Production scheduling strategy adjusting system and method

Abstract

The invention discloses a production scheduling strategy adjustment system and method, and belongs to the technical field of production management, and the method comprises the steps: constructing a multi-source heterogeneous database, setting a semantic association method, generating a demand knowledge graph, setting a prediction and correction method, constructing an order prediction model, generating order quantity fluctuation matrixes with different confidence degrees, and providing quantitative parameters for risk assessment. Constructing a risk assessment basic data set, setting a risk assessment method, dynamically adjusting a risk threshold value to realize adaptive early warning, constructing a risk probability model, and generating a production decision result based on a formulated production decision judgment rule; calculating a factory quantity deviation rate, setting causal analysis, detecting deviation by using a threshold value, triggering an alarm, integrating production data to construct a causal graph, and generating a causal effect report; and calculating a deviation rate of deviation detection, setting a self-optimization method, obtaining an optimal hyper-parameter combination, and optimizing an input feature vector so as to improve the production scheduling prediction accuracy.

Description

Technical Field

[0001] The present invention belongs to the technical field of production management and relates to a production scheduling strategy adjustment system and method. Background Art

[0002] With the deep promotion of industrial digital transformation and intelligent manufacturing, traditional manufacturing industries urgently need to transform towards intelligence and digitization; intelligent workshops have become the core carriers for improving production efficiency and resource allocation through the integration of information technology and data analysis; however, consumer demands show diverse and personalized trends, and traditional workshops rely on manual feedback and limited research data collection methods, making it difficult to obtain comprehensive demand information in real time and lacking efficient data integration and analysis capabilities, resulting in a superficial understanding of demands and insufficient accuracy in production decisions.

[0003] The existing Chinese patent with the publication number CN119623870A discloses a data management method, system, and storage medium for intelligent manufacturing, including calculating prediction data based on historical order data, where the prediction data is the demand data of products in a future predetermined time period; calculating out-of-stock data based on the inventory data and prediction data of products, and defining products with out-of-stock data greater than a first threshold as target products; obtaining the production raw materials and production routes of each target product; if there are the same production equipment in the production routes of multiple target products, defining the target products as twin products and the corresponding production routes as twin routes; obtaining the production duration of each twin product; combining the out-of-stock data and production duration to allocate the occupation duration of each twin product using the twin route, and generating production suggestions according to the occupation duration.

[0004] Although the prior art generates production suggestions by predicting future order situations to assist decision-makers in formulating reasonable production plans, it does not establish a dynamic balance model for customer priorities and production line resource competition. When the order volume of high-priority customers surges, there is a lack of quantitative evaluation of elastic indicators such as equipment overload probability and production capacity buffer rate, and the impact of order volatility and order cancellation probability on the production scheduling plan is not considered. Its static route allocation logic is difficult to meet the multi-objective optimization requirements in emergency scenarios, resulting in production scheduling suggestions being unable to effectively respond to market dynamic changes, easily causing risks such as equipment overload and production capacity backlog when customer demands surge, and at the same time being difficult to balance customer satisfaction and production system stability, reducing the scientificity and flexibility of production decisions. Therefore, the present application provides a production scheduling strategy adjustment system and method. Summary of the Invention

[0005] Aiming at the deficiencies of the prior art, the purpose of the present invention is to provide a production scheduling strategy adjustment system and method, which can make dynamic production ratio decisions, predict real-time risk monitoring probability models, and manage elastic thresholds, achieve the dynamic balance of customer priorities and resource conflicts, quantitatively evaluate the backlog risk and equipment overload probability, and establish a multi-objective collaborative optimization mechanism.

[0006] To achieve the above object, the present invention provides the following technical solutions:

[0007] A production scheduling strategy adjustment system, comprising:

[0008] Construct a multi-source heterogeneous database, set a semantic association method, construct a four-dimensional relationship network, convert behavioral semantics into quantifiable weight parameters, generate a demand knowledge graph, set a prediction and correction method, construct an order prediction model, output the predicted order quantity, predicted order time and confidence interval, and generate an order quantity fluctuation path through normal distribution to generate an order quantity fluctuation matrix with different confidence levels;

[0009] Construct a risk assessment basic data set, set a risk assessment method, dynamically adjust the risk threshold, construct a risk probability model, and generate a production decision result based on the formulated production decision judgment rules;

[0010] Calculate the ex-factory quantity deviation rate, set causal analysis, detect deviations using a threshold and trigger an alarm, integrate production data to construct a causal diagram, and generate a causal effect report;

[0011] Calculate the deviation rate of deviation detection, set a self-optimization method, perform iterative operations through random search combined with Gaussian process regression to obtain the optimal hyperparameter combination, and at the same time perform iterative feature engineering to optimize the input feature vector.

[0012] Furthermore, when generating the demand knowledge graph, the semantic association method includes:

[0013] Obtain text data from the multi-source heterogeneous database, calculate the word frequency and inverse document frequency, screen high-frequency professional vocabulary through TF-IDF values, and construct a dynamic domain dictionary;

[0014] Use the Skip-Gram model to perform iterative training on high-frequency vocabulary, generate a word vector matrix and embed it into the word embedding layer of the basic BERT model to generate an identification model;

[0015] Use the RoBERTa model to perform sentiment analysis on customer service conversation records, output sentiment scores and map them to weight parameters to generate an entity set with sentiment weights;

[0016] Structured data is directly used as knowledge graph entities, unstructured data is sequence-labeled through an identification model, and combined with the entity set with sentiment weights to generate a unified entity list;

[0017] Perform logarithmic smoothing processing on the entity co-occurrence times to generate a basic weight, and combine customer behavior events to adjust the basic weight to construct a four-dimensional relationship network;

[0018] Build a graph database instance based on Neo4j to store a knowledge graph containing several nodes and several edges.

[0019] Further, when building an order prediction model, the prediction and correction method includes:

[0020] According to the customer ID, locate the customer node in the graph database instance, obtain the attribute information of the customer, and calculate the moving window average of the order volume in different time periods to generate customer portrait features;

[0021] Obtain the attribute information of the equipment in the graph database instance, calculate the theoretical production capacity of the equipment, and generate production resource feature information;

[0022] Integrate the customer portrait features, production resource features extracted from the knowledge graph with traditional order data features to generate an input feature vector;

[0023] Build an order prediction model, the input layer receives the constructed input feature vector, and use the Huber loss function and AdamW optimizer to train the order prediction model;

[0024] Based on the multi-source heterogeneous database, generate a real-time vector based on the latest data, and input the real-time vector into the order prediction model to obtain the predicted order volume and predicted order time at the customer level.

