Procurement supply chain integrated optimization method and equipment based on industrial Internet

Abstract

The present invention discloses a method and device for optimizing the integrated procurement supply chain based on the industrial internet. The method comprises the following steps: predicting the market price of raw materials; establishing a multi-period dynamic procurement model with the minimum unit raw material cost, taking into account the predicted market price of raw materials, and optimizing and solving the optimal raw material procurement quantity for each procurement cycle; obtaining the historical supply data of each supplier, constructing a supplier automatic evaluation model based on stacked autoencoders and bagging integrated learning, and obtaining the scores of each supplier; comprehensively considering the optimal raw material procurement quantity and supplier scores, establishing a multi-objective optimization model that minimizes procurement costs and maximizes the comprehensive effects of suppliers, and solving and obtaining the optimal procurement strategy from each supplier. The present invention comprehensively acquires, centrally processes and analyzes the entire procurement supply chain process data through the industrial internet platform, comprehensively considers the correlation between different businesses, and can effectively reduce procurement costs and improve procurement efficiency.

Description

Technical Field

[0001] The present invention belongs to the field of procurement supply chain, and specifically relates to a procurement supply chain integration optimization method based on the industrial Internet. Background Art

[0002] With the rapid advancement of next-generation information technologies such as industrial networks, big data, and artificial intelligence, intelligent manufacturing has become a common theme in the development of the manufacturing industry. As the key infrastructure for intelligent manufacturing and intelligent applications, the Industrial Internet, through the comprehensive interconnection of people, machines, and objects, builds a foundational network connecting devices / equipment, materials, people, and information systems. This enables comprehensive perception, dynamic transmission, and real-time analysis of industrial data, enabling scientific decision-making and intelligent control, and improving the efficiency of manufacturing resource allocation. Considered one of the most important technological fields of the future, the Industrial Internet is attracting widespread attention from all walks of life.

[0003] From an application perspective, procurement and supply chain optimization is one of the key areas where the Industrial Internet can bring significant improvements and benefits. In industrial manufacturing, raw materials are the source of the production and processing process. The quality and supply continuity of raw materials directly impact the stability of production and processing operations. Furthermore, in smelting and processing companies, raw material procurement costs account for approximately 70% of production and processing costs, tying up significant capital. Optimizing the procurement supply chain can significantly reduce production costs and ensure raw material quality, which is crucial for improving the overall profitability of enterprises.

[0004] The procurement supply chain primarily encompasses processes such as market price analysis, total purchase volume decisions, supplier evaluation, and supplier order allocation. Numerous scholars have conducted extensive research on each of these areas. While existing work has achieved promising results in their respective areas, it generally focuses on specific aspects of the procurement supply chain, specifically price forecasting or order allocation. In most companies, different procurement processes are typically performed by different departments or individuals. These isolated departments can easily create information silos, and the in-the-loop approach can lead to significant delays, inefficient information transmission, and poor decision-making. Furthermore, manual decision-making is highly subjective and arbitrary, and cannot comprehensively assess massive amounts of information, resulting in high raw material procurement costs and low efficiency. Therefore, traditional supply chains are no longer able to meet the agility and high-quality demands of intelligent manufacturing, and urgently need to be transformed into highly intelligent digital supply chains.

[0005] The Industrial Internet provides a collaborative and efficient application platform for the supply chain, potentially reshaping its operations. Based on the Industrial Internet, information from disparate systems, departments, and granularity within the supply chain can be integrated, taking into account business activities across the entire supply chain and making comprehensive, well-informed decisions in real time and dynamically. Thanks to the development of the Industrial Internet, it can integrate multi-source data across both horizontal and vertical dimensions, providing a powerful platform for big data analysis and processing, complex modeling, and optimization. This presents new opportunities for agile and efficient procurement supply chain operations. Compared to traditional procurement supply chains, the automated processing capabilities of procurement supply chains based on the Industrial Internet significantly free up labor. A comprehensive, coordinated approach and the application of intelligent methods enable better decision-making, reduce procurement costs, and improve operational efficiency. In short, the Industrial Internet presents new opportunities for agile and efficient procurement supply chain operations. Summary of the Invention

[0006] The present invention provides a procurement supply chain integration optimization method based on the industrial Internet. Through the industrial Internet platform, the entire process data of the procurement supply chain is comprehensively acquired, centrally processed and analyzed, and the correlation between different businesses is comprehensively considered, which can effectively reduce procurement costs and improve procurement efficiency.

[0007] In order to achieve the above technical objectives, the present invention adopts the following technical solutions:

[0008] A procurement supply chain integration optimization method based on the Industrial Internet, including:

[0009] Market price forecast: forecast the market price of raw materials;

[0010] Raw material procurement decision-making: Considering the market forecast price of raw materials, a multi-period dynamic procurement model is established to minimize the unit raw material cost, and the optimal raw material procurement quantity for each procurement cycle is optimized;

[0011] Supplier evaluation: We obtain historical supply data from each supplier and build an automatic supplier evaluation model based on stacked autoencoders and bagging ensemble learning to obtain a score for each supplier.

[0012] Supplier order allocation: Taking into account the optimal raw material purchase volume and supplier ratings, a multi-objective optimization model is established to minimize procurement costs and maximize the overall supplier effect, and the optimal procurement strategy from each supplier is obtained.

[0013] The data required for the market price forecast, raw material procurement decision, supplier evaluation and supplier order allocation are automatically obtained through the industrial Internet platform.

[0014] Furthermore, variational mode decomposition and long short-term memory network are used to predict the market price of raw materials, specifically:

[0015] Step A1: Obtain the market price sequence f(t) of the raw material and decompose it into K modal components u using the variational mode decomposition method. i (t),i=1,2,…K;

[0016] Step A2: For each modal component u, i (t) Split to obtain input and output for model training, which are then used to train LSTM to obtain the prediction model corresponding to each component;

[0017] In step A3, the most recent market price sequence is obtained, decomposed into K modal components according to step A1, and rolling forecasts are performed using various forecast models. Finally, the K forecast sequences are summed to obtain the market forecast price sequence of the raw material.

