An intelligent manufacturing industry data management method and system based on big data processing

Abstract

This invention relates to the field of manufacturing data analysis and processing technology, and discloses an intelligent manufacturing data management method and system based on big data processing. The method includes multi-source data acquisition and semantic annotation, dynamic data governance based on value density assessment, construction and evolution of a manufacturing knowledge graph, hybrid reasoning to generate optimized decisions, execution feedback collection, and closed-loop iterative updates. Through a data value scoring model and a double-buffered knowledge graph architecture, it achieves the identification of high-value data and the autonomous evolution of knowledge. It combines rule-based reasoning and graph neural networks for decision fusion, and uses feedback samples to synchronously optimize data governance strategies and knowledge models, forming a closed-loop optimization management mechanism from data to knowledge graph construction, then to optimization strategies, and finally to optimized new data. This application can improve the utilization rate of manufacturing data and the intelligence of decision-making.

Description

Technical Field

[0001] This invention relates to the field of manufacturing data analysis and processing technology, and more specifically, to an intelligent manufacturing data management method and system based on big data processing. Background Technology

[0002] With the deep integration of industrial big data and artificial intelligence technologies, the intelligent manufacturing industry faces higher demands for intelligent, adaptive, and optimized management capabilities. Currently, manufacturing data management is gradually evolving from static data aggregation to dynamic intelligent decision-making. While traditional data management methods have made some progress in multi-source heterogeneous data acquisition, basic feature extraction, and model reasoning, they still have significant shortcomings in the closed-loop evolution mechanism from data acquisition to knowledge graph construction, optimization strategies, and finally, the resulting optimized new data. Therefore, they cannot effectively support the collaborative evolution and intelligent upgrading of all elements in the manufacturing industry.

[0003] Specifically, existing solutions largely rely on predefined rule bases or fixed structures, lacking effective mechanisms for identifying and activating high-value but unutilized potential data. Furthermore, after decision-making is implemented, actual operational feedback is rarely used for reverse correction or dynamic adjustment of data governance. This results in a lag in response speed and decision-making accuracy when dealing with complex changes in manufacturing site conditions, sudden anomalies, or emerging relationships, limiting the depth and breadth of data-driven intelligent optimization.

[0004] Therefore, how to build an intelligent data management method that can transform data into knowledge graph construction, then into optimization strategies, and finally into optimized new data, so as to continuously mine new knowledge from operational data and dynamically optimize data collection, intelligent decision-making, and manufacturing execution strategies accordingly, has become a technical problem that urgently needs to be solved in the current intelligent manufacturing industry. Summary of the Invention

[0005] This invention provides a data management method and system for intelligent manufacturing based on big data processing. It solves the problem in existing technologies that it is difficult to achieve intelligent data management capabilities from data to knowledge graph construction, to optimization strategies, and finally to optimized new data. It realizes the technical problem of continuously mining new knowledge from operational data and dynamically optimizing data collection, intelligent decision-making, and manufacturing execution strategies accordingly.

[0006] This invention provides a data management method and system for intelligent manufacturing based on big data processing, comprising: Firstly, a data management method for intelligent manufacturing based on big data processing includes the following steps: Collect multi-source heterogeneous data streams from the manufacturing site to generate an initial dataset; Perform dynamic data governance operations based on value density assessment on each data item in the initial dataset to obtain a high-value dataset; The high-value dataset is transferred to the knowledge graph construction layer to generate an initial manufacturing knowledge graph. The nodes and edges in the initial manufacturing knowledge graph are dynamically added, deleted, weighted, and their topology is adjusted to form the evolved manufacturing knowledge graph. Based on the evolved manufacturing knowledge graph, the hybrid inference engine is invoked to generate an optimized decision instruction set for the current manufacturing conditions; The optimized decision instruction set is executed, and the actual running results after execution are collected synchronously as feedback data. The feedback data is associated and bound with the corresponding original decision context to form a feedback sample set, which is then used to trigger the next round of knowledge evolution and iterative updates of data governance strategies.

[0007] Furthermore, dynamic data governance operations based on value density assessment are performed on each data item in the initial dataset to obtain a high-value dataset, including: Through data value assessment function Each data item in the initial dataset is scored in real time; where V(d) represents the value score of data item d, R(d) represents the reference frequency of the data item in the historical decision log, F(d) represents the feature importance score of the feature associated with the data item in the currently deployed prediction model, U(d) represents the usage popularity index of the data item in the current sliding time window, and α, β, and γ are preset weighting coefficients that satisfy α+β+γ=1. The weighting coefficients are dynamically adjusted according to the actual contribution of the feedback sample set. When V(d) is higher than the first threshold T1, the original precision is preserved and all computing resources are allocated to enter the full-precision processing pipeline. When V(d) is between the first threshold T1 and the second threshold T2, perform compressed sampling and allocate moderate computing resources to enter the downsampling pipeline; When V(d) is below the second threshold T2, only metadata is extracted and the original payload is discarded, and the metadata extraction pipeline is entered. Three processing pipelines are run in parallel, and the data items selected after value density assessment are aggregated to form a high-value dataset; The high-value dataset contains hierarchical data that has undergone full-precision processing, downsampling, and metadata extraction. The high-value dataset is sent to the knowledge graph construction layer via a data transmission interface as high-quality input data for entity recognition, relation extraction, and graph generation.

