consumption intensity is calculated and denoted as the order allocation priority score. The priority score reflects the value and efficiency of an order per unit of resource consumption and is a comprehensive indicator for judging the importance of an order. The calculation formula is as follows: ; in, For orders The allocation priority score, According to the corresponding urgency of the demand, This corresponds to the resource consumption intensity. This calculation method prioritizes urgent and resource-efficient orders, achieving efficient resource allocation. The ratio form of priority scores ensures that orders of different sizes and types can be compared and ranked under a unified standard, providing a scientific basis for constructing a globally optimal allocation scheme.

[0031] All orders are sorted in descending order of their priority scores to generate an order allocation priority sequence. Orders with higher priority scores receive capacity allocation first in the hierarchical constraint projection transformation. The construction of this priority sequence enables structured processing of order allocation, ensuring that resources are prioritized for high-priority orders. The sorting process employs efficient sorting algorithms, such as quicksort or heapsort, to handle large-scale order sets. For orders with similar priority scores, secondary sorting rules are introduced, such as considering customer importance or historical cooperation relationships, to further refine the sorting results. The final priority sequence provides a clear processing order for subsequent hierarchical constraint projection transformations, ensuring the rationality and effectiveness of resource allocation.

[0032] In this embodiment of the invention, the detailed implementation steps for establishing a capacity margin buffer for each supplier in the three-segment constraint space include: The process involves statistically analyzing the deviation series between the actual and predicted capacity utilization rates of each supplier in historical allocations. This deviation series reflects the accuracy and stability of supplier capacity forecasts and serves as the foundational data for establishing reasonable buffer zones. The statistical process collects historical allocation data over a specific period (typically the last 6-12 months), calculates the difference between the predicted and actual values ​​for each allocation, and forms a time series. To improve statistical reliability, outliers are identified and processed, such as removing extreme data points affected by unforeseen circumstances or standardizing seasonal fluctuations. The processed deviation series provides a reliable statistical sample for subsequent analysis, reflecting the true volatility characteristics of supplier capacity execution.

[0033] Quantile analysis is performed on the deviation sequence to extract the deviation value corresponding to the 95th quantile, which is denoted as the one-sided safety margin. Quantile analysis is a statistical method for quantifying the distribution of deviations, capturing the distribution characteristics and extreme values ​​of the data. Choosing the 95th quantile as the basis for the safety margin means that actual production capacity can be guaranteed with 95% confidence that it will not exceed the reserved buffer. The analysis process first sorts the deviation sequence, and then calculates the deviation value corresponding to the 95th quantile using interpolation methods. For suppliers with insufficient data, resampling techniques such as Bootstrap are used to enhance statistical reliability, or Bayesian inference methods are introduced to integrate prior knowledge to ensure the accuracy of the safety margin estimation. The extraction of the one-sided safety margin provides a data basis for determining the buffer boundary, reflecting the tolerance for production capacity fluctuations.

[0034] Based on the unilateral safety margin, the upper boundary value of the capacity margin buffer is determined by offsetting downwards from the rigid capacity ceiling. The upper boundary value is a key parameter of the buffer, directly affecting the conservatism of capacity constraints. The calculation formula is: ; in, The upper boundary value of the buffer. This is a hard capacity ceiling. The safety margin ratio is determined based on the one-sided safety margin, typically the ratio of the one-sided safety margin to the capacity ceiling plus a small safety factor (e.g., 0.01-0.02). Setting the upper boundary value ensures that, under normal fluctuation conditions, the actual capacity utilization rate rarely exceeds the hard limit, providing the necessary safety margin while maintaining efficient utilization of capacity resources.

[0035] Calculate the midpoint of the transition zone and set it as the lower boundary value of the capacity margin buffer. The lower boundary value marks the starting position of the buffer; capacity utilization below this value is considered a safe state. The midpoint position is typically defined as: ; in, The lower boundary value of the buffer. This is the upper limit of the safe zone (first percentage threshold). This is a hard upper limit for capacity. This setting ensures that the buffer zone covers the upper half of the transition zone, providing early warning and adjustment space for situations approaching the capacity limit, while not excessively restricting capacity utilization and maintaining efficient system operation. A reasonable lower limit allows the system to identify potential capacity pressure early, taking preventative measures before problems escalate, thus improving overall operational stability and reliability.

[0036] Within the capacity buffer zone, the contraction coefficient of the upper boundary value is dynamically adjusted based on the ratio of the total demand of currently pending orders to the total capacity of suppliers. The contraction coefficient reflects the system's adaptive adjustment to capacity constraints, enhancing constraint protection under high demand pressure. The calculation formula is: ; in, The shrinkage coefficient, This is the minimum shrinkage coefficient (usually set to 0.8-0.9). The sensitivity parameter represents the strength of the control system's response to demand pressure. This is the demand-capacity ratio (total demand divided by total capacity). The tension threshold is typically set between 0.85 and 0.95. Dynamic adjustment of the contraction coefficient enables the system to adapt to changes in market demand, enhancing capacity protection under high-pressure conditions and relaxing constraints under low-pressure conditions, thus improving the flexibility and adaptability of resource allocation.

[0037] When the ratio exceeds the preset tension threshold, the contraction coefficient is set to a value less than 1, causing the effective upper boundary of the capacity reserve buffer to contract downwards, thus strengthening the constraint protection. The formula for calculating the effective upper boundary is: ; in, The adjusted effective upper boundary, The lower boundary of the buffer zone. The original upper boundary, This is the contraction coefficient. When market demand is tight ( ), shrinkage coefficient A value less than 1 causes the effective upper boundary to shift downward, reducing the allocable space and enhancing capacity protection for suppliers. This dynamic adjustment mechanism ensures that the system adopts a more conservative allocation strategy under high demand pressure, preventing capacity collapse due to over-allocation and improving the system's robustness and reliability under extreme conditions.