[0018] Furthermore, the method for determining the number K of variational mode decomposition is: gradually increase the number of decompositions and calculate the corresponding residuals. When the residual is less than the preset value and there is no obvious downward trend, the current number of decompositions is determined to be the optimal decomposition number K; the residual calculation formula is:

[0019]

[0020] Where r res is the residual, and M is the number of training samples obtained by splitting f(t).

[0021] Furthermore, the multi-period dynamic procurement model established in the raw material procurement decision is:

[0022]

[0023] k=1,2,3,...,L(19e)

[0024] Among them, C is the raw material cost of the procurement planning cycle, including the procurement cost C p and inventory cost C w The duration of the procurement planning cycle T consists of L separate decision cycles, and the duration of the kth decision cycle is T k , the raw material purchasing point t in the kth decision cycle k At the midpoint of the current decision cycle; Q k is the total raw material purchase plan for the kth decision cycle;

[0025] Purchase cost C p Including raw material costs and handling fees, P k is the market price forecast obtained in the kth decision cycle; C S is the handling fee for each decision cycle;

[0026] Inventory costs R0 is the initial raw material inventory of the procurement planning cycle; D k T k The production plan corresponds to the raw material demand D k / η, η is the production conversion rate; C I is the annual loan interest rate, C M is the daily inventory management cost per unit of raw material;

[0027] B k Indicates that in the procurement cycle T k Capital budget for raw material procurement, ss is the safety stock, and V is the maximum inventory capacity of the warehouse.

[0028] Furthermore, the particle swarm optimization algorithm is used to solve the multi-period dynamic procurement model, and the penalty function is used to deal with the constraints in the procurement model during the solution, specifically:

[0029] Step B1: Convert the constraints (19b), (19c) and (19d) into the following penalty functions P1, P2 and P3 respectively, and add them together to obtain the total penalty function P:

[0030]

[0031] P=P1+P2+P3 (23)

[0032] Step B2: According to the penalty function P, the multi-period dynamic procurement model is converted into an unconstrained optimization problem:

[0033]

[0034] Where α is the penalty factor;

[0035] Step B3, initialize a particle population, and select the function shown in formula (24) as the fitness function of the particle; wherein the particle dimension is equal to the number of decision cycles L included in the procurement plan cycle, and the position of each particle is represented by the raw material purchase quantity (Q1, Q2, ..., Q L );

[0036] Step B4: Particle swarm optimization is used to find the optimal particle, and the optimal raw material purchase quantity for each decision cycle within the purchase plan cycle is obtained based on the position of the optimal particle.

[0037] Furthermore, the specific process of supplier evaluation is as follows:

[0038] Step C1: Collect historical data of suppliers These include Each supplier has n indicator data for supplier evaluation; experts evaluate suppliers based on each indicator data and obtain the historical score of each supplier; the original historical data and scores to form a new dataset

[0039] Step C2, based on the bagging method, Randomly select N S samples with replacement to form a sub-dataset;

[0040] Step C3, repeat step C2 to get N D Independent sub-datasets;

[0041] Step C4: In each sub-dataset, the n-dimensional evaluation index data of each supplier is used as input, and the supplier's score is used as output. A stacked autoencoder model with n-dimensional input and 1-dimensional output is trained to obtain a mapping relationship between the supplier evaluation index data and the supplier score.

[0042] Step C5, repeat step C4 to obtain N D An independent evaluation model based on stacked autoencoders;

[0043] Step C6, when the new supplier n-dimensional indicator data is input into the stacked autoencoder model, N D The stacked autoencoder model correspondingly obtains N D scoring results, and then take N D The average of the rating results is used as the final rating for the supplier.

[0044] Furthermore, minimizing the procurement cost means minimizing the difference between the raw material procurement cost and the recycling value, which is expressed as the objective function f1(a):

[0045]

[0046] Where a ij is the purchase quantity of raw materials of grade j from supplier i, p ij is the price of the jth grade raw material from the i-th supplier, e ij N is the additional revenue per ton of raw materials of grade j from supplier i; s is the number of raw material suppliers, N g The number of grades of raw materials;

[0047] Maximizing the comprehensive supplier effect means maximizing the overall utility of all suppliers. The goal is to purchase from suppliers with high scores as much as possible, which is expressed as the objective function f2(a):

[0048]

[0049] Where, is the n indicator data of the i-th supplier, N is the score of the dth stacked autoencoder for the i-th supplier in the bagging method; D The number of stacked autoencoders to use for bagging.

[0050] Furthermore, the constraints in establishing the multi-objective optimization model include supply capacity constraints, demand constraints, order quantity constraints, and inventory constraints;

[0051] The supply capacity constraint refers to the raw material procurement quantity a ij The supply capacity of the jth grade raw material of the i-th supplier should not exceed r ij , expressed as:

[0052] a ij ≤r ij ,i=1,2,…,N s ,j=1,2,…,N g (31)

[0053] The demand constraint means that the sum of the main components in the raw materials should meet the production demand, which is expressed as:

[0054]

[0055] Among them, θ ij The main component content of the raw material;

[0056] The quantity constraint means that the total order purchase quantity must meet the purchase plan G and not exceed the preset proportion of the total purchase plan quantity;

[0057] The inventory constraint means that the total order purchase quantity should not exceed the remaining capacity of the warehouse.

[0058] Furthermore, when solving the multi-objective optimization model for supplier order allocation, the multi-objective problem is transformed into a single-objective problem using function transformation and linear weighting method, specifically:

[0059] First, each optimization objective is scaled so that the values ​​of each objective after the transformation are in the same order of magnitude, as shown in the following formula:

[0060]

[0061] where f i * (a) Solve the i-th objective function f by considering the constraints separately i (a) The optimal value obtained;

[0062] Then assign weights ξ to different objective functions i ;

[0063] Finally, the compromise single-objective optimization problem is obtained through linear combination:

[0064]

[0065] An electronic device includes a memory and a processor, wherein the memory stores a computer program, and when the computer program is executed by the processor, the processor implements any of the above-mentioned procurement supply chain integrated optimization methods based on the industrial Internet.

