Makeup supply chain intelligent bid evaluation optimization method based on deep learning

Abstract

The invention discloses an intelligent bid evaluation optimization method for a beauty supply chain based on deep learning. The method comprises the following steps: S1, constructing a historical performance behavior data set of beauty suppliers; s2, utilizing a NeRF neural radiation field technology to map historical reputation data and performance capabilities of the beauty makeup suppliers to a high-dimensional feature space, and constructing continuous reputation distribution of the beauty makeup suppliers; s3, performing interpolation prediction on the continuous reputation distribution of the beauty makeup suppliers by adopting a volume rendering technology, and dynamically generating future reputation trend characteristics of the beauty makeup suppliers; s4, modeling a supplier bidding process into a path search problem by using a GlowNet generation flow network; s5, constructing a comprehensive reward mechanism based on the quotation, performance ability and profit expectation of the beauty makeup suppliers; and S6, in combination with an improved GFlowNet probability flow matching mechanism, dynamic optimization is carried out among different quotation strategies, and a global optimal quotation path is generated. According to the method, NeRF high-dimensional reputation modeling and GlowNet bidding optimization technologies are fully utilized, and the bid evaluation process of the beauty supply chain is intelligently optimized.

Description

Technical Field

[0001] The present invention relates to the technical fields of deep learning and supply chain management, and particularly to an intelligent bid evaluation optimization method for the beauty supply chain based on deep learning. Background Art

[0002] In modern beauty supply chain management, bid evaluation optimization is a key link in the procurement decision-making process, involving the intelligent evaluation and optimization of suppliers' credibility, performance capabilities, pricing strategies, and the bidding process. Traditional bid evaluation methods mainly rely on manual experience, static scoring systems, or rule-based decision models. These methods have many limitations when facing complex and changing market environments and are difficult to meet the needs of intelligent supply chain management. In recent years, the development of artificial intelligence technology has provided new solutions for supply chain bid evaluation optimization. Among them, the application of deep learning technology in supplier credibility modeling, bidding path optimization, and pricing strategy optimization has gradually attracted attention. However, existing deep learning-based methods still face many challenges when dealing with supply chain bid evaluation problems, especially in aspects such as dynamic modeling of supplier credibility, global optimization of bidding strategies, matching of pricing and performance capabilities, and the overall stability of the supply chain.

[0003] In terms of supplier credibility modeling, existing methods mainly evaluate suppliers' performance capabilities based on historical scores and statistical analysis. Such methods usually use simple weighted averages or time-decay-based scoring models, which cannot accurately capture the time-varying dynamics of supplier credibility and cannot predict future performance capabilities. For example, some traditional models use long short-term memory networks (LSTMs) or graph neural networks (GNNs) to simulate supplier credibility changes. However, due to the discrete nature of these models, it is difficult for them to achieve continuous modeling of supplier credibility, resulting in unstable evaluation results in the time dimension. In addition, traditional scoring mechanisms often rely on past performance records and ignore changes in suppliers' market competitiveness at different times, making it impossible to effectively predict future credibility trends and thus affecting the accuracy of procurement decisions.

[0004] In terms of optimizing the bidding strategy, traditional reinforcement learning methods, such as Deep Q-Network (DQN) or Proximal Policy Optimization (PPO), have local optimum problems when optimizing the bidding quotation path of suppliers. These methods often converge to certain specific quotation strategies during the training process and cannot comprehensively explore all possible bidding paths, resulting in a lack of flexibility in the bidding strategies of suppliers. In addition, traditional reinforcement learning methods rely on single-step decision optimization, while the bidding behavior of suppliers is actually a multi-step decision-making process involving complex factors such as quotation adjustment, market response, and competitor strategies. Single-step optimization is difficult to capture the best quotation path of suppliers throughout the bidding process. Therefore, in the optimization of bidding strategies, how to effectively explore the globally optimal quotation path and ensure that suppliers have a competitive advantage in different market environments remains an urgent problem to be solved.

[0005] In terms of the matching between quotation and fulfillment ability, existing methods usually adopt procurement strategies based on fixed rules, such as the lowest quotation first, weight allocation of historical fulfillment records, etc. These methods lack dynamic adjustment of the fulfillment ability of suppliers and may affect the stability of the supply chain due to winning the bid only because of the low price in the case of insufficient fulfillment ability of suppliers. For example, in some cases, suppliers with lower quotations may be unable to deliver on time due to weaker fulfillment ability, resulting in inventory shortages or supply chain breaks. On the other hand, if only evaluating based on the historical fulfillment records of suppliers and ignoring their current market competitiveness and price strategies, it may lead to an increase in procurement costs.

[0006] Therefore, how to provide an intelligent bid evaluation optimization method for the beauty supply chain based on deep learning is an urgent problem to be solved by those skilled in the art. Summary of the Invention

[0007] An object of the present invention is to propose an intelligent bid evaluation optimization method for the beauty supply chain based on deep learning. The present invention makes full use of NeRF high-dimensional reputation modeling and GFlowNet bidding optimization technology to construct a continuous feature distribution of supplier reputation, and optimizes the bidding path through an improved GFlowNet probability flow matching mechanism to ensure the global optimality of the quotation decision. This method can not only accurately predict the future fulfillment ability of suppliers but also dynamically adjust the quotation strategy, making the bidding process more intelligent and flexible. In addition, the present invention combines a comprehensive reward mechanism to achieve the accurate matching of supply chain procurement and fulfillment ability, improving the overall stability of the supply chain and the scientific nature of procurement decisions, and having the advantages of accurate prediction, optimal bidding, and strong supply chain stability.

