Digital production demand prediction method and system based on deep learning

Abstract

The embodiment of the application provides a kind of digital production demand prediction method and system based on deep learning, first, the historical production data set of target production scene is acquired, which contains multiple production related data units with time series labels, then the historical production data set is analyzed, and the demand influence factor graph containing entity node and relationship edge is constructed, the initial demand feature representation is generated based on the demand influence factor graph, then the initial demand feature representation is input into the pre-trained deep learning prediction model for processing, the deep learning prediction model is sequentially connected by graph convolution feature extraction layer, time series dependence modeling layer and multidimensional prediction output layer, and finally the future production demand prediction result of target production scene is generated, including product demand quantity trend description, material consumption fluctuation range estimation and equipment capacity allocation direction guide, which helps to improve the production efficiency of enterprise and reduce production cost.

Description

Technical Field

[0001] This application relates to the field of deep learning technology, and more specifically, to a method and system for predicting digital production demand based on deep learning. Background Technology

[0002] In the field of digital manufacturing, accurate production demand forecasting is crucial for enterprises' production planning, resource allocation optimization, and cost control. Most existing production demand forecasting methods are based on traditional statistical models or simple machine learning algorithms. These methods primarily rely on historical sales data and simple market trend analysis, often neglecting the complex factors involved in the production process. For example, factors such as material consumption, equipment operating status, market demand feedback, and changes in product inventory all significantly impact production demand. However, existing methods struggle to comprehensively and effectively integrate these diverse production-related data and fail to delve into the inherent correlations and dependencies between these data. Furthermore, traditional methods struggle to capture the dynamic changes in production demand over time when processing time-series production data, resulting in low accuracy and reliability of forecasts, failing to meet the precise forecasting requirements of digital manufacturing. Summary of the Invention

[0003] In view of this, the purpose of this application is to provide a method and system for predicting digital production demand based on deep learning.

[0004] According to a first aspect of this application, a method for predicting digital production demand based on deep learning is provided, the method comprising:

[0005] Obtain a set of historical production data for the target production scenario. The set of historical production data includes multiple types of production-related data units with time-series labels. These multiple types of production-related data units include production material consumption data, equipment operating status data, market demand feedback data, and product inventory change data.

[0006] The historical production data set is analyzed to construct a demand influencing factor map containing entity nodes and relational edges. The entity nodes correspond to key influencing factors in production-related data units, and the relational edges represent the degree of correlation and dependence between key influencing factors.

[0007] An initial demand feature representation is generated based on the demand influencing factor map. The initial demand feature representation includes a set of node feature vectors and a set of relation feature matrices. The set of node feature vectors corresponds to the attribute features of entity nodes, and the set of relation feature matrices corresponds to the association strength features of relation edges.

[0008] The initial demand features are input into a pre-trained deep learning prediction model for processing, generating a future production demand prediction result for the target production scenario. The deep learning prediction model includes a graph convolutional feature extraction layer, a temporal dependency modeling layer, and a multi-dimensional prediction output layer connected in sequence. The future production demand prediction result includes a description of product demand quantity trends, an estimate of material consumption fluctuation range, and guidance on equipment capacity allocation direction.

[0009] According to a second aspect of this application, a deep learning-based digital production demand forecasting system is provided. The deep learning-based digital production demand forecasting system includes a processor and a readable storage medium. The readable storage medium stores a program that, when executed by the processor, implements the aforementioned deep learning-based digital production demand forecasting method.

[0010] Based on any of the above aspects, by acquiring a historical production data set containing multiple types of production-related data units with time-series labels, and conducting in-depth analysis of this historical production data set to construct a demand influencing factor map, it is possible to comprehensively and systematically integrate various key factors in the production process, uncover the complex correlations and dependencies hidden behind the data, and generate an initial demand feature representation based on this demand influencing factor map. This initial demand feature representation includes a set of node feature vectors and a set of relation feature matrices. The initial demand feature representation is then input into a pre-trained deep learning prediction model for processing. This model includes a graph convolution feature extraction layer, a temporal dependency modeling layer, and a multi-dimensional prediction output layer connected in sequence. This model can effectively capture the dynamic changes in production demand in time and space, and achieve accurate prediction of future production demand in the target production scenario. The final future production demand prediction results include a description of product demand quantity trends, an estimate of material consumption fluctuation ranges, and guidance on equipment capacity allocation, which helps improve enterprise production efficiency and reduce production costs. Attached Figure Description

[0011] Figure 1 A flowchart illustrating the deep learning-based digital production demand forecasting method provided in this application embodiment is shown.

[0012] Figure 2 This illustration shows a schematic diagram of the component structure of a deep learning-based digital production demand forecasting system provided in an embodiment of this application. Detailed Implementation

[0013] Figure 1The diagram illustrates a flowchart of a deep learning-based digital production demand forecasting method provided in an embodiment of this application. It should be understood that in other embodiments, the order of some steps in the deep learning-based digital production demand forecasting method may be interchanged according to actual needs, or some steps may be omitted or deleted. The detailed steps of this deep learning-based digital production demand forecasting method are described below.

[0014] Step S110: Obtain the historical production data set of the target production scenario. The historical production data set contains multiple types of production-related data units with time series labels. The multiple types of production-related data units include production material consumption data, equipment operating status data, market demand feedback data, and product inventory change data.

[0015] In this embodiment, the target production scenario is a shoe production scenario. To obtain a historical production data set, information needs to be collected from multiple data sources. Production material consumption data can be obtained from the factory's material management system. This material management system records the various materials consumed in each batch of shoe production, such as leather, fabric, shoelaces, and glue, and marks the production time for each record. For example, the area of ​​leather and the length of fabric used in producing athletic shoes within a specific time period are all arranged in chronological order of production time.

[0016] The equipment operation status data comes from the equipment monitoring system, which monitors various operating parameters of the shoemaking equipment in real time. For example, the sewing speed and temperature of the sewing machine, and the pressure and number of cuts of the cutting machine. The operation status data of different equipment at different production times are accurately recorded, such as the running time and temperature fluctuations of the injection molding machine within a certain production cycle.

[0017] Market demand feedback data is collected through multiple channels, including market research and customer feedback platforms. Market research gathers information on changes in consumer demand for shoe styles, functions, and prices over different time periods. Customer feedback platforms record customer reviews, suggestions, and complaints about shoes, reflecting the actual market demand. For example, consumer evaluations of a particular athletic shoe's breathability, comfort, and price acceptance.

[0018] Product inventory change data is obtained from the inventory management system, which records the quantity and time of shoe receipts and departures, as well as the real-time inventory quantity. For example, it records the daily receipt quantity of newly produced shoes and the quantity shipped to distributors, thus displaying the dynamic changes in product inventory.