[0066] Compared with the prior art, the present invention has the following beneficial effects:

[0067] (1) The present invention comprehensively considers factors such as market price changes, production plans, inventory conditions, and capital constraints, and establishes a multi-period dynamic procurement model with the minimum unit cost of raw materials. It makes more comprehensive and refined decisions from a long-term perspective, solves the problems of partial information and extensive models in traditional models, and effectively reduces procurement costs.

[0068] (2) The present invention comprehensively considers various indicators of supplier supply, explores the correlation between various indicators and supplier ratings, and automatically gives a comprehensive rating for each supplier, thus solving the problem that manual supplier evaluation is highly subjective and labor-intensive, resulting in non-objective and low-efficiency supplier evaluation management.

[0069] (3) The present invention starts from the entire process of raw material supply chain business, analyzes the correlation between different businesses, organically connects different businesses, and automatically completes raw material procurement decisions from a global perspective, solving the problems of low decision-making efficiency and poor results caused by serious information islands and isolated businesses in the traditional model.

[0070] (4) In response to the problem that the number of modes in variational mode decomposition (VMD) cannot be reasonably determined, the present invention proposes a method for determining the number of modes based on decomposition residuals, which effectively avoids the problems of under-decomposition and over-decomposition, and provides subsequent predictors with sequences with more distinct features and smoother changes, thereby improving the overall prediction effect. BRIEF DESCRIPTION OF THE DRAWINGS

[0071] Figure 1 It is an overall framework for optimizing the raw material supply chain based on the Industrial Internet;

[0072] Figure 2 This is the overall architecture of the method described in the embodiments of the present application;

[0073] Figure 3 This is a schematic diagram of the VMD-LSTM model described in the embodiment of the present application;

[0074] Figure 4 This is a schematic diagram of rolling purchase in the method described in the embodiment of the present application;

[0075] Figure 5 This is a schematic diagram of raw material inventory changes in the method described in the embodiment of the present application;

[0076] Figure 6 It is the autoencoder structure;

[0077] Figure 7 Schematic diagram of the structure and training process of SAE;

[0078] Figure 8 Schematic diagram of the supplier evaluation model based on Bagging and SAE. DETAILED DESCRIPTION

[0079] The following is a detailed description of an embodiment of the present invention. This embodiment is based on the technical solution of the present invention, provides a detailed implementation method and a specific operation process, and further explains the technical solution of the present invention.

[0080] This embodiment adopts the four-layer architecture proposed in the "Industrial Internet White Paper" of the China Industrial Internet Industry Alliance, including the edge layer, resource layer, platform layer and application layer. This structure gives a clear definition of the structure and function of each unit of the industrial Internet system and is widely recognized. In this patent, the overall architecture of the integrated optimization of the raw material supply chain based on the industrial Internet is as follows: Figure 1 As shown in the figure. First, based on the Industrial Internet platform, internal and external data such as various business decision-making data, production process data, raw material testing data, and market change data are collected and acquired, and centrally stored in the big data platform data warehouse. Next, data governance tasks such as outlier detection and missing value repair are performed on the data, as well as data fusion tasks such as multi-table aggregation. Then, intelligent decision-making algorithms for the raw material supply chain are developed within the integrated development environment of the Industrial Internet platform, and they are standardized and packaged, and deployed using lightweight APIs. Finally, during application development, the algorithm model is called through the API, and a variety of C / S and B / S industrial apps are developed.

[0081] The integrated optimization method for procurement and supply chain based on the Industrial Internet proposed in this embodiment mainly includes four parts: market price forecasting, procurement optimization decision-making, supplier evaluation and order allocation. The overall technical architecture is as follows: Figure 2As shown in the figure, to address the dramatic market price fluctuations, variational mode decomposition and long short-term memory networks are used to predict market prices. Next, to address the short planning cycles and limited information in traditional procurement models, a multi-period dynamic procurement model is established to minimize unit raw material costs, taking into account market forecast prices, inventory information, and production plans. A particle swarm optimization algorithm is then used to optimize the solution and obtain the optimal raw material procurement quantity for each procurement period. Furthermore, to address the subjectivity and inefficiency of manual supplier evaluation, an automatic supplier evaluation model based on stacked autoencoders and bagging ensemble learning is proposed, leveraging data automatically acquired from the Industrial Internet platform, including raw material testing information, contract information, and weighing logs, enabling efficient supplier evaluation. Finally, to address the significant influence of personal preferences and the inability to adapt to massive amounts of information when manually assigning orders, a multi-objective optimization model is developed to minimize procurement costs and maximize supplier comprehensive benefits, taking into account supplier ratings, supplier supply capacity, and raw material quality. The model then obtains the optimal procurement strategy for each supplier. The Industrial Internet can automatically collect and analyze data throughout the procurement supply chain. By considering the relationships between different business tasks, these intelligent methods can be integrated to achieve comprehensive, real-time decision-making from a holistic perspective.

[0082] 1. Market price forecast

[0083] In zinc smelting enterprises, raw material costs usually account for about 70% of total production costs. If raw material prices can be accurately predicted and analyzed, the timing and quantity of raw material procurement can be guided, thereby reducing procurement costs. However, in the real world, market prices are affected by multiple factors such as the global economy, supply and demand, exchange rates, and the coupling correlation between different factors is strong, resulting in frequent market price fluctuations, large fluctuations, and obvious nonlinear and non-steady-state characteristics. Traditional methods and manual analysis often fail to achieve good results. Therefore, accurate and robust price forecasting technology is crucial for corporate procurement and scheduling. In this embodiment, a price forecasting method based on the "decomposition-integration" framework is adopted to achieve short-term accurate forecasting of raw material prices.

[0084] (1) Variational mode decomposition

[0085] Variational mode decomposition is an advanced multi-resolution technique for adaptive non-recursive signal decomposition. It transforms the signal decomposition process into the construction and solution of a variational optimization problem. It can decompose a true signal into a series of bandwidth-limited intrinsic mode functions (IMFs) u k Assume that each IMF u k are all concentrated at a central frequency ω k The center frequency is determined by the decomposition process.