[0038] In this embodiment of the invention, the detailed implementation steps for training and generating a constraint-aware order allocation probability model by embedding the gradient features of the piecewise softening mapping function into the loss function of the deep learning model include: A deep neural network consisting of an encoding layer, a constraint perception layer, and an output layer is constructed. The encoding layer extracts high-dimensional representations of order and supplier features. The network architecture adopts a multilayer perceptron structure, with the encoding layer typically containing 2-3 hidden layers, each using non-linear activation functions such as ReLU to enhance expressive power. Order feature encoding includes order attributes (such as quantity, amount, and deadline) and computational features (such as urgency and resource intensity), while supplier feature encoding includes static attributes (such as capacity limit and professional field) and dynamic states (such as current load and historical performance). The feature encoding process uses embedding techniques to process categorical features and standardizes numerical features to ensure scale consistency across different dimensions. The output of the encoding layer is a high-dimensional vector representation of orders and suppliers, capturing the complex relationships between multidimensional features and providing rich semantic information for subsequent matching decisions.

[0039] In the constraint-aware layer, the allocation state of the current iteration is input into a piecewise softening mapping function to calculate the constraint penalty value corresponding to the capacity utilization of each supplier. The constraint-aware layer is the core innovative part of the network, achieving seamless integration of capacity constraints and the deep learning model. The calculation process first summarizes the cumulative order quantity allocated to each supplier in the current iteration, and then converts it into the corresponding penalty value through the piecewise softening mapping function, reflecting the degree of constraint violation by the current allocation state. The calculation of the penalty value maintains the continuity of the gradient, ensuring that the network can learn the constraint boundary through backpropagation, achieving soft enforcement of constraints. The introduction of the constraint-aware layer enables the model to adaptively adjust the allocation strategy during the optimization process, balancing the relationship between maximizing benefits and satisfying constraints, thus improving the practicality and feasibility of the solution.

[0040] The constraint penalty value is gradientd with respect to the allocation amount, generating a constraint gradient tensor. This tensor characterizes the sensitivity of the allocation amount change to the degree of constraint violation. The constraint gradient is a key signal guiding the model to learn constraint boundaries, intuitively reflecting the degree of influence of the allocation decision on the constraint state. The calculation process employs automatic differentiation, tracing the derivative of the penalty value with respect to the allocation amount through a computational graph to form the gradient tensor. For the piecewise softening mapping function, the gradient characteristics differ in different regions: the safe region exhibits a constant low gradient, the transition region exhibits a smoothly increasing medium gradient, and the forbidden region exhibits a steep high gradient. This gradient characteristic enables the model to accurately perceive constraint boundaries and adjust the allocation strategy in a timely manner when approaching the upper limit, avoiding severe constraint violations. The generation of the constraint gradient tensor provides important information on the constraint dimension for feature fusion and is the core mechanism for achieving constraint-aware learning.

[0041] The constraint gradient tensor and the feature tensor output by the encoding layer are fused element-wise with weights to generate constraint-aware features. The weighting coefficients are dynamically adjusted according to the constraint space region currently occupied by each supplier. Feature fusion is a key step in integrating constraint information and semantic features, achieving adaptive adjustment of constraint strength through a weighting mechanism. The fusion formula is: ; in, The characteristics after fusion These are the original encoded features. To constrain the gradient, This is a weighting factor, the value of which depends on the supplier. The constraint region is dynamically determined: smaller values ​​are used in the safe zone, medium values ​​in the transition zone, and larger values ​​in the prohibited zone. This dynamic weighting mechanism ensures the appropriate integration of constraint information, neither excessively interfering with normal allocation nor failing to provide strong adjustment signals when approaching constraint boundaries, thus achieving refined control of constraint perception. The fused features retain the semantic information of the original features while increasing the sensitivity of the constraint dimension, providing a comprehensive decision-making basis for subsequent allocation decisions.

[0042] Based on constraint-aware features, the output layer outputs the initial allocation probability matrix of each order to each supplier through the Softmax function. The output layer maps high-dimensional features to allocation probabilities, realizing the transformation from feature representation to decision result. The calculation process first maps the constraint-aware features to raw scores through a fully connected layer, and then applies the Softmax function for normalization to ensure that the sum of the allocation probabilities of each order to all suppliers is 1. The formula for calculating the Softmax function is: ; in, For orders Assigned to supplier The probability, For the corresponding raw score, Iterate through all suppliers. The initial allocation probability matrix reflects the model's comprehensive evaluation of order-supplier matching, taking into account order characteristics, supplier characteristics, and constraints, thus providing a probabilistic decision-making basis for subsequent deterministic allocation.

[0043] A multi-objective loss function is constructed, which includes an allocation benefit term, a load balancing term, and a constraint violation penalty term. The constraint violation penalty term is obtained by weighted summation of the penalty values ​​of the cumulative allocation amounts from each supplier using a piecewise softening mapping function. The multi-objective loss function is the core objective guiding model optimization, comprehensively evaluating the multi-dimensional performance of the allocation scheme. The calculation formula is as follows: ; in, For the total loss, To assign probabilities, For orders Assigned to supplier The value and benefits For load balancing metrics, For piecewise softening mapping functions, For orders The amount of resources required, , , The weights are used to control the relative importance of each objective. The design of the multi-objective loss function enables the model to consider multiple optimization objectives simultaneously, finding the optimal balance between maximizing benefits, balancing load, and satisfying constraints, thus achieving comprehensive optimization of the allocation problem.

[0044] The gradient descent algorithm is used to iteratively optimize the parameters of the deep neural network. Network parameters are saved when the multi-objective loss function converges and the constraint violation rate on the validation set is below a preset threshold. Model training employs mini-batch stochastic gradient descent combined with adaptive optimizers such as Adam to accelerate the convergence process. Training data includes historical order allocation cases and simulated multi-scenario data, covering diverse situations with varying demand pressures and constraints. An early stopping mechanism is implemented during training to prevent overfitting; learning rate scheduling strategies, such as cosine annealing or stage decay, are used to optimize the convergence path and final performance. Validation and evaluation consider multiple metrics, including allocation efficiency, constraint violation rate, and load balancing, ensuring the model performs well in all aspects. The saved model parameters represent the optimal state of the system in terms of multi-objective balance and constraint satisfaction, providing a stable and reliable decision-making basis for practical applications.

[0045] The parameter-stabilized deep neural network serves as a constraint-aware order allocation probabilistic model. The stabilization process includes model quantization, inference optimization, and interface encapsulation to ensure efficient model operation in real-world deployment environments. Quantization techniques such as weight quantization and activation quantization reduce the model's memory footprint and computational burden; inference optimization techniques such as operator fusion and parallel computing improve execution efficiency; and interface encapsulation ensures seamless integration of the model with upstream and downstream systems. The stabilized model provides standardized input / output interfaces, accepting order features and supplier features as input and outputting a normalized allocation probability matrix, providing a reliable probabilistic foundation for subsequent deterministic allocation. Model stabilization enables a smooth transition from the training environment to the production environment, ensuring high performance and high reliability of the system in practical applications.