[0025] Further, the prediction and correction method further includes:

[0026] Obtain the normal distribution characteristics of the order volume fluctuation, combine the normal distribution characteristics with the predicted order volume to generate an order volume path;

[0027] Exclude the order volume beyond the production capacity according to the equipment production capacity and raw material supply constraints to generate an effective order volume fluctuation path;

[0028] Build a simulation sample based on the order time, sort the simulation sample in ascending order to generate a confidence interval for the predicted order time;

[0029] Obtain the customer level weight information from the knowledge graph and calculate the weighted standard deviation of different customer levels;

[0030] Combine the predicted value of the order prediction model and the weighted standard deviation, and calculate the fluctuation range of the order volume according to the quantiles of different confidence levels to generate an order time fluctuation matrix.

[0031] Further, when formulating the production decision result, the risk assessment method includes:

[0032] Build a data processing pipeline using the Apache Flink real-time computing engine and connect to the predicted order volume stream and the device status stream;

[0033] Perform a join operation on the two data streams through the unique identifier on the production line to associate the data of the predicted order volume with the data of the real-time operating status of the device;

[0034] Define a risk event window, group the data, apply a custom function to calculate the device load rate, and generate a risk event object;

[0035] Based on a preset time interval, extract the difference between the predicted order volume and the actual order volume, calculate the average prediction error, and update the risk threshold in real time.

[0036] Furthermore, the risk assessment method further includes:

[0037] Determine the nodes of the risk probability model based on the Bayesian network, and construct a risk probability model according to the logical relationship between risk indicators and business experience;

[0038] Collect historical production data and risk event data, and train the risk probability model using historical data through the maximum likelihood estimation method to determine the conditional probability relationship between nodes;

[0039] Based on the time interval, collect new data to retrain the model and update the conditional probability between nodes;

[0040] When a new predicted order is input, use the risk probability model to calculate the risk probability, and generate a production decision result based on the constructed production decision judgment rule.

[0041] Furthermore, when constructing a causal graph, the causal analysis method includes:

[0042] Collect actual production volume data and predicted production volume data in real time according to a preset analysis time interval, and calculate the production volume deviation rate;

[0043] Obtain the actual delivery date of the order, and calculate the delivery date deviation rate in combination with the planned delivery date and the standard production cycle in the production schedule.

[0044] Set a production deviation threshold and a delivery date deviation threshold for deviation detection;

[0045] Once the production volume deviation rate exceeds the production deviation threshold, the production deviation is abnormal, and an alarm is immediately sent;

[0046] Once the delivery date deviation rate exceeds the delivery date deviation threshold, the delivery date deviation is abnormal, and an emergency response mechanism is immediately activated;

[0047] Construct a causal graph containing multiple nodes and edges based on the internal relationship between the production process and data, calculate the causal effect of each factor on the deviation, and generate a causal effect report.

[0048] Furthermore, when optimizing the input feature vector, the self-optimization method includes:

[0049] Statistically count the number of abnormal deviations and the total number of detections within a preset time period, and calculate the deviation rate;

[0050] When the deviation rate is less than or equal to the preset self-optimization threshold, perform the next time order prediction; when the deviation rate is greater than the self-optimization threshold, proceed to the next step to optimize the order prediction model;

[0051] Based on historical data and model characteristics, determine the key hyperparameters affecting performance and the search range;

[0052] Adopt a method combining random sampling and Gaussian process regression, iteratively select hyperparameter combinations, train the model and evaluate the performance;

[0053] Avoid local optima by adding exploration noise until the number of iterations is reached, and determine the optimal hyperparameter combination.

[0054] Furthermore, the self-optimization method also includes:

[0055] Use the SHAP value analysis method to calculate the SHAP values of the features in the input feature vector;

[0056] Remove the features with SHAP values less than the preset contribution threshold, combine business experience and domain knowledge, mine new features, and generate an optimized input feature vector;

[0057] Retrain the order prediction model based on the new input feature vector, evaluate the performance through the validation dataset, and determine whether to retain the new features;

[0058] If the addition of the new features improves the order prediction model, retain the new features; if the new features do not improve the performance of the order prediction model, remove the new features;

[0059] Embed the final input feature vector into the order prediction model for production scheduling prediction, and apply the prediction results to actual production scheduling.

[0060] The production scheduling strategy adjustment method includes:

[0061] Based on a multi-source heterogeneous database, use the BERT model to extract demand semantic keywords, construct a four-dimensional relationship network, and generate a demand knowledge graph;

[0062] Build an order prediction model to output the predicted order volume, time, and confidence interval. Generate the order volume fluctuation path based on the normal distribution, filter the effective paths, and generate a multi-confidence fluctuation matrix by calculating the weighted standard deviation through the customer level weight;

[0063] Dynamically adjust the risk threshold, update the production decision rule through the conditional probability model, and quantify the fluctuation matrix parameters to output the production decision result;

[0064] Calculate the ex-factory quantity deviation rate in real time, construct a causal graph to analyze the causal effect of the deviation, iteratively optimize the order prediction model based on the dynamic threshold, eliminate the features with SHAP values less than the preset contribution threshold, and optimize the feature vector combined with business knowledge.

[0065] The beneficial effects of the present invention:

[0066] Through the construction of the domain-enhanced BERT model and the knowledge graph, accurately capture the manufacturing industry-specific terms and customer sentiment tendencies, form a four-dimensional relationship network, effectively quantify the demand priority, and integrate the customer portrait, production resources, and historical data. Combine the LSTM neural network and Monte Carlo simulation to generate the confidence interval and filter the effective paths, significantly improving the prediction accuracy and risk resistance ability; at the same time, update the conditional probability in real time through the Bayesian network, combine the adaptive threshold adjustment and the three-level decision rule to achieve the accurate identification and hierarchical response of production risks, and use the causal analysis and self-optimization mechanism to analyze through the DoWhy model and SHAP values, continuously optimize the input features and hyperparameters, iteratively improve the performance of the prediction model, reduce production deviations, and significantly improve the accuracy, flexibility, and resource utilization rate of production scheduling through the closed-loop architecture of data credibility, demand explicitness, prediction intelligence, risk controllability, and model self-optimization, reduce the supply chain risk, and ultimately achieve the improvement of production efficiency and the optimization of customer satisfaction. Description of the Drawings

[0067] Figure 1 It is the structure diagram of the production scheduling strategy adjustment system;

[0068] Figure 2 It is the flowchart of the semantic association method of the present invention;

[0069] Figure 3 It is the flowchart of the prediction and correction method of the present invention;

[0070] Figure 4 It is the flowchart of the risk assessment method of the present invention;

[0071] Figure 5 It is the flowchart of the self-optimization method of the present invention;

[0072] Figure 6 It is the flowchart of the production scheduling strategy adjustment method. Detailed Embodiments

[0073] The technical solution of the present invention will be described in detail below with reference to the accompanying drawings and specific embodiments. It should be understood that the embodiments of the present invention and the specific features in the embodiments are detailed descriptions of the technical solution of the present invention, rather than limitations on the technical solution of the present invention. Without conflict, the technical features in the embodiments of the present invention and the embodiments can be combined with each other.