[0086] (2) Long Short-Term Memory Network: The Long Short-Term Memory (LSTM) network is an extended version of the Recurrent Neural Network (RNN) and has been widely used in many fields. Compared with the Recurrent Neural Network, the Long Short-Term Memory Network contains not only external loops but also internal loops of memory cells. Unlike traditional neural networks, LSTM uses memory blocks to maintain the internal state of the network. It can explore the internal abstract features and underlying structure of the data and has a strong predictive ability for time series. Currently, LSTM has been widely used in various time series processing tasks.

[0087] (3) Hybrid prediction method based on VMD and LSTM

[0088] In view of the advantages of VMD and LSTM, this embodiment combines the VMD method with the LSTM model to establish a hybrid prediction model based on the "decomposition-integration" concept. This model is referred to as VMD-LSTM in this article. Assuming f(t) is the original time series, VMD decomposes the original series into multiple independent components u i (t), i = 1, 2, ... K, where K is the total number of components. This decomposition helps to accurately model internal features and improve prediction accuracy. LSTM, as the prediction core of this prediction engine, is used to model and predict each component obtained by decomposition. The schematic diagram of the VMD-LSTM model is shown below. Figure 3 shown.

[0089] In the market price prediction part of this embodiment, the specific process of using variational mode decomposition and long short-term memory network to predict the market price of raw materials is as follows:

[0090] Step A1: Obtain the market price sequence f(t) of the raw material and decompose it into K modal components u using the variational mode decomposition method. i (t),i=1,2,…K;

[0091] Before performing variational mode decomposition, the number of independent components K needs to be set in advance. If the K value is set too small, the original sequence may not be fully decomposed, and complex subsequences still cannot obtain satisfactory prediction results; if the K value is set too large, the original signal will be over-decomposed, and the gap between each sequence will become very small, which cannot reflect the change law of different features, and also leads to a decrease in prediction accuracy and unnecessary computational consumption. In order to identify the optimal number of independent components K, this embodiment proposes a method based on the original sequence decomposition residual r res Method for determining the number of VMD decomposition modes.

[0092]

[0093] Where M is the number of samples, that is, the number of training samples obtained by splitting f(t) into input and output according to the step size. Obviously, r res As the residual of the original sequence decomposition, when it is small to a certain extent, it can be considered that the original sequence has been fully decomposed and each independent component can basically reconstruct the original sequence; when it does not show an obvious downward trend, it can be considered that the number of independent components has reached saturation and no additional independent components are needed. Therefore, according to experience, when the decomposition residual r res When it is less than 1% and there is no obvious downward trend, it can be considered that the optimal number of components is currently obtained.

[0094] In this step A1, the original price sequence f(t) is decomposed into K independent components u by VMD i (t), compared with the original sequence, the independent component subsequence is smoother, more regular, and has a more obvious change pattern, which greatly reduces the burden of the market price prediction model.

[0095] Step A2: For each modal component u, i (t) is split to obtain the input and output for model training, which is then used to train LSTM to obtain the prediction model corresponding to each component

[0096] In step A3, the most recent market price sequence is obtained, decomposed into K modal components according to step A1, and rolling forecasts are performed using various forecast models. Finally, the K forecast sequences are summed to obtain the market forecast price sequence of the raw material.

[0097] The final prediction result is obtained by summing the predicted values ​​of each subsequence. Therefore, this embodiment significantly improves the accuracy of price prediction by combining the advantages of VMD and LSTM.

[0098] In addition, for LSTM neural networks, it is necessary to pre-set model hyperparameters in actual application, but there are many hyperparameters and the search space is large. Traditional manual search and grid search methods are difficult to traverse the parameter space and are time-consuming. Therefore, after manually determining the basic range of hyperparameters based on experience, this embodiment uses a random search method to refine the hyperparameters to obtain better modeling results.

[0099] 2. Purchasing decision optimization

[0100] Currently, most companies make purchasing decisions without considering future fluctuations in raw material prices and demand, or by assuming that future prices and demand are fixed, thus only considering costs during the current purchasing period. This purchasing approach is short-sighted, making it difficult to adapt to market fluctuations and easily leading to high-priced inventory and excessive procurement costs. To address this, this patent establishes a multi-period dynamic procurement model that considers future fluctuations in raw material prices and demand, thereby enabling dynamic and flexible procurement.

[0101] (1) Procurement model

[0102] Procurement strategies such as Figure 4 shown.

[0103] At the first purchase, an overall decision is made for the purchase planning cycle of length T, which includes L individual purchase decision cycles T k , its inventory changes as follows Figure 5 As shown, R0 is the initial raw material inventory of the purchase planning cycle T, ss is the safety stock, Q k is the total raw material purchase plan for the kth purchase cycle, t k The raw material procurement point for the kth decision cycle is at the midpoint of the current decision cycle. The procurement model is used to determine the optimal multi-period procurement quantity and implement the current cycle's procurement quantity, Q1. At the next procurement point, the new information is updated and the forecast is completed. The multi-period procurement model is rebuilt, and the optimal procurement quantity, Q'2, for the corresponding decision cycle is implemented. As time progresses, the decision is made on a rolling basis to determine the optimal procurement quantity for each decision cycle.

[0104] The relevant explanations and assumptions of this procurement model are as follows:

[0105] ① The company purchases raw materials on a monthly basis and makes purchases at the midpoint of each decision period.

[0106] ②Since most production processes require continuous and uninterrupted production, assuming that the raw materials can be 100% supplied and there is no shortage, there is no need for emergency orders.

[0107] ③Purchasing decision cycle T k The purchased raw materials will arrive at the beginning of the next decision period.

[0108] ④ There are no time constraints in the enterprise's production process, that is, continuous production is maintained before and after the procurement period.

[0109] ⑤ The remaining inventory in each decision period will be postponed to the next decision period, which means that current purchasing decisions will affect future inventory levels.