[0046] In this embodiment of the invention, the detailed implementation steps for performing a hierarchical constrained projection transformation on the initial allocation probability matrix include: Normalization is performed on each row of the initial allocation probability matrix to ensure that the sum of the allocation probabilities for each order across all suppliers is 1. Normalization is a fundamental step in standardizing allocation probabilities, ensuring the validity and consistency of the probability distribution. The process uses a simple scaling method, dividing the probability value of each row by the row sum, so that the total allocation probability for each order is exactly equal to 1. For numerical stability, a small constant (e.g., 1e-10) is added to prevent division by zero errors, while also handling potential outliers such as negative values. The normalized probability matrix maintains the relative relationships of the original model outputs while ensuring the rigorous mathematical properties of the probabilistic interpretation, providing a normalized probabilistic foundation for subsequent deterministic allocations.

[0047] The initial allocation probability matrix is ​​multiplied by the demand quantity of each order to obtain the initial allocation quantity matrix. This step transforms abstract probability values ​​into concrete resource allocation quantities, realizing a mapping from probability space to physical space. The calculation process directly applies element-wise multiplication; the row corresponding to each order is multiplied by its demand quantity to obtain a preliminary resource allocation scheme. This transformation maintains the relative relationship of the original probability distribution while introducing the actual impact of order size, making the allocation result more in line with actual business needs. The initial allocation quantity matrix reflects the model's original allocation intention, but may not yet meet all constraints, requiring further optimization and adjustment through subsequent projection transformations.

[0048] According to the order allocation priority sequence, constraint projection operations are performed on each order sequentially. Layered processing is a key mechanism to ensure that high-priority orders receive resources first, realizing priority differentiation among orders. The processing strictly follows the previously constructed priority sequence, starting with the highest priority order and executing projection operations one by one. This sequential processing ensures that the resource needs of high-value, high-urgency orders are met first, guaranteeing the normal execution of critical orders even under resource constraints. The design of layered constraint projection reflects the actual priority considerations in the business, enabling the system to make resource allocation decisions that conform to business logic under constraints.

[0049] In the constraint projection operation, for the current order, the cumulative capacity utilization rate of each supplier is calculated after allocating it to each supplier according to the initial allocation matrix. The cumulative capacity utilization rate is a key indicator for evaluating constraint satisfaction, directly reflecting the actual impact of allocation decisions on supplier capacity. The calculation process maintains a dynamically updated cumulative allocation vector, recording the total amount currently allocated to each supplier, and adding the allocation amount for new orders to the total utilization rate calculation. The utilization rate is obtained by dividing the cumulative amount by the supplier's capacity limit, reflecting the relative level of resource utilization. This calculation step enables the system to monitor the load status of each supplier in real time, providing accurate data for constraint checking and adjustment.

[0050] Suppliers whose cumulative capacity utilization exceeds the upper boundary of the capacity reserve buffer are identified and denoted as the over-limit supplier set. Constraint checking is the core step of projection transformation, ensuring that the allocation scheme does not violate the set safety boundaries. The checking process compares each supplier's cumulative utilization with its individual buffer upper boundary to identify all suppliers exceeding the safety range. The over-limit supplier set represents the target objects that need to be adjusted in allocation and is the starting point for subsequent resource reallocation. Constraint checking uses strict boundary comparisons to ensure that no supplier is assigned orders exceeding its safety capacity, maintaining the stability and reliability of system operation.

[0051] For each supplier in the over-limit supplier set, calculate its over-limit amount and deduct it from the current order quantity allocated to that supplier. Over-limit adjustment is a direct measure to resolve constraint violations, bringing suppliers back to a safe load range by reducing the allocated amount. The calculation process first determines the specific over-limit amount for each over-limit supplier, then deducts the corresponding amount from the currently processed orders, ensuring that the adjusted cumulative allocation amount strictly does not exceed the upper boundary of the buffer. When multiple orders have been allocated to the same over-limit supplier, the adjustment prioritizes the currently processed orders, minimizing impact on the allocation of already determined high-priority orders. This targeted adjustment mechanism ensures that constraints are satisfied while minimizing interference with the original allocation scheme, maintaining the overall optimization property of the allocation decision.

[0052] The deducted excess capacity is redistributed according to the remaining suppliers' capacity redundancy ratio. Capacity redundancy is the difference between the upper boundary of the capacity buffer and the current accumulated capacity utilization. Resource reallocation is a crucial step in balancing the use of remaining capacity, achieving rational reuse of resources through proportional allocation. The calculation process first identifies all non-excessive suppliers as receiving candidates, then calculates the capacity redundancy of each candidate supplier, and determines the reallocation weight based on the redundancy ratio. The reallocation formula is: ; in, To redistribute from excess quantities to suppliers The amount, For the total over-limit, For suppliers The upper boundary value of the buffer, Its current cumulative usage, The system iterates through all non-over-limit suppliers. This proportional allocation mechanism ensures full utilization of surplus capacity while avoiding new over-limit situations caused by redistribution, thus achieving optimal adjustment of resource allocation.

[0053] The initial allocation matrix is ​​updated to generate a deterministic allocation matrix that satisfies the capacity margin buffer constraint. The update process integrates the adjustment results from the preceding steps, forming the final deterministic allocation scheme. The update operation directly modifies the row corresponding to the current order in the initial allocation matrix, filling the adjusted allocation amount into the corresponding position to ensure the consistency and integrity of the matrix. The updated matrix strictly satisfies the capacity constraints of all suppliers, while preserving the allocation intent of the original model to the greatest extent possible within the constraints, achieving a balance between constraint satisfaction and optimization objectives. The deterministic allocation matrix is ​​the final output of the projection transformation, providing a foundational scheme for subsequent risk assessment and possible reallocation.