[0008] Furthermore, the nodes and edges in the initial manufacturing knowledge graph are dynamically added to, deleted from, weighted, and their topology adjusted to form an evolved manufacturing knowledge graph, including: Read the equipment nodes, process nodes, quality nodes, causal relationship edges, and visual attribute nodes from the initial manufacturing knowledge graph; Based on the initial manufacturing knowledge graph, a dual-buffered knowledge graph architecture is constructed. The main graph is used to support the current real-time inference task, and the secondary graph serves as an update buffer to receive new observation data and feedback information. Perform local subgraph reconstruction in the subgraph, including dynamic node addition and deletion operations, edge weight correction operations, and topology adjustment operations, to obtain the reconstruction result; Based on the reconstruction results, the logical output differences of the main and secondary graphs on the critical path from the equipment node through the process node to the quality node are obtained, and the difference measurement value is obtained. Based on the comparison result between the difference metric and the preset tolerance threshold, perform a sub-map switching operation or a conflict resolution operation to obtain the operation result; Based on the operation results, an evolved manufacturing knowledge graph is output, which includes node weights, edge relationships, and topological structure updated with feedback information.

[0009] Furthermore, local subgraph reconstruction is performed in the subgraph, including dynamic node addition and deletion operations, edge weight correction operations, and topology adjustment operations, including: Analyze new observation data to extract entity information, relationship information, and attribute information; Traverse the existing node set in the sub-map to determine whether the entities in the new observation data already exist in the map; For entities that do not exist in the subgraph, add corresponding entity nodes in the subgraph and initialize node attribute values ​​and node weights; For entities that already exist in the sub-map, update the node attribute values ​​based on the new observation data, and calculate the corrected values ​​of the node weights based on the node's historical behavior data and the current observation data; Traverse the existing edge set in the subgraph and determine whether the relationship in the new observation data already exists in the graph; For relations that do not exist in the subgraph, create new relation edges in the subgraph and initialize the edge weights; For existing relationships in the subgraph, the edge weights are adjusted based on the causal strength or correlation strength in the new observation data. Detect changes in node degree distribution and connectivity in the subgraph, and perform node deletion or edge deletion operations when isolated nodes or redundant edges are detected. Record the timestamps and operation types of all add, delete, and modify operations, maintain the time version information of the subgraph, and form the subgraph after the local subgraph reconstruction is completed.

[0010] Furthermore, based on the comparison results between the difference metric and the preset tolerance threshold, a sub-map switching operation or a conflict resolution operation is performed, including: Obtain the difference metric value between the primary and secondary maps, which is defined as the weighted Euclidean distance between the output results of the primary and secondary maps on the critical path; Determine whether the difference metric is lower than a preset tolerance threshold; If the difference metric is lower than the preset tolerance threshold, a subgraph switching operation is performed, atomically replacing all nodes, edges and attribute data of the subgraph with the main graph, and releasing the memory space occupied by the original main graph. If the difference metric value is higher than the preset tolerance threshold, the conflict resolution operation is initiated, including: tracing back the historical decision chain to locate the key node or key edge that caused the abnormal difference metric value. Analyze the attribute and weight differences of key nodes or key edges in the main graph and subgraph; generate conflict resolution suggestions, including node weight correction schemes or edge weight correction schemes. Submit conflict resolution suggestions to the manual verification interface or the automatic verification system for verification; If the verification passes, the subgraph is corrected according to the conflict resolution recommendations, and the local subgraph reconstruction is re-executed; if the verification fails, the subgraph is rolled back to its pre-reconstruction state, and the conflict event log is recorded.

[0011] Furthermore, based on the evolved manufacturing knowledge graph, the hybrid inference engine is invoked to generate a set of optimization decision instructions for the current manufacturing condition, including: Read the node weights, edge relationships, and topology of the evolved manufacturing knowledge graph after updating with feedback information; Based on the node weights and edge relationships, the Rete algorithm is used to perform pattern matching between the node states and rule conditions in the knowledge graph, execute deterministic rule reasoning, and output the rule reasoning results. The evolved manufacturing knowledge graph is input into a three-layer graph attention network. Based on the topology, neighborhood aggregation, attention weighting calculation and ReLU nonlinear transformation are performed in each layer. The hidden state vectors of equipment nodes and process nodes are obtained through multi-layer message passing. The hidden state vectors of the equipment nodes and process nodes are classified and calculated, and the action probability distribution for different optimization objectives is output. The action probability distribution includes the probability values ​​of candidate solutions for parameter adjustment, process reordering or equipment switching. Obtain the uncertainty metric value of the current manufacturing condition, and dynamically adjust the fusion weight of the rule inference result and the action probability distribution based on the uncertainty metric value; Based on the fusion weight, the probability values ​​of the candidate solutions for parameter adjustment, process reordering or equipment switching are weighted and fused with the rule reasoning results to obtain the candidate solution with the highest probability value after fusion. The candidate solutions are generated into corresponding optimization decision instruction sets, which include parameter adjustment instructions, process reordering instructions, or equipment switching instructions. The optimized decision instruction set is output, which includes the fusion result of deterministic decision based on rule reasoning and optimized decision based on action probability distribution.

