Production scheduling strategy adjustment system and method

Abstract

The present invention discloses a production scheduling strategy adjustment system and method, which belongs to the field of production management technology, and includes: constructing a multi-source heterogeneous database, setting a semantic association method, generating a demand knowledge graph, and setting a prediction and correction method, constructing an order prediction model, generating an order quantity fluctuation matrix with different confidence levels, and providing quantitative parameters for risk assessment; constructing a risk assessment basic data set, setting a risk assessment method, dynamically adjusting the risk threshold to achieve adaptive early warning, constructing a risk probability model, and generating production decision results based on the formulated production decision judgment rules; calculating the factory quantity deviation rate, setting causal analysis, using thresholds to detect deviations and trigger alarms, integrating production data to construct a causal graph, and generating a causal effect report; calculating the deviation rate of deviation detection, setting a self-optimization method, obtaining the optimal hyperparameter combination, and optimizing the input feature vector to improve the accuracy of production scheduling prediction.

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 deepening of industrial digital transformation and intelligent manufacturing, traditional manufacturing urgently needs to transform towards intelligence and digitalization; smart workshops have become the core carrier for improving production efficiency and resource allocation by integrating information technology and data analysis; however, consumer demand is showing a trend of diversification and personalization. Traditional workshops rely on manual feedback and limited survey data collection methods, making it difficult to obtain comprehensive demand information in real time, and lack efficient data integration and analysis capabilities, resulting in superficial understanding of demand and insufficient accuracy in production decisions.

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

[0004] Although the existing technology generates production suggestions by predicting future order situations, thereby assisting decision makers in formulating reasonable production plans, it has not established a dynamic balance model between customer priority and competition for production line resources. When the number of high-priority customer orders surges, there is a lack of quantitative evaluation of elastic indicators such as equipment overload probability and production capacity buffer rate. The impact of order volatility and order cancellation probability on production scheduling is not considered. Its static route allocation logic is difficult to cope with the multi-objective optimization needs in emergency scenarios, resulting in production scheduling suggestions that cannot effectively respond to market dynamics. When customer demand surges, it is easy to cause risks such as equipment overload and production capacity backlog. At the same time, it is difficult to balance customer satisfaction and production system stability, reducing the scientific nature and flexibility of production decisions. Therefore, the present application provides a production scheduling strategy adjustment system and method. Summary of the Invention

[0005] In response to the shortcomings of the existing technology, the purpose of the present invention is to provide a production scheduling strategy adjustment system and method, dynamic production ratio decision-making, real-time risk monitoring probability model prediction and elastic threshold management, to achieve a dynamic balance between customer priority and resource conflicts, quantitatively evaluate backlog risks and equipment overload probabilities, and establish a multi-objective collaborative optimization mechanism.

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

[0007] Production scheduling strategy adjustment system, including:

[0008] Build a multi-source heterogeneous database, set up semantic association methods, construct a four-dimensional relationship network, and convert behavioral semantics into quantifiable weight parameters to generate a demand knowledge graph. Set up prediction and correction methods, build an order prediction model, output predicted order volume, predicted order time and confidence interval, and generate order volume fluctuation paths through normal distribution, generating order volume fluctuation matrices with different confidence levels.

[0009] Construct a basic data set for risk assessment, set risk assessment methods, dynamically adjust risk thresholds, build a risk probability model, and generate production decision results based on the established production decision judgment rules;

[0010] Calculate the factory deviation rate, set up cause-and-effect analysis, use thresholds to detect deviations and trigger alarms, integrate production data to build cause-and-effect diagrams, and generate cause-and-effect effect reports;

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

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

[0013] Acquire text data from the multi-source heterogeneous database, calculate word frequency and inverse document frequency, filter high-frequency professional vocabulary by TF-IDF value, and build a dynamic domain dictionary;

[0014] The Skip-Gram model is used to iteratively train high-frequency words to generate a word vector matrix and embed it into the word embedding layer of the basic BERT model to generate a recognition model.

[0015] Use the RoBERTa model to analyze the sentiment tendency of customer service conversation records, output the sentiment score and map it to a weight parameter, and generate an entity set with sentiment weights;

[0016] Structured data is directly used as knowledge graph entities, and unstructured data is sequence-labeled through recognition models. Combined with entity sets with sentiment weights, a unified entity list is generated.

[0017] Perform logarithmic smoothing on the number of entity co-occurrences to generate basic weights, and adjust the basic weights based on customer behavior events to construct a four-dimensional relationship network.

[0018] Build a Neo4j-based graph database instance to store a knowledge graph consisting of several nodes and several edges.

[0019] Furthermore, when constructing an order prediction model, the prediction and correction method includes:

[0020] Based on the customer ID, locate the customer node in the graph database instance, obtain the customer's attribute information, and calculate the sliding window mean of the order volume in different time periods to generate customer profile features;

[0021] Obtaining attribute information of the equipment in the graph database instance, calculating the theoretical production capacity of the equipment, and generating production resource feature information;

[0022] Integrate customer profile features and 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 uses the Huber loss function and AdamW optimizer to train the order prediction model.

