Material supply system index value multi-step prediction method, equipment and medium

Abstract

The invention discloses a material supply system index value multi-step prediction method, equipment and a medium, belongs to the technical field of material supply management and artificial intelligence cross, and aims to solve the technical problem of how to realize accurate prediction and dynamic regulation and control of each link of a supply chain, improve the intelligent level and risk response capability of material industrial management and improve the risk response capability of the supply chain. According to the technical scheme, interfaces are established with a supplier information system, a logistics transportation system and a production department business system, and material supply historical operation data are collected in real time; carrying out anomaly detection and replacement on the historical operation data of material supply by adopting a gate control cycle unit self-encoder GRUAE; constructing an index value multi-step prediction model by adopting a gated loop unit network LAGRU with a local attention mechanism; and performing deep learning increment updating on the index value multi-step prediction model by adopting an elastic weight integration EWC algorithm to obtain the index value multi-step prediction model after model parameter updating.

Description

Technical Field

[0001] The present invention relates to the cross-technical field of material supply management and artificial intelligence, and in particular to a multi-step prediction method, device and medium for index values ​​of a material supply system. Background Art

[0002] In the field of material supply management, the existing technical system still has significant limitations in supplier evaluation, warehouse management and material transportation. Although the traditional ERP system has achieved the integration of procurement, production, sales and other modules, its operation mode based on static data and preset rules has been difficult to adapt to the dynamic market environment. Specifically, supplier management mainly relies on manual evaluation, which has problems such as strong subjectivity and low efficiency, and it is difficult to identify potential risks in a timely manner; the setting of safety stock thresholds in warehouse management is mostly based on experience, which can easily lead to inventory backlogs or supply shortages; the lack of data coordination between modules restricts the company's overall control over the material supply chain.

[0003] At the data analysis level, conventional methods mainly use tools such as Excel and SQL to process structured data, and apply traditional forecasting techniques such as the moving average method. However, such methods can neither effectively process unstructured data nor mine the complex relationships between data, resulting in insufficient forecasting accuracy. Especially when the market environment fluctuates drastically, it is difficult for companies to adjust their material supply strategies in a timely manner, and they face greater operational risks.

[0004] The existing forecasting and early warning mechanisms also have obvious defects. Simple methods such as fixed threshold methods and linear regression models are difficult to deal with complex nonlinear relationships. In the face of emergencies such as natural disasters, early warnings are often delayed or inaccurate, causing companies to miss the best time to respond. In addition, although data-based supplier management technologies (such as supplier relationship management systems) can achieve preliminary evaluation and classification, their evaluation systems rely too much on limited historical data and fixed indicators, ignoring suppliers' innovation capabilities and emergency response levels. In addition, data mining algorithms are relatively simple and difficult to capture market changes in a timely manner, which ultimately affects the quality of corporate procurement decisions and the stability of material supply.

[0005] In summary, the existing technical system has obvious deficiencies in data processing capabilities, prediction accuracy, and system adaptability, and it is difficult to meet the material supply management needs of enterprises in a complex and changing market environment. Therefore, how to achieve accurate prediction and dynamic regulation of each link in the supply chain and improve the intelligent level of material industry management and risk response capabilities are technical issues that need to be solved urgently. Summary of the invention

[0006] The technical task of the present invention is to provide a multi-step prediction method, device and medium for the index values of a material supply system, so as to solve the problem of how to achieve accurate prediction and dynamic regulation of all links in the supply chain, and improve the intelligent level and risk response ability of material industry management.

[0007] The technical task of the present invention is achieved in the following manner. A multi-step prediction method for the index values of a material supply system is as follows: By establishing interfaces with the supplier information system, logistics transportation system and production department business system, the historical operation data of material supply is collected in real time; The gated recurrent unit autoencoder GRU_AE is used to detect and replace anomalies in the historical operation data of material supply; among them, both the encoder and decoder in the gated recurrent unit autoencoder GRU_AE are composed of several GRU layers. The GRU layer compresses the input data into a low-dimensional representation, restores the input data with reduced dimensions to the original data, and then determines whether the input data is abnormal through the reconstruction error of the input data, and completes the construction of the historical operation data reconstruction model of material supply; The gated recurrent unit network with local attention mechanism LA_GRU is used to construct a multi-step prediction model for index values; The elastic weight consolidation EWC algorithm is used to perform deep learning incremental update on the multi-step prediction model for index values, and obtain the multi-step prediction model for index values after model parameter update; The processed historical operation data of material supply is input into the multi-step prediction model for index values after model parameter update, and the future values of different indexes are obtained.

