A Deep Learning-Based Digital Collaborative Management Method and System for MES

Abstract

This invention relates to the field of deep learning technology and discloses a deep learning-based digital collaborative management method and system for MES (Manufacturing Execution System). The method includes: collecting multi-source heterogeneous data to construct a dimensionless multidimensional dataset; then, through node feature embedding and edge relationship learning, analyzing the dynamic correlation between data items to generate a production line operation dependency graph; combining the dynamic correlation weights between data items to perform collaborative anomaly detection and root cause localization, thereby obtaining causal correlation data between abnormal events and data parameters; constructing a digital twin simulation environment to simulate the impact path of heterogeneous data parameters intervening in production disturbances on collaborative anomalies, in order to evaluate the impact results of anomalies; and then, based on a deep learning framework, constructing an iterative scheduling model to iteratively generate a multi-objective optimized collaborative scheduling strategy, realizing dynamic configuration and anomaly prevention among heterogeneous data; thereby enabling collaborative optimization management of production operation parameter configuration and anomaly prevention in large-scale production with high dimensionality, high complexity, and dynamic changes, using multi-source heterogeneous data.

Description

Technical Field

[0001] This invention relates to the field of deep learning technology, and more specifically, to a deep learning-based digital collaborative management method and system for MES (Management Execution System). Background Technology

[0002] With the rapid development of Industry 4.0 and intelligent manufacturing, Manufacturing Execution Systems (MES), as a key link connecting Enterprise Resource Planning (ERP) and Production Control Systems (PCS), play a crucial role in the digital transformation of SMEs. Currently, the digital management systems of SMEs mainly employ discrete point-based analysis methods, viewing the production system as a network structure composed of independent nodes such as production workshops, warehouses, and logistics centers, along with business processes. This heterogeneous data is mostly based on existing production management technologies for optimization decisions, making it difficult to handle high-dimensional, highly complex, and dynamically changing large-scale production management tasks.

[0003] Therefore, how to achieve collaborative optimization management of production operation parameter configuration and anomaly prevention in large-scale production with high dimensionality, high complexity and dynamic changes of multi-source heterogeneous data has become an urgent problem to be solved. Summary of the Invention

[0004] This invention provides a deep learning-based digital collaborative management method and system for MES, which solves the technical problem of how to achieve collaborative optimization management of production operation parameter configuration and anomaly prevention in large-scale production with high-dimensionality, high complexity and dynamic changes of multi-source heterogeneous data.

[0005] This invention provides a deep learning-based digital collaborative management method and system for MES (Manufacturing Execution System), comprising:

[0006] Firstly, a deep learning-based digital collaborative management method for MES includes the following steps:

[0007] Collect heterogeneous data from multiple sources, and construct a dimensionless multidimensional dataset through preprocessing and normalization.

[0008] The dimensionless multidimensional dataset is mapped to graph structure nodes to obtain node features, and the connection exchange, information flow and dependency relationship between parameters in the dimensionless multidimensional dataset are defined as edge relationships.

[0009] By using node feature embedding and edge relationship learning, the dynamic relationship between data items in the dimensionless multidimensional dataset is analyzed to generate a production line operation dependency graph.

[0010] A collaborative analysis mechanism for multi-source heterogeneous data is constructed to obtain the dynamic correlation weights among the collaborative modes of multi-source heterogeneous data. Anomaly detection and root cause localization are performed by combining the dependency relationship graph to obtain causal relationship information between abnormal events and data modes.

[0011] A digital twin simulation environment is constructed based on causal relationship information to simulate the impact path of heterogeneous data parameters on production disturbances on collaborative anomalies, so as to evaluate the abnormal impact results of parameter adjustments.

[0012] An iterative scheduling model is constructed based on dimensionless multidimensional datasets, dependency graphs, and causal relationship information. By combining the results of anomaly impact assessment with real-time production status, a multi-objective optimized collaborative scheduling strategy is generated iteratively to achieve dynamic configuration and anomaly prevention among multi-source heterogeneous data.

[0013] Furthermore, multi-source heterogeneous data are collected and preprocessed and normalized to construct a dimensionless multidimensional dataset, including:

[0014] Real-time acquisition of multi-source heterogeneous data during production and manufacturing operations and data preprocessing to obtain a preprocessed dataset;

[0015] The preprocessed dataset is standardized and cleaned to obtain a clean dataset.

[0016] The clear dataset is time-aligned and normalized after being collected at a preset period to obtain a dimensionless multidimensional dataset.

[0017] Furthermore, through node feature embedding and edge relationship learning, the dynamic relationships between data items in the dimensionless multidimensional dataset are analyzed to generate a production line operation dependency graph, including:

[0018] The dimensionless multidimensional dataset is mapped to graph structure nodes, and the parameters of multi-source heterogeneous data are used as node features.

[0019] Define the connection, exchange, information flow, and dependency relationships between parameters of multi-source heterogeneous data as edge relationships;

[0020] Node feature embedding is performed using a graph attention network structure to capture the local correlation of edge relationships;

[0021] The interaction strength between multi-source heterogeneous data modalities is quantified through a side relation feature learning mechanism.

[0022] Based on local correlation and interaction strength, a multi-layer graph network is used for iterative propagation to analyze the dynamic relationships between data items in a dimensionless multidimensional dataset, generating a production line operation dependency graph that includes the relationship status between heterogeneous data, production operation parameters, and parameter quality indicators.

[0023] Furthermore, the dynamic changes in the parameter states of multi-source heterogeneous data are processed through a temporal graph convolutional network structure to fuse temporal features with graph network topology features to obtain a fusion result.

[0024] Based on the fusion results, the static attributes of each data item parameter in the dimensionless multidimensional dataset are concatenated with the dynamic time-series data to reduce the dimensionality and generate a fusion feature vector of a single node, which serves as the basic input for constructing the production line operation dependency graph.

[0025] Explicit edges are defined by interaction strength connection, while implicit edges are generated by obtaining the cosine similarity between the fused feature vectors of each node. After fusion, a weighted adjacency matrix is ​​formed, which, together with the node features, outputs dynamic edge weights.

[0026] By learning the dynamic edge weights between node features, the dynamic edge weights are updated by aggregating the multi-attention results, and the updated node features are generated by fusing the features of neighboring nodes.

[0027] The dynamic edge weights and updated node features together form a production line operation dependency graph, revealing the explicit associations and implicit dynamic dependencies between multi-source heterogeneous data items during production operation, in order to support production line operation decisions.

[0028] Furthermore, a multi-source heterogeneous data collaborative analysis mechanism is constructed to obtain the dynamic correlation weights among the collaborative modes of multi-source heterogeneous data. This is combined with dependency graphs for anomaly detection and root cause localization to obtain causal correlation information between anomalous events and data modes, including:

[0029] Feature extraction and representation learning are performed on the states of each data item in a dimensionless multidimensional dataset to obtain the correlation matrix between multi-source heterogeneous data modes.

[0030] The contribution of multi-source heterogeneous data modes is dynamically adjusted by using dynamic edge weights to obtain the dynamic correlation weights for collaboration among multi-source heterogeneous data modes;

[0031] Capture the correlation influence between modalities of multi-source heterogeneous data based on dynamic correlation weights;

[0032] pass: To generate a dynamic correlation weight matrix;

[0033] in, This represents the dynamic association weight matrix, and Q represents the query matrix. Let V represent the key matrix at time T, V represent the value matrix, d represent the feature dimension, and softmax represent the normalization function.

[0034] By combining the dynamic association weight matrix with the production line operation dependency graph, an enhanced graph structure topology is constructed.

[0035] By identifying node features and edge relationships that deviate from the normal pattern in the enhanced graph structure topology;

[0036] pass: Perform anomaly detection to obtain anomaly identification results;

[0037] in, Let x represent the anomaly score, and x represent the input feature. Represents the reconstructed features, Represents the square of the L2 norm;

[0038] Based on the anomaly identification results, the root cause node of the anomaly is located through backpropagation and subgraph extraction to obtain root cause localization data;

[0039] Causal inference is used to quantify the causal association strength between anomaly identification results and root cause localization data, and causal association information is generated. The causal association strength represents the degree of direct influence of anomaly identification results on root cause localization data, which is determined by obtaining the difference between mutual information and conditional mutual information.

[0040] The causal strength of causal association information is:

[0041] ;

[0042] Where C(XY) represents the causal correlation strength between anomaly identification result X and location data Y, I(X,Y) represents the interaction information between anomaly identification result X and location data Y, I(X,Y|Z) represents the conditional interaction information between anomaly identification result X and location data Y under given constraint Z, PA(Y) represents the set of all potential parent nodes of location data Y, and {X} represents the quantized data of anomaly identification result.

[0043] Furthermore, a digital twin simulation environment is constructed based on causal relationship information to simulate the impact path of heterogeneous data parameters intervening in production disturbances on collaborative anomalies, in order to evaluate the anomaly impact results of parameter adjustments, including:

[0044] The acquired causal relationship information is mapped to the digital twin simulation environment to establish the correspondence between physical entities and digital models;

[0045] Based on the production line operation dependency graph, define the interaction rules in the digital twin simulation environment to ensure that the digital twin simulation environment accurately reflects the dynamic characteristics data of actual production operation;

[0046] Based on the dynamic characteristics data of production operation, a parameter intervention experiment plan is designed. By adjusting key parameters in the digital twin simulation environment at fixed points, a production disturbance scenario is simulated.

[0047] Based on production disturbance scenarios, the propagation path of production disturbances is tracked, and the state changes of each node are recorded to analyze the impact of production disturbance scenarios on the collaborative anomalies among multi-source heterogeneous data modalities.

[0048] A comprehensive evaluation of the effects of various parameter adjustment schemes on the suppression or exacerbation of coordinated anomalies in simulated production disturbance schemes is conducted to form a quantitative assessment result of the anomaly impact.

[0049] Furthermore, an iterative scheduling model is constructed based on the dimensionless multidimensional dataset, dependency graph, and causal relationship information, including:

[0050] A state representation space is constructed based on a dimensionless multidimensional dataset, which maps the current state of production operation to a high-dimensional feature vector, containing multidimensional information data such as equipment operating parameters, quality indicators, energy consumption data and safety monitoring data.

[0051] The constraint matrix is ​​defined using the production line operation dependency graph. The constraint matrix represents the prior dependencies, resource sharing and conflict relationships between each production unit, so as to generate a scheduling relationship matrix that meets the production operation constraints.

[0052] An iterative scheduling model is constructed based on multidimensional information data, a scheduling relationship matrix, and causal correlation information. The iterative scheduling model specifically includes:

[0053] Multidimensional information data, scheduling relationship matrix and causal association information are used to generate structured data, and the structured data is encoded into sequence data;

[0054] The sequence data is input into the iterative scheduling model; the iterative scheduling model includes an input layer, a first hidden layer, a second hidden layer, a third hidden layer and an output layer, and transmits the intermediate representation data of multiple hidden layers to the output layer. The output layer iteratively outputs the scheduling strategy prediction results corresponding to the state changes between the parameters of the multi-source heterogeneous data.