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

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

[0026] Obtaining normal distribution characteristics of order volume fluctuations, combining the normal distribution characteristics with the predicted order volume, and generating an order volume path;

[0027] Eliminate orders that exceed production capacity based on equipment capacity and raw material supply constraints, and generate an effective order volume fluctuation path;

[0028] Constructing simulated samples based on order time, sorting the simulated samples in ascending order order time, and generating a confidence interval for the predicted order time;

[0029] Obtaining customer grade weight information from the knowledge graph and calculating weighted standard deviations of different customer grades;

[0030] Combining the predicted value of the order forecasting model and the weighted standard deviation, the fluctuation range of the order quantity is calculated according to the quantiles of different confidence levels to generate an order time fluctuation matrix.

[0031] Furthermore, when formulating production decision results, the risk assessment method includes:

[0032] Use the Apache Flink real-time computing engine to build a data processing pipeline, accessing the predicted order volume stream and device status stream;

[0033] Joining the two data streams using the unique identifier of the production line to correlate the predicted order quantity data with the real-time operating status of the equipment;

[0034] Define risk event windows, group data, apply custom functions to calculate device load rates, and generate risk event objects;

[0035] Based on a preset time interval, the difference between the predicted order quantity and the actual order quantity is extracted, the mean of the prediction error is calculated, and the risk threshold is updated 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 build the risk probability model based on the logical relationship between risk indicators and business experience;

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

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

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

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

[0042] Collect actual factory output data and forecast factory output data in real time according to the preset analysis time interval, and calculate the factory output deviation rate;

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

[0044] Set factory deviation thresholds and delivery deviation thresholds to conduct deviation detection;

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

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

[0047] Based on the intrinsic relationship between production processes and data, a causal graph containing multiple nodes and edges is constructed, the causal effect of each factor on the deviation is calculated, and a causal effect report is generated.

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

[0049] Count the number of deviation anomalies and the total number of detections within the preset time period and calculate the deviation rate;

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

[0051] Determine the key hyperparameters and search ranges that affect performance based on historical data and model characteristics;

[0052] We use random sampling combined with Gaussian process regression to iteratively select hyperparameter combinations, train the model, and evaluate performance.

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

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

[0055] Calculating the SHAP value of the feature in the input feature vector using a SHAP value analysis method;

[0056] Eliminate features whose SHAP values ​​are less than a preset contribution threshold, combine business experience and domain knowledge to mine new features, and generate an optimized input feature vector;

[0057] Retraining the order prediction model based on the new input feature vector, evaluating the performance using a validation dataset, and determining whether to retain the new feature;

[0058] If the addition of the new feature improves the order prediction model, the new feature is retained; if the new feature fails to improve the performance of the order prediction model, the new feature is removed;

[0059] The final input feature vector is embedded into the order prediction model for production scheduling prediction, and the prediction results are applied to actual production scheduling.

[0060] Production scheduling strategy adjustment methods include:

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

[0062] Establish an order forecasting model, output the predicted order volume, time and confidence interval, generate order volume fluctuation paths based on normal distribution, screen effective paths, and calculate weighted standard deviations using customer grade weights to generate a multi-confidence fluctuation matrix;

[0063] Dynamically adjust risk thresholds, update production decision rules through conditional probability models, and quantify fluctuation matrix parameters to output production decision results;

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

[0065] Beneficial effects of the present invention:

[0066] Through the construction of domain-enhanced BERT models and knowledge graphs, it accurately captures manufacturing professional terms and customer sentiment tendencies, forms a four-dimensional relationship network, effectively quantifies demand priorities, and integrates customer portraits, production resources and historical data. Combined with LSTM neural networks and Monte Carlo simulations, it generates confidence intervals and screens effective paths, significantly improving prediction accuracy and risk resistance. At the same time, through real-time updates of conditional probabilities through Bayesian networks, combined with adaptive threshold adjustment and three-level decision rules, it achieves accurate identification and graded response of production risks. Utilizing causal analysis and self-optimization mechanisms through DoWhy models and SHAP value analysis, it continuously optimizes input features and hyperparameters, it iteratively improves prediction model performance, and reduces production deviations. Through a closed-loop architecture of data credibility, demand explicitness, intelligent prediction, risk controllability, and model self-optimization, it significantly improves the accuracy, flexibility and resource utilization of production scheduling, reduces supply chain risks, and ultimately achieves improved production efficiency and optimized customer satisfaction. BRIEF DESCRIPTION OF THE DRAWINGS

[0067] Figure 1 Adjust the system structure diagram for production scheduling strategy;

[0068] Figure 2 is a flow chart of the semantic association method of the present invention;

[0069] Figure 3 is a flow chart of the prediction and correction method of the present invention;

[0070] Figure 4 is a flow chart of the risk assessment method of the present invention;

[0071] Figure 5 is a flow chart of the self-optimization method of the present invention;

[0072] Figure 6 Flowchart of the production scheduling strategy adjustment method. DETAILED DESCRIPTION

[0073] The technical solution of the present invention is described in detail below through 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. In the absence of conflict, the embodiments of the present invention and the technical features in the embodiments can be combined with each other.

[0074] Example 1

[0075] refer to Figures 1 to 5 As shown, this embodiment introduces a production scheduling strategy adjustment system, including: a data acquisition 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, and use blockchain technology to solidify key data, such as VIP order commitments and production scheduling agreements, on the chain, and combine hash algorithms to detect data tampering, ensure data authenticity and traceability, and pre-process the collected data. Edge computing is used for anomaly detection and data cleaning to achieve trusted storage and efficient transmission, and finally generate a multi-source heterogeneous database. Each row in the database represents a time point, and the columns contain data from different sources and types, which is convenient for time series analysis. In order to improve query efficiency, timestamps and data category identification fields are built. Establishing indexes enables rapid retrieval of data at specific times or of specific types. Client-side data is connected to the customer database via an 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). Production-side data is obtained through the deployed IoT sensor network, capturing equipment operating parameters (power, temperature, vibration), production line capacity utilization, and fault logs. Supply chain data is synchronized with supplier inventory, logistics timeliness data, and raw material price fluctuation curves via third-party APIs. Market-side data utilizes network technology to capture hot topics on social media.