[0008] Preferably, the historical operation data of material supply includes supplier performance data, logistics transportation data, warehousing management data, production plan data, marketing demand data and environmental data (such as weather); Among them, the supplier performance data includes the on-time delivery rate and the product quality qualification rate; the logistics transportation data includes the transportation time, cost and real-time location tracking information; the warehousing management data includes the inventory level, turnover rate and loss rate; the production plan data includes the production task volume and equipment utilization rate; the marketing demand data includes the historical sales volume and customer demand prediction information.

[0009] Preferably, the construction of the historical operation data reconstruction model of material supply is as follows: The calculation formula of the GRU layer is as follows: ; Among them, represents the update gate; represents the reset gate; represents the new candidate hidden state; represents the updated hidden state; denote the input vector at a moment, where the input vector refers to within a time range, a data matrix obtained by splicing and combining multi-dimensional time series data, with the dimension of ; denote the feature dimension, such as the on-time delivery rate, inventory level, etc.; denote the time step; denote the hidden state at the previous time step; denote the sigmoid function; denote the dot product operation; respectively denote the input weight matrix, hidden state weight matrix, and bias vector of the update gate; respectively denote the input weight matrix, hidden state weight matrix, and bias vector of the reset gate; respectively denote the input weight matrix, hidden state weight matrix, and bias vector at the previous time step; The encoder formula of the gated recurrent unit autoencoder GRU-AE is as follows: ; where denote the hidden state at the last time step of the encoder; denote the hyperbolic tangent function; and respectively denote the weight and bias of the last hidden layer of the encoder; denote the output result of the encoder; The decoder formula of the gated recurrent unit autoencoder GRU-AE is as follows: ; where and respectively denote the weight and bias of the last hidden layer of the decoder; denote the reconstructed historical operation data of material supply. The goal of the gated recurrent unit network autoencoder GRU-AE is to minimize the original input data and the GRU-AE reconstructed input data the error between; and respectively denote and the output data of the hidden layer of the neural network at the moment; The data reconstruction error threshold is set to: ; where denote the reconstruction error threshold interval of the feature in the historical operation data of material supply; denote the mean value of the feature in the historical operation data of material supply; Represents the standard deviation of the features in the historical operation data of material supply ; The data reconstruction model is: ; where Represents the reconstructed feature at time t; Represents the data reconstruction model based on the gated recurrent unit autoencoder GRU-AE; Represents the feature after normalization at time t; The judgment of abnormal conditions in the historical operation data of material supply is as follows: ; Among them, in the construction process of the data reconstruction model of the historical operation data of material supply, the number of neuron nodes in the input layer and the output layer of the network structure is made equal, and the number of neuron nodes in the internal hidden layer is less than or equal to the number of nodes in other layers.

[0010] More preferably, the encoder of the gated recurrent unit autoencoder GRU-AE includes an encoder input layer, an encoder stacked GRU layer, and an encoder attention pooling layer; Among them, the encoder input layer receives the collected multi-dimensional time series data and inputs it with a dimension of ; The encoder stacked GRU layer includes three layers of unidirectional GRU layers for extracting time series features layer by layer; the number of hidden units in the first GRU layer of the encoder is 128, and the input dimension is , and the output dimension ; the number of hidden units in the second GRU layer of the encoder is 64, and the output dimension , further compressing the time series features; the number of hidden units in the third GRU layer of the encoder is 32, and the output dimension , generating a low-dimensional representation ; The encoder attention pooling layer performs weighted summation on along the time dimension to generate a fixed-length encoding for capturing the global time series pattern.

[0011] More preferably, the decoder of the gated recurrent unit autoencoder GRU-AE includes a decoder latent vector expansion layer, a decoder stacked GRU layer, and a decoder fully connected reconstruction layer; Among them, the decoder latent vector expansion layer is used to reconstruct the time step structure, and is repeated times to form data with the same structure as ; The structure of the decoder stacked GRU layer is symmetric to that of the encoder stacked GRU layer. A three-layer unidirectional GRU is adopted to gradually restore the time series dimension. The number of hidden units in the first GRU layer of the decoder is 32, and the input dimension , and the output dimension ; the number of hidden units in the second GRU layer of the decoder is 64, and the output dimension ; the number of hidden units in the third GRU layer of the decoder is 128, and the output dimension .