[0054] In this embodiment of the invention, the detailed implementation steps for detecting whether the deterministic allocation matrix touches the warning boundary of the capacity margin buffer include: The deterministic allocation matrix is ​​traversed to calculate the total allocation for each supplier. This total allocation calculation is a fundamental step in risk assessment, providing a holistic view of the supplier load status. The calculation process sums each column of the matrix to obtain the total number of orders allocated to each supplier. This calculation considers all allocated orders, including those of different priorities and sources, comprehensively reflecting the actual load situation of the suppliers. The total allocation data provides a quantitative basis for subsequent safety margin analysis and is a key input variable for assessing the risk level of the allocation scheme.

[0055] The total allocation for each supplier is compared to the upper boundary of the capacity margin buffer to calculate the capacity safety margin. The capacity safety margin is a key indicator for quantifying supplier load risk, directly reflecting the safe distance a supplier is from its capacity limit. The calculation formula is as follows: ; in, For suppliers The safety margin of production capacity The upper boundary value of the buffer. This represents the total allocation. The ratio form of the safety margin eliminates differences between suppliers of different sizes, providing a standardized risk metric that facilitates horizontal comparisons and threshold setting. A larger safety margin indicates lower risk and a more robust allocation scheme; conversely, a smaller safety margin indicates that the supplier is approaching its capacity limit, posing a potential execution risk.

[0056] A warning margin threshold is set. When the capacity safety margin falls below the warning margin threshold, the supplier is deemed to have reached the warning boundary. The warning threshold setting is a key parameter for risk early warning, determining the system's risk sensitivity and warning timing. The specific value of the threshold is set according to industry characteristics and risk appetite, typically between 5% and 15%. Higher thresholds are set for high-risk-averse businesses, while lower thresholds can be set for highly resilient businesses. The judgment process directly compares the safety margin with the warning threshold. When the margin falls below the threshold, the corresponding supplier is marked as reaching the warning state, requiring further monitoring and possible adjustments. The warning judgment mechanism enables the system to detect potential problems early before risks escalate, providing a foundation for proactive risk management.

[0057] The number of suppliers reaching the warning threshold is counted, and the proportion of the total order value undertaken by these suppliers to the total order value is calculated, denoted as risk exposure. Risk exposure is a comprehensive indicator for assessing the overall risk level, reflecting the proportion of value in a high-risk state within the allocation scheme. High risk exposure indicates that a large number of high-value orders are in a potentially risky state, requiring active intervention and adjustment by the system.

[0058] The system compares the risk exposure level with a preset risk threshold. When the risk exposure level exceeds the preset threshold, a rollback-based reallocation process is triggered. This risk threshold comparison is the mechanism for determining whether to initiate reallocation, ensuring the system takes proactive action when risk reaches a certain level. Threshold settings are typically based on historical data analysis and business risk tolerance, reflecting the system's risk tolerance. When the risk exposure level exceeds the threshold, it indicates that the current allocation scheme's risk level exceeds an acceptable range, requiring risk mitigation and optimization adjustments through rollback-based reallocation. This threshold-based triggering mechanism ensures that the system only initiates resource-intensive reallocation processes when necessary, balancing the needs of risk control and computational efficiency.

[0059] In this embodiment of the invention, the detailed implementation steps for performing rollback-based reallocation according to the order allocation priority sequence include: Based on the set of suppliers that have reached the alert boundary, all orders allocated to these suppliers are extracted to construct a set of orders to be adjusted. The construction of this set is the starting point for rollback and reallocation, defining the potential scope of adjustments. The construction process traverses the deterministic allocation matrix, extracting order information for all suppliers that have triggered the alert, including key attributes such as order ID, allocation amount, and priority score. To improve processing efficiency, an index structure is used to organize the order data, facilitating subsequent priority sorting and selection operations. The completeness of the set to be adjusted ensures that the system can comprehensively evaluate all potential adjustment options, providing ample candidate space for formulating the optimal reallocation strategy.

[0060] When adjusting orders in the order set, the system iterates through the order allocation priority sequence in reverse order, prioritizing the adjustment of lower-priority orders. This reverse iteration is a key strategy to protect high-value orders, ensuring that the system prioritizes sacrificing the allocation stability of lower-priority orders during necessary adjustments. The sorting process directly utilizes the previously constructed priority sequence, but in reverse order from low to high, ensuring that adjustments begin with the least important orders. This strategy aligns with the value protection principle in business operations, maximizing the execution stability of critical orders while providing sufficient adjustment flexibility for the system, achieving a balance between risk control and value protection.

[0061] For orders currently awaiting adjustment, the allocation amounts distributed to suppliers reaching the warning boundary in the deterministic allocation matrix are obtained. Allocation extraction is a necessary step in determining the basis for adjustment, clarifying the specific allocation of each order among high-risk suppliers. The extraction process directly reads the corresponding element values ​​from the deterministic allocation matrix, ensuring data accuracy and consistency. For orders allocated to multiple warning suppliers, the allocation amount for each supplier is extracted separately, providing a basis for subsequent differentiated adjustments. Allocation data is the direct object of adjustment decisions, and its accurate extraction ensures the precision and controllability of the reallocation process.

[0062] Calculate the urgency of a supplier's capacity overrun when it reaches the warning boundary. The urgency is the amount by which the supplier's total allocated capacity exceeds the lower boundary value of the capacity buffer zone. The urgency is a quantitative indicator assessing the urgency of adjustments, reflecting the supplier's load pressure level. The calculation formula is: ; in, For suppliers The excessive urgency, For the total allocation, This represents the lower boundary value of the buffer zone. The definition of over-limit urgency highlights the portion of the supplier's load that exceeds the safe capacity, providing a direct reference for determining the adjustment amount. A higher urgency indicates greater pressure on the supplier, requiring more load shifting to mitigate potential capacity risks.

[0063] Based on the urgency of capacity overruns, the order allocation amount that needs to be migrated from suppliers reaching the warning boundary is determined. Determining the migration amount is the core decision of reallocation, directly impacting the scope and effectiveness of the adjustment. The calculation process considers the current order allocation among alarm-triggered suppliers and the supplier's overrun urgency, taking the smaller of the two as the initial migration amount. For the same order allocated to multiple alarm-triggered suppliers, a proportional allocation strategy is adopted, determining the migration amount from each supplier based on their respective urgency ratios. This urgency-based migration amount determination method ensures the targeting and effectiveness of the adjustment, concentrating resources to address the most pressing capacity pressure issues.