[0055] The at least two multidimensional information data, scheduling relationship matrix and causal association information data items acquired in real time are input into the iterative scheduling model, and the iterative output represents the scheduling strategy prediction result corresponding to the state changes among the multi-source heterogeneous data parameters acquired in real time.

[0056] Furthermore, by combining the anomaly impact assessment results with real-time production status, a multi-objective optimized collaborative scheduling strategy is iteratively generated to achieve dynamic configuration and anomaly prevention among multi-source heterogeneous data, including:

[0057] Collect key parameters of real-time production operation status, including equipment operation status, quality inspection data, energy consumption indicators and safety monitoring information, to form a multi-dimensional real-time status vector; and extract the current topology and constraints from the production line operation dependency graph, and combine them with the historical anomaly pattern library to capture anomaly risks in real-time status.

[0058] Based on the captured abnormal risks, multi-step predictions of future state evolution are made to identify potential abnormal development trends; for the detected abnormal risk points, an intervention effect scoring matrix of design parameters is generated based on the abnormal impact assessment results.

[0059] The multi-objective optimization decision-making mechanism is initiated based on the intervention effect scoring matrix of the design parameters, setting key indicators such as production efficiency, quality stability, energy utilization rate and safety risk as optimization objectives, and dynamically adjusting the weight of each objective according to the current production task priority.

[0060] And search for the optimal scheduling strategy under constraints, by:

[0061] ;

[0062] To balance the effects of short-term intervention with long-term stable scheduling strategies;

[0063] Where J(θ) represents the comprehensive optimization objective, Indicates production efficiency targets. Indicates the quality stability target. Indicates the energy utilization rate target, Indicates security risk objectives. , , and These represent the dynamic weighting coefficients for the four objectives: production efficiency, quality stability, energy utilization rate, and safety risk, respectively, and satisfy the following conditions: ;

[0064] Based on the prediction results of the iterative scheduling model, a collaborative scheduling strategy is deployed, and the optimized scheduling instructions are decomposed into specific execution instructions at the device level, process level, and resource level. The specific execution instructions are securely transmitted to each execution unit through the middleware layer, and an instruction execution confirmation mechanism is established to achieve dynamic configuration and anomaly prevention among multi-source heterogeneous data.

[0065] Establish a real-time monitoring and feedback mechanism to continuously track the execution effect of scheduling strategies and collect data on changes in key performance indicators; when execution deviations or new abnormal events are detected, trigger a rapid response mechanism to select fine-tuning or replanning strategies based on the degree of deviation.

[0066] Furthermore, the cooperative scheduling strategy includes:

[0067] The scheduling problem is decomposed into three levels: strategy, tactics, and operation. These levels are used to handle abnormal situations between heterogeneous data in long-term planning, medium-term scheduling, and real-time response, forming a complete hierarchical collaborative decision-making framework.

[0068] By achieving bidirectional integration of constraint scenarios and guidance scenarios, and using a differentiable programming layer to unify hard constraints and soft constraints into the same framework, the constraint weights are dynamically adjusted according to real-time production conditions to maximize the comprehensive response to production operation requirements.

[0069] Based on the company's strategic priorities, changes in market demand, and actual production conditions, dynamically adjust the priorities of multiple objectives such as production efficiency, quality stability, energy utilization, and safety risks to achieve a dynamic balance among these objectives.

[0070] The design combines anomaly prediction with proactive scheduling. By using a time-series prediction model to predict production status in multiple steps and combining it with a digital twin simulation environment for cyclic simulation and evaluation, the mechanism can proactively adjust the state between heterogeneous data parameters before anomalies actually occur, transforming passive response into proactive prevention.

[0071] Establish a knowledge-driven and data-driven approach, and use a distributed collaborative execution framework to make the collaborative scheduling strategy process transparent and efficient.

[0072] Secondly, a deep learning-based MES digital collaborative management system is provided for executing a deep learning-based MES digital collaborative management method, including:

[0073] Data acquisition module: It is used to collect multi-source heterogeneous data, and construct a dimensionless multidimensional dataset through preprocessing and normalization.

[0074] Dynamic Association Module: It is used to map the dimensionless multidimensional dataset into graph structure nodes to obtain node features, and define the connection exchange, information flow and dependency relationship between each parameter in the dimensionless multidimensional dataset as edge relationship; through node feature embedding and edge relationship learning, it analyzes the dynamic association relationship between each data item in the dimensionless multidimensional dataset and generates the production line operation dependency relationship map.

[0075] Detection and localization module: It is used to build a collaborative analysis mechanism for multi-source heterogeneous data to obtain the dynamic correlation weights between multi-source heterogeneous data modes, and combine the dependency relationship map to perform anomaly detection and root cause localization in order to obtain causal relationship information between abnormal events and data modes.

[0076] Anomaly Impact Module: It is used to construct a digital twin simulation environment based on causal relationship information, simulate the impact path of heterogeneous data parameters intervening in production disturbances on coordinated anomalies, and evaluate the anomaly impact results of parameter adjustments;

[0077] Model training module: It is used to build an iterative scheduling model based on a dimensionless multidimensional dataset, dependency graph, and causal relationship information;

[0078] The collaborative management module combines the results of anomaly impact assessment with real-time production status to iteratively generate multi-objective optimized collaborative scheduling strategies, enabling dynamic configuration and anomaly prevention among multi-source heterogeneous data.

[0079] The beneficial effects of this invention are as follows: By constructing a deep learning-based MES digital collaborative management method, this invention effectively solves the problems faced by SMEs in digital transformation, such as data silos, decentralized management of constrained scenarios, lack of intelligent analysis capabilities, and poor system scalability. Furthermore, it breaks through the technical bottleneck of collaborative optimization in constrained scenarios. By expressing the dynamic correlation between data items through a production line operation dependency graph, and constructing a digital twin simulation environment using causal correlation information, it realizes a collaborative scheduling strategy for multi-objective optimization. This enables collaborative optimization management of production operation parameter configuration and anomaly prevention in large-scale production with high dimensionality, high complexity, and dynamic changes, involving multi-source heterogeneous data. It also improves the average level of production efficiency, reduces product defect rates, reduces operating costs, and decreases safety incidents, while promoting data value mining to support innovative decision-making. In terms of economic benefits, the investment return cycle is shortened, and the overall economic benefits are improved. This effectively lowers the threshold for digital transformation for SMEs, enabling them to achieve digital upgrades gradually and significantly enhancing their competitiveness. Attached Figure Description

[0080] Figure 1 This is a schematic diagram of a deep learning-based MES digital collaborative management method provided in an embodiment of the present invention;

[0081] Figure 2 This is a schematic diagram of the module flow of a deep learning-based MES digital collaborative management system provided in an embodiment of the present invention. Detailed Implementation

[0082] The subject matter described herein will now be discussed with reference to exemplary embodiments. It should be understood that these embodiments are discussed only to enable those skilled in the art to better understand and implement the subject matter described herein, and changes may be made to the function and arrangement of the elements discussed without departing from the scope of this specification. Various processes or components may be omitted, substituted, or added as needed in the examples. Furthermore, some features described in the examples may be combined in other examples.

[0083] At least one embodiment of the present invention discloses a deep learning-based digital collaborative management method and system for MES, such as... Figure 1 Shown, including:

[0084] Step 1: Collect heterogeneous data from multiple sources, and construct a dimensionless multidimensional dataset through preprocessing and normalization.

[0085] Step 2: Map the dimensionless cube to graph structure nodes to obtain node features, and define the connection exchange, information flow and dependency relationship between parameters in the dimensionless cube as edge relationship;

[0086] Step 3: Through node feature embedding and edge relationship learning, analyze the dynamic relationship between data items in the dimensionless multidimensional dataset and generate a production line operation dependency graph;

[0087] Step 4: Construct a multi-source heterogeneous data collaborative analysis mechanism to obtain the dynamic correlation weights between multi-source heterogeneous data modalities, and combine the dependency relationship map to perform anomaly detection and root cause localization to obtain causal relationship information between abnormal events and data modalities;

[0088] Step 5: Construct a digital twin simulation environment based on causal relationship information to simulate the impact path of heterogeneous data parameters on production disturbances on collaborative anomalies, so as to evaluate the abnormal impact results of parameter adjustments;

[0089] Step 6: Construct an iterative scheduling model based on the dimensionless multidimensional dataset, dependency graph, and causal relationship information. Combine the anomaly impact assessment results with real-time production status to iteratively generate a multi-objective optimized collaborative scheduling strategy, thereby realizing dynamic configuration and anomaly prevention among multi-source heterogeneous data.

[0090] In this embodiment, step 1 collects structured or unstructured data from multiple data sources (such as sensors, MES, ERP, PLC, and logs), and performs preprocessing steps such as format cleaning, missing value handling, unit unification, and noise filtering for data from different sources. Subsequently, normalization (such as Min-Max normalization and Z-score standardization) is used to unify the differences in numerical ranges, constructing a dimensionless multidimensional dataset with a unified structure; achieving semantic and dimensional unification of heterogeneous data; reducing noise and interference in model input, improving the accuracy of subsequent analysis; and providing a consistent data foundation for subsequent mapping, analysis, and modeling.

[0091] Step 2 maps each data parameter in the dimensionless multidimensional dataset to a node in the graph, and uses its connections, signal transmission paths, or logical dependencies in actual business / production lines as edges to construct a graph structure model. Nodes contain the original parameter features, while edges represent the mutual influence relationships between data. Transforming the originally planar data structure into a graph structure is beneficial for expressing complex dependencies; it facilitates the capture of interaction relationships and non-linear transmission paths between data items; and it lays the structural foundation for subsequent deep feature mining using graph neural networks.

[0092] Step 3 uses graph embedding methods (such as GNN, GraphSAGE, GCN) to encode node features, and obtains dynamic semantic relationships between nodes through edge weight learning and message passing mechanisms, thereby generating a dependency graph that accurately depicts the actual production line operation; realizing the visualization and structured expression of production line data relationships; capturing the dynamic dependencies between parameters that change over time in actual operation; and laying the foundation for accurate anomaly identification and prediction.

[0093] Step 4 establishes a cross-fusion mechanism between different modalities (e.g., image, text, temperature, pressure) and extracts dynamic weights between modalities through modal collaborative learning. Combining this with the previously obtained dependency graph, graph matching and anomaly score calculation (e.g., GAD, GraphAE) are applied to identify potential anomalies. Furthermore, causal reasoning techniques (e.g., Granger causality, Do-Calculus) are used to locate root causes. This improves the anomaly detection capability in complex multi-source data scenarios, accurately uncovers root causes of problems, shortens investigation time, and supports starting from causal relationships rather than just correlations, enhancing interpretability.

