CN115619033A

CN115619033A - Industrial internet-based purchasing supply chain integrated optimization method and equipment

Info

Publication number: CN115619033A
Application number: CN202211361346.5A
Authority: CN
Inventors: 阳春华; 刘一顺; 黄科科; 李勇刚; 桂卫华
Original assignee: Central South University
Current assignee: Central South University
Priority date: 2022-11-02
Filing date: 2022-11-02
Publication date: 2023-01-17

Abstract

The invention discloses a purchasing supply chain integrated optimization method and device based on an industrial internet, wherein the method comprises the following steps: predicting the market price of the raw materials; considering the market predicted price of the raw materials, establishing a multi-period dynamic purchasing model with the minimum unit raw material cost, and optimally solving the optimal raw material purchasing quantity of each purchasing period; acquiring historical supply data of each provider, constructing an automatic evaluation model of the provider based on stack autoencoder and bagging ensemble learning, and acquiring scores of each provider; and comprehensively considering the optimal raw material purchasing quantity and the grade of the suppliers, establishing a multi-objective optimization model for minimizing purchasing cost and maximizing the comprehensive effect of the suppliers, and solving to obtain the optimal purchasing strategy of each supplier. According to the invention, the data of the whole process of the purchasing supply chain is comprehensively acquired, processed and analyzed in a centralized manner through the industrial internet platform, and the relevance among different services is comprehensively considered, so that the purchasing cost can be effectively reduced, and the purchasing efficiency is improved.

Description

Industrial Internet-based integrated optimization method and equipment for purchasing supply chain

Technical Field

The invention belongs to the field of purchasing supply chains, and particularly relates to a purchasing supply chain integrated optimization method based on an industrial internet.

Background

With the change of new generation information technologies such as industrial networks, big data, artificial intelligence, etc., intelligent manufacturing has become a common theme for the development of manufacturing industry. As a key infrastructure for intelligent manufacturing and intelligent application, the industrial Internet establishes a basic network for connecting devices/equipment, materials, people and information systems through comprehensive interconnection of people, machines and objects, realizes comprehensive perception, dynamic transmission and real-time analysis of industrial data, forms scientific decision and intelligent control, improves manufacturing resource allocation efficiency, is considered to be one of the most important technical fields in the future, and is widely concerned by various industries.

From an application perspective, purchasing supply chain optimization is one of the important areas where industry internet can bring huge promotion and benefits. In industrial manufacturing, raw materials are the source of the production and processing process, and the quality and the supply continuity of the raw materials directly influence the running stability of the production and processing; on the other hand, in smelting processing enterprises, the raw material purchase cost accounts for about 70% of the production and processing cost, and the capital occupation is huge. The optimization of the purchase supply chain can greatly reduce the production cost and ensure the quality of raw materials, and has very important significance for improving the comprehensive benefits of enterprises.

The purchasing supply chain mainly comprises the processes of market price analysis, purchasing total quantity decision, supplier evaluation, supplier order distribution and the like, and at present, a great number of scholars perform a great deal of research on each part. Although existing work has worked well in the corresponding field, it is essential that the work only addresses local problems of the purchasing supply chain, i.e. only the price prediction problem or the order allocation problem is studied. In most of actual enterprises, different purchasing business processes are usually completed by different departments or different personnel, the isolated departments are easy to cause information isolated islands, the mode that the personnel are in the rings also causes great delay, the information transmission efficiency is low, and the decision effect is poor. Further, the manual decision making has strong subjectivity and great randomness, and cannot give comprehensive decision in the face of massive information, so that the actual raw material purchasing cost is high and the efficiency is low. Therefore, the conventional supply chain has been unable to meet the demand of smart manufacturing for agility and high quality, and the need for the evolution into a digital supply chain with high intelligence is urgent.

The industrial internet provides a collaborative and efficient application platform for the supply chain, and the operation mode of the supply chain can be remodeled. Based on the industrial internet, the information of different systems, different departments and different granularities in the supply chain can be integrated, the business activity of the whole supply chain is considered, and comprehensive and reasonable decisions are dynamically made in real time. Benefiting from the development of the industrial internet, the system can integrate multi-source data in the horizontal and vertical fields, and provide a powerful platform to complete the functions of large data analysis processing, complex modeling optimization and the like. This brings new opportunities for agile and efficient operation of the procurement supply chain. Compare in traditional purchase supply chain, the labour can greatly be liberated to its automatic processing mode of purchase supply chain based on industrial internet, and mode and the application of intelligent method that comprehensively orchestrates can obtain more excellent decision-making, reduces purchasing cost, promotes operation efficiency. All in all, the industrial internet brings new opportunities for the agile and efficient operation of the procurement supply chain.

Disclosure of Invention

The invention provides an industrial internet-based integrated optimization method for a purchasing supply chain, which is used for comprehensively acquiring, centrally processing and analyzing the whole flow data of the purchasing supply chain through an industrial internet platform, comprehensively considering the relevance among different services, effectively reducing the purchasing cost and improving the purchasing efficiency.

In order to achieve the technical purpose, the invention adopts the following technical scheme:

an industrial Internet-based integrated optimization method for a purchasing supply chain comprises the following steps:

and (3) market price prediction: predicting the market price of the raw materials;

raw material purchasing decision: considering the market predicted price of the raw materials, establishing a multi-period dynamic purchasing model with the minimum unit raw material cost, and optimally solving the optimal raw material purchasing quantity of each purchasing period;

evaluation by the supplier: acquiring historical goods supply data of each provider, constructing an automatic evaluation model of the provider based on stack self-encoder and bag-sleeving integrated learning, and acquiring grades of each provider;

supplier order distribution: comprehensively considering the optimal raw material purchasing quantity and the grade of the suppliers, establishing a multi-objective optimization model for minimizing purchasing cost and maximizing the comprehensive effect of the suppliers, and solving to obtain the optimal purchasing strategy of each supplier;

the market price prediction, the raw material purchasing decision, the supplier evaluation and the supplier order distribution are carried out, and the required data are automatically obtained through an industrial internet platform.