[0064] Among the remaining suppliers, those with a capacity safety margin greater than a preset redundancy threshold are selected to construct a candidate supplier set for receiving orders. Constructing this candidate set is a crucial step in ensuring that adjustments do not introduce new risks. The selection process ensures that only suppliers with sufficient capacity can accept new allocations. The selection process compares the safety margin of each non-alarm supplier with a redundancy threshold (typically set at 15%-25%), and only suppliers exceeding the threshold are included in the candidate set. This conservative selection strategy ensures that adjustments do not lead to a simple transfer of risk, but rather achieve an effective reduction in the overall risk level. The redundancy threshold setting also considers the capacity requirements after adjustments, ensuring that the receiving party maintains a sufficient safety distance after accepting new orders, thus maintaining the overall stability of the system.

[0065] Based on the matching scores between each supplier in the receiving candidate supplier set and the current orders to be adjusted, the migrated order allocation is redistributed to the receiving candidate supplier set. The matching score is a key indicator guiding the redistribution decision, ensuring orders are assigned to the most suitable supplier. The score calculation comprehensively considers multiple factors, including supplier expertise, historical cooperation quality, geographical proximity, and current capacity margin, and derives a comprehensive score through a weighted average. The redistribution employs a proportional allocation strategy, with the formula as follows: ; in, For orders Assigned to supplier The amount of redistribution, For the total migration, To score the matching degree, It iterates through all candidate suppliers. This matching-based proportional allocation ensures the rationality and efficiency of reallocation, assigning orders to the most suitable suppliers with sufficient capacity, and maximizing the adjusted overall execution quality.

[0066] The deterministic allocation matrix is ​​updated, and the system re-checks whether the warning boundary has been reached until the risk exposure drops below the preset risk threshold. Iterative updates are a necessary mechanism to ensure sufficient risk mitigation, ultimately achieving the risk control objective through multiple rounds of adjustments. The update process directly modifies the corresponding elements in the deterministic allocation matrix, reducing the allocation of alarm-triggered suppliers and increasing the allocation of receiving suppliers, while keeping the total allocation constant. After each update, the system re-executes the risk detection process, calculates the adjusted risk exposure, and compares it with the preset threshold. If the risk still exceeds the limit, the system continues with the next round of rollback and redistribution until the risk drops to an acceptable level or cannot be further adjusted. This iterative optimization mechanism ensures that the system can fully utilize the adjustment space to achieve effective risk control, while avoiding excessive disruption of the original scheme by drastic changes through multiple rounds of fine-tuning, maintaining the overall optimal nature of the allocation scheme.

[0067] In this embodiment of the invention, the detailed implementation steps for outputting the final order allocation scheme after rollback-based redistribution verification include: Hard constraint verification is performed on the deterministic allocation matrix after backtracking redistribution, verifying one by one whether the total allocation for each supplier is strictly less than or equal to the hard capacity limit. Hard constraint verification is a crucial step in ensuring the absolute feasibility of the final solution, ensuring through rigorous numerical comparison that no supplier is allocated an order quantity exceeding its capacity. The verification process first calculates the final total allocation for each supplier and then compares it with its hard capacity limit to ensure strict compliance with constraints. For suppliers whose allocations are close to but do not exceed the limit, the system also calculates a safety margin ratio to assess the robustness of the solution. Hard constraint verification uses precise calculations rather than estimates or approximations to ensure the rigorous accuracy of the results, providing a solid guarantee for the executability of the final solution. This verification serves as the system's last line of defense, ensuring that even with minor deviations in all the aforementioned optimization steps, the final output solution strictly satisfies all hard constraints.

[0068] When the total allocation to any supplier exceeds the hard capacity limit, a forced reduction operation is performed on that supplier, reducing the allocation in reverse order of order priority until the constraint is met. Forced reduction is an emergency mechanism to handle extreme cases, ensuring the system outputs a feasible solution under any circumstances. The reduction process strictly follows the priority protection principle, adjusting orders one by one starting with the lowest priority to maximize the stability of critical orders. Specifically, the excess amount is first calculated, then the orders allocated to that supplier are iterated in reverse order of priority, and the allocation is reduced sequentially until the total allocation falls below the hard capacity limit. The reduction calculation formula is: ; in, To order The amount of the reduced allocation, For orders The amount currently allocated to suppliers exceeding the limit, This represents the current excess quantity from the supplier. This formula ensures that each reduction does not exceed the original allocated quantity of the order and does not exceed the total required reduction, achieving minimal necessary adjustments. While the forced reduction mechanism may result in some low-priority orders not being fully allocated, it ensures that the system can maintain the normal operation of core business even in extreme circumstances, demonstrating the system's high level of guarantee for business continuity.

[0069] The reduced allocation is marked as unallocated demand and recorded in a queue awaiting secondary allocation. Handling unallocated demand is crucial for system integrity, ensuring all order demands are properly tracked and processed. The recording process creates a detailed status record for each reduced order component, including key information such as order ID, original supplier, reduction amount, and reason for reduction, and adds it to a dedicated queue awaiting secondary allocation. This queue employs a priority queue structure, ensuring that high-priority unallocated demands are given priority consideration in subsequent processing. Explicit marking and recording of unallocated demand allows the system to clearly identify and quantify allocation gaps, providing accurate data for subsequent capacity planning, supplier development, or customer communication. This transparent approach avoids the "hidden loss" of demand, enhancing system traceability and integrity.

[0070] The deterministic allocation matrix, validated by hard constraints, is transformed into an order-supplier pairing instruction set, generating and outputting the final order allocation scheme. This instruction set conversion is a crucial step in transforming the abstract allocation matrix into executable business instructions, ensuring seamless integration of the theoretical solution with the actual execution system. The conversion process first iterates through the non-zero elements of the deterministic allocation matrix, extracting valid order-supplier pairing relationships and corresponding allocation quantities. Then, this raw data is converted into a standardized instruction format, including unique identifiers, execution time windows, allocation quantities, special processing requirements, and other complete information needed for operation. Finally, the instruction set is output to the execution system via a system interface. The instruction format design considers the compatibility requirements of different supply chain systems, supporting multiple standard protocols such as EDI, XML, or JSON, ensuring that the instructions can be correctly interpreted and executed by downstream systems. The final output allocation scheme not only satisfies all hard constraints but also achieves multi-objective optimization within the feasible space, providing scientific, efficient, and executable decision support for supply chain management.