[0008] The intelligent bid evaluation optimization method for the beauty supply chain based on deep learning according to the embodiments of the present invention includes the following steps:

[0009] S1. Use the sensor network to collect the historical credit data, performance ability data, and bidding history data of beauty suppliers in real time, and construct a dataset of the historical performance behaviors of beauty suppliers;

[0010] S2. Based on the dataset of the historical performance behaviors of beauty suppliers, use the NeRF (Neural Radiance Field) technology to map the historical credit data and performance ability of beauty suppliers to a high-dimensional feature space, and construct a continuous credit distribution of beauty suppliers;

[0011] S3. Adopt volume rendering technology to interpolate and predict the continuous credit distribution of beauty suppliers, and generate future credit trend features of beauty suppliers;

[0012] S4. Based on the future credit trend features and bidding history data of beauty suppliers, use the GFlowNet (Generative Flow Networks) to model the supplier bidding process as a path search problem, where each supplier's quotation decision corresponds to a state node in the path;

[0013] S5. Construct a comprehensive reward mechanism based on the quotation, performance ability, and profit expectation of beauty suppliers;

[0014] S6. Based on the comprehensive reward mechanism, combined with the improved GFlowNet probability flow matching mechanism, dynamically optimize among different quotation strategies, and finally determine the globally optimal quotation path in the bidding process.

[0015] Optionally, the S2 specifically includes:

[0016] S21. Preprocess the dataset of the historical performance behaviors of beauty suppliers, remove the missing values and outliers in the data, and normalize the historical credit data and performance ability data of suppliers. Use the mean normalization method to make the data distributed within the same numerical range;

[0017] S22. Construct a NeRF neural radiance field model, use the preprocessed historical credit data and performance ability data of beauty suppliers as input features, and set the credit data feature vector of the i-th supplier at time t as F i (t):

[0018] F i (t) = [R i (t), C i (t), B i (t)];

[0019] Among them, R i (t) represents the credit score of the i-th supplier at time t, C i (t) represents the performance ability data of the i-th supplier at time t, B i(t) represents the bidding historical data of the i-th supplier at time t, and is mapped to a high-dimensional space through a multi-layer perceptron to form a high-dimensional feature field of the supplier's reputation;

[0020] S23. Perform positional encoding on the reputation data feature vector F i (t) to enhance the expression ability of the supplier's reputation data in the high-dimensional feature field by using extended Fourier positional encoding:

[0021]

[0022] where γ(F i ) represents the reputation data feature vector after positional encoding, L represents the encoding dimension, and λ l represents the frequency parameter of the first scale;

[0023] S24. Build a NeRF implicit mapping model based on the reputation data feature vector after positional encoding, and define the mapping function as:

[0024] H i (t) = σ(W2·ReLU(W1·γ(F i )) + b1)b2);

[0025] where H i (t) represents the high-dimensional reputation data feature vector of the i-th supplier at time t, σ represents the activation function, ReLU represents the rectified linear unit function, W1 and W2 represent weight matrices, and b1 and b2 represent bias terms;

[0026] S25. Use gradient descent to optimize the parameters of the NeRF implicit mapping model, and the optimization objective function is:

[0027]

[0028] where N represents the total number of suppliers, T represents the time step, α i,t represents the error weighting coefficient of the i-th supplier at time t, λ1 represents the weight regularization coefficient, λ2 represents the time smoothing regularization coefficient, ∥·∥ represents the Euclidean norm, and ∥·∥ F represents the Frobenius norm, and H i (t - 1) represents the high-dimensional reputation data feature vector of the i-th supplier at time t - 1;

[0029] S26. Based on the optimized NeRF implicit mapping model, use the high-dimensional reputation data feature vector in the high-dimensional feature field to generate the continuous reputation distribution of the beauty suppliers.

[0030] Optionally, the specific content of S3 includes:

[0031] S31. Define a reputation density function to characterize the probability distribution of suppliers' reputations:

[0032] σ(R i ) = exp(-β1|R i -μ R |);

[0033] Where σ(R i ) represents the reputation density function, exp represents the exponential function, R i represents the reputation score of the i-th supplier at the current moment, μ R represents the historical reputation mean of the supplier, and β1 represents the decay coefficient of the reputation distribution, such that the reputation density varies with the historical reputation deviation;

[0034] S32. Define a reputation color mapping function to establish a non-linear mapping relationship of reputation in a high-dimensional space:

[0035]

[0036] Where c(R i ) represents the reputation color mapping function, β2 represents the steepness parameter of the reputation mapping curve, and R th represents the threshold of the reputation score;

[0037] Map the supplier reputation score to a normalized color space through the reputation color mapping function, so that the reputation interpolation prediction remains continuous and stable in the high-dimensional space;

[0038] S33. Based on the reputation density function σ(R i ) and the reputation color mapping function c(R i ), use the NeRF volume rendering method to calculate the reputation prediction value and generate the future reputation trend characteristics of beauty suppliers:

[0039]

[0040] Where represents the reputation prediction value of the i-th supplier at time t, τ max represents the end time of the time window for reputation calculation, τ min represents the start time of the time window for reputation calculation, τ represents the time parameter of the integration variable, σ(s) represents the reputation density of the supplier at time s, represents the reputation cumulative decay factor, ensuring the time continuity of the reputation interpolation prediction.