Further, the market price of the raw materials is predicted by adopting variational modal decomposition and a long-short term memory network, and the method specifically comprises the following steps:

a1, obtaining a market price sequence f (t) of raw materials, and decomposing the market price sequence into K modal components u by adopting a variational modal decomposition method _i (t),i＝1,2,…K；

Step A2, according to step length, each modal component u _i (t) splitting to obtain input and output for model training, and further training the LSTM to obtain a prediction model corresponding to each component;

and step A3, acquiring a current and latest market price sequence, decomposing the current and latest market price sequence into K modal components according to the step A1, performing rolling prediction by using each prediction model, and finally summing the K prediction sequences to obtain a market prediction price sequence of the raw materials.

Further, the method for determining the number K of the variational modal decomposition comprises the following steps: gradually increasing the decomposition number and calculating corresponding residual errors, and when the residual errors are smaller than a preset value and have no obvious descending trend, determining the current decomposition number as the optimal decomposition number K; the residual error is calculated as:

in the formula, r _res For residual, M is the number of training samples into which f (t) is split.

Further, the multi-period dynamic purchasing model established in the raw material purchasing decision is as follows:

k＝1,2,3,...,L(19e)

wherein C is the raw material cost of the purchase planning period, including the purchase cost C _p And inventory cost C _w (ii) a The duration T of the purchase planning period comprises L independent decision periods, and the duration of the kth decision period is T _k Material procurement point t of kth decision cycle _k At the midpoint of the current decision cycle; q _k Planning the total amount of raw material purchasing in the kth decision period;

procurement cost C _p Including the raw material costs and the procedural costs,

P _k is the market forecast price obtained by the k decision period of the market price forecast; c _S Is the procedure cost for each decision cycle;

cost of inventory

R ₀ The initial raw material inventory for the purchase planning period; d _k Is T _k The corresponding raw material demand is D _k Eta, eta is the production conversion rate; c _I Is annual loan interest rate, C _M Daily inventory management cost per unit of raw material;

B _k is shown in the procurement period T _k The capital budget for raw material procurement is carried out, ss is the safe stock quantity, and V is the maximum stock capacity of the warehouse.

Further, a particle swarm optimization algorithm is adopted to solve the multi-period dynamic purchasing model, and a penalty function is used to process a constraint item in the purchasing model during solving, specifically:

step B1, converting the constraints (19B), (19 c) and (19 d) into the following penalty functions P ₁ 、P ₂ And P ₃ The sum yields the total penalty function P:

P＝P ₁ +P ₂ +P ₃ (23)

step B2, converting the multi-period dynamic purchasing model into an unconstrained optimization problem according to the penalty function P, wherein the unconstrained optimization problem is as follows:

in the formula, alpha is a penalty factor;

step B3, initializing a particle population, selecting the function shown in formula (24) as the fitness of the particleA responsiveness function; wherein, the dimension of the particle is equal to the number L of decision periods included in the procurement planning period, and the position of each particle is expressed as the procurement quantity (Q) of the raw materials in the L decision periods in the procurement planning period ₁ ,Q ₂ ,…,Q _L )；

And B4, optimizing and solving the optimal particles by adopting a particle swarm optimization algorithm, and obtaining the optimal raw material purchasing quantity of each decision period in the purchasing plan period according to the position of the optimal particles.

Further, the specific process of supplier evaluation is as follows:

step C1, collecting historical data of suppliers

Which comprises

N suppliers are used for evaluating the index data of the suppliers; evaluating the suppliers by experts according to the index data to obtain the historical scores of the suppliers; raw historical data

And scoring to form a new data set

Step C2, based on the bagging method, from the data set

In the random selection

Carrying out substitution on the samples to form a subdata set;

step C3, repeating the step C2 to obtain N _D Mutually independent subdata sets;

step C4, in each subdata set, taking n-dimensional evaluation index data of each supplier as input, taking the grade of the supplier as output, training a stack self-encoder model comprising n-dimensional input and 1-dimensional output, and obtaining the mapping relation between the evaluation index data of the supplier and the grade of the supplier;

step C5, repeating the step C4 to obtain N _D The evaluation models are independent from each other and based on a stack self-encoder;

step C6, when new supplier N-dimensional index data is input into the stack self-encoder model, N _D Each stack self-encoder model correspondingly obtains N _D The result of scoring, then take N _D The average of the individual scoring results is taken as the final score for that vendor.

Further, minimizing the procurement cost means that the difference between the procurement cost and the recovery value of the raw material is minimized, and is expressed as an objective function f ₁ (a)：

In the formula, a _ij Raw material purchase quantity, p, for jth grade of ith supplier _ij Price of material for jth grade of ith supplier, e _ij Additional revenue per ton obtained for purchasing an ith supplier's jth grade of feedstock; n is a radical of _s Number of raw Material suppliers, N _g The grade number of the raw materials;

maximizing the combined effect of suppliers means maximizing the overall utility of all suppliers, and the goal is to purchase suppliers with high scores as much as possible, expressed as an objective function f ₂ (a)：

In the formula (I), the compound is shown in the specification,

n index data for the ith supplier,

for bagging methodScoring of the ith supplier by d stacked autoencoders; n is a radical of _D The number of stacked self-encoders used for the bagging method.