[0094] Step 5 involves inputting key causal relationship information into the digital twin system to simulate the system's response behavior under various parameter disturbances, analyze the specific impact of intervention measures on the abnormal propagation path, and predict "what consequences would occur if a certain abnormal situation occurred"; conducting virtual simulation experiments without interrupting the real production line; assessing the risks of parameter adjustment in advance to avoid blind intervention; and providing a more realistic reference for optimization strategies.

[0095] By integrating standardized data, operational dependency graphs, and causal information, a scheduling optimization model with multiple constraints and objectives is constructed. This model iteratively evolves scheduling strategies through reinforcement learning, evolutionary algorithms, or optimizers, and dynamically adjusts them according to actual production conditions to achieve closed-loop optimization and anomaly prevention capabilities; it enables real-time scheduling optimization based on a global data perspective; it enhances the robustness and flexible responsiveness of the production line; and it can automatically adjust scheduling when system anomalies occur, improving automatic recovery capabilities.

[0096] In this preferred embodiment of Example 1, step 2 involves collecting multi-source heterogeneous data, and then preprocessing and normalizing it to construct a dimensionless multidimensional dataset, including:

[0097] Step 21: Collect multi-source heterogeneous data during production and manufacturing operations in real time and perform data preprocessing to obtain a preprocessed dataset;

[0098] Step 22: Standardize and clean the preprocessed dataset to obtain a clean dataset;

[0099] Step 23: Perform time-series alignment processing on the clear dataset, and perform normalization processing after collecting data at a preset period to obtain a dimensionless multidimensional dataset.

[0100] In this embodiment, step 21 typically relies on sensors, monitoring systems, or equipment interfaces at the production site to collect multi-source heterogeneous data from different production stages and equipment. Data types may include temperature, humidity, pressure, equipment status, production speed, and quality control indicators. Data preprocessing: Since the data comes from multiple sources and varies in format and precision, the main task of preprocessing is to convert data from different sources into a unified format and resolve common problems such as missing values, noisy data, and outliers. Data cleaning: For example, detecting and repairing missing or erroneous values, and removing extreme outliers. Format standardization: Unifying heterogeneous data from different sources into a format that facilitates subsequent processing (such as timestamps, unit conversion, and data type standardization); thereby ensuring that the data has a unified format for the next step of analysis, facilitating subsequent processing; laying the foundation for subsequent standardization and cleaning, improving data quality and usability; and avoiding the impact of data noise or inconsistency on the analysis results.

[0101] Step 22, data standardization, transforms the numerical values ​​of different features to the same scale, ensuring consistent measurement standards for different data items. Common standardization methods include: Z-score standardization: subtracting the mean and dividing by the standard deviation to achieve zero mean and unit variance; and Min-Max standardization: scaling the data to a specified interval (e.g., [0, 1]), suitable for uneven numerical distributions. Standardized data ensures consistent dimensions for different features (such as temperature and humidity), giving them equal influence in subsequent modeling. Data cleaning further addresses issues like missing values, duplicate data, and erroneous values ​​in the preprocessed dataset. Examples include: missing value imputation: using mean imputation, interpolation methods, or model-based prediction; deduplication: ensuring no duplicate records; and outlier handling: removing or correcting potentially erroneous data points. Standardization guarantees numerical consistency, preventing certain features from causing imbalances in the model due to scale issues; cleaning removes erroneous and inconsistent data, making subsequent data analysis more reliable and reducing interference factors in model training.

[0102] Step 23: Different production equipment and sensors may have some time discrepancies or different data acquisition frequencies. The purpose of time alignment is to align data from different sources by time point, ensuring that each dataset has the same timestamp. For example: Interpolation: If the data acquisition times are inconsistent, use linear interpolation or other interpolation methods to align the data to a unified time axis. Data synchronization: Adjust the data acquisition cycle to ensure all data points are within a uniform time window. Acquiring data at a preset cycle: Some data may be acquired at irregular time intervals. To reduce data redundancy and overfitting risks, sample the data at a fixed time period (e.g., per second, per minute). Periodic resampling: Ensure data uniformity and remove overly frequent data fluctuations.

[0103] Normalization: Normalization scales data proportionally to give different features the same units. Min-max normalization is typically used to scale the data to the range [0, 1] or [-1, 1]. Dimensionless Datasets: Normalized datasets are unaffected by the original units of the features, allowing subsequent analysis models to more effectively understand the relative relationships between data. Consistency on the time axis avoids the impact of deviations from different data collection periods on subsequent analysis. Periodic acquisition reduces data redundancy, making the system more efficient and adaptable to different acquisition frequencies.

[0104] Normalization eliminates differences in units or scales across various features of the data, improving the accuracy of subsequent analysis and preventing the model from being overly sensitive to specific units.

[0105] In this preferred embodiment of Example 1, step 3 involves parsing the dynamic relationships between data items in the dimensionless multidimensional dataset through node feature embedding and edge relationship learning, generating a production line operation dependency graph, including:

[0106] Step 31: Map the dimensionless multidimensional dataset into graph structure nodes, and use the parameters of the multi-source heterogeneous data as node features;

[0107] Step 32: Define the connection exchange, information flow, and dependency relationships between parameters of multi-source heterogeneous data as edge relationships;

[0108] Step 33: Use a graph attention network structure to embed node features and capture the local correlation of edge relationships;

[0109] Step 34: Quantify the interaction strength between multi-source heterogeneous data modalities through the edge relation feature learning mechanism;

[0110] Step 35: Based on local correlation and interaction strength, iterative propagation is carried out through a multi-layer graph network to analyze the dynamic relationship between data items in the dimensionless multidimensional dataset, and generate a production line operation dependency graph containing the relationship status between heterogeneous data, production operation parameters and parameter quality indicators.

[0111] In this embodiment, step 31 first transforms the standardized multidimensional dataset into a graph structure and maps different data items to nodes in the graph. These data items can come from different sources (such as sensor data, production parameters, and environmental data), hence they are referred to as "multi-source heterogeneous data." This heterogeneous data is embedded in the graph structure as node features, meaning that each individual node carries corresponding data feature information. This allows a single node to not only represent a data point but also contain the specific parameters or values ​​of that data. By mapping multidimensional data to a graph structure and assigning features to individual nodes, it facilitates the application of subsequent graph algorithms, especially graph neural network (GNN) models. This structure provides a foundation for capturing complex relationships between data and lays a solid foundation for information flow and processing in subsequent steps.

[0112] Step 32 defines the connections, exchanges, information flows, or dependencies between different nodes as "edge relationships." These edges reflect the interactions or relationships between nodes. For example, there may be a certain dependency between sensor data and control parameters in a production line, or there may be connections and exchanges between the operating status of certain equipment and environmental parameters. The definition of edge relationships can be based on business logic, physical models, and data dependency factors; defining edge relationships can effectively reflect the interdependence and dynamic changes between data in the graph structure. This helps in subsequent analysis and learning of the interactive effects between nodes, and edge relationships can capture the physical, temporal, or functional correlations between different parameters, improving the system's adaptability to complex business scenarios.

[0113] Step 33 employs a graph attention network (GAT) to embed node features. GAT uses an attention mechanism to weight the influence of neighboring nodes, making important nodes have a greater impact on the final representation. This mechanism automatically focuses on neighboring nodes that are more important in the current task, effectively capturing the local correlation of edge relationships. The graph attention network, through its local attention mechanism, can capture the different influence strengths between nodes in edge relationships. This allows the model to flexibly adjust its focus among different nodes and edges according to task requirements, thereby enhancing its ability to model complex relationships. This helps improve the data model's sensitivity to subtle dependencies between nodes.

[0114] Step 34 uses an edge relationship feature learning mechanism to learn the features of each edge in the graph, quantifying the interaction strength between different data items (modalities). Different data modalities (such as temperature, humidity, and voltage) may have varying degrees of impact on the system's operation. By learning the features of each edge, the model can identify which data items have strong interactions and which have weak interactions, thereby quantifying the strength of their influence. This quantification of edge relationships can provide more accurate input for subsequent decision-making. By identifying and quantifying the interaction strength between different data items, the model can optimize the system's resource scheduling and operating strategies, improving the overall system's operating efficiency and reliability.

[0115] Step 35: Based on local correlation and interaction strength, iterative propagation through a multi-layer graph network is used to analyze the dynamic relationships between data items in the dimensionless multidimensional dataset, generating a production line operation dependency graph that includes the correlation status between heterogeneous data, production operation parameters, and parameter quality indicators. Through iterative propagation via a multi-layer graph network, based on the local correlation and edge interaction strength learned in previous steps, the dynamic relationships between data items in the dimensionless multidimensional dataset are analyzed. This process, through multi-layer propagation of the graph network, allows for the full fusion of feature information from each node, thereby generating a production line operation dependency graph that includes the correlation status between heterogeneous data, production operation parameters, and quality indicators. This graph clearly shows how each stage of the production process interacts and their impact on the overall production process; it not only visually displays the dynamic relationships between different data items but also reveals potential bottlenecks and optimization opportunities in the production process. Using this graph, production line managers can more effectively monitor the production process and make timely adjustments and optimizations to ensure high production efficiency and quality control.

[0116] Specifically, non-numerical attributes such as equipment type (e.g., "machine tool type A") are mapped to computer-understandable numerical vectors through an embedding layer (e.g., converting categories into 128-dimensional vectors).

[0117] Dynamic feature extraction: A bidirectional LSTM model is used on time-series data (such as temperature records for 10 consecutive minutes) to capture the dependencies between data and extract dynamic feature vectors (e.g., 256-dimensional vectors).

[0118] Feature fusion: The static and dynamic vectors are concatenated into a single long vector, which is then compressed and reduced in dimensionality using a fully connected layer to obtain the initial feature representation for each node. Output data: The fused feature vector for each node serves as the input for subsequent graph structure construction.

[0119] Explicit edge definition: Define the initial edge directly based on the physical connection between devices (such as the direction of the conveyor belt) or the process flow diagram (for example, the initial weight of the edge from device A to device B is 1).

[0120] Implicit edge discovery: Calculate the cosine similarity between all node features. If the similarity exceeds a threshold (e.g., 0.8), then generate an implicit edge (e.g., sensor X and device Y are associated because of their similar vibration patterns).

[0121] Adjacency matrix fusion: Explicit edges (with fixed weights) and implicit edges (with cosine similarity as weights) are merged to form an initial weighted adjacency matrix. Output data: The adjacency matrix containing explicit and implicit edges, along with node feature vectors, are used as input to the graph attention network.

[0122] For each node and its neighbors, dynamic weights are calculated using an attention mechanism. For example:

[0123] Use multiple attention heads (e.g., 8), each head independently learns the correlation strength between nodes.

[0124] The attention weight between node A and its neighbor B may represent "the degree of influence of device B on device A when the current fluctuates".

[0125] Edge weight update: Aggregate the calculation results of multiple attention heads and update the original edge weights (for example, the weight of device A → device B is increased from 0.8 to 1.2).