[0012] Preferably, a gated recurrent unit network with local attention mechanism LA_GRU is used to construct the multi-step prediction model of the index value as follows: A GRU network is used to construct the multi-step prediction model of the index value, which is expressed as follows: ; Among them, represents the predicted value of each index of the material supply at time t; is the multi-step prediction model of the index value constructed by the GRU network model, is the characteristic of the historical operation data of the material supply at time t; A local attention mechanism is added to the GRU network. The local attention mechanism is as follows: ; Among them, represents the attention weight; represents the local context vector; represents the input vector, represents the window size; and respectively represent the model weight and model bias of the previous hidden layer; represents the activation function; Taking the hidden state of the previous time step as the input, the attention weight is calculated through a fully connected layer; Applying the attention weight to the input vector sequence from to window, the local context vector is calculated by weighted summation; among them, represents the th input vector at time step attention weight; The local context vector and the hidden state They are input into the GRU network together to generate the hidden state at the current time step , that is, the updated hidden state.

[0013] Preferably, the elastic weight consolidation (EWC) algorithm is used to perform deep learning incremental update on the multi-step prediction model of the index value as follows: The calculation method of the importance weight of each parameter in the multi-step prediction model of the index value is as follows: ; Wherein, represents the Fisher matrix, which is used to describe the influence degree of the change of the parameters of the multi-step prediction model of the index value on the output of the multi-step prediction model of the index value; represents the training data set of the historical operation data of the material supply; represents the likelihood function of the multi-step prediction model of the index value; represents the expected value calculation operation; represents the regularization parameter; represents the parameters of the multi-step prediction model of the index value importance degree; Update each model parameter. For each model parameter , the update method is as follows: ; Wherein, represents the parameters in the multi-step prediction model of the index value obtained by training with the historical operation data of the material supply; represents the identity matrix; represents the learning rate; represents the new loss function; represents the regularization parameter; represents the updated parameters of the multi-step prediction model of the index value.

[0014] Preferably, the multi-step prediction model of the index value after updating the model parameters is expressed as follows: ; Wherein, represents the multi-step prediction value of the index value at time t; represents the multi-step prediction model of the index value constructed by using the GRU network model with a local attention mechanism updated regularly; represents the characteristics of the historical operation data of the material supply at time t.

[0015] An electronic device, comprising: a memory and at least one processor; Wherein, a computer program is stored on the memory; The at least one processor executes the computer program stored in the memory, such that the at least one processor executes the multi-step prediction method for the material supply system index values as described above.

[0016] A computer-readable storage medium stores a computer program, and the computer program can be executed by a processor to implement the multi-step prediction method for the material supply system index values as described above.

[0017] The multi-step prediction method, device and medium of the material supply system of the present invention have the following advantages: (1) The present invention is used to accurately predict the future multi-step index values of multiple links such as suppliers, transportation, warehousing, and marketing in the supply chain, so as to guide enterprises to dynamically regulate each link of the supply chain and improve the scientificity and foresight of material supply management; (2) Aiming at the problem that the traditional material supply management system is difficult to adapt to the dynamic changes of the market and the supply chain, the present invention analyzes a large amount of historical data through deep learning technology and constructs a multi-link index prediction model, which can predict the future trends of key indicators such as supplier performance, transportation efficiency, inventory level, and marketing demand, and provide accurate decision-making support for enterprises; for example, during the peak season of liquor sales, the model can predict the supply demand in the future multi-steps, help enterprises adjust the supply strategy in advance, and avoid out-of-stock or overstock; (3) In terms of supplier management, the present invention analyzes multi-dimensional data such as the historical performance, delivery ability, and risk factors of suppliers, predicts their future performance and identifies potential risks. Once it is found that a supplier may have problems, an early warning is given in advance, which is convenient for enterprises to adjust the procurement strategy in time to ensure the stability of material supply; in the transportation link, by predicting indicators such as transportation efficiency and cost fluctuations, it helps enterprises optimize the transportation plan, reduce logistics costs and improve the response speed; (4) For warehouse management, the present invention comprehensively considers factors such as market demand, inventory turnover rate, and supply chain fluctuations, predicts the future inventory level, and dynamically adjusts the safety inventory threshold and replenishment strategy; for example, when it is predicted that the raw material price will rise or the inventory turnover rate will decline, it can suggest that enterprises purchase in advance, optimize the inventory structure, reduce costs and improve the capital turnover efficiency; (5) In the marketing link, the present invention analyzes historical sales data, market trends and the effects of marketing activities, predicts future demand changes and their impacts on material supply. Enterprises can plan material supply in advance according to the prediction results, improve the market response speed and competitiveness; in addition, the present invention also supports multi-source data fusion, integrates data from multiple links such as suppliers, transportation, warehousing, and marketing, mines the complex correlations between data, and further improves the prediction accuracy; (6) By constructing a multi-link index prediction model, the present invention realizes accurate prediction and dynamic regulation of each link in the supply chain, provides a scientific and efficient material supply management solution for enterprises, and significantly improves the risk response ability and market competitiveness of enterprises; (7) The present invention uses deep learning to learn historical data to realize multi-step prediction of index values. Whether it is the fluctuation of raw material supply or the seasonal change of market demand, the index values can be predicted in a timely manner, which helps enterprises clearly grasp the actual situation of each link in the material supply and provides solid data support for subsequent decision-making; (8) The present invention uses an autoencoder with a gated recurrent unit network GRU layer for abnormal data detection and processing, and at the same time uses a GRU architecture to establish a multi-step prediction model for index values, and adds a local attention mechanism to the prediction model structure to ensure smaller computational complexity and higher interpretability of the model; moreover, the multi-step prediction model for index values is based on the elastic weight consolidation EWC algorithm for fast incremental learning, which improves the adaptability of the model to new data while protecting old knowledge. The actual application results show that the incremental multi-step prediction model for index values can predict index values in advance according to different seasons and different periods, so that enterprises can adjust the relevant links that may issue supply warnings in advance, thereby stabilizing the supply chain; (9) From the perspective of personalized customization and continuous optimization, the present invention can customize exclusive index evaluation models for different departments to meet differentiated needs; at the same time, the EWC incremental learning algorithm enables the model to continuously optimize with the accumulation of new data, always maintaining adaptability to business changes, which not only improves the work efficiency of each department, but also promotes the development of enterprise material supply management towards intelligence and refinement, and enhances the overall competitiveness of the enterprise. BRIEF DESCRIPTION OF THE DRAWINGS