[0110] The goal of raw material procurement is to reduce the cost of raw materials while ensuring normal production. The cost mainly includes procurement cost C p and inventory cost Cw .

[0111] Procurement cost C during the procurement planning period p Mainly includes raw material costs and handling fees:

[0112]

[0113] Among them, P k is the raw material price of the kth decision cycle predicted by the VMD-LSTM model, C S is the handling fee for each decision period.

[0114] Inventory cost C w Mainly related to the average inventory level. k , the average inventory level is:

[0115]

[0116] The total inventory cost for the purchase plan period T is:

[0117]

[0118] Where η is the production conversion rate, C I is the annual loan interest rate, C M Daily inventory management cost of raw materials per unit, D k T k The production plan corresponds to the raw material demand D k / η.

[0119] With minimizing unit raw material procurement costs within the procurement planning period T as the objective function, a multi-period procurement model is established under fluctuating raw material prices and demand. It is important to note that the unit cost of raw materials during period T is considered here, not the total cost. Clearly, in a procurement model focused on minimizing total cost, the procurement volume during each decision period will be just enough to meet production demand and safety stock, thereby reducing total cost. This model cannot adapt to fluctuations in raw material market prices. If the objective is to minimize unit cost, price fluctuations are fully considered and procurement quantities are adjusted, aiming to purchase as much raw material as possible at a lower price while ensuring safe production. In the long run, this model can effectively leverage the benefits of price fluctuations and reduce overall procurement costs.

[0120] The embodiment of the present invention establishes a multi-period dynamic procurement model that takes into account changes in raw material prices and demand:

[0121]

[0122] k=1,2,3,...,L(19e)

[0123] Among them B k Indicates that during the purchase period T k The capital budget for raw material procurement is V, which is the maximum inventory capacity of the warehouse. The above model takes into account the constraints of budget, safety stock and storage capacity. Formula (19b) indicates that the raw material procurement cost cannot be greater than the capital budget, where is the set of modal components obtained by variational mode decomposition; represents the prediction function of the i-th modal component, that is, the prediction model obtained based on LSTM training. The sum of the predicted values ​​of K modal components is equal to the predicted value of the final price, that is, the price P of the k-th decision cycle. k Formula (19c) indicates that the quantity of purchased raw materials must meet production requirements, and the remaining inventory at the end of the period must not be less than the safety stock to withstand the impact of uncertain factors. Formula (19d) indicates that the raw material inventory cannot exceed the maximum inventory capacity of the warehouse.

[0124] The above multi-period dynamic procurement model takes into account the changes in future raw material prices and demand, and combines various practical constraints to make more comprehensive decisions from a long-term perspective, which can further save costs.

[0125] (2) Optimization method:

[0126] The present invention adopts the particle swarm optimization algorithm (PSO) to solve the multi-period procurement model, and the penalty function is used to process the constraint terms. For formula (19), it is obvious that it is a typical constrained optimization problem. The penalty function method adds the constraint conditions to the objective function in the form of penalty terms. If the current solution is not within the feasible domain, the value of the penalty term in the objective function will become extremely large, making it impossible for the current solution to become the optimal solution; if the current solution is within the feasible domain, the penalty term value is 0, and the change in the value of the penalty term plays a role in constraining the current solution. The penalty function method can transform the original constrained optimization problem into an unconstrained optimization problem, which is convenient for solving the optimization problem.

[0127] The penalty function corresponding to formula (19b) is:

[0128]

[0129] The penalty function corresponding to formula (19c) is:

[0130]

[0131] The penalty function corresponding to formula (19d) is:

[0132]

[0133] The total penalty function is:

[0134] P=P1+P2+P3 (23)

[0135] The unconstrained optimization problem processed by the penalty function method is:

[0136]

[0137] Where α is the penalty factor, which takes a large positive number.

[0138] This example uses a particle swarm algorithm to solve the above unconstrained optimization problem:

[0139] ① Set the particle dimension to be equal to the number of decision cycles L included in the procurement planning cycle, and represent the position of each particle as the raw material procurement quantity (Q1, Q2, ..., Q L ), select the function shown in formula (24) as the fitness function of the particle;

[0140] ② Initialize to get a p An initial population of particles, each with a random position and velocity;

[0141] ③For each particle, compare the current fitness with the individual optimal fitness pb; if the current fitness is better than pb, update pb to the current fitness, and the L-dimensional position of pb is also updated to the current position;

[0142] ④ For each particle, compare the current fitness with the historical optimal fitness gb of the group; if the current fitness is better than gb, update gb to the current particle fitness and record the particle number and position;

[0143] ⑤ Update the speed and position of each particle according to the following formula:

[0144] v i =w in v i +c1rand()(pb i -s i )+c2rand()(gb i -s i ) (25)

[0145] s i =s i +v i (26)

[0146] Among them, ω in is the inertia factor, c1 and c2 are learning factors;

[0147] ⑥ Return to ② and continue the iterative loop until a good enough fitness is obtained or the maximum number of iterations is reached.

[0148] 3. Supplier evaluation management

[0149] Supplier evaluation is typically done manually. Relevant managers design questionnaires based on established rules and experience and distribute them to relevant departments for completion. All questionnaire responses are then tallied to determine a final score, and high-quality suppliers are selected based on the scores. However, the subjectivity and knowledge limitations of these judgments significantly impact the final evaluation results. This method requires manual collection and analysis of large amounts of data, resulting in high workload and low efficiency. Supplier delivery records reflect the true state of supply and provide a true reflection of supplier ratings. The Industrial Internet enables the automatic collection and storage of supplier delivery information, and the use of various algorithms for automated calculations, enabling scientific and efficient supplier evaluation and management. Before evaluating suppliers, it is necessary to determine the evaluation criteria—the specific aspects to be considered. However, different industries have different focus areas, making it difficult to propose a universal evaluation standard that applies to all industries. For specific industries, evaluation criteria need to be analyzed and defined based on the specific context.