Further, the constraint conditions in the process of establishing the multi-objective optimization model comprise supply capacity constraint, demand constraint, order quantity constraint and inventory constraint;

the supply capacity constraint refers to the raw material purchasing amount a _ij Should not exceed the supply capacity r of the jth grade of raw material of the ith supplier _ij Expressed as:

a _ij ≤r _ij ,i＝1,2,…,N _s ,j＝1,2,…,N _g (31)

the requirement constraint means that the sum of the main components in the raw materials should meet the production requirement, and is expressed as:

wherein, theta _ij Is the main component content of the raw material;

the quantity constraint means that the total purchase quantity of the order should meet the purchase plan G and does not exceed the preset proportion of the total quantity of the purchase plan;

the inventory constraint means that the total purchase amount of the order does not exceed the residual capacity of the warehouse.

Further, when the supplier order distribution solves the multi-objective optimization model, the multi-objective problem is converted into the single-objective problem by using a function transformation and linear weighting method, specifically:

firstly, carrying out scale transformation on each optimization target to enable the value of each target after transformation to be in the same order of magnitude, as shown in the following formula:

wherein f is _i ^* (a) Is to solve the ith objective function f by respectively considering the constraint conditions _i (a) The obtained optimal value;

then to different targetsFunction weighting xi _i ；

And finally, obtaining a compromise single-target optimization problem through linear combination:

an electronic device, comprising a memory and a processor, wherein the memory stores a computer program, and when the computer program is executed by the processor, the processor is enabled to implement any one of the above-mentioned methods for optimizing industrial internet-based procurement and supply chain integration.

Compared with the prior art, the invention has the following beneficial effects:

(1) The method comprehensively considers factors such as market price change, production plan, inventory condition, capital limitation and the like, establishes the multi-period dynamic purchasing model with the minimum unit cost of raw materials, makes more comprehensive and fine decisions in the long term, solves the problems of local information and extensive mode of the traditional mode, and effectively reduces purchasing cost.

(2) The invention comprehensively considers each index of supplier supply, excavates the incidence relation of each index to the grade of the supplier, automatically gives the comprehensive grade of each supplier, and solves the problems of the low efficiency and the non-objective evaluation management of the supplier caused by the strong subjectivity and the large workload of the artificial evaluation supplier.

(3) The invention starts from the whole process of the raw material supply chain business, analyzes the incidence relation among different businesses, organically connects the different businesses, and automatically completes the raw material purchasing decision from the global view, thereby solving the problems of low decision efficiency and poor effect caused by serious information isolated island and isolated business in the traditional mode.

(4) Aiming at the problem that the mode quantity in the variable mode decomposition VMD can not be reasonably determined, the invention provides a mode quantity determination method based on decomposition residual errors, which effectively avoids the problems of under-decomposition and over-decomposition, and provides a sequence with more vivid characteristics and more gradual change for a subsequent predictor, thereby improving the comprehensive prediction effect.

Drawings

FIG. 1 is an industrial Internet-based feedstock supply chain optimization overall architecture;

FIG. 2 is an overall architecture of a method according to an embodiment of the present application;

FIG. 3 is a diagram of a VMD-LSTM model according to an embodiment of the present application;

FIG. 4 is a schematic diagram of a rolling procurement process in the method according to an embodiment of the application;

FIG. 5 is a schematic illustration of a raw material inventory change in a method according to an embodiment of the present application;

FIG. 6 is an auto encoder configuration;

FIG. 7 is a diagram of the structure and training process of SAE;

fig. 8 is a schematic diagram of a vendor evaluation model based on Bagging and SAE.

Detailed Description

The following describes embodiments of the present invention in detail, which are developed based on the technical solutions of the present invention, and give detailed implementation manners and specific operation procedures to further explain the technical solutions of the present invention.

In the embodiment, a four-layer architecture proposed in the white paper of industrial internet of the china industrial internet industrial alliance is adopted, and the architecture comprises an edge layer, a resource layer, a platform layer and an application layer, and the architecture clearly defines the structure and the function of each unit of an industrial internet system and is generally accepted. In the patent, the overall architecture of the industrial internet-based raw material supply chain integrated optimization is shown in fig. 1. Firstly, acquiring internal and external data such as various operation decision data, production process data, raw material assay data, market change data and the like based on an industrial internet platform, and uniformly storing the data in a large data platform data warehouse in a centralized manner; then, carrying out data governance work such as abnormal value detection, missing value restoration and the like on the data, and data fusion work such as multi-table aggregation and the like; then developing an intelligent decision algorithm facing to a raw material supply chain in an integrated development environment of an industrial internet platform, and carrying out standardized packaging and lightweight API deployment on the algorithm; and finally, calling the algorithm model through an API (application programming interface) during application development, and developing to form C/S and B/S multi-type industrial APPs.

The integrated optimization method for the industrial internet-based purchasing and supply chain mainly comprises four parts, namely market price prediction, purchasing optimization decision, supplier evaluation and order allocation, and the overall technical architecture is shown in fig. 2. Aiming at the severe fluctuation of market price, the market price is predicted by adopting variational modal decomposition and a long-short term memory network; then, aiming at the problems of short planning period, local information and the like of the traditional purchasing mode, comprehensively considering market forecast price, inventory information, production plan and the like, establishing a multi-period dynamic purchasing model with the minimum unit raw material cost, and performing optimization solution by using a particle swarm algorithm to obtain the optimal raw material purchasing quantity in each purchasing period; then, aiming at the problems of strong subjectivity, low efficiency and the like of a manual evaluation supplier, based on the data of raw material assay information, contract information, weight checking logs and the like automatically acquired by an industrial internet platform, a supplier automatic evaluation model based on stack self-encoder and bag-sleeving method integrated learning is provided, and the efficient evaluation of the supplier is realized; and finally, aiming at the problems that when the order distribution is manually completed, the personal preference influence is large and the large-scale information cannot be adapted to, the grading of suppliers, the supply capacity of the suppliers, the raw material grade and the like are comprehensively considered, a multi-objective optimization model which minimizes the purchasing cost and maximizes the comprehensive utility of the suppliers is established, and the optimal purchasing strategy of each supplier is obtained. The industrial internet is capable of automatically collecting and analyzing data throughout the procurement supply chain. The intelligent method can be organically integrated in consideration of the relationship among different business tasks, and complete comprehensive real-time decision from the integral perspective.