[0019] Step S120: Analyze the historical production data set and construct a demand influencing factor map containing entity nodes and relational edges. The entity nodes correspond to key influencing factors in production-related data units, and the relational edges represent the degree of correlation and dependence between key influencing factors.

[0020] In this embodiment, a comprehensive analysis is performed on the acquired historical production data set to construct a demand influencing factor map. By analyzing the historical production data, key influencing factors are identified as entity nodes, and the degree of correlation and dependence between these key influencing factors is determined, represented by relation edges.

[0021] Step S121: Extract key factors from various production-related data units in the historical production data set to identify core data elements that affect production demand as entity nodes. The core data elements include production order volume, material procurement cycle, equipment failure rate, market price fluctuation range, and customer order delivery cycle.

[0022] Step S1211: Perform sliding window analysis on the production material consumption data in chronological order, calculate the average rate and standard deviation of material consumption within each sliding window, and take the consumption data corresponding to the material type whose average rate and standard deviation of fluctuation exceed the preset range as the key factor of material consumption.

[0023] In this embodiment, a sliding window analysis method is used for production material consumption data. A fixed-size time window is set, and the window moves along the time series. Within each window, the average rate of material consumption and the standard deviation of fluctuation are calculated. For example, a month is set as a sliding window, and the consumption of a certain type of leather is statistically analyzed within this month. The average daily consumption quantity of the leather is calculated to obtain the average rate. At the same time, the fluctuation of the daily consumption quantity relative to the average rate is calculated to obtain the standard deviation of fluctuation. If the average rate and the standard deviation of fluctuation exceed a preset range, the consumption data corresponding to that leather is regarded as a key factor in material consumption. This is because when the average rate and the standard deviation of fluctuation exceed the preset range, it indicates that the consumption of the material has changed abnormally, which may have a significant impact on production demand.

[0024] Step S1212: Perform fault mode identification processing on the equipment operation status data, statistically analyze the fault occurrence frequency and fault repair time of different equipment in the historical production process, and take the equipment operation parameters with a fault occurrence frequency higher than the set frequency threshold and a longer fault repair time as key factors of equipment status.

[0025] In this embodiment, fault mode identification is performed on the equipment operating status data. First, the fault records of different equipment during historical production are analyzed to statistically analyze the fault frequency and repair time for each piece of equipment. For example, for sewing machines, the number of faults within a set time period is counted to calculate the fault frequency. Simultaneously, the time taken from the occurrence of each fault to its repair is recorded. If the fault frequency of a sewing machine exceeds a set frequency threshold and the repair time is long, the operating parameters of that sewing machine, such as sewing speed and temperature, are considered key factors in the equipment status. This is because equipment with a high fault frequency and long repair time will affect the continuity and efficiency of production, thereby impacting production demands.

[0026] Step S1213: Analyze the customer evaluation text in the market demand feedback data, extract the key evaluation indicators of customers on product functions, quality and price through natural language processing technology, and take the evaluation indicators that appear more frequently than the set frequency and have significant polarity matching of emotional polarity as key factors of market demand.

[0027] In this embodiment, natural language processing (NLP) technology is used to analyze customer review text in the market demand feedback data. First, preprocessing operations such as word segmentation and part-of-speech tagging are performed on the customer review text. Then, customer evaluation indicators regarding the shoe's function, quality, and price are extracted. For example, customers might evaluate functional aspects such as breathability, comfort, and durability, as well as quality aspects such as workmanship and materials, and their opinions on whether the price is reasonable. The frequency of these evaluation indicators is statistically analyzed, along with the sentiment polarity of the evaluations—whether they are positive, negative, or neutral. If the frequency of a certain evaluation indicator exceeds a set frequency, and the sentiment polarity matches a significant polarity condition (e.g., very strong positive reviews or very prominent negative reviews), then that evaluation indicator is considered a key factor in market demand. For example, if many customers frequently mention the poor breathability of a certain shoe and give it a negative evaluation, then "breathability" can be considered a key factor in market demand.

[0028] Step S1214: Monitor the inventory turnover rate and safety stock level in the product inventory change data, and take the inventory influencing factors corresponding to the time points when the fluctuation of the inventory turnover rate exceeds the set fluctuation threshold and the safety stock level exceeds the limit threshold as the key factors of inventory change.

[0029] In this embodiment, product inventory change data is monitored, with a focus on inventory turnover rate and safety stock level. Inventory turnover rate reflects the speed at which products move through inventory, calculated as the ratio of the number of products shipped out to the average inventory level over a set time period. Safety stock level is the minimum inventory level set to cope with uncertainties. Monitoring fluctuations in inventory turnover rate is crucial. If, at a certain point in time, the fluctuation exceeds a set fluctuation threshold, it indicates a significant change in product sales. Simultaneously, if the safety stock level exceeds a certain threshold, it may signify inventory backlog. Inventory influencing factors corresponding to these time points, such as production plans and sales strategies, are considered key factors for inventory changes. For example, if the inventory turnover rate of a certain shoe suddenly drops sharply and the safety stock level exceeds the threshold, it may be due to unreasonable production planning or inaccurate market demand forecasting; these influencing factors are considered key factors for inventory changes.

[0030] Step S1215: Integrate the key factors of material consumption, key factors of equipment status, key factors of market demand, and key factors of inventory change to generate a set of entity nodes containing multi-dimensional features.

[0031] In this embodiment, the previously extracted key factors for material consumption, equipment status, market demand, and inventory changes are integrated. These key factors have different characteristics and dimensions, representing different aspects of production demand. By integrating these key factors, a set of entity nodes containing multi-dimensional features is generated. For example, key factors such as consumption data for a certain type of leather, operating parameters of a sewing machine, customer evaluations of shoe breathability, and production plans when the inventory turnover rate of a certain shoe model is abnormal are integrated to form an entity node. This entity node integrates information from multiple aspects such as production materials, equipment, market, and inventory, reflecting more comprehensively the factors affecting production demand.

[0032] Step S122: Establish association relationship determination rules between entity nodes, and calculate the correlation coefficient of any two entity nodes in the time series based on the association relationship determination rules.

[0033] In this embodiment, rules for determining the relationships between entity nodes are established. These rules are based on the inherent logical connections between the core data elements represented by the entity nodes. For example, there may be a positive correlation between production order volume and material procurement cycle time, as an increase in production order volume may lead to a shorter material procurement cycle. The rules for determining the relationships between entity nodes are determined by analyzing historical production data.