[0150] Regarding supplier evaluation methods, this paper proposes a supplier evaluation model based on stacked autoencoders (SAE) and bootstrap aggregation. SAE is a deep neural network composed of multiple layers of autoencoders. It can extract complex and abstract features from data in a hierarchical manner, has a strong ability to approximate complex nonlinear functions, and possesses strong modeling capabilities. Using SAE, complex nonlinear relationships can be established between supplier supply data and supplier scores, such as the main grade of the supply, the quantity supplied, and the level of impurities exceeding the standard. Once the model is trained, suppliers can be automatically and efficiently evaluated.

[0151] An autoencoder (AE) is an unsupervised single hidden layer neural network consisting of an encoder and a decoder. Its output layer is the same as the input layer. Its main purpose is to reconstruct the original input as accurately as possible. Its structure is as follows: Figure 6 shown.

[0152] The main purpose of AE is to reconstruct the original input as accurately as possible. The encoder transforms the input I = [i1, i2, ..., i n ] T ∈R n Mapping to hidden layer κ=[κ1,κ2,…,κ m ] T ∈R m .

[0153] κ=f(I)=σ f (AI+b) (26)

[0154] Where A is a weight matrix with dimension m×n, b∈R mis the bias vector of the hidden layer, σ f represents a non-linear activation function.

[0155] The decoder transforms the hidden layer κ to the output layer through the mapping function g

[0156]

[0157] Where, is a weight matrix of dimension n×m, is the bias vector of the output layer, is a nonlinear activation function, and κ can be considered as a feature extracted from the input data. The autoencoder uses the backpropagation algorithm to learn a mapping After reconstruction As close as possible to the original input I. The parameters of the model can be obtained by minimizing the average reconstruction error.

[0158]

[0159] N is the number of samples. L is the loss function, which can be the traditional root mean square error It can also be the reconstruction cross entropy.

[0160] Stacked Autoencoder SAE is a layered deep neural network composed of multiple AE layers connected together. Its structure and training process are as follows: Figure 7 As shown in Figure 2, a self-encoder (SAE) is constructed through unsupervised pre-training of individual autoencoders (AEs) and supervised fine-tuning of the overall architecture. During the pre-training phase, the original training data is used as input for training to obtain the hidden layer representation of the first AE. The hidden layer representation of the first AE is then used as input for training the second AE to obtain the hidden layer representation of the second AE. This process continues, pre-training each autoencoder layer by layer. After unsupervised pre-training, the input layer and the hidden layers of each AE are stacked. A supervised model (e.g., a FNN) is added on top of the last AE as the output, and the hidden layer representation of the last AE is used as the input of this model. Finally, a complete SAE is established. Furthermore, all parameters of the SAE are fine-tuned using a training algorithm (e.g., gradient descent). By training a multi-layer neural network, the SAE achieves a mapping between basic supplier data and their ratings. In supplier evaluation management, a dataset is constructed using the evaluation indicator values ​​and comprehensive ratings of different suppliers. The model is trained using the supplier's evaluation indicator values ​​as input and the final supplier rating as the output. After the model is trained, the supplier's supply data is fed into the model to automatically calculate the supplier's score. By sorting the comprehensive scores of all suppliers within a specific time period, it is possible to screen out raw material suppliers with higher overall quality.

[0161] In real-world supplier evaluation scenarios, some companies may conduct quarterly or annual supplier evaluations, which can mean insufficient training samples. Considering this, bootstrap aggregation (bagging) is used to improve the performance of SAE-based supplier evaluation models. Bootstrap aggregation (bagging) is an ensemble learning algorithm. It randomly samples raw data to obtain multiple datasets, trains multiple models separately, and aggregates these models to achieve more robust results.

[0162] The supplier automatic evaluation model based on stacked autoencoders and bagging method ensemble learning in this embodiment is as follows: Figure 8 As shown, the process of constructing and using the evaluation method for suppliers is as follows:

[0163] ①Collect historical data of suppliers These include Each supplier has n indicator data for supplier evaluation; experts evaluate suppliers based on each indicator data and obtain the historical score of each supplier; the original historical data and scores to form a new dataset

[0164] ②Based on the bagging method, from the data set Randomly select samples with replacement to form a sub-dataset;

[0165] ③ Repeat ② to get N D Independent sub-datasets;

[0166] ④ In each sub-dataset, the n-dimensional evaluation index data of each supplier is used as input, and the supplier score is used as output. The stacked autoencoder model with n-dimensional input and 1-dimensional output is trained to obtain the mapping relationship between the supplier evaluation index data and the supplier score, that is, score = Λ d (index1,index2,…,index n ),d=1,2,…,N D ;

[0167] ⑤Repeat ④ to get N D An independent evaluation model based on stacked autoencoders;

[0168] ⑥ When new supplier n-dimensional indicator data is input into the stacked autoencoder model, N D The stacked autoencoder model correspondingly obtains N D scoring results, and then take N DThe average of the rating results is used as the final rating of the supplier, namely:

[0169]

[0170] By establishing an integrated model that is more robust than a single SAE model, the method of this embodiment can improve the accuracy and stability of the evaluation results.

[0171] 4. Supplier order allocation

[0172] Order allocation involves determining the raw material purchase quantity for each supplier based on specific targets. In most companies, order allocation is primarily the responsibility of the procurement department. Due to the isolation between departments, procurement departments may focus solely on reducing costs while ignoring the differences in raw material quality and supply stability among suppliers with different ratings. This can have a significant impact on the production department. Furthermore, with a large number of suppliers, the procurement process requires a vast amount of information to consider. Faced with this overwhelming amount of information, it is difficult for decision-makers to synthesize and make rational decisions. Furthermore, manual decision-making is easily influenced by the decision-maker's personal preferences and relationships with suppliers. Therefore, in practice, manual decision-making can easily lead to problems such as high procurement costs and unstable raw material quality. With the Industrial Internet, relevant supplier information can be automatically collected and processed, and intelligent algorithms can be applied to objectively make decisions. Adopting a scientific and rational order allocation method can help improve the stability and quality of raw material supply and reduce procurement costs.