[0077] The prediction module is used to read data from multi-source heterogeneous databases, set semantic association methods, extract demand keywords through the BERT model, combine the production process mapping table with supplier raw material association data, build a four-dimensional relationship network, and convert behavioral semantics into quantifiable weight parameters to generate a demand knowledge graph. At the same time, it sets prediction and correction methods, builds an order prediction model, outputs predicted order quantity, predicted order time and confidence interval, and generates order quantity fluctuation paths through normal distribution. Effective paths are screened by combining equipment capacity and raw material constraints, and weighted standard deviations are calculated based on customer grade weights. Order quantity fluctuation matrices with different confidence levels are generated, providing quantitative parameters for risk assessment, including fluctuation range, time confidence band and customer grade priority.

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

[0079] The optimization module is used to collect data to calculate the deviation rate of factory output, set up causal analysis, use thresholds to detect deviations and trigger alarms, integrate production data to build a causal graph, input the DoWhy model to calculate the causal effect of deviations, generate a heat map to visualize the contribution of factors, realize real-time monitoring of production deviations and root cause analysis, and simultaneously calculate the deviation rate of deviation detection. It also sets a self-optimization method, performs iterative operations through random search combined with Gaussian process regression to obtain the optimal hyperparameter combination, and iterates feature engineering, applying SHAP value analysis to quantify feature contributions, screening and eliminating low-contribution features and adding new features, optimizing the input feature vector, and applying it to the order forecasting model to improve the accuracy of production scheduling forecasts.

[0080] The interactive module is used to provide a visual dashboard, build a multi-dimensional visual interface and intelligent interactive functions, and support users to monitor production scheduling status in real time, dynamically adjust strategy parameters, verify model effects, 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, general natural language processing models (basic BERT models) are difficult to accurately identify. By calculating the TF-IDF value, we can highlight professional terms that appear frequently and are distinctive in manufacturing field texts, provide a domain-specific vocabulary set for subsequent model training, and improve the model's ability to recognize manufacturing professional terms; read historical orders and equipment production logs from multi-source heterogeneous databases, divide these text data into text units, count the number of occurrences of each word for each text unit, and obtain the word frequency TF. At the same time, we calculate the inverse document frequency IDF, multiply TF and IDF to obtain the TF-IDF value, and filter out the top TF-IDF values. High-frequency professional vocabulary is used to construct a dynamic domain dictionary; in this embodiment, ;

[0083] Use Skip-Gram model to iteratively train the selected high-frequency professional vocabulary to generate dimensional word vector matrix; merge the word vector matrix into the word embedding layer of the basic BERT model to obtain the domain-enhanced BERT model, and define it as a recognition model; among them, set the training window size to , which means that during the training process, the current word before and after each The context information of words, the number of iterations is The Skip-Gram model used in this embodiment is a standard algorithm and architecture. The model structure is not modified or expanded, and it can effectively learn the semantic relationship between words. By training on professional manufacturing vocabulary, it captures the semantic characteristics of vocabulary in the manufacturing field and obtains , ;

[0084] Customer emotions and needs are often subjective. Semantic sentiment embedding transforms this subjective information into quantifiable features, providing richer and more accurate information for subsequent calculation of customer demand priorities. The RoBERTa model is used to analyze customer service conversation records for sentiment, outputting corresponding sentiment scores. These scores are then mapped to weight parameters and attached to corresponding entity attributes to generate a set of entities with sentiment weights. Each entity in the set contains sentiment information extracted from the customer service conversation records. The RoBERTa model, based on the Transformer architecture and pre-trained on large-scale text, accurately captures the semantic and sentiment information in text and classifies text as positive, negative, or neutral.

[0085] Directly read structured data from multi-source heterogeneous databases as entities in the knowledge graph. For unstructured data, use recognition models for sequence annotation, using the BIOES annotation format. This allows entities in the text to be identified through annotation results. Regular expressions are used to match implicit semantic information in the text, such as time constraints and quantity limits, to add delivery urgency tags to the text. Entities extracted from different data sources are integrated to form a unified entity list, and the entity sources are annotated. The unified entity list includes a set of entities with sentiment weights, a set of structural entities, and a set of unstructured entities.