[0034] Based on this association determination rule, the correlation coefficient between any two entity nodes in the time series is calculated. In time series data, each entity node has its corresponding timestamp and data value. By comparing the data changes of two entity nodes at the same point in time, the correlation coefficient between them is calculated. For example, comparing the monthly data changes of production order volume and material procurement cycle yields the correlation coefficient between them. This correlation coefficient reflects the degree of association between the two entity nodes.

[0035] Step S123: Establish a relationship edge between entity node pairs whose correlation coefficient exceeds the set correlation coefficient threshold, wherein the weight parameter of the relationship edge is determined by the corresponding correlation coefficient value.

[0036] In this embodiment, a correlation coefficient threshold is set. When the correlation coefficient between any two entity nodes exceeds this threshold, it indicates a strong correlation between the two entity nodes. In this case, a relationship edge is established between the two entity nodes. The weight parameter of the relationship edge is determined by the corresponding correlation coefficient value; the larger the correlation coefficient, the higher the weight of the relationship edge. For example, if the correlation coefficient between production order quantity and material procurement cycle exceeds the threshold, a relationship edge is established between the entity node representing production order quantity and the entity node representing material procurement cycle, and the weight of the relationship edge is determined based on the magnitude of the correlation coefficient. A higher weight indicates a stronger degree of dependence between the two entity nodes.

[0037] Step S124: Use a graph database to store the entity nodes and relation edges to generate a dynamic demand influencing factor graph with timestamps.

[0038] In this embodiment, a graph database is used to store entity nodes and relationship edges. A graph database is a database specifically designed for storing graph-structured data, making it ideal for storing demand influencing factor graphs, which contain nodes and edges. The previously generated entity nodes and relationship edges are stored in the graph database, with a timestamp added to each. The timestamp records the creation time or data update time of each node and edge, giving the demand influencing factor graph a dynamic nature. For example, when the production order quantity changes, the data of the entity node representing the production order quantity is updated, and the update time is recorded. Thus, the demand influencing factor graph stored in the graph database can reflect the relationships between production demand influencing factors at different points in time.

[0039] Step S125: Standardize the dynamic demand influencing factor map to unify the attribute feature dimensions of entity nodes and the weight parameter range of relation edges.

[0040] In this embodiment, the dynamic demand influencing factor map is standardized. Since the attribute characteristics of different entity nodes may have different dimensions and value ranges, and the weight parameters of relationship edges may also differ, the dynamic demand influencing factor map needs to be standardized to facilitate subsequent analysis and processing. For the attribute characteristics of entity nodes, their dimensions are unified. For example, the dimensions of different material consumption key factors are adjusted to ensure they have the same feature representation. For the weight parameters of relationship edges, their ranges are unified to a set interval. Through standardization, the entity nodes and relationship edges in the dynamic demand influencing factor map have unified feature representations and weight ranges, improving the comparability and analysis efficiency of the map.

[0041] Step S130: Generate an initial demand feature representation based on the demand influencing factor map. The initial demand feature representation includes a set of node feature vectors and a set of relation feature matrices. The set of node feature vectors corresponds to the attribute features of entity nodes, and the set of relation feature matrices corresponds to the association strength features of relation edges.

[0042] In this embodiment, an initial demand feature representation is generated based on the constructed demand influencing factor map. This initial demand feature representation includes a set of node feature vectors and a set of relationship feature matrices, corresponding to the attribute features of entity nodes and the association strength features of relationship edges, respectively. By transforming the demand influencing factor map into a feature representation, it can be more conveniently input into subsequent deep learning prediction models for processing.

[0043] Step S131: Perform attribute feature encoding processing on each entity node in the demand influencing factor map, and convert the historical time series data of the entity node into a fixed-dimensional numerical feature vector. The numerical feature vector contains the feature values ​​of the entity node at different time points and the slope of its changing trend.

[0044] In this embodiment, attribute feature encoding is performed on each entity node in the demand influencing factor map. Each entity node has its corresponding historical time series data, which records the feature values ​​of the entity node at different points in time. For example, the historical time series data of an entity node representing the consumption of a certain type of leather records the monthly consumption quantity of that leather. This historical time series data is processed and converted into a fixed-dimensional numerical feature vector.

[0045] First, extract the feature values ​​of the entity nodes at different time points. For example, extract the monthly consumption quantity of the leather. Then, calculate the slope of the trend of these feature values. The slope of the trend reflects the speed and direction of the feature value change over time. For example, if the slope of the trend of leather consumption quantity in a certain month is positive, it indicates that the consumption quantity is increasing; if it is negative, it indicates that the consumption quantity is decreasing. Combine the feature values ​​of the entity nodes at different time points and the slope of the trend to form a numerical feature vector. This numerical feature vector has fixed dimensions, which facilitates subsequent analysis and processing.

[0046] Step S132: Use a graph embedding algorithm to extract features from the relation edges of the demand influencing factor graph, and generate a relation feature matrix that can characterize the correlation strength between entity nodes. The element values ​​of the relation feature matrix correspond to the weight parameters and correlation direction attributes of the relation edges.

[0047] In this embodiment, a graph embedding algorithm is used to extract features from the relation edges of the demand influencing factor graph. Graph embedding is a method for converting graph-structured data into a low-dimensional vector representation. Through graph embedding, the relation edges of the demand influencing factor graph are converted into a relation feature matrix that characterizes the strength of the association between entity nodes.

[0048] The elements of the relation feature matrix correspond to the weight parameters and association direction attributes of the relation edges. For example, an element in the relation feature matrix represents the association strength between two entity nodes, and its value is determined by the weight parameters of the relation edge. Meanwhile, the position and sign of the element indicate the association direction. If there is a positive association between two entity nodes, the element's value is positive; if there is a negative association, the element's value is negative. In this way, the relation feature matrix can more accurately reflect the association relationships between entity nodes.

[0049] Step S133: Establish the corresponding mapping relationship between entity nodes and time series, and add timestamp encoding to the feature vector of each entity node.

[0050] In this embodiment, a mapping relationship is established between entity nodes and time series. The feature vector of each entity node is associated with a specific time point. By establishing the above mapping relationship, the position of the feature vector of each entity node in the time series can be clearly defined. For example, the feature vector of an entity node representing the consumption of a certain type of leather is mapped to the time point corresponding to the leather consumption data.

[0051] Add a timestamp encoding to the feature vector of each entity node. The timestamp encoding records the time information corresponding to the feature vector. By adding timestamp encoding, the feature vector contains information in the time dimension, more comprehensively reflecting the feature changes of the entity node. For example, adding month information to the feature vector indicates which month's data the feature vector was generated from.