[0126] Node Feature Update: Based on the new edge weights, neighbor features are aggregated to generate richer node representations (e.g., the features of device A incorporate the dynamic states of upstream and downstream devices). Output Data: The updated adjacency matrix (dynamic edge weights) and node features are used to generate the final dependency graph.

[0127] In this preferred embodiment of the first example, the dynamic changes in the multi-source heterogeneous data parameter states are processed through a temporal graph convolutional network structure, so as to fuse the temporal features with the graph network topology features to obtain the fusion result;

[0128] Based on the fusion results, the static attributes of each data item parameter in the dimensionless multidimensional dataset are concatenated with the dynamic time-series data to reduce the dimensionality and generate a fusion feature vector of a single node, which serves as the basic input for constructing the production line operation dependency graph.

[0129] Explicit edges are defined by interaction strength connection, while implicit edges are generated by obtaining the cosine similarity between the fused feature vectors of each node. After fusion, a weighted adjacency matrix is ​​formed, which, together with the node features, outputs dynamic edge weights.

[0130] By learning the dynamic edge weights between node features, the dynamic edge weights are updated by aggregating the multi-attention results, and the updated node features are generated by fusing the features of neighboring nodes.

[0131] The dynamic edge weights and updated node features together form a production line operation dependency graph, revealing the explicit associations and implicit dynamic dependencies between multi-source heterogeneous data items during production operation, in order to support production line operation decisions.

[0132] In this embodiment, a temporal graph convolutional network (T-GCN) structure is used to process dynamically changing multi-source heterogeneous data parameter states. Multi-source heterogeneous data refers to data from different devices, systems, sensors, or data types (such as temperature, speed, pressure, and text logs). A T-GCN is employed, combined with: time-dimensional modeling (RNN, GRU, LSTM) to capture temporal trends; and graph convolution (GCN) to model the dependencies between data items using a graph structure.

[0133] Simultaneously acquiring dynamic evolution information (time series) and structural relationship information (topology) of the data lays the foundation for subsequent deep fusion; it can adapt to scenarios where data is updated and changed in real time.

[0134] Dimensionality Reduction through Concatenation: Static attributes + dynamic time-series data to generate a fused feature vector: For each data item (e.g., temperature sensor); static attributes (e.g., device number, location, type) + dynamic time-series features (e.g., temperature changes over a past period); after concatenation, dimensionality reduction methods (e.g., PCA, AutoEncoder, MLP) are used to obtain a fused feature vector with unified dimensions. This unifies the representation of multimodal data, resolving the issue of inconsistent dimensionality; it integrates static and dynamic information, enhancing the semantic expressiveness of each node; and dimensionality reduction improves computational efficiency and reduces redundant features.

[0135] Graph construction: Explicit edges + implicit edges, generating a weighted adjacency matrix:

[0136] Explicit edges: Strong connections determined by business rules or process flow (such as process paths and physical connections between equipment).

[0137] Implicit edges: Calculate cosine similarity based on the fused feature vectors of each node;

[0138] Set a threshold or construct k-nearest neighbors and define similarity edges; after merging the two types of edges, a weighted adjacency matrix is ​​formed.

[0139] Simultaneously modeling explicit physical relationships and implicit data similarity; more comprehensively reflecting the data flow logic and dependencies in the production line; and facilitating the capture of potential impact paths that may exist but have not been explicitly modeled.

[0140] Dynamic edge weight learning + multi-attention mechanism for updating node features: Through graph attention mechanism (GAT), it learns the different importance of different neighbors to the target node; the multi-head attention mechanism enhances the model's robustness and expressive power; it aggregates neighbor node features and generates weighted updated node representations; simultaneously, it updates edge weights to reflect changes in the dynamic dependence strength between nodes. This results in more accurate modeling of dynamic interaction patterns between nodes; adaptability to changing production line states, improving the model's perception of unexpected situations; and dynamic updates to the graph structure itself, making the graph more timely and real-time.

[0141] Construct a production line operation dependency graph to support operational decisions; generate the following from the above steps:

[0142] Updated node features; dynamic edge weights (explicit + implicit); forming a complete, time-evolving production line operation dependency graph. Revealing deep, dynamic relationships between multi-source heterogeneous data items; supporting decision-making tasks such as: fault prediction and diagnosis; process optimization; production line adjustment and scheduling; abnormal behavior detection; and constructing a production line-level "digital twin" model.

[0143] In this preferred embodiment of Example 1, step 4 constructs a multi-source heterogeneous data collaborative analysis mechanism to obtain dynamic correlation weights among multi-source heterogeneous data modalities. This is combined with a dependency graph for anomaly detection and root cause localization to obtain causal correlation information between abnormal events and data modalities. This includes:

[0144] Step 41: Perform feature extraction and representation learning on the states of each data item in the dimensionless multidimensional dataset to obtain the correlation matrix between the multi-source heterogeneous data modes;

[0145] Step 42: Dynamically allocate and adjust the contribution of multi-source heterogeneous data modes using dynamic edge weights to obtain the dynamic correlation weights for collaboration among multi-source heterogeneous data modes;

[0146] Step 43: Capture the correlation influence between multi-source heterogeneous data modes based on dynamic correlation weights;

[0147] pass: To generate a dynamic correlation weight matrix;

[0148] in, This represents the dynamic association weight matrix, and Q represents the query matrix. Let V represent the key matrix at time T, V represent the value matrix, d represent the feature dimension, and softmax represent the normalization function.

[0149] Step 44: Combine the dynamic association weight matrix with the production line operation dependency graph to construct an enhanced graph structure topology;

[0150] Step 45: Identify node features and edge relationships that deviate from the normal pattern on the enhanced graph structure topology;

[0151] pass Perform anomaly detection to obtain anomaly identification results;

[0152] in, Let x represent the anomaly score, and x represent the input feature. Represents the reconstructed features, Represents the square of the L2 norm;

[0153] Step 46: Based on the anomaly identification results, locate the root cause node of the anomaly through backpropagation and subgraph extraction to obtain root cause localization data;

[0154] Step 47: Quantify the causal association strength between the anomaly identification results and the root cause localization data using causal inference, and generate causal association information. The causal association strength represents the degree of direct influence of the anomaly identification results on the root cause localization data, and is determined by obtaining the difference between mutual information and conditional mutual information.

[0155] The causal strength of causal association information is:

[0156] ;

[0157] Where C(XY) represents the causal correlation strength between anomaly identification result X and location data Y, I(X,Y) represents the interaction information between anomaly identification result X and location data Y, I(X,Y|Z) represents the conditional interaction information between anomaly identification result X and location data Y under given constraint Z, PA(Y) represents the set of all potential parent nodes of location data Y, and {X} represents the quantized data of anomaly identification result.

[0158] In this embodiment, step 41 performs feature extraction and deep representation learning on dimensionless multidimensional data. Models such as autoencoders, graph neural networks (GNNs), and Transformers can be used to capture hidden relationships between different modalities of data (such as sensor data, logs, and videos). This achieves information alignment and unified representation between different data sources, reduces redundancy, improves data fusion efficiency, and lays a foundation for subsequent graph construction and causal analysis.

[0159] Step 42 uses dynamic modeling methods (such as attention mechanisms and graph attention networks GAT) to dynamically adjust the weights between modalities, reflecting the changes in the importance of each data modality in different scenarios; overcoming the problem of insufficient response of static modeling to changes in complex systems; achieving collaborative perception among multiple modalities; and adaptively enhancing the data weights of key modalities.

[0160] Step 43 aggregates the dynamic weights obtained in Step 42 into a "dynamic association weight matrix" between modalities, which can be regarded as a topological structure of intermodal influence, used to express the strength of the association between each pair of modalities. It explicitly expresses the degree of influence between modalities; facilitates subsequent integration with graph structures; and supports downstream anomaly detection and causal analysis.

[0161] Step 44 integrates the dynamic association weight matrix into the existing production line operation dependency graph to form an enhanced graph structure, which can be understood as superimposing "modal relationships" on "entity dependencies"; constructing a more expressive system structure graph; while retaining physical / logical dependencies and data associations; which is conducive to uncovering hidden anomaly propagation paths.

[0162] Step 45 performs anomaly detection on nodes and edges in the enhanced graph structure. Graph neural networks, graph autoencoders, and other methods can be used to identify local substructures that deviate from normal patterns. This efficiently detects abnormal points and links in the system, improves the accuracy and interpretability of anomaly detection, and accurately locates potential risk areas.

[0163] Step 46, based on the anomaly identification results, traces the path of anomaly impact through backpropagation technology, while simultaneously extracting subgraphs to pinpoint the source node causing the anomaly; quickly and accurately locates the root cause of the anomaly; reduces false alarm rate and false negative rate; and provides maintenance personnel with actionable decision-making support.

[0164] Step 47 quantifies the strength of the causal relationship between anomalies and root causes using causal inference techniques. It employs the difference between mutual information and conditional mutual information to estimate direct causal impact; strengthens the causal explanation between anomalies and root causes; supports fault traceability and prevention; and provides data support for operational optimization and strategy adjustments.

[0165] In this preferred embodiment of Example 1, step 5 involves constructing a digital twin simulation environment based on causal correlation information to simulate the impact path of heterogeneous data parameters intervening in production disturbances on collaborative anomalies, in order to evaluate the anomaly impact results of parameter adjustments, including:

[0166] Step 51: Map the acquired causal relationship information to the digital twin simulation environment to establish the correspondence between physical entities and digital models;

[0167] Step 52: Define the interaction rules in the digital twin simulation environment based on the production line operation dependency graph to ensure that the digital twin simulation environment accurately reflects the dynamic characteristics data of actual production operation;

[0168] Step 53: Design a parameter intervention experiment plan based on the dynamic characteristic data of production operation, and simulate production disturbance scenarios by adjusting key parameters in the digital twin simulation environment at fixed points;

[0169] Step 54: Based on the production disturbance scenario, track the propagation path of the production disturbance, record the state changes of each node's characteristics, and analyze the impact of the production disturbance scenario on the collaborative anomalies among multi-source heterogeneous data modalities.

[0170] Step 55: Comprehensively evaluate the effects of various parameter adjustment schemes on the production disturbance schemes that simulate the suppression or aggravation of coordinated anomalies, so as to form a quantitative assessment result of the anomaly impact.

[0171] In this embodiment, step 51 maps the causal relationship information (such as the causal relationship between anomaly nodes and their cause nodes) obtained from anomaly detection and root cause localization to the digital twin system, corresponding to the synchronization relationship between the physical entity and the digital model.

[0172] Key technologies: Digital twin mapping model; entity-ID matching and attribute annotation; spatiotemporal synchronization mechanism (real-time / near real-time). These technologies achieve consistency between physical and digital systems; provide a reliable foundational model for virtual simulation; and support subsequent disturbance simulation and impact assessment.