[0018] The present invention will be further described below with reference to the accompanying drawings.

[0019] Att Figure 1 is a flowchart of a multi-step prediction method for index values of a material supply system; Att Figure 2 is a schematic diagram of the convergence of the training curve; Att Figure 3 is a schematic diagram of abnormal detection of the target completion rate feature. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0020] The multi-step prediction method, device and medium of the material supply system of the present invention will be described in detail below with reference to the accompanying drawings of the specification and specific embodiments.

[0021] Embodiment 1: As shown in Att Figure 1 This embodiment provides a multi-step prediction method for index values of a material supply system, and the method is as follows: S1. By establishing interfaces with the supplier information system, logistics transportation system, and production department business system, collect the historical operation data of material supply in real time; S2. Use the gated recurrent unit autoencoder GRU_AE to detect and replace anomalies in the historical operation data of material supply; among them, both the encoder and decoder in the gated recurrent unit autoencoder GRU_AE are composed of several GRU layers. The GRU layer compresses the input data into a low-dimensional representation, restores the input data with reduced dimensions to the original data, and then determines whether the input data is abnormal through the reconstruction error of the input data, completing the construction of the historical operation data reconstruction model of material supply; S3. Use the gated recurrent unit network with local attention mechanism LA_GRU to construct a multi-step prediction model for index values; S4. Use the elastic weight consolidation EWC algorithm to perform deep learning incremental update on the multi-step prediction model of index values, and obtain the multi-step prediction model of index values after updating the model parameters; S5. Input the processed historical operation data of material supply into the multi-step prediction model of index values after updating the model parameters to obtain the future values of different indexes.

[0022] As shown in the appendix Figure 2 Shown is the training convergence diagram of the multi-step prediction model of index values in this embodiment. From the curve, it can be seen that in a limited number of iterations, the error of the training set gradually decreases, and at the same time, the error of the validation set also decreases to a level comparable to that of the training set, indicating that the multi-step prediction model of index values has well fitted the data set and has the ability to predict index values.

[0023] As shown in the appendix Figure 3 Shown is the detection effect of the gated recurrent unit - autoencoder (GRU-AE) material supply related data reconstruction model in this embodiment for the target completion rate feature. It can be seen from the figure that the gated recurrent unit - autoencoder (GRU-AE) material supply related data reconstruction model can effectively detect abnormal points in the time series data and process the abnormal data to ensure the quality of the training data of the multi-step prediction model of index values.

[0024] This embodiment ensures the continuous accuracy of the multi-step prediction index values related to material supply through deep learning and dynamic incremental learning, and improves the adaptability to the supply chain. By predicting in advance the index values related to material supply, it can predict in advance the impact of elements such as the market and weather on the supply chain, and make advance adjustments accordingly to ensure the stability of the supply chain, meeting the needs of the current enterprise management system construction.