[0052] Step S134: Concatenate the entity node feature vectors and relation feature matrices with added timestamp encoding to generate an initial feature data block containing spatiotemporal correlation information.

[0053] In this embodiment, the entity node feature vectors with added timestamp encoding and the relationship feature matrix are concatenated. The entity node feature vectors contain the attribute features and time information of the entity nodes, while the relationship feature matrix reflects the strength of the association between entity nodes. By concatenating these two parts, an initial feature data block containing spatiotemporal association information is generated.

[0054] For example, the feature vector of an entity node representing the consumption of a certain type of leather and the feature matrix of the relationship between that entity node and other entity nodes are concatenated. The concatenated initial feature data block contains not only the attribute features and time information of leather consumption, but also the association information with other entity nodes, thus more comprehensively reflecting the spatiotemporal characteristics of production demand.

[0055] Step S135: Standardize the initial feature data block by adjusting the feature values ​​of different dimensions to the same numerical range using a standardization method.

[0056] In this embodiment, the initial feature data block is standardized. Since the initial feature data block contains feature values ​​of different dimensions, the ranges of these feature values ​​may vary significantly. To facilitate subsequent analysis and processing, it is necessary to adjust the feature values ​​of different dimensions to the same numerical range.

[0057] Standardization methods are employed to process each feature value in the initial feature data block. For example, normalization is used to map the feature values ​​to a defined range, such as between 0 and 1. Standardization makes feature values ​​from different dimensions comparable, improving the model's training performance and prediction accuracy.

[0058] Step S140: Input the initial demand feature representation into the pre-trained deep learning prediction model for processing, and generate the future production demand prediction result of the target production scenario. The deep learning prediction model includes a graph convolution feature extraction layer, a temporal dependency modeling layer and a multi-dimensional prediction output layer connected in sequence. The future production demand prediction result includes a description of the product demand quantity trend, an estimate of the fluctuation range of material consumption and guidance on the direction of equipment capacity allocation.

[0059] In this embodiment, the generated initial demand feature representation is input into a pre-trained deep learning prediction model for processing. This deep learning prediction model consists of a graph convolutional feature extraction layer, a temporal dependency modeling layer, and a multi-dimensional prediction output layer connected in sequence. Through model processing, future production demand prediction results for the target production scenario are generated, including a description of product demand trends, an estimate of material consumption fluctuation ranges, and guidance on equipment capacity allocation directions.

[0060] Step S141: The graph convolution feature extraction layer performs graph convolution operation on the set of node feature vectors and the set of relation feature matrices in the initial demand feature representation to extract the production demand association features between entity nodes.

[0061] Step S1411: Construct an adjacency matrix based on factors influencing production demand. The adjacency matrix is ​​generated according to the relationship edge connection of entity nodes, and the element values ​​of the adjacency matrix represent the connection status between factors influencing production demand.

[0062] In this embodiment, an adjacency matrix based on factors influencing production demand is constructed. The adjacency matrix is ​​a two-dimensional matrix where each element represents the connection status between entity nodes. It is generated based on the connection relationships between entity nodes in the demand influencing factor graph. If two entity nodes have a relational edge, the corresponding element in the adjacency matrix is ​​1, indicating a connection; otherwise, the element is 0, indicating no connection. For example, if an entity node representing production order quantity and an entity node representing material procurement cycle have a relational edge, then the corresponding element in the adjacency matrix for these two entity nodes is 1.

[0063] Step S1412: Standardize the adjacency matrix by using a symmetric standardization method to eliminate the influence of differences in node degree due to different production needs on graph convolution operations, and obtain the standardized adjacency matrix.

[0064] In this embodiment, the constructed adjacency matrix is ​​standardized. Different entity nodes may have different degrees, which represent the number of edges connected to that node. Differences in node degree can affect graph convolution operations. A symmetric standardization method is used to process the adjacency matrix, eliminating the impact of differences in node degree. This standardization process makes the adjacency matrix more suitable for graph convolution operations, improving the model's performance.

[0065] Step S1413: Perform matrix multiplication on the set of node feature vectors and the standardized adjacency matrix to obtain a first-order neighborhood feature aggregation result containing information on the correlation between production order quantity and material consumption.

[0066] In this embodiment, the set of node feature vectors obtained after attribute feature encoding is multiplied by the standardized adjacency matrix. Each vector in the set of node feature vectors represents the attribute feature of an entity node, while the standardized adjacency matrix reflects the connection relationship between entity nodes. Through matrix multiplication, the feature information of adjacent nodes can be aggregated.

[0067] Taking a shoe manufacturing scenario as an example, the set of node feature vectors might contain feature vectors representing entity nodes such as production order volume, material procurement cycle, and equipment failure rate. The standardized adjacency matrix indicates the connection status between these entity nodes. When performing matrix multiplication, for each node, the feature vectors of its neighboring nodes are weighted and summed according to the weights in the adjacency matrix. For example, for a node representing production order volume, the feature vectors of its connected nodes representing material procurement cycle can be weighted according to the values ​​of the corresponding elements in the adjacency matrix, and then combined with its own feature vector to obtain a first-order neighborhood feature aggregation result that includes information on the correlation between production order volume and material consumption. This first-order neighborhood feature aggregation result integrates the feature information of neighboring nodes, and can more comprehensively reflect the production demand correlation between nodes.

[0068] Step S1414: Introduce a weight matrix to adjust the parameters of the first-order neighborhood feature aggregation result. The weight matrix is ​​learned and optimized during the training process of the production demand prediction model.

[0069] In this embodiment, a weight matrix is ​​introduced to further adjust the first-order neighborhood feature aggregation results. The weight matrix is ​​a learnable parameter matrix whose element values ​​are learned and optimized during the training process of the production demand prediction model. The role of the weight matrix is ​​to weight different features in the first-order neighborhood feature aggregation results, highlighting important feature information and suppressing unimportant feature information.

[0070] For example, in a shoe manufacturing scenario, for the aggregation results of first-order neighborhood features related to production order quantity and material consumption, the weight matrix assigns different weights to features related to material consumption. If, during training, it is found that certain material consumption features have a more significant impact on production demand, the weight matrix adjusts the weights of these features to give them a greater role in subsequent processing. Through continuous model training, the element values ​​of the weight matrix are gradually optimized, enabling the model to more accurately capture the features related to production demand.

[0071] Step S1415: The feature aggregation result after weight matrix adjustment is concatenated with the original node feature vector to generate a new set of node feature vectors containing multi-level production demand related features.

[0072] In this embodiment, the feature aggregation result adjusted by the weight matrix is ​​concatenated with the original node feature vector. The original node feature vector represents the node's own attribute features, while the feature aggregation result adjusted by the weight matrix contains the association feature information of adjacent nodes. By concatenating these two parts, a new set of node feature vectors containing multi-level production demand association features can be generated.