[0041] Optionally, the specific content of S4 includes:

[0042] S41. Based on the future credit trend characteristics and bidding historical data of beauty suppliers, construct the state space of the supplier's bidding decision-making. The state of each node represents the bidding decision of the supplier at the corresponding moment.

[0043] S42. Based on the transition probability of the supplier's bidding decision among state nodes, use the GFlowNet generation flow network to represent the supplier's bidding process as a probability flow matching problem of multi-state decision paths. The size of the probability flow of each bidding decision node depends on the transition probability among states, and the bidding path forms a directed probability flow graph composed of bidding state nodes.

[0044] Optionally, the specific content of S5 includes:

[0045] S51. Based on the bidding data, performance ability data and profit expectation data of beauty suppliers, construct a comprehensive bidding reward mechanism, and the comprehensive reward mechanism includes a bidding competitiveness index, a performance ability index and a profit expectation index.

[0046] S52. Define the bidding competitiveness index of the supplier:

[0047]

[0048] Among them, represents the bidding competitiveness index of the i-th supplier at time t, P avg (t) represents the average bid of all suppliers at time t, and P i (t) represents the real-time bid of the i-th supplier at time t.

[0049] S53. Define the performance ability index of the supplier:

[0050]

[0051] Among them, C i (t) represents the performance ability index of the i-th supplier at time t, represents the credit prediction value of the i-th supplier at time t, exp represents the exponential function, and γ c represents the sensitivity coefficient of the performance ability mapping, and G i (t) represents the credit trend gradient of the i-th supplier at time t, and G th represents the credit gradient threshold.

[0052] S54. Define the profit expectation index of the supplier:

[0053] U i (t) = Q i (t)[V market (t) - P i (t)] - L i(t);

[0054] Among them, U i (t) represents the profit expectation index of the i-th supplier at time t, Q i (t) represents the expected sales volume of beauty products provided by the i-th supplier at time t, V market (t) represents the market average price of beauty products at time t, L i (t) represents the unit fulfillment cost of the i-th supplier at time t;

[0055] S55. Build a comprehensive reward mechanism and set the comprehensive reward of the supplier in the bidding process as the weighted sum of the bid competitiveness index, fulfillment ability index, and profit expectation index.

[0056] Optionally, the S6 specifically includes:

[0057] S61. Based on the comprehensive reward mechanism, establish an optimization space for the supplier's bidding price strategy, set the conversion relationship between different bidding strategies, and assign an initial reward value to each bidding strategy, so that the optimization of the bidding strategy is dynamically adjusted considering bid competitiveness, fulfillment ability, and profit expectation;

[0058] S62. Build an improved GFlowNet probability flow matching mechanism, and set the probability flow of each bidding price state node to be matched through the probability flow balance constraint between upstream and downstream nodes to ensure that the inflow and outflow of any node are balanced;

[0059] S63. The improved GFlowNet probability flow matching mechanism introduces a dynamic flow adjustment factor, and sets that when the state of each bidding path is transferred each time, the probability flow distribution of the state transfer is adaptively adjusted according to the change of the comprehensive reward of the current node:

[0060] F(s→s') = F0(s→s')·exp(δ·ΔR bid (s,s'));

[0061] Among them, F(s→s') represents the adjusted probability flow from state node s to state node s', F0(s→s') represents the initial probability flow from state node s to state node s', exp represents the exponential function, δ represents the dynamic flow adjustment factor, and ΔR bid (s,s') represents the difference in comprehensive reward between state node s and state node s';

[0062] S64. Use the improved GFlowNet probability flow matching mechanism for multi-round iterative optimization, and update the probability flow distribution between the state nodes of the bidding path in each round of optimization, so that the search for the bidding path gradually converges to the area with the highest comprehensive reward;

[0063] S65. Calculate the cumulative probability flow values of each state node in the bidding path, and identify the cumulative probability flow values of the bidding paths corresponding to different bidding strategies;

[0064] S66. Sort all the bidding paths of the supplier according to the magnitude of the cumulative probability flow values, and finally determine the globally optimal bidding path during the bidding process.

[0065] The beneficial effects of the present invention are as follows:

[0066] First of all, by introducing the NeRF (Neural Radiance Field) technology, the present invention performs high-dimensional feature modeling on the historical credit data and performance capabilities of suppliers, and uses the volume rendering method to achieve continuous interpolation prediction of supplier credit in the time dimension, thus overcoming the limitations of traditional discrete scoring methods. The NeRF model can effectively capture the dynamic change trend of supplier credit, not only provide more stable credit assessment results, but also predict the performance capabilities of suppliers at future moments, enabling the purchaser to identify potential performance risks earlier during the bid evaluation process and optimize the supply chain cooperation strategy.

[0067] Secondly, the present invention uses the GFlowNet (Generative Flow Network) to model the supplier bidding process, mapping the bidding price decision into a path search problem, thus breaking through the local optimality problem of traditional reinforcement learning methods in bidding strategy optimization. By constructing a probability flow matching mechanism based on GFlowNet, each bidding decision of the supplier corresponds to a state node in the bidding path, making the bidding strategy optimization a global path search process rather than a single-step decision optimization. The improved GFlowNet probability flow matching mechanism introduces a dynamic flow adjustment factor during the bidding process, enabling the search direction of the bidding path to be adaptively adjusted according to the bidding competitiveness, credit trend, and historical bidding performance of the supplier, thereby ensuring that the final bidding strategy takes into account both competitiveness and performance capabilities, and avoiding a long-term decline in competitiveness caused by locally optimal bids. This globally optimized method significantly improves the rationality of the supplier bidding strategy, enabling the purchaser to select better-quality suppliers during the bidding process and enhancing the overall stability of the supply chain.