[0086] In the original data, the distribution of co-occurrence times between entities is unbalanced. If the co-occurrence times are directly used as weights, a few relationships with extremely high co-occurrence times will dominate, resulting in an unreasonable weight distribution. Logarithmic smoothing makes the weight distribution more balanced and reasonable, reducing the impact of extreme values. To avoid the overly linear effect of co-occurrence times on weights, a logarithmic function is used to smooth the co-occurrence times and calculate the basic weights. Customer needs and market environment are dynamic, and static relationship weights cannot reflect these changes. Based on customer behavior events, such as order cancellations and expedited requests, the basic weights are adjusted to construct a four-dimensional relationship network of "customer-order-equipment-supplier", thereby generating a weighted entity relationship network. Each relationship edge in the network has a corresponding weight; the expression is as follows:

[0087]

[0088]

[0089] Where, is the basic weight, is the number of collinearity, is the adjusted weight, is the influencing 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 weighted entity relationship network, nodes and edges are created in Neo4j to generate a Neo4j-based graph database instance, which stores a knowledge graph containing several nodes and several edges. Through the attribute graph model, entities, relationships and attributes are integrated together to store and display the knowledge graph in the form of a graph, realize the semantic association between customer needs and production resources, and support subsequent forecasting and scheduling. Among them, a customer node is created, whose attributes include customer level, sentiment score, and years of cooperation; a product node is created, whose attributes include production process and equipment dependency; an equipment node is created, whose attributes include rated power and maintenance cycle; relationship types are defined, such as ordering, dependency, and production, and each edge contains a weight value and an update timestamp.

[0091] Furthermore, when building an order forecasting model, the forecasting and correction methods include:

[0092] Customer delivery and customization preferences, customer level, and historical order volume trends have a significant impact on predicting future order volume and order timing. Extracting these features enables the prediction model to better learn customer demand patterns, improving the accuracy and pertinence of predictions. Using Neo4j's Cypher query statement, we locate the customer node based on the customer ID and obtain customer attribute information, such as delivery weight, customization weight, and customer level. At the same time, we obtain the customer's historical order records from the knowledge graph and calculate the order volume and time of different time periods. 、 、 The sliding window mean of the order volume within the time window is aggregated to generate a customer profile feature set, including customer transaction weight, customization weight, customer level, and the sliding window mean of the order volume in different time windows;

[0093] Product production relies on equipment and raw materials. The capacity and status of equipment, as well as the availability of raw materials, directly impact order delivery capabilities. Considering these production resource characteristics when forecasting order volume and order time ensures that forecasts align with actual production conditions and avoids unrealistic predictions. Using Cypher queries, we find the specified product node in the knowledge graph, locate the associated device node, and obtain device attribute information, such as rated power, maintenance cycle, and current load. Based on the device's rated power and standard production efficiency, we calculate the device's theoretical capacity. We then assess the device's available capacity based on its maintenance cycle and current maintenance status. We also obtain the supplier-raw material supply relationship from the knowledge graph and obtain real-time supplier inventory data through an API. We calculate the available raw material quantity based on the safety stock factor and determine the replenishment time for the raw material based on the procurement cycle and transportation time. This generates production resource characteristic information, including the theoretical capacity, available capacity, current load rate, and available raw material quantity and replenishment time of the product-related equipment. The safety stock factor is the quantile of the standard normal distribution corresponding to the target service level. The safety stock is calculated by multiplying the standard deviation of historical demand fluctuations, the procurement lead time, and the safety stock factor. The difference between the supplier's real-time inventory and the safety stock is the available raw material quantity.

[0094] Integrate customer profile features and production resource features extracted from the knowledge graph with traditional order data features (such as historical order volume and order time interval), encode order time data, convert dates into timestamps, and perform normalization to generate input feature vectors.

[0095] An order prediction model was constructed based on an LSTM neural network. The input layer received the constructed input feature vector and trained the order prediction model. The order prediction model was used to predict order quantity and order time based on the input feature vector. A two-layer LSTM unit was used as the hidden layer, with each layer containing 128 neurons to learn complex patterns in time series data. A Dropout layer (Dropout = 0.2) was used to prevent overfitting. A Dense layer was used in the output layer, and the output dimension was determined based on the prediction target (order quantity and order time). The Huber loss function was selected to balance the robustness of MAE and MSE. The AdamW optimizer (learning rate = 0.001, weight decay = 0.01) was used for model training. An early stopping mechanism was also implemented. Training was terminated if the validation set loss did not decrease for five consecutive epochs.

[0096] The time period corresponding to the latest time point in multi-source heterogeneous databases After collecting data, the same steps as above are followed to generate real-time vectors, which are then input into the order prediction model to obtain the customer-level predicted order quantity and predicted order time.

[0097] By deeply analyzing past order data, we can obtain the normal distribution characteristics of order volume fluctuations. , based on historical data, iterative simulation is performed. In each simulation, a random number is generated with the help of a random number generation function in a computer programming language to represent a hypothetical order volume fluctuation value. Combined with the predicted order volume, the order volume path in the future time period is constructed, thereby generating several order volume paths; among them, , is the standard deviation of the prediction error, is the standard deviation of historical order volume; where the forecast error is the difference between the forecasted order volume and the actual order volume;

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

[0099] Since order time is affected by multiple production links and has a certain degree of uncertainty, the confidence interval of order time calculated through Monte Carlo simulation can more comprehensively reflect the possible range of order time and provide a more accurate time reference for production scheduling. Assuming that the fluctuation of order time follows a certain probability distribution, combined with the time uncertainty of each link in the production process, such as equipment processing time and raw material transportation time, Monte Carlo simulation is performed to generate a large number of simulated samples of order time, among which the number of simulations is 1. ;

[0100] Sort the simulated samples in ascending order. According to the statistical principle, The confidence interval of the sorted sample is The quantile is used as the lower bound, The quantile is used as the upper limit to obtain the predicted order time Confidence interval of ;in, 、 are the lower and upper limits of the confidence interval respectively;

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

[0102]

[0103] Where, The weight of each level of customers, is the standard deviation of historical order volume of customers at each level, is the number of customers;

[0104] Combined with the predicted value and weighted standard deviation of the order forecast model, the fluctuation range of the order volume is calculated according to the quantiles of different confidence levels, and the order time fluctuation matrix of different customer levels at different confidence levels is generated. Different confidence levels correspond to different quantile values. For example, the corresponding quantile and , at this time the lower limit of the order volume fluctuation range and upper limit The expression is as follows:

[0105]

[0106]

[0107] Where, The predicted order quantity of the order prediction module.