[0025] The historical operation data of material supply in step S1 of this embodiment includes multi-dimensional information such as supplier performance data (such as on-time delivery rate, product quality qualification rate), logistics transportation data (such as transportation time, cost, real-time location tracking), warehousing management data (such as inventory level, turnover rate, loss rate), production plan data (such as production task volume, equipment utilization rate), marketing demand data (such as historical sales volume, customer demand forecast), and environment (such as weather). By establishing interfaces with the supplier information system, logistics transportation system, production department business system, etc., real-time data collection is realized to ensure the comprehensiveness, real-time nature, and consistency of the data, providing a high-quality data foundation for subsequent anomaly detection, model construction, and predictive analysis.

[0026] The construction of the reconstruction model for the historical operation data of material supply in step S2 of this embodiment is specifically as follows: S201. The calculation formula of the GRU layer is specifically as follows: ; Among them, represents the update gate; represents the reset gate; represents the new candidate hidden state; represents the updated hidden state; represents the input vector at time The input vector refers to the data matrix obtained by splicing and combining multi-dimensional time series data within the time range of , with a dimension of ; represents the feature dimension, such as on-time delivery rate, inventory level, etc.; represents the time step; represents the hidden state of the previous time step; represents the sigmoid function; represents the dot product operation; respectively represent the input weight matrix, hidden state weight matrix, and bias vector of the update gate; respectively represent the input weight matrix, hidden state weight matrix, and bias vector of the reset gate; respectively represent the input weight matrix, hidden state weight matrix, and bias vector of the previous time step; S202. The encoder formula of the gated recurrent unit autoencoder GRU-AE is as follows: ; Among them, represents the hidden state of the last time step of the encoder; represents the hyperbolic tangent function; and respectively represent the weight and bias of the last hidden layer of the encoder; Represents the output result of the encoder; S203. The decoder formula of the gated recurrent unit autoencoder GRU-AE is as follows: The decoder formula of the gated recurrent unit autoencoder GRU-AE is as follows: ; Wherein, and respectively represent the weight and bias of the last hidden layer of the decoder; Represents the reconstructed historical operation data of material supply. The goal of the gated recurrent unit network autoencoder GRU-AE is to minimize the original input data and the GRU-AE reconstructed input data The error between; and respectively represent and The output data of the hidden layer of the neural network at time; S204. The data reconstruction error threshold is set to: ; Wherein, Represents the reconstruction error threshold interval of the feature in the historical operation data of material supply; Represents the mean value of the feature in the historical operation data of material supply; Represents the standard deviation of the feature in the historical operation data of material supply; S205. The data reconstruction model is: ; Wherein, Represents the data of the reconstructed feature at time t; Represents the data reconstruction model based on the gated recurrent unit autoencoder GRU-AE; Represents the data of the feature after normalization at time t; S206. The judgment of abnormal conditions of the historical operation data of material supply is as follows: ; Wherein, in the process of constructing the data reconstruction model of the historical operation data of material supply, the number of neuron nodes in the input layer and output layer of the network structure is made equal, and the number of neuron nodes in the internal hidden layer is less than or equal to the number of nodes in other layers.

[0027] The encoder of the gated recurrent unit autoencoder GRU-AE in step S2 of this embodiment includes an encoder input layer, an encoder stacked GRU layer, and an encoder attention pooling layer; Among them, the encoder input layer receives the collected multi-dimensional time series data and inputs it with a dimension of ; The encoder stacked GRU layer includes three layers of unidirectional GRU layers, which are used to extract time series features layer by layer; the number of hidden units in the first GRU layer of the encoder is 128, and the input dimension is , and the output dimension ; the number of hidden units in the second GRU layer of the encoder is 64, and the output dimension , further compressing the time series features; the number of hidden units in the third GRU layer of the encoder is 32, and the output dimension , generating a low-dimensional representation ; The encoder attention pooling layer sums along the time dimension with weights to generate a fixed-length encoding , which is used to capture the global time series pattern.

[0028] The decoder of the gated recurrent unit autoencoder GRU-AE in step S2 of this embodiment includes a decoder latent vector expansion layer, a decoder stacked GRU layer, and a decoder fully connected reconstruction layer; Among them, the decoder latent vector expansion layer is used to reconstruct the time step structure, and is repeated times to form data with the same structure as ; The structure of the decoder stacked GRU layer is symmetric to that of the encoder stacked GRU layer, and three layers of unidirectional GRU are used to gradually restore the time series dimension; the number of hidden units in the first GRU layer of the decoder is 32, and the input dimension , and the output dimension ; the number of hidden units in the second GRU layer of the decoder is 64, and the output dimension ; the number of hidden units in the third GRU layer of the decoder is 128, and the output dimension .