[0074] Embodiment 1

[0075] Refer to Figures 1 to 5 As shown, this embodiment introduces a production scheduling strategy adjustment system, including: a data collection module, a prediction module, a risk assessment module, an optimization module, and an interaction module;

[0076] The data collection module is used to collect multi-source heterogeneous data from the customer side, production side, supply chain side, and market side through various channels. Using blockchain technology, key data such as VIP order commitments and production scheduling agreements are chained and solidified. Combining the hash algorithm to achieve data tampering detection, ensuring data authenticity and traceability, and preprocessing the collected data. Using edge computing for anomaly detection and data cleaning to achieve trusted storage and efficient transmission, and finally generating a multi-source heterogeneous database. Each row in the database represents a time point, and the columns contain data from different sources and types, facilitating time series analysis. To improve query efficiency, indexes are established for the time stamp and data category identification fields to achieve fast retrieval of specific time or specific type of data; among them, the data on the customer side is connected to the customer database through the API interface to capture historical orders (order volume, cancellation rate, payment cycle), customer level tags (VIP / core / ordinary), and real-time interaction behaviors (online consultation logs, contract terms). The data on the production side is obtained through the deployed IoT sensor network to obtain equipment operation parameters (power, temperature, vibration), production line capacity utilization rate, and fault logs. The data on the supply chain side is synchronized with the supplier inventory, logistics timeliness data, and raw material price fluctuation curves through a third-party API. The data on the market side is obtained through network technology to obtain social media hot topics;

[0077] The prediction module is used to read the data in the multi-source heterogeneous database, set the semantic association method, extract demand keywords through the BERT model, combine the production process mapping table and supplier raw material association data to construct a four-dimensional relationship network, and convert the behavioral semantics into quantifiable weight parameters to generate a demand knowledge graph. At the same time, set the prediction and correction methods, construct an order prediction model, output the predicted order volume, predicted order time, and confidence interval, and generate an order volume fluctuation path through the normal distribution. Combine equipment capacity and raw material constraints to screen effective paths, calculate the weighted standard deviation according to the customer level weight, and generate an order volume fluctuation matrix with different confidence levels to provide quantitative parameters including fluctuation range, time confidence band, and customer level priority for risk assessment;

[0078] The risk assessment module is used to construct a basic data set for risk assessment, set risk assessment methods, and dynamically adjust risk thresholds to achieve adaptive early warning, construct a risk probability model to update conditional probabilities, and formulate production decision judgment rules to output risk levels and production decision results;

[0079] The optimization module is used to collect data to calculate the deviation rate of the ex-factory quantity, set causal analysis, detect deviations using thresholds and trigger alarms, integrate production data to construct a causal diagram, input the DoWhy model to calculate the causal effect of deviations, generate a heat map to visualize the contribution degree of factors, realize real-time monitoring and root cause analysis of production deviations, and at the same time calculate the deviation rate of deviation detection, set self-optimization methods, perform iterative operations through random search combined with Gaussian process regression to obtain the optimal hyperparameter combination, and at the same time iterate feature engineering, apply SHAP value analysis to quantify feature contributions, screen and eliminate low-contribution features and add new features, optimize the input feature vector, and apply it to the order prediction model to improve the accuracy of production scheduling prediction;

[0080] The interaction module is used to provide a visual dashboard, construct a multi-dimensional visual interface and intelligent interaction functions, and support users to monitor the production scheduling status in real time, dynamically adjust strategy parameters, verify the model effect, and quickly respond to abnormal events.

[0081] Furthermore, when generating the demand knowledge graph, the semantic association methods include:

[0082] Since manufacturing text data contains a large number of professional terms and specific contexts, it is difficult for general natural language processing models (basic BERT models) to accurately identify. By calculating TF-IDF values, professional vocabulary that appears frequently and has distinctiveness in manufacturing domain texts is highlighted, providing a domain-specific vocabulary set for subsequent model training and enhancing the model's ability to identify manufacturing professional terms; Read historical orders and equipment production logs from multi-source heterogeneous databases, divide these text data into text units, for each text unit, count the number of occurrences of each vocabulary to obtain the term frequency TF, and at the same time, calculate the inverse document frequency IDF, multiply TF and IDF to get the TF-IDF value, and screen out the top high-frequency professional vocabulary, and construct a dynamic domain dictionary based on this; In this embodiment, take ;

[0083] Use the Skip - Gram model to perform iterative training on the selected high-frequency professional vocabulary to generate a dimensional word vector matrix; Merge the word vector matrix into the word embedding layer of the basic BERT model to obtain a domain-enhanced BERT model, which is defined as the recognition model; Among them, set the training window size to , which means that during the training process, the context information of words before and after the current word is considered, and the number of iterations is . The Skip - Gram model used in this embodiment is a standard algorithm and architecture, without modifying or expanding the model structure, and can effectively learn the semantic relationships between words. By training on manufacturing professional vocabulary, it captures the semantic features of words in the manufacturing field, taking , ;

[0084] Customers' emotions and demands are often subjective. Through semantic emotion embedding, these subjective information are transformed into quantifiable features, providing richer and more accurate information for subsequent calculation of customer demand priorities. Use the RoBERTa model to analyze the sentiment tendency of customer service conversation records, output corresponding sentiment scores, map the sentiment scores to weight parameters, and attach them to the corresponding entity attributes to generate an entity set with sentiment weights. Each entity in the set contains the sentiment information extracted from the customer service conversation records. Among them, the RoBERTa model is based on the Transformer architecture, pre - trained on a large - scale text, accurately captures the semantic and sentiment information in the text, and classifies the text into positive, negative or neutral sentiment;

[0085] Directly read the structured data in the multi - source heterogeneous database as the entities of the knowledge graph. For unstructured data, use an identification model for sequence labeling, adopting the BIOES labeling format to identify entities in the text through the labeling results, match the implicit semantic information in the text, such as time constraints and quantity limits, add lead - time urgency labels to the text, integrate the entities extracted from different data sources to form a unified entity list, and label the entity sources. Among them, the unified entity list contains the entity set with sentiment weights, the structured entity set and the unstructured entity set;