[0108] Furthermore, when formulating production decision outcomes, risk assessment methods include:

[0109] Read the prediction and correction methods and related data in the knowledge graph to generate a basic risk assessment data set, including order volume fluctuations, equipment load, raw material inventory, and customer demand priorities. This provides a comprehensive data foundation for the construction of a risk indicator system, and builds a risk indicator system to achieve quantitative risk assessment.

[0110] We used the Apache Flink real-time computing engine to build a real-time data processing pipeline. We first defined the input source, connecting it to the predicted order volume stream and the equipment status stream. In Flink, we used the join operation to link the order volume data and equipment status data based on the unique identifier of the production line. The predicted order volume stream contains information about future order quantities, while the equipment status stream provides real-time feedback on the operating status of production equipment, such as temperature, pressure, and vibration.

[0111] Set the risk event window to , the data is grouped and processed. Within the window, a custom function is applied to calculate the equipment load rate and generate a risk event object. The equipment load rate is an important indicator used to measure the production pressure of the equipment. By calculating the equipment load rate, it is possible to timely discover whether the equipment has an overload risk. The risk event object contains key information related to the risk, 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.

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

[0113]

[0114] Where, 、 are the risk thresholds before and after the update, is the learning rate, is the recent deviation rate, and is also the mean of the prediction error. In this embodiment, ;

[0115] Determine the nodes of the risk probability model based on the Bayesian network. Based on the logical relationship between risk indicators and business experience, build 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, and the time and type of risk events. Preprocess this data and convert it into a format suitable for model training. Use historical data to train the risk probability model. Using maximum likelihood estimation, determine the conditional probability relationship between each node, reflecting the mutual influence and dependency between different risk indicators.

[0117] During the operation of the system, as new production data and risk events continue to occur, , 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 changes in the production environment in a timely manner and improve the accuracy of risk probability prediction;

[0118] Formulate production decision-making rules based on risk probability, based on the company's risk tolerance and production strategy. The company's risk tolerance is determined by its financial situation, market position, and goals, while the production strategy takes into account production efficiency, cost control, and customer satisfaction.

[0119] When new forecast order data is input, the risk probability model calculates the risk probability of the forecast order based on the current risk index value using the Bayesian formula, and generates the final production decision result based on the production decision judgment rules, including whether to arrange the forecast order for production, corresponding decision recommendations, and related information of the forecast 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 primary indicators, secondary indicators, and tertiary indicators;

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

[0123] At the secondary indicator level, for supply risk, by calculating the risk of raw material shortage , determine whether the raw material supply meets the order requirements; for production risks, calculate the risk of equipment overload , evaluate whether the equipment capacity can bear the order production task; for delivery risk, calculate the delivery delay risk , which measures whether there is a risk of delay in order delivery time; the expression is as follows:

[0124]

[0125]

[0126]

[0127] Where, is the raw material inventory, To predict order volume, To predict the demand capacity of order quantity, is the available capacity of the equipment, The upper limit of the order time interval, is the standard production cycle;

[0128] At the third-level indicator level, for raw material shortages, calculate the procurement cycle volatility , analyze changes in the procurement cycle; for equipment overload, calculate the failure downtime rate , evaluates the stability of the device; the expression is as follows:

[0129]

[0130]

[0131] Where, is the actual procurement cycle, is the historical purchasing cycle average, is the historical failure time, is the total running time.

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

[0133] Based on the company's risk tolerance and production strategy, a three-level risk decision-making rule is formulated, dividing the risk probability range into low-risk, medium-risk and high-risk areas, corresponding to different production decision-making logics;

[0134] When the risk probability of a predicted order is lower than the set low-risk threshold, the predicted order is judged to be able to be arranged for production normally, and the production scheduling process is automatically triggered to synchronize the order demand to the production execution system;

[0135] When the risk probability is higher than the set high-risk threshold, the forecast order is judged to require careful consideration for production, triggering the risk warning mechanism and sending a red alert to the production scheduling department, which will take additional risk response measures, 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 weighted coefficient, and the customer level weight is obtained from the knowledge graph. and order profit margin , calculate the comprehensive decision coefficient , determine the production ratio 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 visualized on the digital twin platform, with production proportions marked with different colors.

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

[0141] With the help of the production execution terminal, actual factory output data is collected at fixed time intervals, the corresponding forecast factory output data is obtained from the production scheduling terminal, and the factory output deviation rate is calculated;

[0142] Set the factory deviation threshold to detect factory quantity deviation. Once the factory quantity deviation rate exceeds the factory deviation threshold, an alarm will be immediately sent to the production scheduling center and relevant persons in charge.

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

[0144] Set a delivery deviation threshold to detect delivery deviations. Once the delivery deviation rate exceeds the threshold, the delivery deviation is abnormal and the emergency response mechanism is immediately activated, triggering a series of preset emergency handling processes, such as adjusting production priorities and coordinating resources.