[0029] The construction of the multi-step prediction model of the index value by using the gated recurrent unit network LA_GRU with a local attention mechanism in step S3 of this embodiment is specifically as follows: S301. Construct a multi-step prediction model of the index value by using a GRU network, which is expressed as follows: ; Among them, represents the predicted value of each index of the material supply at time t; is the multi-step prediction model of the index value constructed by using the GRU network model, is the characteristic of the historical operation data of the material supply at time t; S302. Incorporate a local attention mechanism into the GRU network. The local attention mechanism is as follows: ; Among them, represents the attention weight; represents the local context vector; represents the input vector, represents the window size; and respectively represent the model weight and model bias of the previous hidden layer; represents the activation function; S303. Use the hidden state of the previous time step as the input, and calculate the attention weight through a fully connected layer; S304. Apply the attention weight to the input vector sequence from to in the window, and calculate the local context vector by weighted summation. Among them, represents the attention weight of the th input vector at time step ; S305. Input the local context vector and the hidden state of the previous time step into the GRU network to generate the hidden state of the current time step, that is, the updated hidden state.

[0030] The specific process of using the Elastic Weight Consolidation (EWC) algorithm to perform deep learning incremental update on the multi-step prediction model of the index value in step S4 of this embodiment is as follows: S401. The calculation method of the importance weight of each parameter in the multi-step prediction model of the index value is as follows: ; Among them, represents the Fisher matrix, which is used to describe the influence degree of the change of the parameters of the multi-step prediction model of the index value on the output of the multi-step prediction model of the index value; represents the training data set of the historical operation data of the material supply; represents the likelihood function of the multi-step prediction model of the index value; represents the operation of calculating the expected value; represents the regularization parameter; represents the importance degree of the parameter of the multi-step prediction model of the index value; S402. Update each model parameter. For each model parameter , the update method is as follows: ; wherein represents the parameter in the multi-step prediction model of the index value trained using the historical operation data of material supply; represents the identity matrix; represents the learning rate; represents the new loss function; represents the regularization parameter; represents the updated parameter of the multi-step prediction model of the index value.

[0031] The multi-step prediction model of the index value after the model parameter update in step S5 of this embodiment is expressed as follows: ; wherein represents the multi-step prediction value of the index value at time t; represents the multi-step prediction model of the index value constructed by using the GRU network model with a local attention mechanism and updated regularly; represents the characteristics of the historical operation data of material supply at time t.

[0032] Embodiment 2: This embodiment also provides an electronic device, including: a memory and a processor; wherein, the memory stores computer execution instructions; The processor executes the computer execution instructions stored in the memory, so that the processor executes the multi-step prediction method of the material supply system index value in any embodiment of the present invention.

[0033] The processor may be a central processing unit (CPU), or may also be other general-purpose processors, digital signal processors (DSPs), application-specific integrated circuits (ASICs), off-the-shelf programmable gate arrays (FPGAs), or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc. The processor may be a microprocessor or the processor may also be any conventional processor, etc.

[0034] The memory can be used to store computer programs and / or modules. By running or executing the computer programs and / or modules stored in the memory, and by invoking the data stored in the memory, the processor can implement various functions of the electronic device. The memory mainly includes a program storage area and a data storage area. Among them, the program storage area can store the operating system, application programs required for at least one function, etc.; the data storage area can store the data created according to the use of the terminal, etc. In addition, the memory may further include high-speed random access memory, and may also include non-volatile memory, such as a hard disk, a memory, a plug-in hard disk, a smart media card (SMC), a secure digital (SD) card, a flash memory card, at least one magnetic disk storage period, a flash memory device, or other volatile solid-state storage devices.

[0035] Embodiment 3: This embodiment also provides a computer-readable storage medium, in which multiple instructions are stored. The instructions are loaded by the processor to cause the processor to execute the multi-step prediction method for the material supply system index values in any embodiment of the present invention. Specifically, a system or device equipped with a storage medium can be provided. On this storage medium, software program codes for implementing the functions in any one of the above embodiments are stored, and the computer (or CPU or MPU) of the system or device is caused to read and execute the program codes stored in the storage medium.

[0036] In this case, the program code read from the storage medium itself can implement the functions in any one of the above embodiments. Therefore, the program code and the storage medium storing the program code constitute a part of the present invention.

[0037] Examples of the storage medium for providing the program code include a floppy disk, a hard disk, a magneto-optical disk, an optical disk (such as CD-ROM, CD-R, CD-RW, DVD-ROM, DVD-RAM, DVD-RW, DVD+RW), a magnetic tape, a non-volatile memory card, and a ROM. Optionally, the program code can be downloaded from a server computer via a communication network.

[0038] In addition, it should be clear that not only can the functions in any one of the above embodiments be implemented by executing the program code read by the computer, but also by causing the operating system etc. operating on the computer based on the instructions of the program code to complete part or all of the actual operations.