[0073] Taking shoe production as an example, the original node feature vector may contain historical data on production order volume and basic information on material procurement cycles. The feature aggregation result after weight matrix adjustment includes the correlation feature information of nodes adjacent to production order volume (such as material procurement cycles, equipment failure rates, etc.). By concatenating these two parts, the new node feature vector set not only includes the node's own attribute features but also first-order and higher-order neighborhood correlation features, which can more comprehensively reflect the correlation of production demand.

[0074] Step S142: Input the production demand association features obtained by graph convolution operation into the time-series dependency modeling layer, and capture the periodic fluctuation pattern of production demand in the time dimension through the recurrent neural network structure to obtain feature vectors at different production cycle time points.

[0075] Step S1421: Input the production demand association features into the input layer of the recurrent neural network in the order of the production cycle time series. The input at each time step corresponds to the demand association feature vector of a production cycle.

[0076] In this embodiment, the production demand correlation features obtained by graph convolution operations are input into the input layer of a recurrent neural network in the time series order of the production cycle. In the shoe production scenario, the production cycle can be one month or one quarter, etc. The input at each time step is a demand correlation feature vector of a production cycle. This demand correlation feature vector contains various correlation information of production demand within that production cycle, such as the correlation between production order quantity and material consumption, and the correlation between equipment failure rate and production efficiency.

[0077] For example, the first time step input is the demand-related feature vector of the first production cycle, which integrates related information such as production order volume, material procurement status, and equipment operating status within that cycle. Subsequent time steps input demand-related feature vectors of subsequent production cycles, and the recurrent neural network processes these input vectors sequentially according to time order.

[0078] Step S1422: In the hidden layer of the recurrent neural network, the demand information of the historical production cycle is selectively retained or forgotten using the gating mechanism of the long short-term memory unit.

[0079] In this embodiment, the hidden layer of the recurrent neural network uses Long Short-Term Memory (LSTM) units. LSTM has a gating mechanism, including an input gate, a forget gate, and an output gate. The forget gate selectively forgets demand information from historical production cycles. In the shoe production scenario, some demand information from historical production cycles may no longer be important for predicting the current production cycle. The forget gate determines whether to forget this historical information based on the current input demand-related feature vector and the hidden state from the previous time step.

[0080] For example, if market demand in a past production cycle fluctuated abnormally due to a special holiday, but the current production cycle does not have similar special circumstances, the forget gate might choose to forget the demand information for that cycle. The input gate is responsible for deciding which information from the current input to add to the cell state to update it. The output gate decides which information to output as the hidden state at the current moment based on the cell state and the current input. Through this gating mechanism, LSTM can effectively process long-sequence data and capture the long-term dependencies of production demand over time.

[0081] Step S1423: Calculate the hidden state of the current production cycle. The hidden state is determined by the current input demand-related feature vector, the hidden state of the previous production cycle, and the output of the forget gate.

[0082] In this embodiment, the calculation of the hidden state of the current production cycle is jointly determined by the currently input demand-related feature vector, the hidden state of the previous production cycle, and the output of the forget gate. In the shoe production scenario, the currently input demand-related feature vector contains related information such as the production order quantity, material consumption, and equipment operation within this production cycle. The hidden state of the previous production cycle is the result of the recurrent neural network output at the previous moment, which contains historical information from previous production cycles.

[0083] The output of the forget gate determines which historical information is retained and which is forgotten. The demand-related feature vector of the current input interacts with the cell state through the input gate, updating the cell state. Then, the output gate determines which information to output as the hidden state of the current production cycle based on the updated cell state and the current input. For example, if the production order volume of the current production cycle increases significantly, the input gate will add this information to the cell state, while the forget gate may forget some historical information related to low order volume. The final hidden state of the current production cycle combines the current input and historical information.

[0084] Step S1424: Output the hidden state of each production cycle to generate a hidden state sequence containing time series dynamic information.

[0085] In this embodiment, the recurrent neural network processes the input for each production cycle to obtain the corresponding hidden state. Outputting the hidden states of each production cycle sequentially generates a hidden state sequence containing dynamic time-series information. In the shoe production scenario, this hidden state sequence records the dynamic changes in production demand-related information across different production cycles.

[0086] For example, the hidden state of the first production cycle reflects the correlation between production order volume, material consumption, and other factors within that cycle, while the hidden state of the second production cycle reflects the changes in these factors within that cycle. Through the hidden state sequence, the periodic fluctuations in production demand over time can be observed, such as whether production order volume exhibits seasonal fluctuations and whether material consumption changes regularly with changes in production scale.

[0087] Step S1425: In the output layer of the recurrent neural network, the hidden state sequence is weighted and summed to obtain the feature vectors at different production cycle time points. The weight parameters are dynamically adjusted according to the importance of different production cycles to future demand prediction.

[0088] In this embodiment, the generated hidden state sequence is weighted and summed in the output layer of the recurrent neural network. The weight parameters are dynamically adjusted based on the importance of different production cycles to future demand forecasting. In the shoe production scenario, the importance of different production cycles to future demand forecasting may vary. For example, production cycles closer to the current time may be more important for future demand forecasting because they better reflect current market trends and production conditions.

[0089] Therefore, hidden states in production cycles closer to the current time can be assigned higher weights, while hidden states in production cycles further back in time can be assigned lower weights. By weighted summation, different hidden states in the hidden state sequence are combined according to their weights to obtain feature vectors for different production cycle time points. These feature vectors integrate information from different production cycles and highlight the characteristics of production cycles that are important for predicting future demand, thus more accurately reflecting the patterns of production demand over time.

[0090] Step S143: Through the attention mechanism module set at the output end of the time-series dependency modeling layer, the feature vectors at different production cycle time points are weighted and feature fusion processed to obtain the time-series feature vectors processed by the attention mechanism.

[0091] In this embodiment, an attention mechanism module is set at the output of the temporal dependency modeling layer. The role of the attention mechanism is to assign weights and fuse features from feature vectors at different production cycle time points. In the shoe production scenario, feature vectors from different production cycles may contain different important information, and the attention mechanism can weight the feature vectors according to the importance of this information.

[0092] For example, some feature vectors from production cycles might contain important information about changes in market trends, while others might contain information about improvements in production processes. The attention mechanism automatically learns the importance weights of these feature vectors, assigning higher weights to important feature vectors and lower weights to less important ones. These weighted feature vectors are then fused to obtain a time-series feature vector processed by the attention mechanism. This time-series feature vector highlights the most important information for production demand forecasting and can be used more effectively for subsequent forecasts.