1. Market price forecasting

In zinc smelting enterprises, the raw material cost usually accounts for about 70% of the total production cost. If the price of the raw material can be accurately predicted and analyzed, the purchasing time and the purchasing quantity of the raw material can be guided, and the purchasing cost is reduced. However, in the real world, market prices are influenced by various factors such as global economy, supply-demand relationship, exchange rate and the like, and coupling relevance between different factors is strong, so that the market prices fluctuate frequently, change greatly, and have obvious nonlinear and unsteady characteristics. Conventional methods and manual analysis often fail to achieve good results. Therefore, accurate and robust price forecasting techniques are critical to the procurement and scheduling of enterprises. In the embodiment, a price prediction method based on a decomposition-integration frame is adopted, so that the short-term accurate prediction of the price of the raw material is realized.

(1) Variational modal decomposition

Variational modal decomposition is an advanced multi-resolution technique for adaptive non-recursive signal decomposition. The method converts the signal decomposition process into the construction and solution of the variation optimization problem. It can decompose a true value signal into a series of Intrinsic Mode Functions (IMF) u with bandwidth limitation _k . Suppose each IMFu _k Are all centered on a central frequency ω _k Nearby, the center frequency is determined by a decomposition process.

(2) Long and short term memory network: long Short Term Memory (LSTM) networks are an extended version of Recurrent Neural Networks (RNNs) and are widely used in many areas. In contrast to recurrent neural networks, long-short term memory networks contain not only external circulation but also internal circulation of memory cells. Unlike traditional neural networks, LSTM uses memory blocks to maintain the internal state of the network. The method can explore internal abstract features and an underlying structure of data, and has strong prediction capability on time series. Currently, LSTM has been widely used in various time series processing tasks.

(3) Mixed prediction method based on VMD and LSTM

In view of the advantages of VMD and LSTM, the present embodiment combines the VMD method and LSTM model to create a hybrid prediction model based on the idea of "decomposition-integration", which is referred to herein as VMD-LSTM. Assuming f (t) is the original time sequence, the VMD decomposes the original sequence into a plurality of independent components u _i (t), i =1,2, \ 8230, K, where K is the total number of components. The decomposition is helpful for accurately modeling the internal features and improving the prediction accuracy. LSTM is used as the prediction core of the prediction engine to model and predict the components obtained by decomposition. A schematic diagram of the VMD-LSTM model is shown in FIG. 3.

In the market price prediction part of this embodiment, the specific process of predicting the market price of the raw material by using variational modal decomposition and long-short term memory network is as follows:

Before the metamorphic modal decomposition is carried out, the number K of independent components needs to be set in advance. If the K value is set to be too small, the original sequence may not be decomposed sufficiently, and the complex subsequence still cannot obtain a satisfactory prediction result; if the K value is set to be too large, the original signal will be excessively decomposed, the difference between the sequences will become small, the change rule of different characteristics cannot be reflected, the prediction accuracy is also reduced, and unnecessary calculation consumption is brought. In order to identify the optimal number of independent components K, the present embodiment proposes a method for decomposing residual r based on the original sequence _res The VMD decomposition mode number determination method of (1).

Where M is the number of samples, i.e., the number of training samples resulting from splitting f (t) into input and output by step size. Is obvious of r _res As the residual error of the original sequence decomposition, when the residual error is small to a certain degree, the original sequence can be considered to have been sufficiently decomposed, and each independent component can basically reconstruct the original sequence; when it does not have a significant downward trend, it can be considered that the number of independent components has reached saturation, and no extra independent components are needed. Therefore, the residual r is empirically decomposed _res Below 1% and without a significant downward trend, it is considered that the optimum number of components is currently obtained.

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 and more regular, the change mode is more obvious, and the burden of a market price prediction model is greatly reduced.

Step A2, according to the step length, aiming at each modal component u _i (t) splitting to obtain input and output for model training, and further training LSTM to obtain prediction model corresponding to each component

And A3, acquiring the current latest market price sequence, decomposing the current latest market price sequence into K modal components according to the step A1, performing rolling prediction by using each prediction model, and finally summing the K prediction sequences to obtain the market prediction price sequence of the raw materials.

And adding the predicted values of each subsequence to obtain a final predicted result. Therefore, the present embodiment significantly improves the accuracy of price prediction by combining the advantages of VMD and LSTM.

In addition, for the LSTM neural network, model hyper-parameters need to be preset in the practical application process, but the hyper-parameters are more, and the search space is larger. The traditional manual searching and grid searching method is difficult to traverse the parameter space and consumes long time. Therefore, after the basic range of the hyper-parameter is manually determined empirically, the embodiment adopts a random search method to refine and determine the hyper-parameter so as to obtain a better modeling result.

2. Procurement decision optimization

Currently, most enterprises assume that future changes of prices and demands of raw materials are not considered when making purchasing decisions, or the future prices and future demands of the raw materials are fixed values, so that only the cost of the current purchasing period is considered. The purchasing mode is too short to deal with market changes, high-price stockings are easy to be caused, and the purchasing cost is too high. Therefore, the patent establishes a multi-period dynamic purchasing model considering the price and demand change of the future raw materials, thereby realizing dynamic and flexible purchasing.

(1) Purchasing model

The procurement strategy is shown in figure 4.