[0039] In addition, it can be understood that the program code read from the storage medium is written into the memory provided in the expansion board inserted into the computer or into the memory provided in the expansion unit connected to the computer. Subsequently, based on the instructions of the program code, the CPU etc. installed on the expansion board or the expansion unit are caused to execute part and all of the actual operations, thereby implementing the functions in any one of the above embodiments.

[0040] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, rather than to limit them; although the present invention has been described in detail with reference to the foregoing embodiments, those of ordinary skill in the art should understand that they can still modify the technical solutions described in the foregoing embodiments, or perform equivalent replacements on some or all of the technical features; and these modifications or replacements do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of the present invention.

Claims

1. A multi-step prediction method for material supply system index values, characterized in that: The method is as follows: By establishing interfaces with supplier information systems, logistics and transportation systems, and production department business systems, historical data on material supply operations can be collected in real time; The gated recurrent unit autoencoder GRU_AE is used to detect and replace anomalies in the historical operation data of material supply. The encoder and decoder in the gated recurrent unit autoencoder GRU_AE are composed of several GRU layers. The GRU layer compresses the input data into a low-dimensional representation and restores the input data with reduced dimensions to the original data. Then, the input data reconstruction error is used to determine whether the input data is abnormal, thus completing the construction of the reconstruction model of the historical operation data of material supply. The gated recurrent unit network LA_GRU with local attention mechanism is used to build a multi-step prediction model for indicator values; The elastic weight integration EWC algorithm is used to perform deep learning incremental update on the multi-step prediction model of the index value, and the multi-step prediction model of the index value after the model parameters are updated is obtained; The processed historical operation data of material supply are input into the multi-step prediction model of indicator values ​​after the model parameters are updated to obtain the future values ​​of different indicators.

2. The multi-step prediction method for material supply system index values ​​according to claim 1 is characterized in that: The historical operation data of material supply includes supplier performance data, logistics and transportation data, warehouse management data, production planning data, marketing demand data and environmental data; Among them, supplier performance data includes delivery on time rate and product quality pass rate; logistics and transportation data includes transportation time, cost and real-time location tracking information; warehouse management data includes inventory level, turnover rate and loss rate; production planning data includes production task volume and equipment utilization rate; marketing demand data includes historical sales volume and customer demand forecast information.

3. The multi-step prediction method for material supply system index values ​​according to claim 1 is characterized in that: The reconstruction model of historical operation data of material supply is constructed as follows: The calculation formula of the GRU layer is as follows: ； in, represents the update gate; Reset gate. represents the new candidate hidden state; Represents the updated hidden state; express The input vector at time instant, the input vector refers to Within the time range, the data matrix obtained by splicing and combining multi-dimensional time series data has a dimension of ; Represents feature dimension; represents the time step; represents the hidden state of the previous time step; Represents the sigmoid function; Represents the dot product operation; Represent the input weight matrix, hidden state weight matrix and bias vector of the update gate respectively; Represent the input weight matrix, hidden state weight matrix and bias vector of the reset gate respectively; Represent the input weight matrix, hidden state weight matrix and bias vector of the previous time step respectively; The encoder formula of the gated recurrent unit autoencoder GRU-AE is as follows: ； in, represents the hidden state of the encoder at the last time step; represents the hyperbolic tangent function; and Represent the weight and bias of the last hidden layer of the encoder respectively; Represents the output result of the encoder; The decoder formula of the gated recurrent unit autoencoder GRU-AE is as follows: ； in, and Represent the weight and bias of the last hidden layer of the decoder respectively; Represents the reconstructed historical data of material supply operation. The goal of the gated recurrent unit network autoencoder GRU-AE is to minimize the original input data And GRU-AE reconstructs the input data The error between and Respectively and The hidden layer output data of the neural network at that moment; The data reconstruction error threshold is set as: ;in, Represents the characteristics of historical operation data of material supply The reconstruction error threshold interval; Represents the characteristics of historical operation data of material supply The mean of Represents the characteristics of historical operation data of material supply The standard deviation of The data reconstruction model is: ;in, Represents the reconstructed features Data at time t; Represents a data reconstruction model based on the gated recurrent unit autoencoder GRU-AE; Representation characteristics Normalized data at time t; The abnormal situation of historical operation data of material supply is judged as follows: ； Among them, in the process of constructing the reconstruction model of historical operation data of material supply, the number of neuron nodes in the input layer and the output layer of the network structure is equal and the number of neuron nodes in the internal hidden layer is less than or equal to the number of nodes in other layers.