[0093] Step S144: Activate the fully connected neural network sub-layer corresponding to the prediction target dimension of the multi-dimensional prediction output layer. The prediction target dimension includes the quantity of product demand, the quantity of material consumption, and the quantity of equipment capacity demand. In this step, an independent fully connected neural network sub-layer is set up in advance for each prediction target dimension. The input of the fully connected neural network sub-layer is the time-series feature vector containing production cycle features output by the time-series dependency modeling layer.

[0094] Step S1441: Set a first fully connected neural network sub-layer for the target dimension of product demand quantity prediction. The number of neurons in the first fully connected neural network sub-layer is determined based on the feature dimensions of historical product order data.

[0095] In this embodiment, a first fully connected neural network sublayer is set up for the target dimension of product demand quantity prediction. The number of neurons in this sublayer is determined based on the feature dimensions of historical product order data. In the shoe production scenario, historical product order data may include features such as the order quantity of different shoe styles, the order distribution in different seasons, and the order preferences of different customer groups.

[0096] For example, if the historical product order data has high feature dimensions, including order information for various styles and different time periods, then the number of neurons in the first fully connected neural network sublayer may be correspondingly large to be able to process and learn this complex feature information. The input to this first fully connected neural network sublayer is a temporal feature vector containing production cycle features output by the temporal dependency modeling layer, which can perform further feature transformation and processing on the temporal feature vector to predict the quantity of product demand.

[0097] Step S1442: Set a second fully connected neural network sub-layer for the target dimension of material consumption quantity prediction. The number of neurons in the second fully connected neural network sub-layer is consistent with the number of main production material types.

[0098] In this embodiment, a second fully connected neural network sublayer is set up for the target dimension of material consumption quantity prediction. The number of neurons in this sublayer is consistent with the number of main production material types. In the shoe production scenario, the main production materials may include leather, fabric, shoelaces, glue, etc.

[0099] Each neuron corresponds to a primary production material and is used to predict the consumption quantity of that material. The input to the second fully connected neural network sublayer is also a temporal feature vector containing production cycle characteristics, output from the temporal dependency modeling layer. The second fully connected neural network sublayer then predicts the consumption quantity of each primary production material based on the temporal feature vector and its own neuron connection weights.

[0100] Step S1443: Set a third fully connected neural network sub-layer for the target dimension of equipment capacity demand prediction. The input features of the third fully connected neural network sub-layer include historical correlation features of equipment operating status.

[0101] In this embodiment, a third fully connected neural network sublayer is set up for the target dimension of predicting equipment production capacity demand. The input features of this sublayer include historical correlation features of equipment operating status. In the shoe production scenario, the historical correlation features of equipment operating status may include information such as equipment failure rate, maintenance time, and operating efficiency.

[0102] The third fully connected neural network sublayer combines these historical correlation features with the temporal feature vectors containing production cycle characteristics output by the temporal dependency modeling layer to predict the required equipment capacity. For example, if a piece of equipment has a high failure rate, more backup equipment or higher equipment capacity may be needed to ensure normal production. The third fully connected neural network sublayer will make a comprehensive judgment and prediction based on this information.

[0103] Step S1444: Initialize the weight parameters of each fully connected neural network sublayer by using a uniform distribution random initialization method to ensure that the initial values ​​of the weight parameters are within a reasonable range.

[0104] In this embodiment, the weight parameters of the first, second, and third fully connected neural network sublayers are initialized. A uniformly distributed random initialization method is used to ensure that the initial values ​​of the weight parameters are within a reasonable range. Uniformly distributed random initialization refers to randomly selecting the initial values ​​of the weight parameters within a pre-defined interval.

[0105] For example, a reasonable interval can be defined, such as [-a, a], where a is a positive number determined empirically or experimentally. Within this interval, weight parameters for each fully connected neural network sublayer are randomly assigned values. This ensures that the initial values ​​of the weight parameters are neither too large nor too small, which is beneficial for model training and convergence. During model training, the weight parameters of each fully connected neural network sublayer are optimized using the backpropagation algorithm, with the mean squared error between the actual production data and the predicted data as the optimization objective.

[0106] Step S145: Perform feature transformation processing on the time-series feature vector through the fully connected neural network sublayer to generate a predictive feature vector that matches product order fluctuations, material procurement plans, and equipment capacity allocation.

[0107] In this embodiment, each fully connected neural network sublayer performs feature transformation processing on the input temporal feature vector. In the shoe production scenario, the first fully connected neural network sublayer processes the temporal feature vector to generate a predicted feature vector that matches the fluctuations in product orders. This predicted feature vector contains information about the future trend of product demand, such as whether the order quantity is increasing or decreasing, and the magnitude of the increase or decrease.

[0108] The second fully connected neural network sublayer transforms the time-series feature vector to generate a predicted feature vector that matches the material procurement plan. This predicted feature vector contains forecast information on the consumption quantities of various major production materials, which can be used to formulate a reasonable material procurement plan. The third fully connected neural network sublayer generates a predicted feature vector that matches equipment capacity allocation. This predicted feature vector contains forecast information on the required equipment capacity, which helps to rationally allocate equipment resources and improve production efficiency.

[0109] Step S146: Apply activation function processing to the predicted feature vector, and use a linear activation function to generate continuous predicted values.

[0110] In this embodiment, the generated predicted feature vector is processed by an activation function. A linear activation function is used, which transforms the predicted feature vector linearly to generate continuous predicted values. In the shoe manufacturing scenario, for predicted feature vectors that match product order fluctuations, material procurement plans, and equipment capacity allocation, continuous predicted values ​​can be obtained after processing with the linear activation function.

[0111] For example, the predictive feature vector of product order fluctuations, after being processed by a linear activation function, yields a predicted value for the specific number of product orders. This predicted value is a continuous numerical value, which can more accurately reflect the actual situation of product demand. Similarly, the predictive feature vectors of material procurement plans and equipment capacity allocation will also yield continuous predicted values, which can be used to guide actual production management decisions.

[0112] Step S147: Integrate the output results of multiple prediction target dimensions and generate a future production demand prediction result containing multi-dimensional prediction information according to the input specifications of the production management system.

[0113] In this embodiment, the prediction results output from the first, second, and third fully connected neural network sublayers, which match product order fluctuations, material procurement plans, and equipment capacity allocation, are integrated. In the shoe production scenario, these prediction results represent forecast information on product demand quantity, material consumption quantity, and equipment capacity demand quantity, respectively.