[0071] In this embodiment of the invention, the detailed implementation steps of the piecewise softening mapping function outputting a smooth transition penalty value based on the Sigmoid function within the transition region include: The transition zone is defined by setting the starting boundary as the first boundary point and the ending boundary as the second boundary point, which corresponds to the hard capacity ceiling. Setting these boundary points is a fundamental step in constructing the transition zone mapping function, clarifying the function's domain. The first boundary point corresponds to the upper bound of the safe zone (i.e., the aforementioned first percentage threshold), typically set to 70%-85% of the hard capacity ceiling. The second boundary point directly corresponds to the hard capacity ceiling, marking the absolute upper limit of acceptable capacity utilization. These two boundary points constitute the range of the transition zone and are key reference points for setting the Sigmoid function parameters. The precise definition of these boundary points ensures seamless integration between the transition zone mapping function and adjacent region functions, maintaining the continuity and smoothness of the overall mapping function and providing a reliable foundation for subsequent gradient calculations and optimization processes.

[0072] Construct a Sigmoid function, whose input variable is the current capacity utilization. The Sigmoid function is ideal for achieving a smooth transition; its S-shaped curve provides a natural transition from low penalty to high penalty. Its basic form is: ; in, The value of the Sigmoid function. This refers to capacity utilization. The center point is located at the center point. The scaling parameter controls the steepness of the curve. The Sigmoid function has a range of (0,1), which can be mapped to the desired penalty value range through a linear transformation. The continuous differentiability of the function ensures the effective transmission of gradient information during the optimization process, enabling the model to accurately perceive the constraint boundaries and achieve smooth execution of soft constraints.

[0073] The center point of the Sigmoid function is set at the midpoint between the first and second boundary points. Setting the center point is a key parameter for adjusting the shape of the Sigmoid function, directly affecting the intermediate position of the transition. The calculation formula is: ; in, The center point is located at the center point. This is the first boundary point (upper boundary of the safe zone). This is the second boundary point (the hard capacity ceiling). Setting the center point midway between the two boundaries ensures the Sigmoid function reaches its maximum rate of change in the middle of the transition region, achieving a symmetrical transition of the penalty value from low to high. This setting ensures the system provides appropriate warnings and suppression for capacity utilization approaching the ceiling, while avoiding prematurely imposing excessively high penalties, thus maintaining a reasonable utilization of the optimization space.

[0074] The slope parameter of the Sigmoid function is adjusted to smoothly connect the output value at the first boundary point with the maximum penalty value of the safe zone, and the output value at the second boundary point with the minimum penalty value of the forbidden zone. Slope adjustment is a crucial step in ensuring function continuity; seamless fusion of the three function segments is achieved through precise control of the transition rate. The adjustment process first determines the objective function values ​​at the two boundary points: the first boundary point corresponds to the maximum penalty value of the safe zone (i.e., λ·T1), and the second boundary point corresponds to the minimum penalty value of the forbidden zone (i.e., β). Then, the slope parameter σ that satisfies these two boundary conditions is solved using numerical optimization methods. The complete expression for the transition zone Sigmoid function is: ; This function ensures that... The output value is ,exist The output value is This achieves a smooth transition with adjacent regions. The smoothness of the function provides a continuous signal for gradient optimization, enabling the model to accurately perceive the soft boundaries of capacity constraints and achieving a balance between the soft execution of constraints and the pursuit of optimization objectives.

[0075] During model training, the distribution density of actual capacity utilization within the transition zone is monitored. Distribution monitoring is fundamental to adaptively adjusting function parameters, optimizing constraint enforcement strategies through a data-driven approach. The monitoring process collects capacity utilization statistics from the model's allocation results during training, calculates the frequency distribution of each sub-interval in the transition zone, and identifies high-frequency utilization areas and patterns. Distribution analysis employs kernel density estimation or histogram methods to visually display the concentration trend and dispersion of capacity utilization. This real-time monitoring mechanism enables the system to understand the model's actual constraint perception effect, providing data support for dynamic parameter adjustment and ensuring the matching and effectiveness of the constraint function with actual optimization behavior.

[0076] When a concentration of distribution density is detected in the upper half of the transition zone, the slope parameter is dynamically increased to accelerate the growth rate of the penalty value and enhance the inhibitory effect on approaching the hard capacity limit. Adaptive adjustment is a key mechanism for optimizing constraint enforcement, responding to changes in model behavior through dynamic parameter updates. The adjustment logic is based on the risk aversion principle. When it is found that the model tends to push a large number of allocations towards the capacity limit, the slope parameter is increased to make the Sigmoid function steeper, applying a stronger penalty signal in advance and guiding the model away from high-risk areas. The adjustment magnitude is proportional to the distribution concentration; the higher the concentration, the more significant the adjustment, ensuring that the system can respond promptly and correct potential risk accumulation. This adaptive mechanism greatly improves the intelligence and effectiveness of constraint enforcement, enabling the system to dynamically adjust its strategy according to actual operating conditions and maintain the optimal constraint-optimization balance.

[0077] The above describes the deep learning-based intelligent allocation method for multi-objective supplier orders in the embodiments of this application. The following describes the deep learning-based intelligent allocation system for multi-objective supplier orders in the embodiments of this application. Please refer to [link / reference]. Figure 2 One embodiment of the deep learning-based multi-objective supplier order intelligent allocation system in this application includes: The data acquisition module is used to acquire supplier capacity constraint data and order demand characteristic data; The constraint softening mapping module is used to construct a piecewise softening mapping function for the supplier's capacity boundary based on the capacity constraint data. The piecewise softening mapping function transforms the hard capacity limit into a three-segment constraint space that includes a safe zone, a transition zone, and a prohibited zone. The priority ranking module is used to calculate the urgency and resource consumption intensity of each order based on order demand characteristic data, and to construct an order allocation priority sequence. The buffer design module is used to establish a capacity margin buffer for each supplier in a three-segment constraint space. The boundary values ​​of the capacity margin buffer are determined by the statistical distribution of historical allocation deviations. The constraint-aware modeling module is used to embed the gradient features of the piecewise softening mapping function into the loss function of the deep learning model to train and generate a constraint-aware order allocation probability model; the initial allocation probability matrix of each order to each supplier is output through the order allocation probability model. The hierarchical projection transformation module is used to perform hierarchical constraint projection transformation on the initial allocation probability matrix. The hierarchical constraint projection transformation transforms the soft probability values ​​in the initial allocation probability matrix into a deterministic allocation matrix that satisfies the capacity margin buffer constraint. The risk detection module is used to detect whether the deterministic allocation matrix has reached the warning boundary of the capacity margin buffer. When the warning boundary is detected, the rollback adjustment module is used to perform rollback-style reallocation according to the order allocation priority sequence, and the final order allocation scheme after rollback-style reallocation verification is output. The modules are connected via wired and / or wireless means to enable data transmission between them.