When purchasing for the first time, making an overall decision on a purchasing plan period with the duration of T, wherein the overall decision comprises L independent purchasing decision periods T _k The inventory variation is shown in FIG. 5, where R ₀ For the initial stock of raw materials of the purchase planning period T, ss is the safety stock, Q _k Planning total quantity of material procurement for the kth procurement period, t _k The material purchasing point of the kth decision period is at the middle point of the current decision period. Obtaining multi-period optimal purchasing quantity by using purchasing model and implementing purchasing quantity Q of current period ₁ . At the next purchasing point, updating the new information, completing prediction, reconstructing a multi-stage purchasing model, and implementing the optimal purchasing quantity Q 'corresponding to the decision period' ₂ . And (4) as the time advances, sequentially rolling to calculate and finish the decision to obtain the best purchasing quantity determined in each decision period.

The relevant explanations and assumptions for this procurement model are as follows:

(1) the enterprise purchases raw materials monthly and at the mid-point of each decision period.

(2) As most production processes are required to be continuous and uninterrupted, the assumption is that the raw materials can be supplied by hundreds of percent, the situation of shortage of goods does not occur, and emergency ordering is not needed.

(3) Procurement decision period T _k The raw material purchased will arrive at the beginning of the next decision period.

(4) There is no time constraint in the enterprise production process, namely continuous production is kept before and after the procurement period.

(5) The remaining inventory for each decision period will be postponed until the next decision period, meaning that the current procurement decision will affect future inventory levels.

The goal of raw material procurement is to reduce raw material costs while ensuring normal production. The cost mainly includes the procurement cost C _p And inventory cost C _w 。

Procurement cost C of procurement planning period _p The method mainly comprises the following steps of raw material cost and procedure cost:

wherein P is _k Is the k decision period obtained by the prediction of the VMD-LSTM modelRaw material price of (C) _S Is the procedure cost for each decision period.

Inventory cost C _w Mainly related to the average inventory level. For a procurement decision period T _k The average inventory level is:

the total inventory cost for the procurement planning period T is:

where eta is the conversion rate of production, C _I Is annual loan interest rate, C _M Daily stock management cost per unit of raw materials, D _k Is T _k The corresponding raw material demand is D _k /η。

And establishing a multi-period purchasing model under the condition of raw material price and demand change by taking the lowest unit raw material purchasing cost in a purchasing plan period T as an objective function. It is worth noting that the raw material unit cost during T is considered here, not the total cost. Obviously, in the procurement mode, which targets the lowest total cost, the procurement volume during each decision period will just meet the production requirements and safety stock to reduce the total cost. This model cannot accommodate changes in raw material market prices. If the aim is to minimize the unit cost, the change of the price is fully considered, the purchase quantity is adjusted, and more raw materials are purchased at a lower price on the premise of meeting the safety production. In the long run, the mode can well utilize benefits brought by price change, and the overall purchasing cost is reduced.

The embodiment of the invention establishes a multi-period dynamic purchasing model considering the price and demand change of raw materials:

k＝1,2,3,...,L (19e)

wherein B is _k Indicating during procurement period T _k The capital budget for making raw material purchases, V is the maximum inventory capacity of the warehouse. The above model takes into account constraints on budget, safety stock and storage capacity. The formula (19 b) indicates that the raw material procurement cost cannot exceed the capital budget, among them

Is a modal component set obtained by variational modal decomposition;

and (3) a prediction function representing the ith modal component, namely a prediction model obtained based on LSTM training. The sum of the predicted values of the K modal components equals the predicted value of the final price, i.e. the price P of the kth decision cycle _k . The formula (19 c) shows that the quantity of the purchased raw materials must meet the production requirement, and the remaining stock at the end of the period must not be lower than the safety stock to resist the influence of uncertain factors. Equation (19 d) indicates that the stock of raw materials cannot exceed the maximum stock capacity of the warehouse.

The multi-period dynamic purchasing model considers the change of the price and the demand of the raw materials in the future, combines various practical constraints, makes a more comprehensive decision in the long term and can further save the cost.

(2) The optimization method comprises the following steps:

the invention adopts Particle Swarm Optimization (PSO) to solve the multi-period purchasing model, and the punishment function is used for processing the constraint item. For equation (19), which is obviously a typical constrained optimization problem, the penalty function method adds constraints to the objective function in the form of penalty terms. If the current solution is not in the feasible region, the value of a penalty term in the objective function becomes extremely large, so that the current solution cannot become the optimal solution; if the current solution is in the feasible region, the value of the penalty item is 0, and the constraint effect on the current solution is achieved through the value change of the penalty item. The original constrained optimization problem can be converted into an unconstrained optimization problem through a penalty function method, so that the solution of the optimization problem is facilitated.

The penalty function for equation (19 b) is:

the penalty function for equation (19 c) is:

the penalty function for equation (19 d) is:

the overall penalty function is:

P＝P ₁ +P ₂ +P ₃ (23)

the unconstrained optimization problem processed by the penalty function method is as follows:

where α is a penalty factor, taking a large positive number.

In this embodiment, a particle swarm algorithm is used to solve the above unconstrained optimization problem:

(1) setting the dimension of the particles to be equal to the number L of decision cycles included in a procurement planning cycle, and tabulating the position of each particleRaw material procurement quantity (Q) shown as L decision periods within a procurement planning period ₁ ,Q ₂ ,…,Q _L ) Selecting the function shown in the formula (24) as a fitness function of the particle;

(2) initializing to obtain a signal containing N _p An initial population of particles, each particle having a random position and velocity;

(3) for each particle, comparing the current fitness with the individual optimal fitness pb; if the current fitness is better than pb, updating pb to the current fitness, and updating the L-dimensional position of pb to the current position;

(4) for each particle, comparing the current fitness with the historical optimal fitness gb for the group; if the current fitness is better than gb, updating gb to the current fitness of the particle, and recording the number and position of the particle;

(5) the velocity and position of each particle is updated separately according to the following formula:

v _i ＝w _in v _i +c ₁ rand()(pb _i -s _i )+c ₂ rand()(gb _i -s _i ) (25)

s _i ＝s _i +v _i (26)

wherein, ω is _in Is an inertia factor, c ₁ And c ₂ Is a learning factor;

(6) and (3) returning to the step (2), and continuing the iterative loop until a good enough fitness is obtained or the maximum iteration number is reached.