[0068] In addition, in terms of matching the quotation with the performance ability, the present invention constructs a comprehensive reward mechanism, taking into account multiple factors such as the quotation competitiveness, performance ability, and profit expectation of the supplier, so that the optimization of the bidding not only focuses on the lowest quotation, but comprehensively considers the long-term cooperation value of the supplier. Traditional methods often struggle to balance the trade-off between low quotations and high performance capabilities during the bidding process. However, the present invention can reduce the performance risk and improve the long-term stability of the supply chain while ensuring the optimization of procurement costs by dynamically optimizing the quotation strategy. At the same time, the present invention further optimizes the execution process of the bidding strategy, enabling the quotation path of the supplier to be flexibly adjusted according to changes in the market environment, ensuring the best bidding state under different market conditions. In addition, through the time series prediction ability of the NeRF model, the present invention can also analyze the historical data to combine the supplier credit trend, thereby dynamically adjusting the procurement strategy to make the supply chain management more intelligent and precise.

[0069] The present invention not only optimizes the single bidding process but also provides a more intelligent control mechanism for the overall optimization of the supply chain. Through the improved GFlowNet model, the present invention can evaluate the long-term impact of different quotation paths from a global perspective, making the optimization of the bidding strategy not limited to the current procurement cycle but capable of affecting the stability of the long-term operation of the supply chain. At the same time, combined with the high-dimensional credit modeling ability of NeRF, the present invention can dynamically optimize the supply chain partnership, enabling the purchaser to form a stable and sustainable cooperation network during the supplier selection process, ultimately realizing the intelligence, precision, and efficiency of supply chain management. Brief Description of the Drawings

[0070] The drawings are used to provide a further understanding of the present invention and constitute a part of the specification. They are used together with the embodiments of the present invention to explain the present invention and do not constitute a limitation to the present invention. In the drawings:

[0071] Figure 1 It is the overall flowchart of the intelligent bid evaluation optimization method for the beauty supply chain based on deep learning proposed by the present invention. Detailed Embodiments

[0072] Now, the present invention will be further described in detail with reference to the drawings. These drawings are all simplified schematic diagrams, only showing the basic structure of the present invention in a schematic way, so they only show the components related to the present invention.

[0073] Refer to Figure 1 , the intelligent bid evaluation optimization method for the beauty supply chain based on deep learning includes the following steps:

[0074] S1. Use the sensor network to collect the historical credit data, performance ability data, and bidding historical data of beauty suppliers in real time, and construct a historical performance behavior dataset of beauty suppliers;

[0075] S2. Based on the historical performance behavior dataset of beauty suppliers, use the NeRF (Neural Radiance Field) technology to map the historical credit data and performance capabilities of beauty suppliers to a high-dimensional feature space, and construct a continuous credit distribution of beauty suppliers;

[0076] S3. Adopt volume rendering technology to interpolate and predict the continuous credit distribution of beauty suppliers, and generate future credit trend features of beauty suppliers;

[0077] S4. Based on the future credit trend features and bidding historical data of beauty suppliers, use the GFlowNet (Generative Flow Network) to model the supplier bidding process as a path search problem, where each supplier's price quotation decision corresponds to a state node in the path;

[0078] S5. Based on the price quotations, performance capabilities, and profit expectations of beauty suppliers, construct a comprehensive reward mechanism;

[0079] S6. Based on the comprehensive reward mechanism, combined with the improved GFlowNet probability flow matching mechanism, dynamically optimize among different price quotation strategies, and finally determine the globally optimal price quotation path in the bidding process.

[0080] In this embodiment, the S2 specifically includes:

[0081] S21. Preprocess the historical performance behavior dataset of beauty suppliers, remove the missing values and outliers in the data, and normalize the historical credit data and performance capability data of the suppliers. Use the mean normalization method to make the data distributed within the same numerical range;

[0082] S22. Construct a NeRF neural radiance field model, use the preprocessed historical credit data and performance capability data of beauty suppliers as input features, and set the credit data feature vector of the i-th supplier at time t as F i (t):

[0083] F i (t) = [R i (t), C i (t), B i (t)];

[0084] Among them, R i (t) represents the credit score of the i-th supplier at time t, C i (t) represents the performance capability data of the i-th supplier at time t, B i (t) represents the bidding historical data of the i-th supplier at time t, and is mapped to a high-dimensional space through a multi-layer perceptron to form a high-dimensional feature field of supplier credit;

[0085] S23. Perform positional encoding on the credit data feature vector F i (t), and use extended Fourier positional encoding to enhance the expressive ability of the supplier credit data in the high-dimensional feature field:

[0086]

[0087] Among them, γ(F i ) represents the credit data feature vector after positional encoding, L represents the encoding dimension, and λ l represents the frequency parameter of the first scale;

[0088] S24. Construct a NeRF implicit mapping model based on the credit data feature vector after positional encoding, and define the mapping function as:

[0089] H i (t) = σ(W2·ReLU(W1·γ(F i )) + b1)b2);

[0090] Among them, H i (t) represents the high-dimensional credit data feature vector of the i-th supplier at time t, σ represents the activation function, ReLU represents the rectified linear unit function, W1 and W2 represent weight matrices, and b1 and b2 represent bias terms;

[0091] S25. Use gradient descent to optimize the parameters of the NeRF implicit mapping model, and the optimization objective function is:

[0092]

[0093] Among them, N represents the total number of suppliers, T represents the time step, α i,t represents the error weighting coefficient of the i-th supplier at time t, λ1 represents the weight regularization coefficient, λ2 represents the time smoothing regularization coefficient, ∥·∥ represents the Euclidean norm, and ∥·∥ F represents the Frobenius norm, and H i (t - 1) represents the high-dimensional credit data feature vector of the i-th supplier at time t - 1;

[0094] S26. Based on the optimized NeRF implicit mapping model, use the high-dimensional credit data feature vector in the high-dimensional feature field to generate the continuous credit distribution of the beauty suppliers.