[0086] In the original data, the distribution of co - occurrence times between entities is uneven. If the co - occurrence times are directly used as weights, a small number of relationships with extremely high co - occurrence times will dominate, resulting in an unreasonable weight distribution. Through logarithmic smoothing processing, the weight distribution becomes more balanced and reasonable, reducing the impact of extreme values. To avoid the influence of co - occurrence times on weights being too linear, use a logarithmic function to smooth the co - occurrence times to calculate the basic weights. And since customer demands and the market environment are dynamically changing, static relationship weights cannot reflect these changes. According to customer behavior events, such as order cancellation and urgent requests, adjust the basic weights to construct a four - dimensional relationship network of "customer - order - device - supplier", thereby generating an entity relationship network with weights. Each relationship edge in the network has a corresponding weight. The expression is as follows:

[0087]

[0088]

[0089] In the formula, is the basic weight, is the number of collinear times, is the adjusted weight, is the influence factor of customer behavior events;

[0090] The structure of the knowledge graph is complex, containing a large number of entities and relationships. The graph database can better adapt to this data structure, provide efficient storage and query performance, and facilitate the analysis and application of the knowledge graph. Neo4j is selected as the graph database. According to the unified entity list and the entity relationship network with weights, nodes and edges are created in Neo4j to generate a graph database instance based on Neo4j, storing a knowledge graph containing several nodes and several edges. Through the property graph model, entities, relationships, and attributes are integrated together, and the knowledge graph is stored and displayed in the form of a graph to realize the semantic association between customer needs and production resources and support subsequent prediction and production scheduling. Among them, customer nodes are created, and their attributes include customer level, sentiment score, and cooperation years. Product nodes are created, and their attributes include production process and equipment dependence. Equipment nodes are created, and their attributes include rated power and maintenance cycle. Relationship types are defined, such as ordering, dependence, and production. Each edge contains a weight value and an update timestamp.

[0091] Furthermore, when constructing the order prediction model, the prediction and correction methods include:

[0092] The delivery date and customization preferences of customers, customer level, and historical order volume trends have an important impact on predicting future order volume and order time. Extracting these features enables the prediction model to better learn the demand patterns of customers and improve the accuracy and pertinence of prediction. Through the Cypher query statement of Neo4j, the customer node is located according to the customer ID, and the attribute information of the customer is obtained, such as delivery option weight, customization weight, and customer level. At the same time, the historical order records of the customer are obtained from the knowledge graph, and the moving window mean of the order volume within different time periods , , is calculated. These data are aggregated to generate a customer portrait feature set, including customer delivery option weight, customization weight, customer level, and the moving window mean of the order volume in different time windows;

[0093] The production of products depends on equipment and raw materials. The production capacity and status of equipment and the supply situation of raw materials directly affect the order delivery ability. When predicting the order quantity and order time, considering these production resource characteristics can ensure that the prediction results conform to the actual production situation and avoid unrealistic predictions. Using Cypher queries, find the specified product node from the knowledge graph, locate the associated equipment nodes, and obtain the attribute information of the equipment, such as rated power, maintenance cycle, and current load. Based on the rated power and standard production efficiency of the equipment, calculate the theoretical production capacity of the equipment. Combine the maintenance cycle and current maintenance status of the equipment to evaluate the available production capacity of the equipment. And obtain the supply relationship between suppliers and raw materials from the knowledge graph, obtain the inventory data of suppliers in real time through the API, calculate the available quantity of raw materials according to the safety stock coefficient, and determine the replenishment time of raw materials in combination with the procurement cycle and transportation time, generating production resource characteristic information, including the theoretical production capacity, available production capacity, current load rate of the equipment associated with the product, and the available quantity and replenishment time of raw materials. Among them, the safety stock coefficient is the quantile of the standard normal distribution corresponding to the target service level, and the safety stock is obtained by multiplying the standard deviation of historical demand fluctuations, the procurement lead time, and the safety stock coefficient. The difference between the real-time inventory of the supplier and the safety stock quantity is the available quantity of raw materials.

[0094] Integrate the customer portrait features and production resource features extracted from the knowledge graph with traditional order data features (such as historical order quantity, order time interval), perform encoding processing on the order time data, convert the date to a timestamp, and perform normalization processing to generate an input feature vector.

[0095] Build an order prediction model based on the LSTM neural network. The input layer receives the constructed input feature vector and trains the order prediction model. Among them, the order prediction model is used to predict the order quantity and order time according to the input feature vector. Set 2 layers of LSTM units as the hidden layer, with 128 neurons in each layer to learn the complex patterns in the time series data. Prevent overfitting through the Dropout layer (Dropout = 0.2). The output layer uses the Dense layer, determines the output dimension according to the prediction targets (order quantity and order time), selects the Huber loss function to balance the robustness of MAE and MSE, and uses the AdamW optimizer (learning rate = 0.001, weight decay = 0.01) to train the model. At the same time, set an early stopping mechanism. If the validation set loss does not decrease for 5 consecutive epochs, terminate the training.

[0096] Based on the time period corresponding to the latest time point in the multi-source heterogeneous database Collect data within it. After going through the same steps as above, generate a real-time vector, input the real-time vector into the order prediction model, and obtain the predicted order quantity and predicted order time at the customer level.

[0097] By deeply analyzing past order data, obtain the normal distribution characteristics of order volume fluctuations , based on historical data for iterative simulation. In each simulation, use the random number generation function in the computer programming language to generate random numbers, representing a hypothetical order volume fluctuation value. Combine with the predicted order volume to construct the order volume path within the future time period, thereby generating several order volume paths; among them, , is the standard deviation of the prediction error, is the standard deviation of historical order volume; among them, the prediction error is the difference between the predicted order volume and the actual order volume;

[0098] Refer to the equipment production capacity constraints and supplier raw material supply constraints in the knowledge graph, exclude the order volume fluctuation paths that exceed the production and supply capabilities, and generate the effective order volume fluctuation paths after constraint screening;

[0099] Since the order time is affected by multiple production links and there is a certain degree of uncertainty, calculating the confidence interval of the order time through Monte Carlo simulation can more comprehensively reflect the possible range of the order time and provide a more accurate time reference for production scheduling; assuming that the fluctuation of the order time follows a certain probability distribution, combine the time uncertainties of each link in the production process, such as equipment processing time and raw material transportation time, to conduct Monte Carlo simulation and generate a large number of simulated samples of the order time. Among them, the number of simulations ;

[0100] Sort the simulated samples in ascending order. According to statistical principles, to obtain confidence interval, take the quantile of the sorted samples as the lower limit, quantile as the upper limit, and obtain the confidence interval of the predicted order time ; among them, ; , are the lower and upper limits of the confidence interval respectively;

[0101] Obtain the customer level weight information from the knowledge graph and calculate the weighted standard deviation of different customer levels , and the expression is as follows:

[0102]

[0103] In the formula, is the weight of each level of customer, is the standard deviation of historical order volume of each level of customer, is the number of customers;

[0104] Combined with the predicted value and weighted standard deviation of the order prediction model, calculate the fluctuation range of the order volume according to the quantiles of different confidence levels, and generate an order time fluctuation matrix for different customer levels at different confidence levels. Among them, different confidence levels correspond to different quantile values. Taking the confidence level as an example, the corresponding quantiles and . At this time, the lower limit and the upper limit of the order volume fluctuation range are expressed as follows:

[0105]

[0106]

[0107] In the formula, is the predicted order volume of the order prediction module.