3. Supplier evaluation management

Supplier evaluation is typically done manually by hand. And related management personnel design questionnaires according to rules and experiences and send the questionnaires to related departments for filling. And then counting the final scores of all questionnaire results, and screening out high-quality suppliers from high scores to low scores. However, the subjectivity and knowledge limitations of these judgments greatly affect the final evaluation results. The method needs to manually collect and analyze a large amount of data, and has large workload and low efficiency. The record data of the supplier supply can reflect the real condition of the supply and is the real photo of the grading of the supplier. Under the industrial internet, various information of the goods arriving from the suppliers can be automatically collected and stored, and various algorithms can be used for automatically completing calculation, so that scientific and efficient evaluation management of the suppliers is realized. Before evaluating a supplier, it is necessary to determine an evaluation index, i.e. from which aspects the supplier is evaluated. However, different industries have different emphasis points, and it is difficult to provide a common set of evaluation criteria for all industries. For a particular industry, evaluation criteria need to be analyzed and defined according to specific scenarios.

In the supplier evaluation method, the invention provides a supplier evaluation model based on Stacking Automatic Encoder (SAE) and bootstrap aggregation. SAE is a deep neural network composed of multiple layers of self-encoders. The method can extract complex and abstract features in data in a layering mode, and has strong capability of approximating complex nonlinear functions and strong modeling capability. Complex non-linear relationships between supplier supply data and supplier scores, such as supply quality, supply quantity, impurity over-standard conditions, etc., can be established using SAE. After the model training is completed, the efficient evaluation can be automatically carried out on the suppliers.

An Automatic Encoder (AE) is an unsupervised single-hidden-layer neural network, consisting of an encoder and a decoder. The output layer is the same as the input layer, the main purpose is that the output layer reconstructs the original input as accurately as possible, and the structure is shown in fig. 6.

The main purpose of AE is to reconstruct the original input as accurately as possible. The encoder inputs I = [ I ] through a mapping function f ₁ ,i ₂ ,…,i _n ] ^T ∈R ⁿ Mapping to hidden layer k = [ k ] ₁ ,κ ₂ ,…,κ _m ] ^T ∈R ^m 。

κ＝f(I)＝σ _f (AI+b) (26)

Wherein A is a weight matrix with dimension of m × n, b ∈ R ^m Is the offset vector, σ, of the hidden layer _f Representing a non-linear activation function.

The decoder converts the hidden layer κ to the output layer by a mapping function g

In the formula (I), the compound is shown in the specification,

is a weight matrix of dimension n x m,

is the offset vector of the output layer and,

a non-linear activation function, κ, may be considered a feature extracted from the input data. The automatic encoder uses back propagation algorithm to learn as much as possible to obtain a mapping

After reconstitution

As much as possible equal to the original input I. The parameters of the model can be obtained by minimizing the average reconstruction error.

N is the number of samples. L is a loss function, either a conventional root mean square error

Or reconstructed cross entropy.

The stacked self-encoder SAE is a hierarchical deep neural network connected by a plurality of AE layers, and the structure and training process thereof are shown in fig. 7. SAEs were constructed by unsupervised pre-training of individual AEs and supervised fine-tuning of the overall structure. In the pre-training phase, the original training data is used as input for training to obtain a hidden layer representation of the first AE. The hidden-layer representation of the first AE is then trained as an input to the second AE, resulting in a hidden-layer representation of the second AE. And by analogy, pre-training all the self-encoders layer by layer. After unsupervised pre-training, the input layer, the hidden layers of each AE are stacked, a supervised model (e.g., FNN) is added as output on top of the last AE, and the hidden layer of the last AE is represented as input to the model. Finally, a complete SAE is established. Further, all parameters of the SAE will be fine-tuned as a whole by a training algorithm (e.g. gradient descent). By training the multi-layer neural network, SAE implements associative mapping between a vendor's basic data and its scores. In the supplier evaluation management, data sets are constructed by using evaluation index values and comprehensive grading results of different suppliers. And (4) taking the evaluation index values of the suppliers as model input, and taking the final scores of the suppliers as final output to train the model. After the model is trained, the supplier data of a certain supplier is input into the model, and then the grade of the supplier can be automatically calculated. By sorting the composite scores of all suppliers within a specific time period, raw material suppliers with higher composite quality can be screened out.

For supplier evaluation in real-world scenarios, some companies may evaluate suppliers quarterly or annually, meaning that the sample available for learning may not be sufficient. In this regard, bootstrapping aggregation (Bagging) was used to improve the performance of SAE-based vendor evaluation models. The Bootstrap aggregation (Bagging) Bagging method is an ensemble learning algorithm. The method carries out random sampling on original data to obtain a plurality of data sets, trains a plurality of models respectively, and obtains a result with stronger robustness by aggregating the plurality of models.