[0114] Following the input specifications of the production management system, these forecast results are formatted and combined to generate future production demand forecasts containing multi-dimensional information. For example, the production management system may require forecast information to be input in a set format, such as arranged by time series and forecast target dimension. Integrating the forecast results from different forecast target dimensions according to this input specification yields future production demand forecasts that include descriptions of product demand trends, estimates of material consumption fluctuation ranges, and guidance on equipment capacity allocation, which can be directly used to guide production decisions in shoe manufacturing plants.

[0115] In the above embodiments, during the data acquisition phase, encryption technology is used to process data involving privacy-sensitive information, such as customer personal information and detailed order records. For example, a symmetric encryption algorithm is used to encrypt customer identity information, and the encrypted data is stored in a secure database. During data transmission, secure transmission protocols, such as SSL / TLS, are used to ensure that data is not stolen or tampered with during transmission.

[0116] Regarding data storage, strict access control is implemented on the servers storing the data. Different user roles and permissions are set up, ensuring that only authorized personnel can access and process privacy-sensitive data. Data is also backed up regularly to prevent data loss. During data usage, privacy-sensitive data undergoes anonymization. For example, customer names and phone numbers are replaced or masked, retaining only essential features for production demand forecasting, ensuring data analysis and model training are conducted without compromising privacy.

[0117] Differential privacy technology is employed in the training process of the artificial intelligence model. Differential privacy technology adds noise to the training data to ensure that the model's output does not reveal the privacy information of individual data records. During model training, the amount of noise added is continuously adjusted to balance the model's accuracy and the degree of data privacy protection.

[0118] During the model deployment and usage phases, access to the model is monitored and audited. All access requests and operations to the model are recorded to promptly identify abnormal behavior and take appropriate measures. Simultaneously, the model's output is re-examined to ensure it does not contain any privacy-sensitive information.

[0119] Furthermore, Figure 2 A schematic diagram of the hardware structure of a deep learning-based digital production demand forecasting system 100 for implementing the methods provided in the embodiments of this application is shown. Figure 2 As shown, the deep learning-based digital production demand forecasting system 100 may include at least one processor 102 (the processor 102 may be, but is not limited to, a microprocessor MCU or a programmable logic device FPGA, etc.), a memory 104 for storing data, a transmission device 106 for communication functions, and a controller 108. Those skilled in the art will understand that... Figure 2 The structure shown is for illustrative purposes only and does not limit the structure of the deep learning-based digital production demand forecasting system 100. For example, the deep learning-based digital production demand forecasting system 100 may also include components that are more complex than... Figure 2 The more or fewer components shown, or having the same Figure 2 The different configurations shown.

[0120] The memory 104 can be used to store software programs and modules of the application software, such as the program instructions corresponding to the method embodiments described above in this application. The processor 102 executes various functional applications and data processing by running the software programs and modules stored in the memory 104, thereby realizing the above-described method for predicting digital production demand based on deep learning. The transmission device 106 is used to acquire or send data via a network.

[0121] Those skilled in the art will understand that all or part of the steps of the above embodiments can be implemented by hardware or by a program instructing related hardware. The program can be stored in a computer-readable storage medium, such as a read-only memory, a disk, or an optical disk.

Claims

1. A method for predicting digital production demand based on deep learning, characterized in that, The method includes: Obtain a historical production data set for the target production scenario. The historical production data set contains multiple types of production-related data units with time-series labels. The multiple types of production-related data units include production material consumption data, equipment operating status data, market demand feedback data, and product inventory change data. The historical production data set is analyzed to construct a demand influencing factor map containing entity nodes and relational edges. The entity nodes correspond to key influencing factors in production-related data units, and the relational edges represent the degree of correlation and dependence between key influencing factors. An initial demand feature representation is generated based on the demand influencing factor map. The initial demand feature representation includes a set of node feature vectors and a set of relation feature matrices. The set of node feature vectors corresponds to the attribute features of entity nodes, and the set of relation feature matrices corresponds to the association strength features of relation edges. The initial demand features are input into a pre-trained deep learning prediction model for processing, generating a future production demand prediction result for the target production scenario. The deep learning prediction model includes a graph convolutional feature extraction layer, a temporal dependency modeling layer, and a multi-dimensional prediction output layer connected in sequence. The future production demand prediction result includes a description of the product demand quantity trend, an estimate of the fluctuation range of material consumption, and guidance on the direction of equipment capacity allocation. The step of inputting the initial demand feature representation into a pre-trained deep learning prediction model for processing to generate a future production demand prediction result for the target production scenario includes: The graph convolution feature extraction layer performs graph convolution operations on the set of node feature vectors and the set of relation feature matrices in the initial demand feature representation to extract the production demand association features between entity nodes. The production demand-related features obtained by graph convolution operation are input into the time-series dependency modeling layer. The recurrent neural network structure captures the periodic fluctuation pattern of production demand in the time dimension and obtains feature vectors at different production cycle time points. The attention mechanism module set at the output of the time-series dependency modeling layer performs weight allocation and feature fusion processing on the feature vectors of different production cycle time points to obtain the time-series feature vectors processed by the attention mechanism. Activate the fully connected neural network sub-layer corresponding to the prediction target dimension of the multi-dimensional prediction output layer. The prediction target dimension includes the quantity of product demand, the quantity of material consumption, and the quantity of equipment capacity demand. In this way, an independent fully connected neural network sub-layer is set up in advance for each prediction target dimension. The input of the fully connected neural network sub-layer is the time-series feature vector containing production cycle features output by the time-series dependency modeling layer. The time-series feature vectors are transformed by the fully connected neural network sublayer to generate predictive feature vectors that match product order fluctuations, material procurement plans, and equipment capacity allocation. The predicted feature vector is processed by an activation function, and a continuous predicted value is generated by using a linear activation function; The output results of multiple prediction target dimensions are integrated to generate future production demand prediction results containing multi-dimensional prediction information according to the input specifications of the production management system. The step of performing graph convolution operations on the set of node feature vectors and the set of relation feature matrices in the initial demand feature representation through the graph convolution feature extraction layer to extract the production demand association features between entity nodes includes: Construct an adjacency matrix based on factors influencing production demand. The adjacency matrix is ​​generated according to the connection status of relational edges of entity nodes, and the element values ​​of the adjacency matrix represent the connection status between factors influencing production demand. The adjacency matrix is ​​standardized by using a symmetric standardization method to eliminate the influence of differences in node degree due to different production needs on graph convolution operations, resulting in a standardized adjacency matrix. The set of node feature vectors is multiplied by the standardized adjacency matrix to obtain a first-order neighborhood feature aggregation result containing information on the correlation between production order quantity and material consumption. A weight matrix is ​​introduced to adjust the parameters of the first-order neighborhood feature aggregation result. The weight matrix is ​​learned and optimized during the training process of the production demand prediction model. The feature aggregation result after weight matrix adjustment is concatenated with the original node feature vector to generate a new set of node feature vectors containing multi-level production demand-related features.