[0078] This invention achieves intelligent allocation in multi-supplier, multi-order scenarios through capacity constraint softening mapping, order priority ranking, surplus buffer establishment, constraint-aware deep learning, and backtracking reallocation. The constraint-aware learning method of this invention can accurately balance maximizing efficiency with constraint satisfaction, effectively handle capacity fluctuations and demand changes, and provide a systematic solution for supply chain management.

[0079] The above are merely preferred embodiments of the present invention and are not intended to limit the present invention. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art can still modify the technical solutions described in the foregoing embodiments or make equivalent substitutions for some of the technical features. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

[0080] It should be noted that all formulas in this manual are calculated by removing dimensions and taking their numerical values. The formulas are derived from software simulations based on a large amount of collected data to obtain the most recent real-world results. The preset parameters and thresholds in the formulas are set by those skilled in the art according to the actual situation.

[0081] Although embodiments of the invention have been shown and described, those skilled in the art will understand that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the claims and their equivalents.

Claims

1. A multi-objective supplier order intelligent allocation method based on deep learning, characterized in that, include: Obtain supplier capacity constraints and order demand characteristics data; Based on the capacity constraint data, a segmented softening mapping function for the supplier capacity boundary is constructed. The segmented softening mapping function transforms the hard capacity limit into a three-segment constraint space that includes a safe zone, a transition zone, and a prohibited zone. Based on the order demand characteristic data, the urgency and resource consumption intensity of each order are calculated, and an order allocation priority sequence is constructed. Within the three-segment constraint space, a capacity margin buffer is established for each supplier, including: Statistical analysis of the deviation sequence between the actual capacity utilization rate and the predicted capacity utilization rate of each supplier in historical allocations; Quantile analysis was performed on the deviation sequence, and the deviation value corresponding to the 95th quantile was extracted and denoted as the one-sided safety margin. Based on the single-sided safety margin, the upper boundary value of the capacity margin buffer is determined by shifting downward from the rigid capacity limit value. Calculate the midpoint position of the transition zone and set the midpoint position as the lower boundary value of the capacity margin buffer. Within the capacity buffer zone, the contraction coefficient of the upper boundary value is dynamically adjusted based on the ratio of the total demand of the current orders to be allocated to the total capacity of the suppliers. When the ratio is greater than the preset tension threshold, the contraction coefficient is set to a value less than 1, causing the effective upper boundary of the capacity margin buffer to contract downward. The gradient features of the segmented softening mapping function are embedded into the loss function of the deep learning model to train and generate a constraint-aware order allocation probability model. The initial allocation probability matrix of each order to each supplier is output through the order allocation probability model. A hierarchical constraint projection transformation is performed on the initial allocation probability matrix, which transforms the soft probability values ​​in the initial allocation probability matrix into a deterministic allocation matrix that satisfies the capacity margin buffer constraint. Detect whether the deterministic allocation matrix touches the warning boundary of the capacity margin buffer; When the warning boundary is detected, a fallback reallocation is performed according to the order allocation priority sequence, and the final order allocation scheme verified by the fallback reallocation is output.

2. The intelligent allocation method for multi-objective supplier orders based on deep learning according to claim 1, characterized in that, The step of constructing a piecewise softening mapping function for the supplier's capacity boundary based on the capacity constraint data includes: Extract the hard capacity limit value for each supplier from the capacity constraint data; Based on the aforementioned hard capacity limit, a first percentage threshold is set downwards, and the interval where the capacity utilization rate is lower than the first percentage threshold is marked as a safe zone; A transition zone is set between the first percentage threshold and the hard capacity limit; A prohibited zone shall be established above the aforementioned mandatory production capacity limit. Construct the piecewise softening mapping function, which outputs a linear penalty value in the safe zone, a smooth transition penalty value based on the Sigmoid function in the transition zone, and an exponential penalty value in the forbidden zone.

3. The intelligent allocation method for multi-objective supplier orders based on deep learning according to claim 1, characterized in that, The step of calculating the urgency and resource consumption intensity of each order based on the order demand characteristic data, and constructing an order allocation priority sequence, includes: Extract order delivery deadline, order amount, order quantity, and order complexity from the order demand characteristic data; A time urgency coefficient is calculated based on the difference between the order delivery deadline and the current time. This time urgency coefficient is mapped to the remaining time using an exponential decay function. The product of the order amount and the time urgency coefficient is recorded as the demand urgency. Based on the order quantity and the order complexity, the expected consumption of supplier capacity by a single order is calculated and denoted as the resource consumption intensity. Calculate the ratio of the urgency of the demand to the intensity of resource consumption, and record it as the order allocation priority score; The order allocation priority scores of all orders are sorted in descending order to generate the order allocation priority sequence. The orders ranked higher in the order allocation priority sequence are given priority in capacity allocation during the hierarchical constraint projection transformation.

4. The intelligent allocation method for multi-objective supplier orders based on deep learning according to claim 1, characterized in that, The step of embedding the gradient features of the piecewise softening mapping function into the loss function of a deep learning model to train and generate a constraint-aware order allocation probability model includes: A deep neural network consisting of an encoding layer, a constraint perception layer, and an output layer is constructed, wherein the encoding layer extracts high-dimensional representations of order features and supplier features; In the constraint-aware layer, the allocation state of the current iteration is input into the segmented softening mapping function to calculate the constraint penalty value corresponding to the capacity occupancy of each supplier; The gradient of the constraint penalty value with respect to the allocation amount is calculated to generate a constraint gradient tensor, which characterizes the sensitivity of the allocation amount change to the degree of constraint violation. The constraint gradient tensor and the feature tensor output by the coding layer are fused element-wise with weights to generate constraint-aware features. The weighting coefficients are dynamically adjusted according to the constraint space region where each supplier is currently located. Based on the constraint-aware features, the initial allocation probability matrix of each order to each supplier is output in the output layer through the Softmax function; A multi-objective loss function is constructed, which includes an allocation benefit term, a load balancing term, and a constraint violation penalty term. The constraint violation penalty term is obtained by weighted summation of the penalty values ​​of the cumulative allocation amount of each supplier through the piecewise softening mapping function. The parameters of the deep neural network are iteratively optimized using the gradient descent algorithm. When the multi-objective loss function converges and the constraint violation rate on the validation set is lower than a preset threshold, the network parameters are saved. The deep neural network with its parameters fixed is used as the constraint-aware order allocation probability model.