[0108] Furthermore, when formulating the production decision result, the risk assessment method includes:

[0109] Read the relevant data in the prediction and correction method and the knowledge graph to generate a risk assessment basic data set, including order volume fluctuation, equipment load, raw material inventory, and customer demand priority, providing a comprehensive data basis for the construction of the risk index system, and constructing the risk index system to realize the quantitative assessment of risks;

[0110] Adopt the Apache Flink real-time computing engine to build a real-time data processing pipeline; first define the input source, connect to the predicted order volume stream and the equipment status stream. In Flink, use the join operation to associate the order volume data and the equipment status data according to the unique identifier on the production line. Among them, the predicted order volume stream contains the prediction information of future order quantities, while the equipment status stream real-time feedbacks the operating status of production equipment, such as the temperature, pressure, and vibration of the equipment;

[0111] Set the risk event window to , group and process the data. Within the window, apply a custom function to calculate the equipment load rate and generate a risk event object. Among them, the equipment load rate is an important indicator for measuring the production pressure of the equipment. By calculating the equipment load rate, it is possible to timely discover whether there is an overload risk of the equipment. The risk event object contains key information related to risks, such as risk type, risk occurrence time, and risk severity, providing a clear basis for subsequent risk analysis and processing;

[0112] To dynamically adjust the risk threshold, establish a prediction error data collection mechanism, and every , extract the difference between the predicted order volume and the actual order volume from the log file or database of the prediction module, calculate the mean value of the prediction error, and update the risk threshold in real time; the expression is as follows:

[0113]

[0114] In the formula, and are the risk thresholds before and after the update respectively, is the learning rate, is the recent deviation rate, which is also the mean value of the prediction error. In this embodiment, is taken;

[0115] Determine the nodes of the risk probability model based on the Bayesian network. According to the logical relationship between risk indicators and business experience, construct a risk probability model based on the Bayesian network. The model nodes include secondary risk indicators and tertiary risk indicators;

[0116] Collect historical production data and risk event data, such as order volume, equipment status, raw material inventory, time and type of risk events. Preprocess these data and convert them into a format suitable for model training. Use the historical data to train the risk probability model. Through the maximum likelihood estimation method, determine the conditional probability relationship between each node, which reflects the mutual influence and dependence relationship between different risk indicators;

[0117] During the operation of the system, as new production data and risk events occur continuously, every , collect new data, retrain the risk probability model, and update the conditional probability of each node, so that the risk probability model can adapt to the changes in the production environment in time and improve the accuracy of risk probability prediction;

[0118] According to the enterprise's risk tolerance and production strategy, formulate a production decision-making judgment rule based on risk probability; among them, the enterprise's risk tolerance is determined by the enterprise's financial status, market position, and goals, and the production strategy takes into account production efficiency, cost control, and customer satisfaction;

[0119] When new predicted order data is input, the risk probability model calculates the risk probability of the predicted order according to the current risk indicator values, and generates the final production decision result based on the production decision-making judgment rule, including whether to arrange the predicted order for production, corresponding decision suggestions, and relevant information of the predicted order, such as normal production, proportional production, and prohibited production.

[0120] Furthermore, the risk indicator system is as follows:

[0121] Classify and quantify risks according to the hierarchical structure of first-level indicators, second-level indicators, and third-level indicators.

[0122] For the first-level indicators, three major categories are determined: supply risk, production risk, and delivery risk;

[0123] At the second-level indicator level, for supply risk, by calculating the raw material shortage risk , it is judged whether the raw material supply meets the order demand; for production risk, calculate the equipment overload risk , and evaluate whether the equipment production capacity can carry the order production task; for delivery risk, calculate the delivery delay risk , and measure whether there is a risk of delay in the order delivery time; The expressions are as follows:

[0124]

[0125]

[0126]

[0127] In the formula, is the raw material inventory, is the predicted order quantity, is the required production capacity of the predicted order quantity, is the available production capacity of the equipment, is the upper limit of the order time interval, is the standard production cycle;

[0128] At the third-level indicator level, for raw material shortage, calculate the procurement cycle volatility , and analyze the change of the procurement cycle; for equipment overload, calculate the failure shutdown rate , and evaluate the stability of the equipment; The expressions are as follows:

[0129]

[0130]

[0131] In the formula, is the actual procurement cycle, is the average value of the historical procurement cycle, is the historical failure time, is the total running time.

[0132] Furthermore, the production decision judgment rules are as follows:

[0133] According to the enterprise's risk tolerance and production strategy, formulate the third-level risk decision rules, divide the risk probability interval into low-risk area, medium-risk area and high-risk area, corresponding to different production decision logics;

[0134] When the risk probability of a predicted order is lower than the set low-risk threshold, it is determined that the predicted order can be normally scheduled for production, and the production scheduling process is automatically triggered to synchronize the order requirements to the production execution system;

[0135] When the risk probability is higher than the set high-risk threshold, it is determined that the predicted order requires careful consideration for production, the risk warning mechanism is triggered, a red alert is sent to the production scheduling department, and additional risk response measures are taken, such as adjusting the production plan and finding alternative suppliers;

[0136] When the risk probability is between the low-risk threshold and the high-risk threshold, the production ratio is dynamically calculated according to the weighting coefficient, and the customer level weight is obtained from the knowledge graph and the order profit margin , and the comprehensive decision coefficient is calculated , and the production ratio is determined according to the coefficient ; The expression is as follows:

[0137]

[0138]

[0139] All decision results are synchronized to the production execution terminal through the API interface and visually displayed on the digital twin platform, with different colors marking the production ratio.