2. The deep learning-based digital production demand forecasting method according to claim 1, characterized in that, The analysis of the historical production data set, and the construction of a demand influencing factor graph including entity nodes and relationship edges, includes: Key factors are extracted from various production-related data units in the historical production data set to identify core data elements that affect production demand as entity nodes. These core data elements include production order volume, material procurement cycle, equipment failure rate, market price fluctuation range, and customer order delivery cycle. Establish rules for determining the association relationship between entity nodes, and calculate the correlation coefficient of any two entity nodes in the time series based on the association relationship determination rules; Establish a relationship edge between entity node pairs whose correlation coefficient exceeds a set correlation coefficient threshold, and the weight parameter of the relationship edge is determined by the corresponding correlation coefficient value. The entity nodes and relation edges are stored using a graph database to generate a dynamic demand influencing factor graph with timestamps; The dynamic demand influencing factor map is standardized to unify the attribute feature dimensions of entity nodes and the weight parameter range of relation edges.

3. The deep learning-based digital production demand forecasting method according to claim 2, characterized in that, The process of extracting key factors from various production-related data units in the historical production data set to identify core data elements affecting production demand as entity nodes includes: The production material consumption data is processed by sliding window analysis according to the time series order. The average rate and standard deviation of material consumption within each sliding window are calculated. The consumption data of material types whose average rate and standard deviation of fluctuation exceed the preset range are used as key factors of material consumption. The equipment operation status data is processed for fault mode identification, and the frequency of fault occurrence and fault repair time of different equipment in the historical production process are statistically analyzed. The equipment operation parameters with a fault occurrence frequency higher than a set frequency threshold and a longer fault repair time are used as key factors of equipment status. Analyze the customer evaluation text in the market demand feedback data, and extract key evaluation indicators of customers' product functions, quality and price through natural language processing technology. Evaluation indicators that appear more frequently than the set frequency and have a significant match in emotional polarity are taken as key factors of market demand. Monitor the inventory turnover rate and safety stock level in the product inventory change data, and take the inventory influencing factors corresponding to the time points when the fluctuation of the inventory turnover rate exceeds the set fluctuation threshold and the safety stock level exceeds the limit threshold as the key factors of inventory change. The key factors of material consumption, equipment status, market demand, and inventory changes are integrated to generate a set of entity nodes containing multi-dimensional features.

4. The deep learning-based digital production demand forecasting method according to claim 1, characterized in that, The generation of initial demand feature representation based on the demand influencing factor map includes: Each entity node in the demand influencing factor map is subjected to attribute feature encoding processing, and the historical time series data of the entity node is converted into a fixed-dimensional numerical feature vector. The numerical feature vector contains the feature values ​​of the entity node at different time points and the slope of its changing trend. A graph embedding algorithm is used to extract features from the relation edges of the graph of demand influencing factors to generate a relation feature matrix that can characterize the correlation strength between entity nodes. The element values ​​of the relation feature matrix correspond to the weight parameters and correlation direction attributes of the relation edges. Establish a mapping relationship between entity nodes and time series, and add timestamp encoding to the feature vector of each entity node; The entity node feature vectors and relation feature matrices with added timestamp encoding are concatenated to generate an initial feature data block containing spatiotemporal correlation information. The initial feature data block is standardized to adjust the feature values ​​of different dimensions to the same numerical range.

5. The deep learning-based digital production demand forecasting method according to claim 4, characterized in that, The step of encoding attribute features for each entity node in the demand influencing factor map, converting the historical time series data of the entity nodes into fixed-dimensional numerical feature vectors, includes: For each entity node, extract its time-series data within a preset time period during the historical production process, where the preset time period covers multiple production cycles; The time series data is smoothed and filtered, and a moving average algorithm is used to eliminate noise fluctuations in the data to obtain a smoothed time series curve. On the smoothed time series curve, select time sampling points at equal intervals, and extract the feature value of each sampling point and the difference between adjacent sampling points as basic features; Calculate the statistical characteristics of the time series data, including mean, variance, maximum value, minimum value, and kurtosis. The basic features and statistical features are fused to generate a fixed-dimensional numerical feature vector.

6. The deep learning-based digital production demand forecasting method according to claim 1, characterized in that, The method of capturing the periodic fluctuation pattern of production demand over time by using a recurrent neural network structure to obtain feature vectors at different production cycle time points includes: The production demand-related features are input into the input layer of the recurrent neural network in the order of the production cycle time series. The input at each time step corresponds to a demand-related feature vector of a production cycle. In the hidden layers of a recurrent neural network, the gating mechanism of long short-term memory units is used to selectively retain or forget demand information from historical production cycles. Calculate the hidden state of the current production cycle, which is determined by the current input demand-related feature vector, the hidden state of the previous production cycle, and the output of the forget gate. Output the hidden state of each production cycle to generate a hidden state sequence containing dynamic time-series information; In the output layer of the recurrent neural network, the hidden state sequence is weighted and summed to obtain feature vectors at different production cycle time points. The weight parameters are dynamically adjusted according to the importance of different production cycles to future demand prediction.

7. The deep learning-based digital production demand forecasting method according to claim 1, characterized in that, The process involves pre-setting an independent fully connected neural network sub-layer for each prediction target dimension. The input to the fully connected neural network sub-layer is the temporal feature vector output by the temporal dependency modeling layer, including: A first fully connected neural network sublayer is set for the target dimension of product demand quantity prediction. The number of neurons in the first fully connected neural network sublayer is determined based on the feature dimensions of historical product order data. A second fully connected neural network sublayer is set for the target dimension of material consumption quantity prediction. The number of neurons in the second fully connected neural network sublayer is consistent with the number of major production material types. A third fully connected neural network sublayer is set for the target dimension of equipment capacity demand prediction. The input features of the third fully connected neural network sublayer include historical correlation features of equipment operating status. The weight parameters of each fully connected neural network sublayer are initialized using a uniform distribution random initialization method to ensure that the initial values ​​of the weight parameters are within a reasonable range. During model training, the weight parameters of each fully connected neural network sublayer are optimized using the backpropagation algorithm, with the mean square error between the actual production demand data and the predicted data as the optimization target.

8. A deep learning-based digital production demand forecasting system, characterized in that, It includes a processor and a readable storage medium storing a program that, when executed by the processor, implements the deep learning-based digital production demand forecasting method according to any one of claims 1-7.

Images