4. The multi-step prediction method for material supply system index values ​​according to claim 3 is characterized in that: The encoder of the gated recurrent unit autoencoder GRU-AE includes an encoder input layer, an encoder stacked GRU layer, and an encoder attention pooling layer; The encoder input layer receives the collected multi-dimensional time series data and inputs the dimension ; The encoder stacked GRU layer includes three unidirectional GRU layers, which are used to extract temporal features layer by layer. The number of hidden units in the first GRU layer of the encoder is 128, and the input dimension is , output dimension ; The number of hidden units in the second GRU layer of the encoder is 64, and the output dimension , further compressing the temporal features; the number of hidden units in the third GRU layer of the encoder is 32, and the output dimension is , generating a low-dimensional representation ; Encoder attention pooling layer pair Weighted sum along the time dimension to generate fixed-length code , used to capture global timing patterns.

5. The multi-step prediction method for material supply system index values ​​according to claim 3 is characterized in that: The decoder of the gated recurrent unit autoencoder GRU-AE includes a decoder latent vector expansion layer, a decoder stacked GRU layer, and a decoder fully connected reconstruction layer; Among them, the decoder potential vector expansion layer is used to reconstruct the time step structure. repeat Second, form and Data with the same structure; The structure of the decoder stacked GRU layer is symmetrical with that of the encoder stacked GRU layer. It uses three layers of unidirectional GRU to gradually restore the temporal dimension. The number of hidden units in the first GRU layer of the decoder is 32, and the input dimension is , output dimension ; The number of hidden units in the second GRU layer of the decoder is 64, and the output dimension is ; The number of hidden units in the third GRU layer of the decoder is 128, and the output dimension is .

6. The multi-step prediction method for material supply system index values ​​according to claim 1 is characterized in that: The gated recurrent unit network LA_GRU with a local attention mechanism is used to construct a multi-step prediction model for indicator values ​​as follows: The GRU network is used to construct a multi-step prediction model for indicator values, which is expressed as follows: ； in, It represents the predicted value of each material supply index at time t; It is a multi-step prediction model of indicator value built by GRU network model. is the historical operation data characteristics of material supply at time t; A local attention mechanism is added to the GRU network. The local attention mechanism is as follows: ； in, represents the attention weight; represents the local context vector; represents the input vector, Indicates the window size; and Represent the model weight and model bias of the previous hidden layer respectively; represents the activation function; The hidden state of the previous time step As input, the attention weights are calculated through a fully connected layer ; The attention weight Applied to a sequence of input vectors of arrive Window, calculate the local context vector by weighted summation ;in, Indicates Input vector At time step The attention weight of Local context vector The hidden state at the previous time step Input into the GRU network together to generate the hidden state of the current time step , which is the updated hidden state.

7. The multi-step prediction method for material supply system index values ​​according to claim 1 is characterized in that: The elastic weight integration EWC algorithm is used to perform deep learning incremental updates on the multi-step prediction model of the index value as follows: The importance weight of each parameter in the multi-step prediction model of the indicator value is calculated as follows: ； in, Represents the Fisher matrix, which is used to describe the degree of influence of changes in the parameters of the multi-step prediction model of the indicator value on the output of the multi-step prediction model of the indicator value; A training dataset representing historical operation data of material supply; Represents the likelihood function of the multi-step prediction model of the indicator value; Represents an expected value calculation operation; represents the regularization parameter; Represents the multi-step prediction model parameters of the indicator value degree of importance; Update each model parameter. , the update method is as follows: ； in, Represents the parameters in the multi-step prediction model of the index value obtained by training with the historical operation data of material supply; represents the identity matrix; represents the learning rate; Represents the new loss function; represents the regularization parameter; Represents the updated index value multi-step prediction model parameters.

8. The multi-step prediction method for material supply system index values ​​according to claim 1 is characterized in that: The multi-step prediction model of the index value after the model parameters are updated is expressed as follows: ； in, Represents the multi-step forecast value of the indicator value at time t; Represents a multi-step prediction model of indicator values ​​built using a GRU network model with a local attention mechanism that is updated regularly; Represents the historical operation data characteristics of material supply at time t.

9. An electronic device, characterized in that: include: memory and at least one processor; Wherein, the memory stores a computer program; The at least one processor executes the computer program stored in the memory, so that the at least one processor executes the multi-step prediction method for indicator values ​​of a material supply system according to any one of claims 1 to 8.

10. A computer-readable storage medium, characterized in that: The computer-readable storage medium stores a computer program, which can be executed by a processor to implement the multi-step prediction method for indicator values ​​of a material supply system according to any one of claims 1 to 8.

Images