[0140] Furthermore, when constructing the causal diagram, the causal analysis methods include:

[0141] With the help of the production execution terminal, collect the actual ex-factory quantity data at fixed time intervals, obtain the corresponding predicted ex-factory quantity data from the production scheduling terminal, and calculate the ex-factory quantity deviation rate;

[0142] Set the ex-factory deviation threshold for ex-factory quantity deviation detection. Once the ex-factory quantity deviation rate exceeds the ex-factory deviation threshold, the ex-factory deviation is abnormal, and an alarm is immediately sent to the production scheduling center and relevant responsible persons;

[0143] Use the logistics management terminal and the order tracking terminal to obtain the actual delivery date of the order, extract the planned delivery date and standard production cycle of the order from the production schedule, and calculate the delivery date deviation rate;

[0144] Set the delivery date deviation threshold for delivery date deviation detection. Once the delivery date deviation rate exceeds the delivery date deviation threshold, the delivery date deviation is abnormal, and the emergency response mechanism is immediately activated to trigger a series of preset emergency handling processes, such as adjusting the production priority and coordinating resources;

[0145] When the factory deviation is abnormal, it is necessary to deeply analyze the root causes of the deviation to provide a basis for formulating effective countermeasures; comprehensively integrate various types of data in the production process, including production logs, equipment operation logs, sensor data, personnel operation records, and model prediction records, and construct a causal graph containing multiple nodes and edges based on the internal relationship between the production process and the data to describe the causal relationship between various factors;

[0146] Take the sorted data and the constructed causal graph as parameters and input them into the DoWhy causal model. Use the propensity score matching method to calculate the causal effect of each factor on the deviation, clarify the root cause of the deviation and the contribution degree of each factor, and output a causal effect report generated in PDF format;

[0147] Generate a heat map of the causal path contribution degree through a visualization tool to intuitively display the causal relationship between various factors and the contribution degree to the deviation.

[0148] Furthermore, when optimizing the input feature vector, the self-optimization method includes:

[0149] According to a preset time period , count the number of times the deviation abnormality occurs and the total number of times of deviation detection , and calculate the deviation rate within a certain time period , and the expression is as follows:

[0150]

[0151] When the deviation rate is less than or equal to the preset self-optimization threshold, return to the prediction module to predict the next time order again; when the deviation rate is greater than the self-optimization threshold, enter the next step to optimize the order prediction model;

[0152] To continuously improve the accuracy and adaptability of the production scheduling model, establish a model self-optimization mechanism, and continuously improve the model performance through parameter dynamic adjustment and feature engineering iteration; before starting hyperparameter optimization, deeply analyze the characteristics of the order prediction model and past prediction experience, determine the hyperparameters that have a greater impact on the performance of the order prediction model, and set a reasonable search range; set the search range of the learning rate to , the learning rate determines the step size of parameter update during the model training process. If the value is too large, the model will not converge, and if the value is too small, the training process will be too slow. Set the search range of the number of neurons in the hidden layer to , the number of neurons directly affects the learning ability and complexity of the model. If the number is too small, the model will not be able to learn the complex patterns in the data, and if the number is too large, it is easy to cause overfitting. The range of the dropout rate is set to , Dropout is a technique to prevent overfitting, which reduces the co-adaptation problem between neurons by randomly discarding some neurons;

[0153] Randomly select a set of hyperparameters from the pre-set hyperparameter search space, use this set of hyperparameters to configure the order prediction model, and use the historical order data and equipment status data in the multi-source heterogeneous database to train the order prediction model, and then use the validation data set to verify it again. By calculating the model indicators, such as accuracy, loss function value, and mean square error, the performance of the model is comprehensively evaluated.

[0154] Based on the model performance evaluation results, Gaussian process regression is used to predict the next set of hyperparameters. Based on the existing hyperparameters and model performance data, a probability model is constructed to predict the possible performance of the model under other hyperparameter combinations. In order to prevent the algorithm from falling into the local optimal solution, a certain amount of exploration noise is added to the predicted hyperparameters so that the algorithm can jump out of the local optimal solution and continue to search for the global optimal solution in a larger search space.

[0155] According to the above steps, iterative operations are performed, and hyperparameter selection, model training, performance evaluation and hyperparameter prediction are repeated until the number of iterations is reached, and the optimal hyperparameter combination obtained by the iteration is automatically loaded into the order prediction model. At this time, the model has reached a relatively optimal state in terms of hyperparameter configuration, providing more reliable support for subsequent prediction work;

[0156] Based on the SHAP value analysis method, the contribution of each feature in the input feature vector to the model prediction result is quantitatively evaluated, and the SHAP value of each feature is calculated to determine the contribution of each feature;

[0157] Filter out features whose SHAP values ​​are less than the preset contribution threshold from the input feature vector and remove them. At the same time, combine business experience and domain knowledge to mine new features with potential value and add them to the input feature vector to generate an optimized input feature vector.

[0158] Use the optimized input feature vector and the original training data to retrain the order prediction model. After training, use the new validation data set to evaluate the model performance and compare the model indicators to determine whether the addition of new features has improved the model performance.

[0159] Based on the model performance evaluation results, decide whether to retain the new features; if the addition of the new features improves the model performance, retain the new features; if the new features fail to improve the model performance, remove the new features;

[0160] Embed the final input feature vector into the order prediction model for production scheduling prediction, apply the prediction results to actual production scheduling, and continuously collect feedback data to drive subsequent optimization.

[0161] Embodiment 2

[0162] Please refer to Figure 6 , another embodiment provided by the present invention: a production scheduling strategy adjustment method, including:

[0163] Collect data from the customer side, production side, supply chain side, and market side through multiple channels, use blockchain technology to hash and solidify key data, combine edge computing to achieve anomaly detection and data cleaning, construct a multi-source heterogeneous database with timestamp indexing, and establish indexes for timestamp and data category fields to improve the retrieval efficiency of time series data;

[0164] Use the BERT model to extract demand semantic keywords, combine the production process mapping table and supplier raw material data to construct a four-dimensional relationship network, generate a demand knowledge graph through semantic weight parameters, construct an order prediction model, output the predicted order quantity, time, and confidence interval, generate an order quantity fluctuation path based on the normal distribution, combine equipment capacity and raw material constraints to screen effective paths, calculate the weighted standard deviation through customer level weights, and generate a multi-confidence fluctuation matrix;

[0165] Construct a risk assessment basic data set, dynamically adjust the risk threshold, update the production decision rule through the conditional probability model, and output the production decision result based on the quantization parameter of the fluctuation matrix;

[0166] Real-time calculate the ex-factory quantity deviation rate, integrate production data to construct a causal graph, input the DoWhy model to calculate the deviation causal effect, and generate a heat map to visualize the factor contribution degree; based on the dynamic threshold of the deviation rate, iteratively optimize the hyperparameters through random search combined with Gaussian process regression, use SHAP value analysis to screen low-contribution features, add new features combined with business knowledge, optimize the input feature vector and retrain the order prediction model;

[0167] Provide real-time production scheduling status monitoring, dynamic adjustment of strategy parameters, model effect verification, and abnormal event response functions through a multi-dimensional visualization dashboard.