[0173] This paper aims to minimize procurement costs and maximize overall utility, taking into account constraints such as inventory, supply capacity, and demand, and establishes an order allocation model. The relevant explanations and assumptions are as follows:

[0174] ① Raw materials can be divided into N g There are different grades, and the prices of different grades are different.

[0175] ②There are N s There are raw material suppliers. Different suppliers have different prices and supply capacities for different grades of raw materials.

[0176] ③ The presence of k enriched ingredients in the raw materials can bring additional benefits. The content of enriched ingredients varies from supplier to supplier and grade to grade.

[0177] ④The price of the same rich ingredient is the same for different suppliers.

[0178] ⑤Order allocation only applies to the current decision cycle.

[0179] (1) Objective function:

[0180] Procurement cost is the difference between the purchase cost of raw materials and their recovery value, and should be minimized.

[0181]

[0182] where a ij is the purchase quantity of raw materials of grade j from supplier i, p ij is the price of the jth grade raw material from the i-th supplier, e ij The additional revenue obtained from purchasing the j-th grade raw material from the i-th supplier per ton.

[0183] Overall utility: The overall utility of all suppliers should be maximized, with the goal of purchasing from suppliers with high scores as much as possible.

[0184]

[0185] in is the nth indicator data of the i-th supplier.

[0186] (2) Constraints

[0187] Supply capacity constraint: order quantity a ij The supply capacity of the jth grade raw material of the i-th supplier should not exceed r ij .

[0188] a ij ≤r ij ,i=1,2,…,N s ,j=1,2,…,N g (31)

[0189] Demand constraint: The sum of the main components in the raw materials should meet the requirements of the production plan.

[0190]

[0191] where θ ij The main component content of the raw material.

[0192] Quantity constraint: The total order quantity should meet the purchase plan G and cannot be too much. On the one hand, the total order quantity should meet the purchase plan total quantity; on the other hand, the total order quantity cannot be too large and cannot exceed 2% of the purchase plan total quantity. Therefore

[0193]

[0194] Inventory constraint: The total order quantity should not exceed the remaining capacity S of the warehouse.

[0195]

[0196] (3) Optimization method

[0197] The order allocation model described above is a typical multi-objective constrained optimization problem (CMOP). This paper transforms the multi-objective problem into a single-objective problem using function transformation and linear weighting. First, each optimization objective is scaled so that the values ​​of each objective are on the same order of magnitude, as shown in the following equation:

[0198]

[0199] where f i * (a) is the optimal value obtained by solving the i-th optimization problem by considering the constraints separately.

[0200] Then, different objective functions are assigned weights ξ i The higher the weight, the more important the objective function is.

[0201] Finally, through linear combination, the compromise single-objective optimization problem is obtained:

[0202]

[0203] Through the above processing, the multi-objective constrained optimization problem is greatly simplified, so that the original problem can be solved by conventional optimization methods.

[0204] This example uses a particle swarm optimization (PSO) algorithm to solve the function transformation and the final optimization problem, and obtains the optimal raw material order quantity for each supplier:

[0205] Step 1: Taking f1(a) as the objective function and equations (31) to (34) as constraints, the constrained optimization problem is transformed into an unconstrained optimization problem according to the penalty function method described in the procurement decision optimization process. Then, PSO is used to solve the problem and obtain the optimal value f1 corresponding to the optimal solution. * (a);

[0206] Step 2: Taking f2(a) as the objective function and equations (31) to (34) as constraints, the constrained optimization problem is transformed into an unconstrained optimization problem according to the penalty function method described in the procurement decision optimization process. Then, PSO is used to solve the problem and obtain the optimal value corresponding to the optimal solution.

[0207] Step 3: According to equations (35) and (36), the multi-objective optimization problem is transformed into a single-objective optimization problem (with equations (31) to (34) as constraints);

[0208] Step 4: According to the penalty function method described in the procurement decision optimization process, the constrained optimization problem is transformed into an unconstrained optimization problem, and finally the PSO solution is used to obtain the optimal solution.

[0209] When PSO is used to solve the above Step 1, Step 2 and Step 4, the position of each particle is represented by the raw material purchase quantity of all grades of all suppliers {a ij ; i=1,…,N s ; j = 1,…,N g}, the optimal value is the maximum objective function value, and the optimal solution is the particle with the largest objective function value.

[0210] In the above method, the data required for market price forecasting, raw material procurement decisions, supplier evaluation management and supplier order allocation can all be obtained from the industrial Internet platform, and the above processes can all be completed automatically on the industrial Internet platform, completing integrated decision-making from the perspective of the entire raw material supply chain process, overcoming the problems of high procurement costs and low efficiency caused by information isolation and subjective cognition under manual decision-making, effectively reducing raw material procurement costs and improving supply chain management efficiency.

[0211] The above embodiments are preferred embodiments of the present application. Ordinary technicians in this field can also make various changes or improvements on this basis. Without departing from the overall concept of the present application, these changes or improvements should fall within the scope of protection required by the present application.

Claims

1. A procurement supply chain integration optimization method based on the industrial Internet, characterized in that: include: Market price forecast: forecast the market price of raw materials; Raw material procurement decision-making: Considering the market forecast price of raw materials, a multi-period dynamic procurement model is established to minimize the unit raw material cost, and the optimal raw material procurement quantity for each procurement cycle is optimized; Supplier evaluation: We obtain historical supply data from each supplier and build an automatic supplier evaluation model based on stacked autoencoders and bagging ensemble learning to obtain a score for each supplier. The specific process of supplier evaluation is as follows: Step C1: Collect historical data of suppliers These include Each supplier has n indicator data for supplier evaluation; experts evaluate suppliers based on each indicator data and obtain the historical score of each supplier; the original historical data and scores to form a new dataset Step C2, based on the bagging method, Randomly select samples with replacement to form a sub-dataset; Step C3, repeat step C2 to get N D Independent sub-datasets; Step C4: In each sub-dataset, the n-dimensional evaluation index data of each supplier is used as input, and the supplier's score is used as output. A stacked autoencoder model with n-dimensional input and 1-dimensional output is trained to obtain a mapping relationship between the supplier evaluation index data and the supplier score. Step C5, repeat step C4 to obtain N D An independent evaluation model based on stacked autoencoders; Step C6, when the new supplier n-dimensional indicator data is input into the stacked autoencoder model, N D The stacked autoencoder model correspondingly obtains N D scoring results, and then take N D The average of the rating results is used as the final rating of the supplier; Supplier order allocation: Taking into account the optimal raw material purchase volume and supplier ratings, a multi-objective optimization model is established to minimize procurement costs and maximize the overall supplier effect, and the optimal procurement strategy from each supplier is obtained. The data required for the market price forecast, raw material procurement decision, supplier evaluation and supplier order allocation are automatically obtained through the industrial Internet platform.