5. The intelligent allocation method for multi-objective supplier orders based on deep learning according to claim 1, characterized in that, The step of performing a hierarchical constrained projection transformation on the initial allocation probability matrix includes: Normalization is performed on each row of the initial allocation probability matrix; Multiply the initial allocation probability matrix by the demand quantity of each order to obtain the initial allocation quantity matrix; According to the order allocation priority sequence, the constraint projection operation is performed on each order in sequence. In the constrained projection operation, for the current order, the cumulative capacity utilization rate of each supplier is calculated after it is allocated to each supplier according to the initial allocation matrix; Suppliers whose cumulative capacity utilization rate exceeds the upper boundary value of the capacity margin buffer are identified and denoted as the over-limit supplier set. For each supplier in the set of suppliers exceeding the limit, calculate its excess amount and deduct the excess amount from the current order quantity allocated to that supplier; The deducted excess quantity will be redistributed according to the remaining suppliers' capacity redundancy ratio. Update the initial allocation matrix to generate a deterministic allocation matrix that satisfies the capacity margin buffer constraint.

6. The intelligent allocation method for multi-objective supplier orders based on deep learning according to claim 1, characterized in that, The detection of whether the deterministic allocation matrix touches the warning boundary of the capacity margin buffer includes: Traverse the deterministic allocation matrix and calculate the total allocation for each supplier; The total allocation for each supplier is compared with the upper boundary value of the capacity margin buffer to calculate the capacity safety margin. A warning margin threshold is set. When the capacity safety margin is less than the warning margin threshold, it is determined that the supplier has reached the warning boundary. The number of suppliers that have reached the warning threshold is counted, and the proportion of the total value of orders undertaken by these suppliers to the total value of all orders is calculated, which is recorded as the risk exposure. The risk exposure is compared with a preset risk threshold. When the risk exposure is greater than the preset risk threshold, the rollback redistribution process is triggered.

7. The intelligent allocation method for multi-objective supplier orders based on deep learning according to claim 6, characterized in that, The step of performing a rollback-based reallocation according to the order allocation priority sequence includes: Based on the set of suppliers that have reached the warning boundary, extract all orders allocated to these suppliers and construct a set of orders to be adjusted. For the orders in the set of orders to be adjusted, traverse them in reverse order of the order allocation priority sequence, and adjust the low priority orders first; For the current order to be adjusted, obtain the allocation amount that is assigned to the supplier that has reached the warning boundary in the deterministic allocation matrix; Calculate the urgency of the overcapacity of the suppliers that have reached the warning boundary; Based on the urgency of the overcapacity, determine the order allocation volume that needs to be moved out of the suppliers that have reached the warning boundary; Among the remaining suppliers, those with a capacity safety margin greater than a preset redundancy threshold are selected to construct a set of candidate suppliers for acceptance. Based on the matching score between each supplier in the receiving candidate supplier set and the current order to be adjusted, the migrated order allocation is redistributed to the receiving candidate supplier set; Update the deterministic allocation matrix and re-detect whether the warning boundary has been reached until the risk exposure drops below the preset risk threshold.

8. The intelligent allocation method for multi-objective supplier orders based on deep learning according to claim 2, characterized in that, The piecewise softening mapping function outputs a smooth transition penalty value based on the Sigmoid function within the transition region, including: The starting boundary of the transition zone is set as the first boundary point, and the ending boundary is set as the second boundary point, with the second boundary point corresponding to the hard capacity limit value. Construct a Sigmoid function, the input variable of which is the current capacity utilization; The center point of the Sigmoid function is set at the midpoint between the first boundary point and the second boundary point; Adjust the slope parameter of the Sigmoid function so that the output value of the Sigmoid function at the first boundary point is smoothly connected with the maximum penalty value of the safe zone, and the output value at the second boundary point is smoothly connected with the minimum penalty value of the forbidden zone; During model training, the distribution density of actual capacity utilization within the transition zone is monitored; When the distribution density is detected to be concentrated in the upper half of the transition zone, the slope parameter is dynamically increased to accelerate the growth rate of the penalty value and enhance the suppression effect on approaching the hard capacity limit value.

9. A deep learning-based intelligent multi-objective supplier order allocation system, used to implement the deep learning-based intelligent multi-objective supplier order allocation method according to any one of claims 1 to 8, characterized in that, include: The data acquisition module is used to acquire supplier capacity constraint data and order demand characteristic data; The constraint softening mapping module is used to construct a piecewise softening mapping function for the supplier's capacity boundary based on the capacity constraint data. The piecewise softening mapping function transforms the hard capacity limit into a three-segment constraint space that includes a safety zone, a transition zone, and a prohibited zone. The priority sorting module is used to calculate the urgency and resource consumption intensity of each order based on the order demand characteristic data, and to construct an order allocation priority sequence. The buffer design module is used to establish a capacity margin buffer for each supplier in the three-segment constraint space. The constraint-aware modeling module is used to embed the gradient features of the piecewise softening mapping function into the loss function of the deep learning model, train and generate a constraint-aware order allocation probability model; and output the initial allocation probability matrix of each order to each supplier through the order allocation probability model. The hierarchical projection transformation module is used to perform hierarchical constraint projection transformation on the initial allocation probability matrix. The hierarchical constraint projection transformation transforms the soft probability values ​​in the initial allocation probability matrix into a deterministic allocation matrix that satisfies the capacity margin buffer constraint. The risk detection module is used to detect whether the deterministic allocation matrix touches the warning boundary of the capacity margin buffer. When the warning boundary is detected, the rollback adjustment module is used to perform rollback-style reallocation according to the order allocation priority sequence, and the final order allocation scheme after verification by the rollback-style reallocation is output.

Images