[0095] In this embodiment, the specific content of S3 includes:

[0096] S31. Define a credit density function to describe the probability distribution of the supplier credit:

[0097] σ(R i ) = exp(-β1|Ri -μ R |);

[0098] wherein, σ(R i ) represents the reputation density function, exp represents the exponential function, R i represents the reputation score of the i-th supplier at the current moment, μ R represents the historical reputation mean of the supplier, and β1 represents the attenuation coefficient of the reputation distribution, such that the reputation density changes with the historical reputation deviation;

[0099] S32. Define a reputation color mapping function to establish a non-linear mapping relationship of reputation in a high-dimensional space:

[0100]

[0101] wherein, c(R i ) represents the reputation color mapping function, β2 represents the steepness parameter of the reputation mapping curve, and R th represents the threshold of the reputation score;

[0102] Map the supplier reputation score to a normalized color space through the reputation color mapping function, so that the reputation interpolation prediction remains continuous and stable in the high-dimensional space;

[0103] S33. Based on the reputation density function σ(R i ) and the reputation color mapping function c(R i ), use the NeRF volume rendering method to calculate the reputation prediction value and generate the future reputation trend characteristics of the beauty supplier:

[0104]

[0105] wherein, represents the reputation prediction value of the i-th supplier at time t, τ max represents the end time of the time window for reputation calculation, τ min represents the start time of the time window for reputation calculation, τ represents the time parameter of the integral variable, σ(s) represents the reputation density of the supplier at time s, represents the reputation cumulative attenuation factor to ensure the time continuity of the reputation interpolation prediction.

[0106] In this embodiment, the specific content of S4 includes:

[0107] S41. Based on the future reputation trend characteristics of the beauty supplier and the bidding historical data, construct a supplier bidding decision state space, and the state of each node represents the bidding decision of the supplier at the corresponding moment;

[0108] S42. Based on the transition probability of the supplier's quotation decision among state nodes, the GFlowNet is used to generate a flow network to represent the supplier's bidding quotation process as a probability flow matching problem of a multi-state decision path. The probability flow size of each quotation decision node depends on the transition probability among states, and the bidding path forms a directed probability flow graph composed of quotation state nodes.

[0109] In this embodiment, the specific steps of S5 are as follows:

[0110] S51. Based on the quotation data, performance ability data, and profit expectation data of beauty suppliers, a comprehensive bidding reward mechanism is constructed. The comprehensive reward mechanism includes a quotation competitiveness index, a performance ability index, and a profit expectation index.

[0111] S52. Define the quotation competitiveness index of the supplier:

[0112]

[0113] Among them, represents the quotation competitiveness index of the i-th supplier at time t, P avg (t) represents the average quotation of all suppliers at time t, and P i (t) represents the real-time quotation of the i-th supplier at time t;

[0114] S53. Define the performance ability index of the supplier:

[0115]

[0116] Among them, C i (t) represents the performance ability index of the i-th supplier at time t, represents the reputation prediction value of the i-th supplier at time t, exp represents the exponential function, γ c represents the sensitivity coefficient of the performance ability mapping, G i (t) represents the reputation trend gradient of the i-th supplier at time t, and G th represents the reputation gradient threshold;

[0117] S54. Define the profit expectation index of the supplier:

[0118] U i (t) = Q i (t)[V market (t) - P i (t)] - L i (t);

[0119] Among them, U i (t) represents the profit expectation index of the i-th supplier at time t, and Q i(t) represents the expected sales volume of beauty products provided by the i-th supplier at time t, V market (t) represents the market average price of beauty products at time t, L i (t) represents the unit fulfillment cost of the i-th supplier at time t;

[0120] S55. Build a comprehensive reward mechanism, and set the comprehensive reward of the supplier in the bidding process as the weighted sum of the quotation competitiveness index, fulfillment ability index, and profit expectation index.

[0121] In this embodiment, the S6 specifically includes:

[0122] S61. Based on the comprehensive reward mechanism, establish an optimization space for the supplier's bidding quotation strategy, set the conversion relationship between different quotation strategies, and assign an initial reward value to each quotation strategy, so that the optimization of the quotation strategy is dynamically adjusted considering the quotation competitiveness, fulfillment ability, and profit expectation;

[0123] S62. Build an improved GFlowNet probability flow matching mechanism, and set the probability flow of each bidding quotation state node to be matched through the probability flow balance constraint between upstream and downstream nodes to ensure that the inflow and outflow of any node are balanced;

[0124] S63. The improved GFlowNet probability flow matching mechanism introduces a dynamic flow adjustment factor, and sets that when transferring the state of each bidding path, the probability flow distribution of the state transfer is adaptively adjusted according to the change of the comprehensive reward of the current node:

[0125] F(s→s') = F0(s→s')·exp(δ·ΔR bid (s,s'));

[0126] Among them, F(s→s') represents the adjusted probability flow from state node s to state node s', F0(s→s') represents the initial probability flow from state node s to state node s', exp represents the exponential function, δ represents the dynamic flow adjustment factor, and ΔR bid (s,s') represents the difference in comprehensive rewards between state node s and state node s';

[0127] S64. Use the improved GFlowNet probability flow matching mechanism for multi-round iterative optimization, and update the probability flow distribution between the state nodes of the bidding path in each round of optimization, so that the search for the bidding path gradually converges to the area with the highest comprehensive reward;

[0128] S65. Calculate the cumulative probability flow value of each state node in the bidding path, and identify the cumulative probability flow value of the bidding path corresponding to different quotation strategies;

[0129] S66. Sort all the bidding paths of the suppliers according to the cumulative probability flow value, and finally determine the globally optimal quotation path in the bidding process.