In this embodiment, the automatic evaluation model of a supplier based on stack self-encoder and bag-in-bag method ensemble learning is shown in fig. 8, and the process of the method for constructing and evaluating the supplier is as follows:

(1) collecting historical data of suppliers

Which comprises

N suppliers are used for evaluating index data of the suppliers respectively; evaluating the suppliers by experts according to the index data to obtain the historical scores of the suppliers; raw historical data

And scoring to form a new data set

(2) From data sets based on the bagging method

In the random selection

Carrying out substitution on the samples to form a subdata set;

(3) repeating (2) to obtain N _D Mutually independent subdata sets;

(4) in each subdata set, taking n-dimensional evaluation index data of each supplier as input and the grade of the supplier as output, training a stack self-encoder model comprising n-dimensional input and 1-dimensional output, and obtaining the mapping relation between the supplier evaluation index data and the supplier grade, namely score = Lambda _d (index ₁ ,index ₂ ,…,index _n ),d＝1,2,…,N _D ；

(5) Repeating (4) to obtain N _D The evaluation models are independent from each other and based on a stack self-encoder;

(6) when new supplier N-dimensional index data is input to the stacked self-encoder model, N _D Each stack self-encoder model correspondingly obtains N _D Scoring the result, and then taking N _D The average of the individual scoring results is taken as the final score for the supplier, i.e.:

by establishing an integrated model with higher robustness than a single SAE model, the method can improve the accuracy and stability of the evaluation result.

4. Supplier order distribution

Order allocation refers to determining the raw material purchase quantity of each supplier according to a certain target. In most businesses, order distribution is primarily responsible for the purchasing department. Due to the isolation between the business of the departments, the purchasing department may only focus on reducing costs, and neglect the differences in raw material quality and supply stability of different scoring suppliers, which can have a great impact on the production department. Meanwhile, because the number of suppliers is large, the information to be considered in the purchasing process is large, and in the face of a large amount of information, a decision maker is difficult to make a reasonable decision by integrating various information. Furthermore, manual decisions are susceptible to decision maker personal preferences and their relationship to the supplier. Therefore, in practice, the manual decision mode easily causes problems of high purchase cost, unstable raw material quality and the like. Under the industrial internet, relevant information of suppliers can be automatically collected and comprehensively processed, and relevant intelligent algorithms can be applied to make decisions in an objective form. And a scientific and reasonable order distribution method is adopted, so that the stability and the quality of raw material supply are improved, and the purchasing cost is reduced.

The invention aims at minimizing the purchasing cost and maximizing the overall effectiveness, considers the constraint conditions of inventory, supply capacity, demand and the like, and establishes an order distribution model. The relevant explanations and assumptions are as follows:

(1) the raw material can be divided into N _g The prices of different grades are different from grade to grade.

(2) Now has N _s The price of raw materials of different grades of different suppliers is different, and the supply capacity is different.

(3) The k-rich components contained in the raw material can bring additional benefits. The enriched component content of different suppliers can be different from each other.

(4) The price of the same enriched component is the same for different suppliers.

(5) Order allocation is only applicable for the current decision cycle.

(1) An objective function:

the purchasing cost is the difference between the purchasing cost and the recovery value of the raw materials and should be reduced as much as possible.

Wherein a is _ij Raw material purchase quantity, p, for jth grade of ith supplier _ij Price of material for jth grade of ith supplier, e _ij Additional revenue per ton obtained for purchasing the ith supplier's jth grade of feedstock.

Overall utility: the overall utility of all suppliers should be maximized with the goal of making purchases with the highest scoring suppliers possible.

Wherein

Is the nth index data at the ith supplier.

(2) Constraint conditions

Supply capacity constraints: order quantity a _ij Should not exceed the supply capacity r of the jth grade of raw material of the ith supplier _ij 。

a _ij ≤r _ij ,i＝1,2,…,N _s ,j＝1,2,…,N _g (31)

And (3) requirement constraint: the sum of the main components in the raw materials should meet the requirements of the production plan.

Wherein theta is _ij Principal composition of raw materialsDividing the content.

Quantity constraint: the total order quantity should satisfy the procurement plan G and should not be excessive. On one hand, the total amount of the order should meet the total amount of the purchasing plan; on the other hand, the total amount of orders cannot be too large to exceed 2% of the total amount of the procurement plan, and therefore

And (4) inventory constraint: the total order should not exceed the warehouse remaining capacity S.

(3) Optimization method

The order allocation model is a typical multiobjective constraint optimization problem (CMOP). The method converts the multi-target problem into the single-target problem by using a function transformation and linear weighting method. Firstly, each optimization target is subjected to scale transformation, so that the value of each transformed target is in the same order of magnitude, as shown in the following formula:

wherein f is _i ^* (a) Are the optimal values obtained by solving the ith optimization problem by considering the constraint conditions, respectively.

Then, different objective functions are weighted xi _i . The higher the weight, the more important the representation of the objective function.

Finally, a compromise single-objective optimization problem is obtained by linear combination:

through the processing, the multi-objective constraint optimization problem is greatly simplified, so that the original problem can be solved through a conventional optimization method.

In this embodiment, a Particle Swarm Optimization (PSO) is used to solve the function transformation and the final optimization problem, so as to obtain the optimal raw material order quantity of each supplier:

step1: with f ₁ (a) Taking equations (31) to (34) as constraints for an objective function, converting the constrained optimization problem into an unconstrained optimization problem according to a penalty function method in the purchasing decision optimization process, and solving by using a PSO (particle swarm optimization) to obtain an optimal value f corresponding to the optimal solution ₁ ^* (a)；

Step2: with f ₂ (a) Taking equations (31) to (34) as constraints for an objective function, converting the constrained optimization problem into an unconstrained optimization problem according to a penalty function method in the purchasing decision optimization process, and then solving by using a PSO (particle swarm optimization) to obtain an optimal value corresponding to an optimal solution

Step3: converting the multi-objective optimization problem into a single-objective optimization problem according to the equations (35) and (36) (taking the equations (31) to (34) as constraints);

step4: and (3) according to a penalty function method in the purchasing decision optimization process, converting the constrained optimization problem into an unconstrained optimization problem, and finally solving by using a PSO (particle swarm optimization) to obtain an optimal solution.

When Step1, step2 and Step4 above are solved using PSO, the position of each particle is expressed as the raw material procurement quantity { a ] of all grades of all suppliers _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.