2. The method according to claim 1, characterized in that Variational mode decomposition and long short-term memory network are used to predict the market price of raw materials. Specifically: Step A1: Obtain the market price sequence f(t) of the raw material and decompose it into K modal components u using the variational mode decomposition method. i (t),i=1,2,…K; Step A2: For each modal component u, i (t) Split to obtain input and output for model training, which are then used to train LSTM to obtain the prediction model corresponding to each component; In step A3, the most recent market price sequence is obtained, decomposed into K modal components according to step A1, and rolling forecasts are performed using various forecast models. Finally, the K forecast sequences are summed to obtain the market forecast price sequence of the raw material.

3. The method according to claim 2, characterized in that The method for determining the number of variational mode decompositions K is to gradually increase the number of decompositions and calculate the corresponding residuals. When the residual is less than the preset value and there is no obvious downward trend, the current number of decompositions is determined to be the optimal number of decompositions K. The residual calculation formula is: Where r res is the residual, and M is the number of training samples obtained by splitting f(t).

4. The method according to claim 1, wherein The multi-period dynamic procurement model established in the raw material procurement decision is: Among them, C is the raw material cost of the procurement planning cycle, including the procurement cost C p and inventory cost C w The duration of the procurement planning cycle T consists of L separate decision cycles, and the duration of the kth decision cycle is T k , the raw material purchasing point t in the kth decision cycle k At the midpoint of the current decision cycle; Q k is the total raw material purchase plan for the kth decision cycle; represents the prediction function of the i-th modal component; Purchase cost C p Including raw material costs and handling fees, P k is the market price forecast obtained in the kth decision cycle; C S is the handling fee for each decision cycle; Inventory costs R0 is the initial raw material inventory of the procurement planning cycle; D k T k The production plan corresponds to the raw material demand D k / η, η is the production conversion rate; C I is the annual loan interest rate, C M is the daily inventory management cost per unit of raw material; B k Indicates that in the procurement cycle T k Capital budget for raw material procurement, ss is the safety stock, and V is the maximum inventory capacity of the warehouse.

5. The method according to claim 4, characterized in that The particle swarm optimization algorithm is used to solve the multi-period dynamic procurement model, and the penalty function is used to deal with the constraints in the procurement model during the solution. Specifically: Step B1: Convert the constraints (19b), (19c) and (19d) into the following penalty functions P1, P2 and P3 respectively, and add them together to obtain the total penalty function P: P=P1+P2+P3 (23) Step B2: According to the penalty function P, the multi-period dynamic procurement model is converted into an unconstrained optimization problem: Where α is the penalty factor; Step B3, initialize a particle population, and select the function shown in formula (24) as the fitness function of the particle; wherein the particle dimension is equal to the number of decision cycles L included in the procurement plan cycle, and the position of each particle is represented by the raw material purchase quantity (Q1, Q2, ..., Q L ); Step B4: Particle swarm optimization is used to find the optimal particle, and the optimal raw material purchase quantity for each decision cycle within the purchase plan cycle is obtained based on the position of the optimal particle.

6. The method according to claim 1, wherein Minimizing procurement costs means minimizing the difference between raw material procurement costs and recycling value, which is expressed as the objective function f1(a): Where a ij is the purchase quantity of raw materials of grade j from supplier i, p ij is the price of the jth grade raw material from the i-th supplier, e ij N is the additional revenue per ton of raw materials of grade j from supplier i; s is the number of raw material suppliers, N g The number of grades of raw materials; Maximizing the comprehensive supplier effect means maximizing the overall utility of all suppliers. The goal is to purchase from suppliers with high scores as much as possible, which is expressed as the objective function f2(a): Where, is the n indicator data of the i-th supplier, N is the score of the dth stacked autoencoder for the i-th supplier in the bagging method; D The number of stacked autoencoders to use for bagging.

7. The method according to claim 6, characterized in that The constraints in establishing the multi-objective optimization model include supply capacity constraints, demand constraints, order quantity constraints and inventory constraints; The supply capacity constraint refers to the raw material procurement quantity a ij The supply capacity of the jth grade raw material of the i-th supplier should not exceed r ij , expressed as: a ij ≤r ij ,i=1,2,…,N s ,j=1,2,…,N g (31) The demand constraint means that the sum of the main components in the raw materials should meet the production demand, which is expressed as: Among them, θ ij The main component content of the raw material; The quantity constraint means that the total order purchase quantity must meet the purchase plan G and not exceed the preset proportion of the total purchase plan quantity; The inventory constraint means that the total order purchase quantity should not exceed the remaining capacity of the warehouse.

8. The method according to claim 6, characterized in that When solving the multi-objective optimization model for supplier order allocation, the multi-objective problem is transformed into a single-objective problem using function transformation and linear weighting method. Specifically: First, each optimization objective is scaled so that the values ​​of each objective after the transformation are in the same order of magnitude, as shown in the following formula: where f i * (a) Solve the i-th objective function f by considering the constraints separately i (a) The optimal value obtained; Then assign weights ξ to different objective functions i ; Finally, the compromise single-objective optimization problem is obtained through linear combination:

9. An electronic device comprising a memory and a processor, wherein the memory stores a computer program, wherein: When the computer program is executed by the processor, the processor is caused to implement the method according to any one of claims 1 to 8.

Images