[0130] Example 1:

[0131] To verify the feasibility of the present invention in implementation, the present invention is applied to the supply chain management system of a large beauty e-commerce platform. This platform purchases beauty products worth more than 5 billion yuan annually, covering multiple categories such as skin care, makeup, and perfume, and the number of suppliers exceeds 5,000. However, in actual operation, this platform faces multiple supply chain bid evaluation optimization problems: First, the supplier reputation modeling is inaccurate. The traditional scoring system mainly relies on historical transaction data and lacks the prediction of future performance capabilities, resulting in some suppliers winning the bid at a low price but having insufficient performance capabilities, affecting the overall stability of the supply chain; second, the optimization of the bidding strategy is limited. The supplier quotation strategy often adjusts based on experience, and it is difficult for traditional reinforcement learning methods to find the globally optimal path, resulting in insufficient competition in the bidding process; third, the matching of procurement and inventory is not intelligent, and the dynamic performance capabilities of suppliers are not fully considered, resulting in unreasonable procurement plans, excessive inventory of some products, and insufficient supply of some products; fourth, the optimization of logistics scheduling is insufficient, and the distribution strategy is difficult to adjust dynamically, resulting in low overall operating efficiency of the supply chain.

[0132] Based on the NeRF neural radiance field technology and the GFlowNet generative flow network, the present invention intelligently optimizes the reputation, bidding path, procurement decision, and supply chain optimization of suppliers. First, the system uses the NeRF technology to model the historical reputation data, performance ability data, and bidding history data of 5,000 suppliers, constructs a high-dimensional reputation feature field of suppliers, and performs interpolation prediction through volume rendering technology to dynamically generate the future reputation trend of suppliers. Compared with the traditional reputation scoring system based on weighted average, this method can realize the continuous modeling of supplier reputation in the time dimension and predict the future performance capabilities of suppliers through high-dimensional interpolation methods, effectively avoiding procurement decision-making mistakes caused by the lag of historical scores.

[0133] In the bidding link, the system uses GFlowNet to model the bidding process, regards the supplier quotation process as a path search problem, and constructs a bidding state transition probability flow model. By introducing an improved GFlowNet probability flow matching mechanism, the supplier bidding strategy is optimized, enabling the quotation decision to achieve global optimality in various market environments. In the experiment, the performance of 500 suppliers under different bidding strategies was compared. The bidding path optimized by the method of the present invention increased the bidding success rate by 15.6% compared with the traditional reinforcement learning method, and at the same time, the overall score of the supplier performance ability increased by 22.8%. This result shows that the present invention can not only improve the bidding success rate of suppliers but also effectively reduce the risk of winning the bid at a low price but failing to meet the performance requirements.

[0134] In terms of procurement optimization, the present invention combines NeRF reputation modeling with GFlowNet bidding optimization to construct a comprehensive reward mechanism, and calculates the optimal procurement plan based on the supplier's quotation, performance ability and profit expectation. 50 beauty products were selected for the experiment to compare the performance of the method of the present invention and the traditional procurement method in terms of inventory matching. The results show that after adopting the method of the present invention, the inventory turnover rate of the supply chain has increased by 18.2%, the inventory out-of-stock rate has decreased by 27.4%, and the supplier performance delay rate has decreased by 30.5%. These data indicate that the present invention can effectively improve the accuracy of the procurement plan, reduce the risk of supply chain disruption, while reducing inventory backlogs and improving the overall supply chain operation efficiency.

[0135] In terms of logistics optimization, the present invention combines the prediction of the supplier's reputation trend with the bidding optimization result to dynamically adjust the distribution strategy, and combines the smart contract mechanism to ensure the transparency of the supplier's performance. The test results show that after optimizing with the method of the present invention, the on-time delivery rate of logistics distribution has increased to 96.3%, which is 12.8% higher than the traditional scheduling strategy, while the logistics cost has decreased by 9.7%. This result shows that the present invention has significant advantages in improving the stability of the supply chain and reducing the operating cost.

[0136] Through the above tests and data analysis, the application of the present invention in the evaluation and optimization of beauty supply chain bidding has achieved good results, which can accurately model the supplier's reputation, optimize the bidding strategy, realize intelligent matching of procurement and inventory, and improve the logistics scheduling efficiency.