[0168] Furthermore, the specific steps for outputting the production decision result include:

[0169] Read the relevant data in the prediction and correction method and the knowledge graph to generate a risk assessment basic data set, including order quantity fluctuation, equipment load, raw material inventory, and customer demand priority, provide a comprehensive data basis for the construction of the risk index system, and construct the risk index system to achieve quantitative assessment of risks;

[0170] Use the Apache Flink real-time computing engine to build a real-time data processing pipeline; first, define the input source to access the predicted order volume stream and the device status stream. In Flink, use the join operation to associate the order volume data and the device status data based on the unique identifier on the production line. Among them, the predicted order volume stream contains the predicted information of future order quantities, while the device status stream provides real-time feedback on the operating status of production equipment, such as the temperature, pressure, and vibration of the equipment.

[0171] Set the risk event window to , process the data in groups. Within the window, apply a custom function to calculate the device load rate and generate a risk event object. Among them, the device load rate is an important indicator for measuring the production pressure of the device. By calculating the device load rate, it is possible to promptly detect whether there is an overload risk for the device. The risk event object contains key information related to risks, such as the risk type, the time when the risk occurs, and the severity of the risk, providing a clear basis for subsequent risk analysis and processing.

[0172] To dynamically adjust the risk threshold, establish a prediction error data collection mechanism. Every , extract the difference between the predicted order volume and the actual order volume from the log file or database of the prediction module, calculate the mean value of the prediction error, and update the risk threshold in real time. The expression is as follows:

[0173]

[0174] In the formula, and are the risk thresholds before and after the update respectively, is the learning rate, is the recent deviation rate, which is also the mean value of the prediction error. In this embodiment, it takes ;

[0175] Determine the nodes of the risk probability model based on the Bayesian network. According to the logical relationship between risk indicators and business experience, construct a risk probability model based on the Bayesian network. The model nodes include secondary risk indicators and tertiary risk indicators.

[0176] Collect historical production data and risk event data, such as order volume, device status, raw material inventory, the time and type of risk events. Preprocess these data and convert them into a format suitable for model training. Use the historical data to train the risk probability model. Through the maximum likelihood estimation method, determine the conditional probability relationship between each node, which reflects the mutual influence and dependence relationship between different risk indicators.

[0177] During the operation of the system, as new production data and risk events continue to occur, every , collect new data, retrain the risk probability model, and update the conditional probabilities of each node to enable the risk probability model to adapt to changes in the production environment in a timely manner and improve the accuracy of risk probability prediction;

[0178] According to the enterprise's risk tolerance and production strategy, formulate production decision-making judgment rules based on risk probability; among them, the enterprise's risk tolerance is determined by the enterprise's financial status, market position, and goals, and the production strategy takes into account production efficiency, cost control, and customer satisfaction;

[0179] When new predicted order data is input, the risk probability model calculates the risk probability of the predicted order using Bayes' formula based on the current risk indicator values, and generates the final production decision result based on the production decision-making judgment rules, including whether to arrange the predicted order for production, corresponding decision suggestions, and relevant information about the predicted order, such as normal production, proportional production, and prohibited production.

[0180] In summary, in the above embodiments, the present invention collects multi-source heterogeneous data from customers, production, supply chain, and market sides through multiple channels, uses blockchain technology to solidify key data and combines the hash algorithm to ensure data authenticity, and stores it in a multi-source heterogeneous database after preprocessing through edge computing; then constructs a four-dimensional relationship network through a domain-enhanced BERT model to generate a demand knowledge graph, integrates the customer portrait and production resource characteristics to train an order prediction model, combines the normal distribution to simulate the order fluctuation path and filters effective paths to generate an order volume fluctuation matrix with different confidence levels; then dynamically adjusts the risk threshold based on the Bayesian network, constructs a risk indicator system for real-time risk assessment, formulates a three-level decision rule to output production decisions, triggers an alarm when the deviation rate exceeds the threshold, integrates production data to construct a causal graph and inputs it into the DoWhy model to analyze the root cause, optimizes the hyperparameters of the prediction model through random search combined with Gaussian process regression, and applies SHAP values to iteratively screen feature vectors; finally, displays the risk heat map and production status through a visualization dashboard, provides an artificial correction interface and generates a production scheduling report to achieve multi-terminal interactive warning.

[0181] The above are only the preferred embodiments of the present invention, and the protection scope of the present invention is not limited to the above embodiments. All technical solutions falling within the idea of the present invention belong to the protection scope of the present invention. It should be noted that for those of ordinary skill in the art in this technical field, several improvements and refinements made without departing from the principle of the present invention should also be regarded as the protection scope of the present invention.

Claims

1. A production scheduling strategy adjustment system, characterized in that, Including: Construct a multi-source heterogeneous database, set a semantic association method, construct a four-dimensional relationship network, transform behavioral semantics into quantifiable weight parameters, generate a demand knowledge graph, set a prediction and correction method, construct an order prediction model, output the predicted order quantity, predicted order time and confidence interval, and generate an order quantity fluctuation path through normal distribution to generate an order quantity fluctuation matrix with different confidence levels; Construct a risk assessment basic data set, set a risk assessment method, dynamically adjust the risk threshold, construct a risk probability model, and generate a production decision result based on the formulated production decision judgment rules; Calculate the ex-factory quantity deviation rate, set causal analysis, detect deviations using thresholds and trigger alarms, integrate production data to construct a causal diagram, and generate a causal effect report; Calculate the deviation rate of deviation detection, set a self-optimization method, perform iterative operations through random search combined with Gaussian process regression to obtain the optimal hyperparameter combination, and at the same time iterate feature engineering to optimize the input feature vector.