[0173] Step 52 establishes interaction rules between entities in the digital twin system based on the existing production line operation dependency graph, such as logistics paths, equipment interlocks, process sequences, and energy / information flow paths. Key technologies include: graph-based interaction modeling; dynamic rule engine (such as a state machine or event-driven system); and real-time data-driven rule updates. This ensures that the simulation environment possesses realistic interactive behavior; accurately reproduces changes in system operating status; and achieves dynamic feedback and prediction capabilities.

[0174] Step 53 involves adjusting key operating parameters (such as temperature, voltage, speed, and feeding cycle time) in the digital twin system to construct anomaly / disturbance scenarios that may actually occur. Key technologies include: sensitivity analysis for selecting intervention points; Design of Experiments (DoE); and repeatable disturbance simulation mechanisms. This allows for rapid simulation of various possible operational risks; risk prediction without affecting the actual production line; and support for systematic trial and error and optimal solution selection.

[0175] Step 54 involves tracking the propagation path of the disturbance from its source to other nodes in the system during the intervention simulation, and observing the changes in the state of each node (such as equipment efficiency, data anomaly amplitude, modal coupling strength, etc.). Key technologies include: graph neural network path tracing; dynamic system response modeling; and multimodal feature fusion analysis. This reveals the anomaly propagation mechanism within the system; analyzes the anomaly co-evolutionary relationship between modes; and provides theoretical support for risk control point identification.

[0176] Step 55 compares the system's response to anomalies under different parameter intervention schemes to assess which schemes inhibit anomalies and which may exacerbate them, achieving quantitative evaluation. Key technologies include: multi-scenario simulation comparison; construction of anomaly impact indicator system (e.g., collaborative anomaly scoring, response time, impact range); and visualization result analysis. This leads to scientific and effective anomaly response strategies; supports proactive system optimization and preventative maintenance; and enhances the overall resilience and reliability of the system.

[0177] In this preferred embodiment of Example 1, step 6, which involves constructing an iterative scheduling model based on a dimensionless multidimensional dataset, a dependency graph, and causal association information, includes:

[0178] Step 61: Construct a state representation space based on the dimensionless multidimensional dataset, and map the current state of production operation into a high-dimensional feature vector, which includes multidimensional information data such as equipment operating parameters, quality indicators, energy consumption data and safety monitoring data;

[0179] Step 62: Define a constraint matrix using the production line operation dependency graph. The constraint matrix represents the preorder dependencies, resource sharing, and conflict relationships between each production unit to generate a scheduling relationship matrix that meets the production operation constraints.

[0180] Step 63: Construct an iterative scheduling model based on multidimensional information data, scheduling relationship matrix, and causal correlation information. The iterative scheduling model specifically includes:

[0181] Step 64: Generate structured data from the multidimensional information data, scheduling relationship matrix, and causal association information, and encode the structured data into sequence data;

[0182] Step 65: Input the sequence data into the iterative scheduling model; the iterative scheduling model includes an input layer, a first hidden layer, a second hidden layer, a third hidden layer and an output layer, and transmits the intermediate representation data of multiple hidden layers to the output layer. The output layer iteratively outputs the scheduling strategy prediction results corresponding to the state changes between the parameters of the multi-source heterogeneous data.

[0183] Step 66: Input at least two multidimensional information data, scheduling relationship matrix and causal association information data items acquired in real time into the iterative scheduling model, and iteratively output the scheduling strategy prediction results corresponding to the state changes among the multi-source heterogeneous data parameters acquired in real time.

[0184] In this embodiment, step 61 standardizes and dimensionlessly processes various dynamic operational information (equipment parameters, energy consumption, quality, safety, etc.) involved in the production system, mapping it to a unified high-dimensional representation space to form feature vectors with engineering semantics. These vectors contain data such as: equipment operating parameters (e.g., rotational speed, temperature, vibration); product quality indicators (e.g., dimensional error, yield); energy consumption data (e.g., energy consumption per process, current, voltage, etc.); and safety monitoring data (e.g., alarm count, risk score, etc.). By establishing a complete state profile, the scheduling model can easily obtain a global view; dimensionless processing solves the problem of inconsistent unit dimensions, improving the model's generalization ability; and high-dimensional data fusion enables fine-grained characterization of complex systems.

[0185] Step 62 extracts the constraint relationships between various production units through the production line operation dependency graph, constructs a constraint condition matrix, and clarifies: process sequence (precedence dependency); resource sharing (such as the same robot cannot serve multiple units simultaneously); and conflict relationships (such as process concurrency conflicts and physical space conflicts). This ensures that the scheduling model generates feasible solutions without violating actual constraints; supports complex scheduling dependency modeling in flexible manufacturing systems; and avoids deadlocks, resource conflicts, and other problems during scheduling.

[0186] Step 63 combines multidimensional information data (state features), scheduling constraint matrices, and causal information to establish a deep iterative model for scheduling prediction. Historical data is used to drive the training of the model to learn the mapping relationship between task states and optimal scheduling strategies. Causal information is utilized to enhance the model's ability to perceive anomaly propagation paths; improve the adaptability of the scheduling system to complex dynamic scenarios; and support online dynamic optimization of scheduling strategies.

[0187] Step 64 transforms static structural data (such as spectrogram structure, constraint matrix, and eigenvectors) into time-series data that can be processed by neural network models. This facilitates the use of time-series modeling structures (such as RNNs and Transformers) to capture the evolutionary relationship between history, current state, and prediction. The sequence modeling structure is used to extract state evolution trends; enhance the scheduling model's ability to identify time delay effects; and achieve the time-series prediction capability of scheduling strategies (non-one-time decision-making).

[0188] Step 65 uses a neural network structure (typically a multi-layer structure: input layer → multiple hidden layers → output layer) to iteratively learn the complex nonlinear relationship between intermediate states and scheduling policies. Finally, the output layer generates the scheduling policy prediction result. Model structure:

[0189] Input layer: Receive state vector + scheduling matrix + causal vector;

[0190] Hidden layers: Each layer abstracts and extracts feature information, which is then passed down layer by layer;

[0191] Output layer: Outputs the predicted scheduling strategy (such as priority list, process-time allocation table, etc.).

[0192] Deep learning architectures enhance the ability to model nonlinear relationships; multi-hidden-layer designs improve the ability to express feature interactions; and they provide interpretable scheduling suggestions or visual graphs.

[0193] Step 66, during the deployment phase, uses multi-dimensional state information (at least two dimensions, such as energy consumption and quality) and structural constraints as input in real time, iteratively outputting the prediction result of the optimal scheduling strategy matching the current state. This enables real-time dynamic scheduling (Rescheduling); rapid response to sudden events or production anomalies; and supports an online optimization mechanism.

[0194] In this preferred embodiment of Example 1, step 6 combines the anomaly impact assessment results with the real-time production status to iteratively generate a multi-objective optimized collaborative scheduling strategy, realizing dynamic configuration and anomaly prevention among multi-source heterogeneous data, including:

[0195] Step 67: Collect key parameters of real-time production operation status, including equipment operation status, quality inspection data, energy consumption indicators and safety monitoring information, to form a multi-dimensional real-time status vector; and extract the current topology and constraints from the production line operation dependency graph, and combine them with the historical anomaly pattern library to capture anomaly risks in the real-time status.

[0196] Step 68: Based on the captured abnormal risks, perform multi-step prediction of future state evolution and identify potential abnormal development trends; for the detected abnormal risk points, generate a design parameter intervention effect scoring matrix based on the abnormal impact assessment results.

[0197] Step 69: Activate the multi-objective optimization decision-making mechanism based on the intervention effect scoring matrix of the design parameters, set the key indicators of production efficiency, quality stability, energy utilization rate and safety risk as optimization objectives, and dynamically adjust the weight of each objective according to the current production task priority;

[0198] And search for the optimal scheduling strategy under constraints, by:

[0199] ;

[0200] To balance the effects of short-term intervention with long-term stable scheduling strategies;

[0201] Where J(θ) represents the comprehensive optimization objective, Indicates production efficiency targets. Indicates the quality stability target. Indicates the energy utilization rate target, Indicates security risk objectives. , , and These represent the dynamic weighting coefficients for the four objectives: production efficiency, quality stability, energy utilization rate, and safety risk, respectively, and satisfy the following conditions: ;

[0202] Step 610: Based on the prediction results of the iterative scheduling model, execute the collaborative scheduling strategy deployment, decompose the optimized scheduling instructions into specific execution instructions at the device level, process level, and resource level; securely transmit the specific execution instructions to each execution unit through the middleware layer, and establish an instruction execution confirmation mechanism to achieve dynamic configuration and anomaly prevention among multi-source heterogeneous data;

[0203] Step 611: Establish a real-time monitoring and feedback mechanism to continuously track the execution effect of the scheduling strategy and collect data on changes in key performance indicators; when execution deviations or new abnormal events are detected, trigger a rapid response mechanism to select fine-tuning or replanning strategies based on the degree of deviation.

[0204] In this embodiment, step 67 acquires equipment operating status, quality inspection results, energy consumption data, and safety monitoring information in real time. State vector construction: This information is uniformly constructed into a high-dimensional state vector, representing the current overall operating status of the system. Structural modeling: The current production line topology and operational constraints are obtained through a "production line operation dependency graph." Pattern recognition: Combining a historical anomaly pattern library, pattern matching and anomaly detection algorithms (such as clustering, isolated forest, Bayesian methods, etc.) are used to determine whether there are potential anomaly risks. This achieves comprehensive perception of the operating status of complex industrial systems. It can identify potential faults or quality problems in advance, preventing the spread of problems.

[0205] Step 68, based on the identified abnormal risks, uses time series forecasting models (such as LSTM, Transformer, etc.) to predict future state change trends, constructing a "design parameter intervention effect scoring matrix" to simulate the effect of different parameter adjustments on anomaly mitigation; a higher score indicates more effective intervention. Based on the prediction results, it identifies which aspects may worsen and require priority handling. This provides forward-looking early warning and response references for the scheduling system; it also pre-selects the most influential control parameters to improve response efficiency and effectiveness.

[0206] Step 69 uses production efficiency, quality stability, energy utilization rate, and safety risk as objectives. Based on the priority of production tasks and real-time requirements, the above objectives are weighted and adapted to task changes. Evolutionary algorithms, particle swarm optimization, and NSGA-II algorithms are used to search for the optimal scheduling strategy under constraints, taking into account both short-term effects and long-term stability. This achieves a personalized, task-oriented scheduling strategy and improves the overall performance and resource utilization efficiency of the system.

[0207] Step 610 refines the overall scheduling scheme into executable equipment-level, process-level, and resource-level tasks; uses industrial middleware (such as OPC UA, MQTT, etc.) to securely and accurately transmit instructions to different control units; the feedback mechanism ensures that each instruction is correctly received and executed, and can automatically adjust when a fault is detected; ensures that the scheduling strategy is implemented without deviating from the design goal; and improves the system's self-healing ability and robustness to execution anomalies.