[0137] Table 1 Comparison table of the intelligent bid evaluation optimization effects of the traditional method and the method of the present invention in the beauty supply chain

[0138] Evaluation Index Traditional Method After Optimization of the Invention Improvement Rate Success Rate of Supplier Bidding 58.2％ 73.8％ +15.6％ Score of Supplier Performance Ability 6.5 / 10 8.0 / 10 +22.8％ Inventory Turnover Rate of Supply Chain 4.3 times / year 5.1 times / year +18.2％ Inventory Stock-Out Rate 14.2％ 10.3％ -27.4％ Supplier Performance Delay Rate 9.8％ 6.8％ -30.5％ Logistics Distribution Punctuality Rate 85.5％ 96.3％ +12.8％ Decrease in Logistics Distribution Cost - -9.7％ -9.7％

[0139] The data comparison results in Table 1 of this embodiment show that the present invention has significant advantages in the evaluation and optimization of supply chain bidding. The success rate of supplier bidding has increased from 58.2% to 73.8%, an increase of 15.6%, effectively reducing the situation of winning the bid at a low price but lacking performance ability. The supplier performance ability score has increased by 22.8%, indicating that the NeRF reputation modeling method can more accurately evaluate the long-term stability of suppliers. The inventory turnover rate of the supply chain has increased by 18.2%, the inventory out-of-stock rate has decreased by 27.4%, and the performance delay rate has decreased by 30.5%, showing that the present invention can optimize the procurement plan and improve the reliability of the supply chain. The on-time delivery rate of logistics distribution has increased to 96.3%, an increase of 12.8%, while the logistics cost has decreased by 9.7%, proving that the improved GFlowNet bidding optimization mechanism can dynamically adjust the supply chain strategy and improve the overall operation efficiency. Overall, the present invention has achieved obvious optimization in supplier bid evaluation, bidding optimization, procurement matching and logistics management, making the supply chain more intelligent and stable.

[0140] As described above, it is only a preferred specific embodiment of the present invention, but the protection scope of the present invention is not limited thereto. Any person skilled in the art within the technical scope disclosed by the present invention, according to the technical solution and inventive concept of the present invention, making equivalent substitutions or changes, shall be covered by the protection scope of the present invention.

Claims

1. An intelligent bid evaluation optimization method for the beauty supply chain based on deep learning, characterized in that The steps include: S1. Use sensor networks to collect historical reputation data, contract performance data, and bidding history data of beauty suppliers in real time to build a data set of historical contract performance behaviors of beauty suppliers; S2. Based on the historical fulfillment behavior dataset of beauty suppliers, NeRF neural radiation field technology is used to map the historical reputation data and fulfillment capabilities of beauty suppliers into a high-dimensional feature space to construct a continuous reputation distribution of beauty suppliers. S3. Use volume rendering technology to interpolate and predict the continuous reputation distribution of beauty suppliers and generate the future reputation trend characteristics of beauty suppliers; S4. Based on the future reputation trend characteristics and bidding history data of beauty suppliers, GFlowNet is used to generate a flow network to model the supplier bidding process as a path search problem, where each supplier's bidding decision corresponds to a state node in the path; S5. Build a comprehensive reward mechanism based on the quotations, fulfillment capabilities and profit expectations of beauty suppliers; S6. Based on the comprehensive reward mechanism and the improved GFlowNet probabilistic flow matching mechanism, dynamic optimization is performed among different bidding strategies to finally determine the global optimal bidding path in the bidding process.

2. The intelligent bid evaluation optimization method for the beauty supply chain based on deep learning according to claim 1, wherein The S2 specifically includes: S21. Preprocess the data set of historical performance behavior of cosmetics suppliers, remove missing values and outliers in the data, and normalize the historical reputation data and performance ability data of suppliers, using the mean normalization method to distribute the data within the same numerical range; S22. Build a NeRF neural radiance field model, use the preprocessed historical credit data and performance ability data of beauty suppliers as input features, and set the credit data feature vector of the i-th supplier at time t as F i (t): F i f(t)=[R i (t), C i (t), B i (t)]; Among them, R i (t) represents the credit score of the i-th supplier at time t, and C i (t) represents the performance ability data of the i-th supplier at time t, and B i (t) represents the bidding history data of the i-th supplier at time t, and is mapped to a high-dimensional space through a multi-layer perceptron to form a high-dimensional feature field of supplier credit; S23. Perform positional encoding on the credit data feature vector F i (t), and adopt extended Fourier positional encoding to enhance the expression ability of supplier credit data in the high-dimensional feature field: Among them, γ(F i ) represents the reputation data feature vector after position encoding, L represents the encoding dimension, and λ l represents the frequency parameter of the first scale; S24. Construct a NeRF implicit mapping model based on the position-encoded reputation data feature vector, and define the mapping function as: H i (t) = σ(W2 · ReLU(W1 · γ(F i )) + b1)b2); Among them, H i (t) represents the high-dimensional credit data feature vector of the i-th supplier at time t, σ represents the activation function, ReLU represents the rectified linear unit function, W1 and W2 represent the weight matrices, and b1 and b2 represent the bias terms; S25. Use gradient descent to optimize the parameters of the NeRF implicit mapping model, and the optimization objective function is: Among them, N represents the total number of suppliers, T represents the time step, and α i,t represents the error weighting coefficient of the i-th supplier at time t, λ1 represents the weight regularization coefficient, λ2 represents the time smoothing regularization coefficient, ∥·∥ represents the Euclidean norm, and ∥·∥ F represents the Frobenius norm, and H i (t - 1) represents the high-dimensional reputation data feature vector of the i-th supplier at time t - 1; S26. Based on the optimized NeRF implicit mapping model, the high-dimensional reputation data feature vector in the high-dimensional feature field is used to generate the continuous reputation distribution of beauty suppliers.