2. The production scheduling strategy adjustment system according to claim 1, characterized in that: When generating the demand knowledge graph, the semantic association method includes: Obtain text data from the multi-source heterogeneous database, calculate the term frequency and inverse document frequency, screen high-frequency professional vocabulary through TF-IDF values, and construct a dynamic domain dictionary; Use the Skip-Gram model to iteratively train high-frequency vocabulary, generate a word vector matrix and embed it into the word embedding layer of the basic BERT model to generate an identification model; Use the RoBERTa model to perform sentiment analysis on customer service conversation records, output sentiment scores and map them to weight parameters to generate an entity set with sentiment weights; Structured data is directly used as knowledge graph entities, and unstructured data is sequence-labeled through the identification model and combined with the entity set with sentiment weights to generate a unified entity list; Perform logarithmic smoothing processing on the entity co-occurrence times to generate a basic weight, combine customer behavior events, adjust the basic weight, and construct a four-dimensional relationship network; Construct a graph database instance based on Neo4j to store the knowledge graph containing several nodes and several edges.

3. The production scheduling strategy adjustment system according to claim 2, characterized in that: When constructing the order prediction model, the prediction and correction method includes: According to the customer ID, locate the customer node in the graph database instance, obtain the attribute information of the customer, and calculate the moving window average of the order quantity in different time periods to generate customer portrait features; Obtain the attribute information of the equipment in the graph database instance, calculate the theoretical production capacity of the equipment, and generate production resource feature information; Integrate the customer portrait features, production resource features extracted from the knowledge graph and traditional order data features to generate an input feature vector; Construct an order prediction model, the input layer receives the constructed input feature vector, and use the Huber loss function and AdamW optimizer to train the order prediction model; Based on the multi-source heterogeneous database, generate a real-time vector based on the latest data, input the real-time vector into the order prediction model, and obtain the predicted order quantity and predicted order time at the customer level.

4. The production scheduling strategy adjustment system according to claim 3, wherein The prediction and correction method further includes: Obtaining the normal distribution characteristics of the order volume fluctuation, combining the normal distribution characteristics with the predicted order volume to generate an order volume path; Excluding the order volume exceeding the production capacity according to the equipment production capacity and raw material supply constraints to generate an effective order volume fluctuation path; Constructing a simulation sample based on the order time, sorting the simulation sample in ascending order to generate a confidence interval for the predicted order time; Obtaining customer level weight information from the knowledge graph and calculating the weighted standard deviation of different customer levels; Combining the predicted value of the order prediction model and the weighted standard deviation, calculating the fluctuation range of the order volume according to the quantiles of different confidence levels, and generating an order time fluctuation matrix.

5. The production scheduling strategy adjustment system according to claim 4, wherein: When formulating the production decision result, the risk assessment method includes: Using the Apache Flink real-time computing engine to construct a data processing pipeline and accessing the predicted order volume stream and the equipment status stream; Performing a join operation on the two data streams through the unique identifier on the production line to associate the data of the predicted order volume with the data of the real-time operation status of the equipment; Defining a risk event window, grouping the data, and applying a custom function to calculate the equipment load rate to generate a risk event object; Based on a preset time interval, extracting the difference between the predicted order volume and the actual order volume, calculating the average prediction error, and updating the risk threshold in real time.

6. The production scheduling strategy adjustment system according to claim 5, characterized in that, The risk assessment method further includes: Determining the nodes of the risk probability model based on the Bayesian network, and constructing a risk probability model according to the logical relationship between risk indicators and business experience; Collecting historical production data and risk event data, and training the risk probability model using historical data through the maximum likelihood estimation method to determine the conditional probability relationship between nodes; Based on the time interval, collecting new data to retrain the model and updating the conditional probability between nodes; When a new predicted order is input, using the risk probability model to calculate the risk probability, and generating a production decision result based on the constructed production decision judgment rule.

7. The production scheduling strategy adjustment system according to claim 6, wherein: When constructing the causal diagram, the causal analysis method includes: Real-time collecting the actual output data and the predicted output data according to a preset analysis time interval, and calculating the output deviation rate; Obtaining the actual delivery date of the order, and calculating the delivery date deviation rate in combination with the planned delivery date and the standard production cycle in the production scheduling plan; Setting an output deviation threshold and a delivery date deviation threshold for deviation detection; Once the output deviation rate exceeds the output deviation threshold, the output deviation is abnormal, and an alarm is immediately sent; Once the delivery date deviation rate exceeds the delivery date deviation threshold, the delivery date deviation is abnormal, and an emergency response mechanism is immediately activated; Based on the internal relationship between the production process and the data, constructing a causal diagram including multiple nodes and edges, calculating the causal effect of each factor on the deviation, and generating a causal effect report.

8. The production scheduling strategy adjustment system according to claim 7, wherein: When optimizing the input feature vector, the self-optimization method includes: Counting the number of deviation anomalies and the total number of detections within a preset time period, and calculating the deviation rate; When the deviation rate is less than or equal to the preset self-optimization threshold, perform the next time order prediction; when the deviation rate is greater than the self-optimization threshold, enter the next step to optimize the order prediction model; Based on historical data and model characteristics, determine the key hyperparameters affecting performance and the search range; Adopt a method combining random sampling and Gaussian process regression to iteratively select hyperparameter combinations, train the model and evaluate the performance; Avoid local optima by adding exploration noise until the iteration number is reached, and determine the optimal hyperparameter combination.

9. The production scheduling strategy adjustment system according to claim 8, wherein The self-optimization method further includes: Using the SHAP value analysis method to calculate the SHAP values of the features in the input feature vector; Eliminate the features with SHAP values less than the preset contribution threshold, combine business experience and domain knowledge, mine new features, and generate an optimized input feature vector; Retrain the order prediction model based on the new input feature vector, evaluate the performance through the validation data set, and judge whether to retain the new features; If the addition of the new features improves the order prediction model, retain the new features; if the new features do not improve the performance of the order prediction model, remove the new features; Embed the final input feature vector into the order prediction model for production scheduling prediction, and apply the prediction results to actual production scheduling.

10. A method for adjusting production scheduling strategy, which is implemented based on the production scheduling strategy adjustment system described in any one of claims 1-9, characterized in that, Including: Based on a multi-source heterogeneous database, use the BERT model to extract demand semantic keywords, construct a four-dimensional relationship network, and generate a demand knowledge graph; Establish an order prediction model, output the predicted order quantity, time and confidence interval, generate an order quantity fluctuation path based on the normal distribution, screen effective paths, and generate a multi-confidence fluctuation matrix by calculating the weighted standard deviation through customer level weights; Dynamically adjust the risk threshold, update the production decision rules through the conditional probability model, and quantify the fluctuation matrix parameters to output the production decision results; Real-time calculate the ex-factory quantity deviation rate, construct a causal graph to analyze the deviation causal effect, iteratively optimize the order prediction model based on the dynamic threshold, eliminate the features with SHAP values less than the preset contribution threshold, and optimize the feature vector in combination with business knowledge.

Images