[0208] Step 611 monitors key performance indicators such as production rate, yield rate, and energy consumption; once an execution deviation or new anomaly is detected, it triggers fine-tuning (such as PID self-tuning) or replans the entire strategy process; it feeds new data and feedback results back into the system to further enhance the model's accuracy and adaptability; it constructs a complete "perception-analysis-execution-feedback" closed loop; and it improves the system's adaptability, stability, and self-optimization capabilities.

[0209] In this preferred embodiment of Example 1, the cooperative scheduling strategy includes:

[0210] The scheduling problem is decomposed into three levels: strategy, tactics, and operation. These levels are used to handle abnormal situations between heterogeneous data in long-term planning, medium-term scheduling, and real-time response, forming a complete hierarchical collaborative decision-making framework.

[0211] By achieving bidirectional integration of constraint scenarios and guidance scenarios, and using a differentiable programming layer to unify hard constraints and soft constraints into the same framework, the constraint weights are dynamically adjusted according to real-time production conditions to maximize the comprehensive response to production operation requirements.

[0212] Based on the company's strategic priorities, changes in market demand, and actual production conditions, dynamically adjust the priorities of multiple objectives such as production efficiency, quality stability, energy utilization, and safety risks to achieve a dynamic balance among these objectives.

[0213] The design combines anomaly prediction with proactive scheduling. By using a time-series prediction model to predict production status in multiple steps and combining it with a digital twin simulation environment for cyclic simulation and evaluation, the mechanism can proactively adjust the state between heterogeneous data parameters before anomalies actually occur, transforming passive response into proactive prevention.

[0214] Establish a knowledge-driven and data-driven approach, and use a distributed collaborative execution framework to make the collaborative scheduling strategy process transparent and efficient.

[0215] In this embodiment, the scheduling problem is decomposed into three levels: strategy, tactics, and operations.

[0216] Strategic level (long-term planning): Analyze macro trends (such as market demand and corporate development strategy), formulate long-term resource allocation and capacity layout plans; provide directional guidance for enterprises, and ensure that the top-level design of the scheduling system is consistent with the development of enterprises.

[0217] Tactical level (mid-term scheduling): Based on the goals provided by the strategic level and combined with the production cycle (such as weekly / monthly plans), formulate optimized production and resource scheduling plans; improve resource utilization, balance production load, and enhance mid-term operational efficiency.

[0218] Operational layer (real-time response): Monitors and adjusts real-time data during the production process, especially handling abnormal information from heterogeneous systems; enhances the system's flexibility and robustness, enables rapid response to emergencies, and ensures production continuity.

[0219] Two-way fusion of constrained and guiding scenarios + differentiable programming;

[0220] Two-way integration:

[0221] Constraint scenarios: hard constraints (such as equipment capabilities, resource limits);

[0222] Guiding scenarios: Soft guidance (such as prioritizing the production of high-profit products).

[0223] By fusing the two through a unified differentiable optimization framework, the scheduling algorithm adheres to rigid rules while preserving optimization space. This achieves "principled flexibility," seeking better scheduling solutions within constraints.

[0224] Differentiable Programming Layer:

[0225] By introducing differentiable optimization modules (such as deep learning + optimization), the scheduling objective is embedded in the neural network training process, and the weights of constraint terms are dynamically adjusted. The model can adapt to changes in production, and the optimized solution is more closely aligned with real-time conditions, thus improving the intelligence level of scheduling.

[0226] Multi-objective dynamic balancing mechanism: Based on the company's strategy, market demand, and actual production line status at different points in time, the priority of the following objectives is dynamically adjusted: production efficiency; product quality stability; energy utilization rate; and safety risk control. This avoids the single-objective deviation of "pursuing efficiency while neglecting safety," achieves refined and comprehensive scheduling management, and maintains a dynamic balance in the company's operations near the optimal point.

[0227] Anomaly prediction + proactive scheduling mechanism:

[0228] Utilize time-series prediction models (such as LSTM and Transformer) to predict production status at multiple future time steps; simulate possible scenarios in a digital twin environment; and proactively adjust production strategies or resource allocation if potential abnormal trends are detected. This shifts the focus from "responding to failures after they occur" to "preventing potential problems before they occur," significantly improving system stability and reducing downtime losses.

[0229] Knowledge-driven and data-driven approaches are integrated for distributed collaborative execution.

[0230] Knowledge-driven: Introducing expert rule bases and industrial knowledge graphs for initial scheduling design;

[0231] Data-driven: Continuously optimize the scheduling model using historical and real-time data;

[0232] Distributed collaborative execution framework: This framework enables multi-system collaboration (such as ERP, MES, SCADA) to execute scheduling strategies in a distributed manner. It leverages human-machine collaboration to enhance the adaptability of the scheduling system; achieves transparency, traceability, and intervention in the scheduling execution process; improves overall enterprise collaboration efficiency; and reduces information silos.

[0233] Multi-scenario Application Examples: The method and system of this invention can be applied to multiple digital transformation application scenarios. The following details application examples for each scenario, based on specific industry characteristics:

[0234] Case 1: Digitalization of the product lifecycle.

[0235] 1. Product design.

[0236] Implementation scenario: A machinery manufacturing company faces problems such as long product design cycles, frequent design changes, and low first-time success rate, and needs to improve design efficiency and quality.

[0237] Implementation Plan: This system will be used to digitize product design, and will optimize designs by analyzing historical design data, market feedback, and production data through a multimodal deep learning model. The system will collect multi-source information, including CAD design data, material parameters, and historical performance test data, to construct a product knowledge graph, supporting designers in parametric design and simulation verification.

[0238] Implementation results: After implementation, the product design cycle was shortened by 35%, the number of design changes was reduced by 40%, and the first-time success rate was increased by 28%.

[0239] 2. Process design.

[0240] Implementation scenario: A certain auto parts company has a complex process design process, process parameter optimization relies on experience, and the first pass rate of products is low, which affects production efficiency.

[0241] Implementation Plan: Utilize this system to optimize process design workflows. The system uses deep learning algorithms to analyze the correlation between historical process parameters and product quality, automatically generating optimal process routes and parameter combinations. The system integrates CAD / CAM data, material properties, and equipment capability parameters to build a process knowledge base, achieving intelligent process design.

[0242] Implementation results: After application, process design efficiency increased by 42%, process parameter optimization accuracy reached 93%, and product first-pass yield increased by 25%.

[0243] 3. Marketing Management.

[0244] Implementation scenario: A consumer electronics company's marketing decisions lack data support, customer needs are not accurately grasped, and marketing costs are high with poor results.

[0245] Implementation Plan: This system will be used to achieve digital transformation in marketing. It integrates sales data, customer feedback, social media data, and market research information, analyzing consumer behavior patterns and market trends through a multimodal Transformer network. The system constructs a customer profile knowledge graph to support precision marketing and personalized recommendations.

[0246] Implementation results: After implementation, marketing conversion rate increased by 32%, customer acquisition cost decreased by 28%, and customer satisfaction increased by 18%.

[0247] 4. After-sales service.

[0248] Implementation scenario: A construction machinery company has slow after-sales service response, and most repairs are reactive. Equipment failures cause long downtime for customers, affecting customer satisfaction.

[0249] Implementation Plan: This system optimizes after-sales service processes. The system collects equipment operation data through IoT devices, combines this data with maintenance records and user feedback, and uses a BiLSTM-Attention network to predict equipment failures and maintenance needs. The system constructs an equipment health status assessment model to achieve predictive maintenance and remote diagnostics.

[0250] Implementation results: After implementation, the accuracy rate of equipment failure prediction reached 87%, unplanned downtime was reduced by 45%, and maintenance response time was shortened by 60%.

[0251] Case 2: Digitalization of Production Execution.

[0252] 1. Planning and Scheduling

[0253] Implementation scenario: An electronics assembly company has a low production plan execution rate, frequent delivery delays, long production cycles, and low resource utilization.

[0254] Implementation Plan: This system will be used to optimize production planning and scheduling. The system integrates order data, inventory information, equipment status, and personnel scheduling information, and generates multi-objective optimized production plans through reinforcement learning algorithms. The system constructs a production constraint relationship graph to achieve dynamic scheduling and real-time optimization.

[0255] Implementation results: After implementation, the production plan execution rate increased by 38%, the on-time delivery rate increased by 45%, and the production cycle was shortened by 30%.

[0256] 2. Production control.

[0257] Implementation scenario: A precision manufacturing company has insufficient monitoring of its production process, slow response to anomalies, low production efficiency, and poor product consistency.

[0258] Implementation Plan: This system will be used to achieve intelligent control of the production process. The system will collect real-time data from the production line through industrial cameras and sensor networks, and use an improved YOLOv5 network for product defect detection and production anomaly identification. The system will construct a digital twin simulation environment for the production process, enabling visualization and real-time monitoring.

[0259] Implementation results: After application, the response time to production anomalies was reduced by 75%, production efficiency was increased by 33%, and product consistency was improved by 29%.

[0260] 3. Quality Management.

[0261] Implementation scenario: A medical device company has a cumbersome quality management process, difficulty in quality traceability, and the root cause analysis of defects relies on manual experience, resulting in high quality costs.

[0262] Implementation Plan: This system will be used to improve quality management. The system integrates online testing data, laboratory test results, and supplier quality information, and analyzes quality influencing factors and root causes of defects through a multimodal deep learning model. The system constructs a quality traceability knowledge graph to achieve full-process quality control.

[0263] Implementation results: After application, the product defect rate decreased by 42%, the quality traceability time was shortened by 85%, and the quality cost was reduced by 31%.

[0264] 4. Equipment management.

[0265] Implementation scenario: A steel company experiences frequent equipment failures, maintenance relies mainly on reactive repairs, resulting in low equipment availability, high maintenance costs, and short equipment lifespan.

[0266] Implementation Plan: This system optimizes equipment management processes. The system collects equipment operating data through vibration and temperature sensors, and uses a time-series graph convolutional network to analyze equipment health status and predict failure risks. The system constructs an equipment knowledge graph to achieve full lifecycle management of equipment.

[0267] Implementation results: After implementation, equipment availability increased by 25%, maintenance costs decreased by 35%, and equipment lifespan was extended by 20%.

[0268] 5. Safe production.

[0269] Implementation scenario: A chemical enterprise has high safety risks, relies on manual inspections to identify safety hazards, has insufficient safety accident prevention, and has a slow emergency response.

[0270] Implementation Plan: This system will be used to improve workplace safety. The system collects safety-related data through gas sensors and video surveillance, and uses a multimodal Transformer network to identify safety risks and provide early warnings of potential accidents. The system constructs a safety knowledge graph to enable intelligent risk assessment and emergency plan recommendations.

[0271] Implementation results: After implementation, the incidence of safety accidents decreased by 65%, the rate of safety hazard identification increased by 78%, and the emergency response time was shortened by 70%.

[0272] 6. Energy management.

[0273] Implementation scenario: A textile company has high energy consumption, low energy utilization efficiency, high energy costs, and lacks effective energy optimization strategies.