3. The intelligent bid evaluation optimization method for the beauty supply chain based on deep learning according to claim 1, wherein The S3 specifically includes: S31. Define the reputation density function to describe the probability distribution of supplier reputation: σ(R i ) = exp(-β1|R i - μ R |); Among them, σ(R i ) represents the reputation density function, exp represents the exponential function, R i represents the reputation score of the i-th supplier at the current moment, μ R represents the historical reputation mean of the supplier, and β1 represents the attenuation coefficient of the reputation distribution, such that the reputation density changes with the historical reputation deviation; S32. Define a reputation color mapping function to establish a nonlinear mapping relationship of reputation in a high-dimensional space: Among them, c(R i ) represents the reputation color mapping function, β2 represents the steepness parameter of the reputation mapping curve, and R th represents the threshold of the reputation score; Mapping the supplier reputation score to a normalized color space through the reputation color mapping function, so that the reputation interpolation prediction remains continuous and stable in the high-dimensional space; S33. Based on the reputation density function σ(R i ) and the reputation color mapping function c(R i ), use the NeRF volume rendering method to calculate the reputation prediction value and generate the future reputation trend characteristics of the beauty product suppliers: Among them, represents the predicted reputation value of the \(i\)-th supplier at time \(t\), and \(\tau\) max represents the end time of the time window for reputation calculation, and \(\tau\) min represents the start time of the time window for reputation calculation. \(\tau\) represents the time parameter of the integration variable, and \(\sigma(s)\) represents the reputation density of the supplier at time \(s\). represents the reputation cumulative decay factor, ensuring the time continuity of reputation interpolation prediction.

4. The intelligent bid evaluation optimization method for the beauty supply chain based on deep learning according to claim 1, wherein, The S4 specifically includes: S41. Based on the future reputation trend characteristics and bidding history data of beauty suppliers, a supplier bidding decision state space is constructed, and the state of each node represents the supplier's bidding decision at the corresponding moment; S42. Based on the transition probability of supplier quotation decisions between state nodes, GFlowNet is used to generate a flow network to represent the supplier's bidding process as a probability flow matching problem of multi-state decision paths. The probability flow of each quotation decision node depends on the transition probability between states, and the bidding path forms a directed probability flow graph composed of quotation state nodes.

5. The intelligent bid evaluation optimization method for the beauty supply chain based on deep learning according to claim 1, wherein The S5 specifically includes: S51. Based on the quotation data, performance ability data, and profit expectation data of beauty suppliers, construct a comprehensive bidding reward mechanism, where the comprehensive reward mechanism includes a quotation competitiveness index, a performance ability index, and a profit expectation index; S52. Define the quotation competitiveness index of the supplier: Among them, represents the quotation competitiveness index of the i-th supplier at time t, and P avg (t) represents the average quotation of all suppliers at time t, and P i (t) represents the real-time quotation of the i-th supplier at time t; S53. Define the performance ability index of the supplier: Among them, C i (t) represents the performance ability index of the i-th supplier at time t, represents the credit prediction value of the i-th supplier at time t, exp represents the exponential function, γ c represents the sensitivity coefficient of the performance ability mapping, G i (t) represents the credit trend gradient of the i-th supplier at time t, G th represents the credit gradient threshold; S54. Define the profit expectation index of the supplier: U i U(t) = Q i U(t)[V market U(t) - P i U(t)] - L i U(t); Among them, U i (t) represents the profit expectation index of the i-th supplier at time t, Q i (t) represents the expected sales volume of beauty products provided by the i-th supplier at time t, V market (t) represents the market average price of beauty products at time t, L i (t) represents the unit fulfillment cost of the i-th supplier at time t; S55. Construct a comprehensive reward mechanism, and set the comprehensive reward of the supplier during the bidding process as the weighted sum of the quotation competitiveness index, the performance ability index, and the profit expectation index.

6. The intelligent bid evaluation optimization method for the beauty supply chain based on deep learning according to claim 1, wherein The specific content of S6 includes: S61. Based on the comprehensive reward mechanism, establish an optimization space for the supplier's bidding quotation strategy, set the conversion relationship between different quotation strategies, and assign an initial reward value to each quotation strategy, so that the optimization of the quotation strategy is dynamically adjusted considering quotation competitiveness, performance ability, and profit expectation; S62. Construct an improved GFlowNet probability flow matching mechanism, and set the probability flow of each bidding quotation status node to be matched through the probability flow balance constraint between upstream and downstream nodes to ensure that the inflow and outflow of any node are balanced; S63. The improved GFlowNet probability flow matching mechanism introduces a dynamic flow adjustment factor, and sets that when the status of each bidding path is transferred each time, the probability flow distribution of the status transfer is adaptively adjusted according to the change of the comprehensive reward of the current node: F(s→s') = F0(s→s')·exp(δ·ΔR bid (s, s')); Among them, F(s→s') represents the adjusted probability flow from state node s to state node s', F0(s→s') represents the initial probability flow from state node s to state node s', exp represents the exponential function, δ represents the dynamic flow adjustment factor, and ΔR bid (s, s') represents the comprehensive reward difference between state node s and state node s'; S64. Use the improved GFlowNet probability flow matching mechanism to perform multiple rounds of iterative optimization, and update the probability flow distribution between the status nodes of the bidding path in each round of optimization, so that the search for the bidding path gradually concentrates on the area with the highest comprehensive reward; S65. Calculate the cumulative probability flow value of each status node in the bidding path, and identify the cumulative probability flow value of the bidding path corresponding to different quotation strategies; S66. Sort all the bidding paths of the supplier according to the magnitude of the cumulative probability flow value, and finally determine the globally optimal bidding path during the bidding process.

Images