[0145] When factory deviations are abnormal, in-depth analysis of the root causes of the deviations is required to provide a basis for formulating effective response strategies. Comprehensively integrate various types of data from the production process, including production logs, equipment operation logs, sensor data, personnel operation records, and model prediction records. Based on the inherent relationship between the production process and data, a causal graph containing multiple nodes and edges is constructed to describe the causal relationship between various factors.

[0146] The organized data and the constructed causal diagram are used as parameters and input into the DoWhy causal model. The propensity score matching method is used to calculate the causal effect of each factor on the deviation, clarify the root cause of the deviation and the contribution of each factor, and output a causal effect report generated in PDF format.

[0147] Through visualization tools, a causal path contribution heat map is generated to intuitively display the causal relationship between factors and the degree of contribution to the deviation.

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

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

[0150]

[0151] When the deviation rate is less than or equal to the preset self-optimization threshold, the system returns to the prediction module and predicts the next order again. When the deviation rate is greater than the self-optimization threshold, the system proceeds to the next step to optimize the order prediction model.

[0152] In order to continuously improve the accuracy and adaptability of the production scheduling model, a model self-optimization mechanism is established to continuously improve the model performance through dynamic parameter adjustment and feature engineering iteration. Before starting hyperparameter optimization, the characteristics of the order forecasting model and past forecasting experience are deeply analyzed to determine the hyperparameters that have a greater impact on the performance of the order forecasting model and set a reasonable search range. The search range of the learning rate is set to The learning rate determines the step size of the model parameter update during the 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. The search range of the number of hidden layer neurons is set 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. If the number is too large, it will easily lead to overfitting. The range of the dropout rate is set to ,Dropout is a technique to prevent overfitting by randomly discarding some neurons to reduce the co-adaptation problem between neurons;

[0153] A set of hyperparameters is randomly selected from a pre-defined hyperparameter search space. This set of hyperparameters is used to configure the order prediction model. The model is then trained using historical order data and device status data from a multi-source heterogeneous database. The model is then validated again using a validation dataset. The model's performance is comprehensively evaluated by calculating model metrics such as accuracy, loss function value, and mean squared error.

[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 probabilistic model is constructed to predict the model's possible performance under other hyperparameter combinations. To prevent the algorithm from falling into a local optimal solution, a certain amount of exploratory noise is added to the predicted hyperparameters, allowing the algorithm to escape the local optimal solution and continue to search for the global optimal solution in a larger search space.

[0155] Following the above steps, iterative operations are performed, repeating hyperparameter selection, model training, performance evaluation, and hyperparameter prediction until the number of iterations is reached. The optimal hyperparameter combination obtained through the iteration is automatically loaded into the order prediction model. At this point, 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] Features with SHAP values ​​less than a preset contribution threshold are filtered out from the input feature vector and removed. At the same time, new features with potential value are mined by combining business experience and domain knowledge. These features are added to the input feature vector to generate an optimized input feature vector.

[0158] Retrain the order prediction model using the optimized input feature vector and the original training data. After training, evaluate the model performance using a new validation dataset. By comparing model metrics, determine whether the addition of the new features has improved 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 do not improve the model performance, remove the new features;

[0160] The final input feature vector is embedded into the order forecasting model for production scheduling forecasting, and the forecast results are applied to actual production scheduling, and feedback data is continuously collected to drive subsequent optimization.

[0161] Example 2

[0162] See also Figure 6 Another embodiment provided by the present invention is a method for adjusting a production scheduling strategy, comprising:

[0163] Through multiple channels, we collect data from the customer side, production side, supply chain side, and market side. We use blockchain technology to hash and solidify key data. Combined with edge computing, we implement anomaly detection and data cleansing. We build a multi-source heterogeneous database with timestamp indexing, and index timestamp and data category fields to improve the efficiency of time series data retrieval.

[0164] The BERT model is used to extract demand semantic keywords. This is combined with the production process mapping table and supplier raw material data to construct a four-dimensional relationship network. A demand knowledge graph is generated using semantic weight parameters. An order forecasting model is constructed to output the predicted order quantity, time, and confidence interval. An order quantity fluctuation path is generated based on the normal distribution. Effective paths are screened by combining equipment capacity and raw material constraints. The weighted standard deviation is calculated using customer grade weights to generate a multi-confidence fluctuation matrix.

[0165] Build a basic data set for risk assessment, dynamically adjust risk thresholds, update production decision rules through conditional probability models, quantify parameters based on fluctuation matrices, and output production decision results;

[0166] Calculate the factory deviation rate in real time, integrate production data to build a causal diagram, input the DoWhy model to calculate the deviation causal effect, and generate a heat map to visualize the factor contribution. Based on the dynamic threshold of the deviation rate, it uses random search combined with Gaussian process regression to iteratively optimize hyperparameters, apply SHAP value analysis to screen low-contribution features, add new features based on business knowledge, optimize the input feature vector, and retrain the order prediction model.

[0167] Through multi-dimensional visual dashboards, it provides real-time production scheduling status monitoring, dynamic adjustment of strategy parameters, model effect verification and abnormal event response functions.

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

[0169] Read the prediction and correction methods and related data in the knowledge graph to generate a basic risk assessment data set, including order volume fluctuations, equipment load, raw material inventory, and customer demand priorities. This provides a comprehensive data foundation for the construction of a risk indicator system, and builds a risk indicator system to achieve quantitative risk assessment.