[0274] Implementation Plan: This system optimizes energy management. It collects energy consumption data through smart meters and flow meters, and uses a BiLSTM-Attention network to analyze energy consumption patterns and predict energy demand. The system constructs an energy flow knowledge graph to achieve optimized energy allocation and generate energy-saving strategies.

[0275] Implementation results: After implementation, energy consumption decreased by 28%, energy utilization efficiency increased by 32%, and energy costs decreased by 25%.

[0276] Case Study 3: Supply Chain Digitalization.

[0277] 1. Procurement Management.

[0278] Implementation scenario: A home appliance company faces high procurement costs, unstable supplier delivery, long procurement cycles, and a lack of a scientific supplier evaluation system.

[0279] Implementation Plan: This system will be used to optimize the procurement management process. The system integrates demand forecasting, inventory data, supplier evaluation, and market price information, and optimizes procurement decisions and supplier selection through deep learning algorithms. The system constructs a supplier knowledge graph to achieve intelligent price comparison and risk assessment.

[0280] Implementation results: After implementation, procurement costs decreased by 22%, on-time delivery rate of suppliers increased by 40%, and procurement cycle was shortened by 35%.

[0281] 2. Warehousing and logistics.

[0282] Implementation scenario: A fast-moving consumer goods (FMCG) company suffers from low warehousing efficiency, high distribution costs, slow inventory turnover, and low utilization of logistics resources.

[0283] Implementation Plan: This system will be used to optimize warehouse logistics management. The system collects warehouse logistics information through RFID and AGV status data, and uses reinforcement learning algorithms to optimize warehouse layout and picking routes. The system constructs a logistics network knowledge graph to achieve intelligent inventory allocation and delivery route optimization.

[0284] Implementation results: After implementation, warehousing efficiency improved by 45%, distribution costs decreased by 30%, and inventory turnover increased by 38%.

[0285] Case 4: Digitalization of Management Decisions.

[0286] 1. Financial Management.

[0287] Implementation scenario: A manufacturing company suffers from inefficient financial analysis, inaccurate cost control, low capital utilization efficiency, and untimely financial risk warnings.

[0288] Implementation Plan: This system will be used to digitize financial management. The system integrates financial data on production costs, sales revenue, and cash flow, and uses deep learning algorithms to analyze cost structure and predict cash flow. The system will also construct a financial knowledge graph to enable cost tracking and early warning of financial risks.

[0289] Implementation results: After implementation, financial analysis efficiency improved by 65%, cost control accuracy improved by 28%, and capital utilization efficiency improved by 32%.

[0290] 2. Human Resources.

[0291] Implementation scenario: A large manufacturing company has an unreasonable staffing structure, low employee satisfaction, high human resource costs, and a poor match between skills and positions.

[0292] Implementation Plan: This system will be used to optimize human resource management. The system integrates employee performance, training records, and work time data, and analyzes employee capabilities and predicts talent demand through a multimodal deep learning model. The system constructs a talent knowledge graph to achieve intelligent scheduling and skills matching.

[0293] Implementation results: After implementation, staffing efficiency increased by 40%, employee satisfaction increased by 25%, and human resource costs decreased by 18%.

[0294] 3. Collaborative work.

[0295] Implementation scenario: A multi-factory manufacturing enterprise faces challenges such as low efficiency in cross-departmental collaboration, difficulties in project progress management, high communication costs, and poor information sharing.

[0296] Implementation Plan: This system will improve collaborative work efficiency. It integrates project progress, document management, and communication records, using deep learning algorithms to analyze collaboration patterns and optimize workflows. The system will also construct a business process knowledge graph to enable cross-departmental collaboration and information sharing.

[0297] Implementation results: After implementation, project collaboration efficiency increased by 50%, decision response time was shortened by 60%, and communication costs were reduced by 35%.

[0298] 4. Decision support.

[0299] Implementation scenario: A certain integrated manufacturing group relies on experience for decision-making, lacks data support, has slow market response, untimely strategic adjustments, and low decision-making accuracy.

[0300] Implementation Plan: This system will be used to improve management decision-making. The system integrates multi-dimensional data from production, sales, finance, and marketing, and uses a multi-modal Transformer network to analyze business relationships and predict development trends. The system constructs an enterprise operational knowledge graph, enabling multi-dimensional data visualization and intelligent decision recommendations.

[0301] Implementation results: After implementation, decision-making accuracy improved by 45%, market response speed improved by 55%, and strategic adjustment efficiency improved by 40%.

[0302] like Figure 2 As shown, a deep learning-based MES digital collaborative management system includes:

[0303] Data acquisition module: It is used to collect multi-source heterogeneous data, and construct a dimensionless multidimensional dataset through preprocessing and normalization.

[0304] Dynamic Association Module: It is used to map the dimensionless multidimensional dataset into graph structure nodes to obtain node features, and define the connection exchange, information flow and dependency relationship between each parameter in the dimensionless multidimensional dataset as edge relationship; through node feature embedding and edge relationship learning, it analyzes the dynamic association relationship between each data item in the dimensionless multidimensional dataset and generates the production line operation dependency relationship map.

[0305] Detection and localization module: It is used to build a collaborative analysis mechanism for multi-source heterogeneous data to obtain the dynamic correlation weights between multi-source heterogeneous data modes, and combine the dependency relationship map to perform anomaly detection and root cause localization in order to obtain causal relationship information between abnormal events and data modes.

[0306] Anomaly Impact Module: It is used to construct a digital twin simulation environment based on causal relationship information, simulate the impact path of heterogeneous data parameters intervening in production disturbances on coordinated anomalies, and evaluate the anomaly impact results of parameter adjustments;

[0307] Model training module: It is used to build an iterative scheduling model based on a dimensionless multidimensional dataset, dependency graph, and causal relationship information;

[0308] The collaborative management module combines the results of anomaly impact assessment with real-time production status to iteratively generate multi-objective optimized collaborative scheduling strategies, enabling dynamic configuration and anomaly prevention among multi-source heterogeneous data.

[0309] The embodiments of the present invention have been described above, but the embodiments are not limited to the specific implementation methods described above. The specific implementation methods described above are merely illustrative and not restrictive. Those skilled in the art can make more similar embodiments based on the guidance of the present embodiments, and all of them are within the protection scope of the present embodiments.

Claims

1. A deep learning-based digital collaborative management method for MES, characterized in that, Includes the following steps: Collect heterogeneous data from multiple sources, and construct a dimensionless multidimensional dataset through preprocessing and normalization. The dimensionless multidimensional dataset is mapped to graph structure nodes to obtain node features, and the connection exchange, information flow and dependency relationship between parameters in the dimensionless multidimensional dataset are defined as edge relationships. By using node feature embedding and edge relationship learning, the dynamic relationship between data items in the dimensionless multidimensional dataset is analyzed to generate a production line operation dependency graph. A collaborative analysis mechanism for multi-source heterogeneous data is constructed to obtain the dynamic correlation weights among the collaborative modes of multi-source heterogeneous data. Anomaly detection and root cause localization are performed by combining the dependency relationship graph to obtain causal relationship information between abnormal events and data modes. A digital twin simulation environment is constructed based on causal relationship information to simulate the impact path of heterogeneous data parameters on production disturbances on collaborative anomalies, so as to evaluate the abnormal impact results of parameter adjustments. In this process, the operating status of dynamically changing multi-source heterogeneous data parameters is processed to fuse temporal features with graph network topology features to obtain a fusion result. Based on the fusion results, the static attributes of the parameters of each data item in the dimensionless multidimensional dataset are concatenated with the dynamic time-series data to reduce the dimensionality, so as to generate the fusion feature vector of a single node, which serves as the basic input for constructing the dependency graph. Obtain the cosine similarity between the fused feature vectors of each node to generate implicit edges, and define explicit edges using interaction strength connections. After fusion, a weighted adjacency matrix is ​​formed, which, together with the node features, outputs dynamic edge weights. By learning the dynamic edge weights between node features, the dynamic edge weights are updated by aggregating the multi-attention results, and the updated node features are generated by fusing the features of neighboring nodes. The dynamic edge weights and updated node features together form a production line operation dependency graph, revealing the explicit associations and implicit dynamic dependencies between multi-source heterogeneous data items during production operation, so as to support production line operation decisions. This includes constructing a multi-source heterogeneous data collaborative analysis mechanism to obtain dynamic correlation weights among multi-source heterogeneous data modalities, and combining dependency graphs for anomaly detection and root cause localization to obtain causal correlation information between abnormal events and data modalities, including: Feature extraction and representation learning are performed on the states of each data item in a dimensionless multidimensional dataset to obtain the correlation matrix between multi-source heterogeneous data modes. The contribution of multi-source heterogeneous data modes is dynamically adjusted by using dynamic edge weights to obtain the dynamic correlation weights for collaboration among multi-source heterogeneous data modes; Based on the dynamic association weights, the correlation influence between multi-source heterogeneous data modes is captured, and a dynamic association weight matrix is ​​generated. By combining the dynamic association weight matrix with the production line operation dependency graph, an enhanced graph structure topology is constructed. Anomaly identification results are obtained by identifying node features and edge relationships that deviate from the normal pattern on the enhanced graph structure topology; Based on the anomaly identification results, the root cause node of the anomaly is located through backpropagation and subgraph extraction to obtain root cause localization data; Causal inference is used to quantify the strength of the causal association between anomaly identification results and root cause localization data, generating causal association information; the causal association strength represents the degree of direct influence of anomaly identification results on root cause localization data, which is determined by obtaining the difference between mutual information and conditional mutual information. Among these, a digital twin simulation environment is constructed based on causal correlation information to simulate the impact path of heterogeneous data parameters intervening in production disturbances on collaborative anomalies, in order to evaluate the anomaly impact results of parameter adjustments, including: The acquired causal relationship information is mapped to the digital twin simulation environment to establish the correspondence between physical entities and digital models; Based on the production line operation dependency graph, define the interaction rules in the digital twin simulation environment to ensure that the digital twin simulation environment accurately reflects the dynamic characteristics data of actual production operation; Based on the dynamic characteristics data of production operation, a parameter intervention experiment plan is designed. By adjusting key parameters in the digital twin simulation environment at fixed points, a production disturbance scenario is simulated. Based on production disturbance scenarios, the propagation path of production disturbances is tracked, and the state changes of each node are recorded to analyze the impact of production disturbance scenarios on the collaborative anomalies among multi-source heterogeneous data modalities. A comprehensive evaluation of the effects of various parameter adjustment schemes on the suppression or exacerbation of coordinated anomalies in simulated production disturbance schemes is conducted to form a quantitative assessment result of the anomaly impact. An iterative scheduling model is constructed based on dimensionless multidimensional datasets, dependency graphs, and causal relationship information. By combining the results of anomaly impact assessment with real-time production status, a multi-objective optimized collaborative scheduling strategy is generated iteratively to achieve dynamic configuration and anomaly self-healing among multi-source heterogeneous data.