According to the method, data required by market price prediction, raw material purchasing decision, supplier evaluation management and supplier order distribution can be obtained from an industrial internet platform, the processes can be automatically completed on the industrial internet platform, and integrated decision is completed from the perspective of the whole process of a raw material supply chain, so that the problems of high purchasing cost and low efficiency caused by information isolation and cognitive subjectivity under artificial decision are solved, the raw material purchasing cost is effectively reduced, and the supply chain management efficiency is improved.

The above embodiments are preferred embodiments of the present application, and those skilled in the art can make various changes or modifications without departing from the general concept of the present application, and such changes or modifications should fall within the scope of the claims of the present application.

Claims

1. An industrial internet-based integrated optimization method for a purchasing supply chain is characterized by comprising the following steps:

evaluation by the supplier: acquiring historical supply data of each provider, constructing an automatic evaluation model of the provider based on stack autoencoder and bagging ensemble learning, and acquiring scores of each provider;

supplier order distribution: comprehensively considering the optimal raw material purchase quantity and the supplier score, establishing a multi-objective optimization model for minimizing purchase cost and maximizing the comprehensive effect of the suppliers, and solving to obtain the optimal purchase strategy of each supplier;

the market price prediction, the raw material purchasing decision, the supplier evaluation and the supplier order distribution are carried out, and the required data is automatically obtained through an industrial internet platform.

2. The method according to claim 1, characterized in that the market price of the raw material is predicted using variational modal decomposition and long-short term memory networks, in particular:

Step A2, according to the step length, aiming at each modal component u _i (t) splitting to obtain inputs and outputs for model training, and further for training the LSTM to obtain the correspondence of each componentThe predictive model of (2);

3. The method according to claim 2, wherein the number K of variational modal decompositions is determined by: gradually increasing the decomposition number and calculating corresponding residual errors, and when the residual errors are smaller than a preset value and have no obvious downward trend, determining the current decomposition number as an optimal decomposition number K; wherein the residual error is calculated as:

4. The method of claim 1, wherein the multi-cycle dynamic procurement model established in the raw material procurement decision is:

k＝1,2,3,...,L (19e)

wherein C is the raw material cost of the purchase planning cycle, including the purchase cost C _p And inventory cost C _w (ii) a The duration T of the purchase planning period comprises L independent decision periods, and the duration of the kth decision period is T _k Material procurement point t of kth decision cycle _k At the midpoint of the current decision cycle; q _k Planning the total quantity of raw material purchase for the kth decision period;

a prediction function representing the ith modal component;

cost of inventory

R ₀ The initial raw material inventory for the procurement planning period; d _k Is T _k The corresponding raw material demand is D _k Eta, eta is the production conversion rate; c _I Is annual loan interest rate, C _M Daily inventory management cost per unit of raw material;

5. The method according to claim 4, wherein a particle swarm optimization algorithm is used for solving the multi-cycle dynamic purchasing model, and a penalty function is used for processing a constraint term in the purchasing model during solving, specifically:

P＝P ₁ +P ₂ +P ₃ (23)

in the formula, alpha is a penalty factor;

b3, initializing a particle population, and selecting a function shown in a formula (24) as a fitness function of the particles; wherein, the dimension of the particle is equal to the number L of decision periods included in the procurement planning period, and the position of each particle is expressed as the procurement quantity (Q) of the raw materials in the L decision periods in the procurement planning period ₁ ,Q ₂ ,…,Q _L )；

And B4, optimizing and solving the optimal particles by adopting a particle swarm optimization, and obtaining the optimal raw material purchase amount of each decision period in the purchase plan period according to the position of the optimal particles.

6. The method according to claim 1, wherein the supplier evaluation is performed by the following specific procedures:

step C1, collecting historical data of suppliers

Which comprises

And scoring to form a new data set

Step C2, based on the bagging method, from the data set

In the random selection

Carrying out substitution on the samples to form a subdata set;

step C5, repeating the step C4 to obtain N _D The evaluation models are independent from each other and are based on a stack self-encoder;

step C6, when new supplier N-dimensional index data is input into the stack self-encoder model, N _D Each stack correspondingly gets N from the coder model _D Scoring the result, and then taking N _D The average of the individual scoring results is taken as the final score for that vendor.

7. The method of claim 1, wherein minimizing the procurement cost means minimizing a difference between the procurement cost and the recycle value of the raw material, expressed as an objective function f ₁ (a)：

In the formula, a _ij For the jth grade of raw material procurement of the ith supplier, p _ij Price of material for jth grade of ith supplier, e _ij Additional revenue per ton obtained for purchasing an ith supplier's jth grade of feedstock; n is a radical of _s Number of raw Material suppliers, N _g The grade number of the raw materials;

In the formula (I), the compound is shown in the specification,

n index data for the ith supplier,

scoring the ith supplier for the ith stacked self-encoder in a bagging process; n is a radical of _D The number of stacked self-encoders used for the bagging method.

8. The method of claim 7, wherein the constraints in building the multiobjective optimization model include supply capacity constraints, demand constraints, order quantity constraints, and inventory constraints;

a _ij ≤r _ij ,i＝1,2,…,N _s ,j＝1,2,…,N _g (31)

wherein, theta _ij Is the main component content of the raw material;

9. The method of claim 7, wherein in solving the multi-objective optimization model for supplier order distribution, the multi-objective problem is transformed into a single-objective problem using a functional transformation and a linear weighting method, specifically:

firstly, carrying out scale transformation on each optimization target to ensure that the value of each target after transformation is in the same order of magnitude, as shown in the following formula:

then different objective functions are weighted xi _i ；

Finally, obtaining a compromise single-target optimization problem through linear combination:

10. an electronic device comprising a memory and a processor, the memory having stored therein a computer program, wherein the computer program, when executed by the processor, causes the processor to carry out the method according to any one of claims 1 to 9.