[0170] We used the Apache Flink real-time computing engine to build a real-time data processing pipeline. We first defined the input source, connecting it to the predicted order volume stream and the equipment status stream. In Flink, we used the join operation to link the order volume data and equipment status data based on the unique identifier of the production line. The predicted order volume stream contains information about future order quantities, while the equipment status stream provides real-time feedback on the operating status of production equipment, such as temperature, pressure, and vibration.

[0171] Set the risk event window to , the data is grouped and processed. Within the window, a custom function is applied to calculate the equipment load rate and generate a risk event object. The equipment load rate is an important indicator used to measure the production pressure of the equipment. By calculating the equipment load rate, it is possible to timely discover whether the equipment has an overload risk. The risk event object contains key information related to the risk, 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] In order to dynamically adjust the risk threshold, a prediction error data collection mechanism is established. , extract the difference between the predicted order quantity and the actual order quantity from the log file or database of the prediction module, calculate the mean of the prediction error, and update the risk threshold in real time; the expression is as follows:

[0173]

[0174] Where, 、 are the risk thresholds before and after the update, is the learning rate, is the recent deviation rate, and is also the mean of the prediction error. In this embodiment, ;

[0175] Determine the nodes of the risk probability model based on the Bayesian network. Based on the logical relationship between risk indicators and business experience, build 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, equipment status, raw material inventory, and the time and type of risk events. Preprocess this data and convert it into a format suitable for model training. Use historical data to train the risk probability model. Using maximum likelihood estimation, determine the conditional probability relationship between each node, reflecting the mutual influence and dependency between different risk indicators.

[0177] During the operation of the system, as new production data and risk events continue to occur, , 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 changes in the production environment in a timely manner and improve the accuracy of risk probability prediction;

[0178] Formulate production decision-making rules based on risk probability, based on the company's risk tolerance and production strategy. The company's risk tolerance is determined by its financial situation, market position, and goals, while the production strategy takes into account production efficiency, cost control, and customer satisfaction.

[0179] When new forecast order data is input, the risk probability model calculates the risk probability of the forecast order based on the current risk index value using the Bayesian formula, and generates the final production decision result based on the production decision judgment rules, including whether to arrange the forecast order for production, corresponding decision recommendations, and related information of the forecast order, such as normal production, proportional production, and prohibited production.

[0180] In summary, the present invention collects multi-source heterogeneous data from customers, production, supply chain and market sides through multiple channels, solidifies key data using blockchain technology and combines hash algorithms to ensure data authenticity, and stores the data in a multi-source heterogeneous database after edge computing preprocessing; then, a four-dimensional relationship network is constructed through a domain-enhanced BERT model to generate a demand knowledge graph, integrates customer portraits and production resource characteristics to train an order prediction model, combines normal distribution to simulate order fluctuation paths and screen effective paths, and generates order quantity fluctuation matrices with different confidence levels; then, the risk threshold is dynamically adjusted based on the Bayesian network, a risk indicator system is constructed for real-time risk assessment, a three-level decision rule is formulated to output production decisions, an alarm is triggered when the deviation rate exceeds the threshold, production data is integrated to construct a causal graph and input into the DoWhy model to analyze the root cause, the prediction model hyperparameters are optimized through random search combined with Gaussian process regression, and SHAP values ​​are applied to iteratively screen feature vectors; finally, the risk heat map and production status are displayed through a visual dashboard, a manual correction interface is provided, and a production scheduling report is generated to achieve multi-terminal interactive early warning.

[0181] The above description is merely a preferred embodiment of the present invention. The scope of protection of the present invention is not limited to the above embodiment. All technical solutions based on the concept of the present invention are within the scope of protection of the present invention. It should be noted that for those skilled in the art, various improvements and modifications that do not depart from the principles of the present invention should also be considered within the scope of protection of the present invention.

Claims

1. A production scheduling strategy adjustment method, characterized in that: include: Build a multi-source heterogeneous database, set up semantic association methods, construct a four-dimensional relationship network, and convert behavioral semantics into quantifiable weight parameters to generate a demand knowledge graph. Set up prediction and correction methods, build an order prediction model, output predicted order volume, predicted order time and confidence interval, and generate order volume fluctuation paths through normal distribution, generating order volume fluctuation matrices with different confidence levels. Construct a basic data set for risk assessment, set risk assessment methods, dynamically adjust risk thresholds, build a risk probability model, and generate production decision results based on the established production decision judgment rules; Calculate the factory deviation rate, set up cause-and-effect analysis, use thresholds to detect deviations and trigger alarms, integrate production data to build cause-and-effect diagrams, and generate cause-and-effect effect reports; Calculate the deviation rate of deviation detection, set up a self-optimization method, perform iterative operations through random search combined with Gaussian process regression to obtain the optimal hyperparameter combination, and iterate feature engineering to optimize the input feature vector; When generating a demand knowledge graph, the semantic association method includes: Acquire text data from the multi-source heterogeneous database, calculate word frequency and inverse document frequency, filter high-frequency professional vocabulary by TF-IDF value, and build a dynamic domain dictionary; The Skip-Gram model is used to iteratively train high-frequency words to generate a word vector matrix and embed it into the word embedding layer of the basic BERT model to generate a recognition model. Use the RoBERTa model to analyze the sentiment tendency of customer service conversation records, output the sentiment score and map it to weight parameters, and generate an entity set with sentiment weights; Structured data is directly used as knowledge graph entities, and unstructured data is sequence-labeled through recognition models. Combined with entity sets with sentiment weights, a unified entity list is generated. Perform logarithmic smoothing on the number of entity co-occurrences to generate basic weights, and adjust the basic weights based on customer behavior events to construct a four-dimensional relationship network. Build a Neo4j-based graph database instance to store a knowledge graph consisting of several nodes and several edges.