2. The deep learning-based MES digital collaborative management method according to claim 1, characterized in that, Collect heterogeneous data from multiple sources, and construct a dimensionless multidimensional dataset through preprocessing and normalization, including: Real-time acquisition of multi-source heterogeneous data during production and manufacturing operations and data preprocessing to obtain a preprocessed dataset; The preprocessed dataset is standardized and cleaned to obtain a clean dataset. The clear dataset is time-aligned and normalized after being collected at a preset period to obtain a dimensionless multidimensional dataset.

3. The deep learning-based MES digital collaborative management method according to claim 1, characterized in that, By using node feature embedding and edge relationship learning, the dynamic relationships between data items in a dimensionless multidimensional dataset are analyzed to generate a production line operation dependency graph, including: The dimensionless multidimensional dataset is mapped to graph structure nodes, and the parameters of multi-source heterogeneous data are used as node features. Define the connection, exchange, information flow, and dependency relationships between parameters of multi-source heterogeneous data as edge relationships; Node feature embedding is performed using a graph attention network structure to capture the local correlation of edge relationships; The interaction strength between multi-source heterogeneous data modalities is quantified through a side relation feature learning mechanism. Based on local correlation and interaction strength, a multi-layer graph network is used for iterative propagation to analyze the dynamic relationships between data items in a dimensionless multidimensional dataset, generating a production line operation dependency graph that includes the relationship status between heterogeneous data, production operation parameters, and parameter quality indicators.

4. The deep learning-based MES digital collaborative management method according to claim 1, characterized in that, An iterative scheduling model is constructed based on a dimensionless multidimensional dataset, dependency graph, and causal relationship information, including: A state representation space is constructed based on a dimensionless multidimensional dataset, which maps the current state of production operation to a high-dimensional feature vector, containing multidimensional information data such as equipment operating parameters, quality indicators, energy consumption data and safety monitoring data. The constraint matrix is ​​defined using the production line operation dependency graph. The constraint matrix represents the prior dependencies, resource sharing and conflict relationships between each production unit, so as to generate a scheduling relationship matrix that meets the production operation constraints. An iterative scheduling model is constructed based on multidimensional information data, a scheduling relationship matrix, and causal correlation information. The iterative scheduling model specifically includes: Multidimensional information data, scheduling relationship matrix and causal association information are used to generate structured data, and the structured data is encoded into sequence data; The sequence data is input into the iterative scheduling model; the iterative scheduling model includes an input layer, a first hidden layer, a second hidden layer, a third hidden layer and an output layer, and transmits the intermediate representation data of multiple hidden layers to the output layer. The output layer iteratively outputs the scheduling strategy prediction results corresponding to the state changes between the parameters of the multi-source heterogeneous data. The at least two multidimensional information data, scheduling relationship matrix and causal association information data items acquired in real time are input into the iterative scheduling model, and the iterative output represents the scheduling strategy prediction result corresponding to the state changes among the multi-source heterogeneous data parameters acquired in real time.

5. The deep learning-based MES digital collaborative management method according to claim 4, characterized in that, By combining the anomaly impact assessment results with real-time production status, a multi-objective optimized collaborative scheduling strategy is iteratively generated to achieve dynamic configuration and anomaly self-healing among multi-source heterogeneous data, including: Collect key parameters of real-time production operation status, including equipment operating status, quality inspection data, energy consumption indicators and safety monitoring information, to form a multi-dimensional real-time status vector; Extract the current topology and constraints from the production line operation dependency graph, and combine them with the historical anomaly pattern library to capture anomaly risks in real time. Based on the captured abnormal risks, multi-step predictions of future state evolution are made to identify potential abnormal development trends. For the detected abnormal risk points, an intervention effect scoring matrix for design parameters is generated based on the abnormal impact assessment results; The multi-objective optimization decision-making mechanism is initiated based on the intervention effect scoring matrix of the design parameters, setting key indicators such as production efficiency, quality stability, energy utilization rate and safety risk as optimization objectives, and dynamically adjusting the weight of each objective according to the current production task priority. And under constraints, search for the optimal scheduling strategy to balance the short-term intervention effect with the long-term stable scheduling strategy; Based on the prediction results of the iterative scheduling model, a collaborative scheduling strategy is deployed, and the optimized scheduling instructions are decomposed into specific execution instructions at the device level, process level, and resource level. The middleware layer securely transmits specific execution instructions to each execution unit, while establishing an instruction execution confirmation mechanism to achieve dynamic configuration and anomaly self-healing among multi-source heterogeneous data. Establish a real-time monitoring and feedback mechanism to continuously track the execution effect of scheduling strategies and collect data on changes in key performance indicators; when execution deviations or new abnormal events are detected, trigger a rapid response mechanism to select fine-tuning or replanning strategies based on the degree of deviation.

6. The deep learning-based MES digital collaborative management method according to claim 5, characterized in that, The collaborative scheduling strategy also includes: The scheduling problem is decomposed into three levels: strategy, tactics, and operation. These levels are used to handle abnormal situations between heterogeneous data in long-term planning, medium-term scheduling, and real-time response, forming a complete hierarchical collaborative decision-making framework. By achieving bidirectional integration of constraint scenarios and guidance scenarios, and using a differentiable programming layer to unify hard constraints and soft constraints into the same framework, the constraint weights are dynamically adjusted according to real-time production conditions to maximize the comprehensive response to production operation requirements. Based on the company's strategic priorities, changes in market demand, and actual production conditions, dynamically adjust the priorities of multiple objectives such as production efficiency, quality stability, energy utilization, and safety risks to achieve a dynamic balance among these objectives. The design combines anomaly prediction with proactive scheduling. By using a time-series prediction model to predict production status in multiple steps and combining it with a digital twin simulation environment for cyclic simulation and evaluation, the mechanism can proactively adjust the state between heterogeneous data parameters before anomalies actually occur, transforming passive response into proactive prevention. Establish a knowledge-driven and data-driven approach, and use a distributed collaborative execution framework to make the collaborative scheduling strategy process transparent and efficient.

7. A deep learning-based MES digital collaborative management system, used to execute a deep learning-based MES digital collaborative management method as described in any one of claims 1-6, characterized in that, include: Data acquisition module: It is used to collect multi-source heterogeneous data, and construct a dimensionless multidimensional dataset through preprocessing and normalization. Dynamic Association Module: It is used to map the dimensionless multidimensional dataset into graph structure nodes to obtain node features, and define the connection exchange, information flow and dependency relationship between each parameter in the dimensionless multidimensional dataset as edge relationship; through node feature embedding and edge relationship learning, it analyzes the dynamic association relationship between each data item in the dimensionless multidimensional dataset and generates the production line operation dependency relationship map. Detection and localization module: It is used to build a collaborative analysis mechanism for multi-source heterogeneous data to obtain the dynamic correlation weights between multi-source heterogeneous data modes, and combine the dependency relationship map to perform anomaly detection and root cause localization in order to obtain causal relationship information between abnormal events and data modes. Anomaly Impact Module: It is used to construct a digital twin simulation environment based on causal relationship information, simulate the impact path of heterogeneous data parameters on production disturbances on collaborative anomalies, and evaluate the anomaly impact results of parameter adjustment; in particular, it processes the operating status between dynamically changing multi-source heterogeneous data parameters to integrate time-series features with graph network topology features to obtain a fusion result; Based on the fusion results, the static attributes of the parameters of each data item in the dimensionless multidimensional dataset are concatenated with the dynamic time-series data to reduce the dimensionality, so as to generate the fusion feature vector of a single node, which serves as the basic input for constructing the dependency graph. Obtain the cosine similarity between the fused feature vectors of each node to generate implicit edges, and define explicit edges using interaction strength connections. After fusion, a weighted adjacency matrix is ​​formed, which, together with the node features, outputs dynamic edge weights. By learning the dynamic edge weights between node features, the dynamic edge weights are updated by aggregating the multi-attention results, and the updated node features are generated by fusing the features of neighboring nodes. The dynamic edge weights and updated node features together form a production line operation dependency graph, revealing the explicit associations and implicit dynamic dependencies between multi-source heterogeneous data items during production operation, so as to support production line operation decisions. This includes constructing a multi-source heterogeneous data collaborative analysis mechanism to obtain dynamic correlation weights among multi-source heterogeneous data modalities, and combining dependency graphs for anomaly detection and root cause localization to obtain causal correlation information between abnormal events and data modalities, including: Feature extraction and representation learning are performed on the states of each data item in a dimensionless multidimensional dataset to obtain the correlation matrix between multi-source heterogeneous data modes. The contribution of multi-source heterogeneous data modes is dynamically adjusted by using dynamic edge weights to obtain the dynamic correlation weights for collaboration among multi-source heterogeneous data modes; Based on the dynamic association weights, the correlation influence between multi-source heterogeneous data modes is captured, and a dynamic association weight matrix is ​​generated. By combining the dynamic association weight matrix with the production line operation dependency graph, an enhanced graph structure topology is constructed. Anomaly identification results are obtained by identifying node features and edge relationships that deviate from the normal pattern on the enhanced graph structure topology; Based on the anomaly identification results, the root cause node of the anomaly is located through backpropagation and subgraph extraction to obtain root cause localization data; Causal inference is used to quantify the strength of the causal association between anomaly identification results and root cause localization data, generating causal association information; the causal association strength represents the degree of direct influence of anomaly identification results on root cause localization data, which is determined by obtaining the difference between mutual information and conditional mutual information. Among these, a digital twin simulation environment is constructed based on causal correlation information to simulate the impact path of heterogeneous data parameters intervening in production disturbances on collaborative anomalies, in order to evaluate the anomaly impact results of parameter adjustments, including: The acquired causal relationship information is mapped to the digital twin simulation environment to establish the correspondence between physical entities and digital models; Based on the production line operation dependency graph, define the interaction rules in the digital twin simulation environment to ensure that the digital twin simulation environment accurately reflects the dynamic characteristics data of actual production operation; Based on the dynamic characteristics data of production operation, a parameter intervention experiment plan is designed. By adjusting key parameters in the digital twin simulation environment at fixed points, a production disturbance scenario is simulated. Based on production disturbance scenarios, the propagation path of production disturbances is tracked, and the state changes of each node are recorded to analyze the impact of production disturbance scenarios on the collaborative anomalies among multi-source heterogeneous data modalities. A comprehensive evaluation of the effects of various parameter adjustment schemes on the suppression or exacerbation of coordinated anomalies in simulated production disturbance schemes is conducted to form a quantitative assessment result of the anomaly impact. Model training module: It is used to build an iterative scheduling model based on a dimensionless multidimensional dataset, dependency graph, and causal relationship information; The collaborative management module combines the results of anomaly impact assessment with real-time production status to iteratively generate multi-objective optimized collaborative scheduling strategies, enabling dynamic configuration and anomaly prevention among multi-source heterogeneous data.

Images