2. The production scheduling strategy adjustment method according to claim 1, characterized in that: When building an order forecasting model, the forecasting and correction method includes: Based on the customer ID, locate the customer node in the graph database instance, obtain the customer's attribute information, and calculate the sliding window mean of the order volume in different time periods to generate customer profile features; Obtaining attribute information of the equipment in the graph database instance, calculating the theoretical production capacity of the equipment, and generating production resource feature information; Integrate customer profile features and production resource features extracted from the knowledge graph with traditional order data features to generate an input feature vector; Build an order prediction model. The input layer receives the constructed input feature vector and uses the Huber loss function and AdamW optimizer to train the order prediction model. Based on the multi-source heterogeneous database, a real-time vector is generated based on the latest data, and the real-time vector is input into the order prediction model to obtain a customer-level predicted order quantity and predicted order time.

3. The production scheduling strategy adjustment method according to claim 2, characterized in that: The prediction and correction method further includes: Obtaining normal distribution characteristics of order volume fluctuations, combining the normal distribution characteristics with the predicted order volume, and generating an order volume path; Eliminate orders that exceed production capacity based on equipment capacity and raw material supply constraints, and generate an effective order volume fluctuation path; Constructing simulated samples based on order time, sorting the simulated samples in ascending order order time, and generating a confidence interval for the predicted order time; Obtaining customer grade weight information from the knowledge graph and calculating weighted standard deviations of different customer grades; The predicted value of the order forecasting model and the weighted standard deviation are combined to calculate the fluctuation range of the order quantity according to quantiles of different confidence levels to generate an order quantity fluctuation matrix.

4. The production scheduling strategy adjustment method according to claim 3, characterized in that: In formulating production decision-making outcomes, the risk assessment approach includes: Use the Apache Flink real-time computing engine to build a data processing pipeline, accessing the predicted order volume stream and device status stream; Joining the two data streams using the unique identifier of the production line to correlate the predicted order quantity data with the real-time operating status of the equipment; Define risk event windows, group data, apply custom functions to calculate device load rates, and generate risk event objects; Based on a preset time interval, the difference between the predicted order quantity and the actual order quantity is extracted, the mean of the prediction error is calculated, and the risk threshold is updated in real time.

5. The production scheduling strategy adjustment method according to claim 4, characterized in that: The risk assessment method further comprises: Determine the nodes of the risk probability model based on the Bayesian network, and build the risk probability model based on the logical relationship between risk indicators and business experience; Collect historical production data and risk event data, use the maximum likelihood estimation method to train the risk probability model using historical data, and determine the conditional probability relationship between each node; Based on the time interval, new data is collected to retrain the model and update the conditional probabilities between nodes; When a new forecast order is input, the risk probability model is used to calculate the risk probability, and a production decision result is generated based on the constructed production decision judgment rule.

6. The production scheduling strategy adjustment method according to claim 5, characterized in that: When constructing a causal graph, the causal analysis method includes: Collect actual factory output data and forecast factory output data in real time according to the preset analysis time interval, and calculate the factory output deviation rate; Obtain the actual delivery date of the order, combine it with the planned delivery date in the production schedule and the standard production cycle, and calculate the delivery deviation rate; Set factory deviation thresholds and delivery deviation thresholds to conduct deviation detection; Once the factory quantity deviation rate exceeds the factory deviation threshold, the factory deviation is abnormal, and an alarm is immediately sent; Once the delivery deviation rate exceeds the delivery deviation threshold, the delivery deviation is abnormal, and the emergency response mechanism is immediately activated; Based on the intrinsic relationship between production processes and data, a causal graph containing multiple nodes and edges is constructed, the causal effect of each factor on the deviation is calculated, and a causal effect report is generated.

7. The production scheduling strategy adjustment method according to claim 6, characterized in that: When optimizing the input feature vector, the self-optimization method includes: Count the number of deviation anomalies and the total number of detections within the preset time period and calculate the deviation rate; When the deviation rate is less than or equal to the preset self-optimization threshold, the next time order forecast is performed; when the deviation rate is greater than the self-optimization threshold, the next step is to optimize the order forecast model; Determine the key hyperparameters and search ranges that affect performance based on historical data and model characteristics; We use random sampling combined with Gaussian process regression to iteratively select hyperparameter combinations, train the model, and evaluate performance. Avoid local optimality by adding exploration noise until the number of iterations is reached and the optimal hyperparameter combination is determined.

8. The production scheduling strategy adjustment method according to claim 7, characterized in that: The self-optimization method further comprises: Calculating the SHAP value of the feature in the input feature vector using a SHAP value analysis method; Eliminate features whose SHAP values ​​are less than a preset contribution threshold, combine business experience and domain knowledge to mine new features, and generate an optimized input feature vector; Retraining the order prediction model based on the new input feature vector, evaluating the performance using a validation dataset, and determining whether to retain the new feature; If the addition of the new feature improves the order prediction model, the new feature is retained; if the new feature fails to improve the performance of the order prediction model, the new feature is removed; The final input feature vector is embedded into the order prediction model for production scheduling prediction, and the prediction results are applied to actual production scheduling.

Images