[0012] Furthermore, the optimized decision instruction set is executed, and the actual operating results after execution are collected synchronously as feedback data, including: Read parameter adjustment instructions, process reordering instructions, or equipment switching instructions generated based on the fusion of rule reasoning and action probability distribution from the optimization decision instruction set; The parameter adjustment instructions, process rescheduling instructions, or equipment switching instructions are sent to the PLC controller or HMI terminal via a RESTful API that conforms to the ISA-95 Level 3 specification. During the execution of parameter adjustment instructions, process rescheduling instructions, or equipment switching instructions, multi-dimensional execution data, including actual equipment operating parameters, product quality inspection results, energy consumption metering values, and operator confirmation feedback, are collected in real time. The multidimensional execution data is appended with a unique tag identical to the original decision instruction, and the execution start and end times, environmental conditions, and operation context are recorded to form a structured feedback data stream and write it into the InfluxDB time series database. The feedback data stream is output, which includes the actual operating parameters of the equipment, product quality test results, energy consumption measurement values ​​and their corresponding original decision contexts, corresponding to the parameter adjustment instruction, process rescheduling instruction or equipment switching instruction.

[0013] Furthermore, the feedback data is associated and bound with the corresponding original decision context to form a feedback sample set, which is then used to trigger the next round of knowledge evolution and iterative updates to the data governance strategy, including: A time alignment operation is performed on the actual operating parameters of the equipment, the product quality test results, and the energy consumption metering values ​​to obtain the alignment operation results. Based on the alignment operation results, the indicators scattered in different collection periods are unified into the decision execution time window, and the missing values ​​are filled in by linear interpolation method to generate time-aligned feedback data. The indicators of different dimensions in the time-aligned feedback data are normalized and converted into dimensionless values ​​in the range [0,1] to generate normalized feedback data. The deviation between the normalized feedback data and the expected target is obtained based on the weighted Euclidean distance. A deviation feedback label is generated based on the deviation value. The deviation feedback label includes positive reinforcement, negative correction or neutral ignore types. The normalized feedback data is associated and bound with the deviation feedback label to form a feedback sample set; The feedback sample set is stored in the Redis feedback memory pool in the form of key-value pairs, where the key is a unique identifier for the decision instruction and the value is a composite object containing the execution context, expected target, actual result and the deviation value. The updated value set of the relevant node weights and edge relationships in the knowledge graph is calculated based on the deviation feedback labels in the feedback sample set. The updated value set includes node weight increments, edge weight correction coefficients, and topology adjustment parameters. The adjusted set of weighting coefficients in the data value assessment function is calculated based on the actual contribution of the normalized feedback data items in the feedback sample set. This adjusted set includes the increment of the citation frequency weight. Increment of feature importance weights and the increment of the heat weight. ; Based on the updated value set and the adjusted value set, a dual-path parallel update operation is performed.

[0014] Further, based on the updated value set and the adjusted value set, a dual-path parallel update operation is performed, including: In the knowledge graph evolution path, the node weight increment is accumulated to the current node weight value to obtain the updated node weight value; the edge weight correction coefficient is multiplied by the current relation edge weight value to obtain the corrected edge weight value; the topology adjustment parameter is applied to the node addition / deletion and edge connection relationship adjustment of the graph structure to complete the update of the knowledge graph from the current state to the evolved state; In the data governance path, the incremental reference frequency weight is added to the current reference frequency weight to obtain the updated reference frequency weight; the incremental feature importance weight is added to the current feature importance weight to obtain the updated feature importance weight; the incremental usage popularity weight is added to the current usage popularity weight to obtain the updated usage popularity weight; and the updated data value assessment function is recalculated using the three updated weight values, so that the knowledge graph evolution and data governance strategy are updated in synergy.

[0015] Secondly, a smart manufacturing data management system based on big data processing includes: Data acquisition module: used to collect multi-source heterogeneous data streams from the manufacturing site and generate an initial dataset; Data processing module: used to perform dynamic data governance operations based on value density assessment on each data item in the initial dataset to obtain a high-value dataset; The graph construction module is used to transmit the high-value dataset to the knowledge graph construction layer to generate an initial manufacturing knowledge graph; and to dynamically add, delete, correct weights, and adjust the topology of nodes and edges in the initial manufacturing knowledge graph to form an evolved manufacturing knowledge graph. Instruction generation module: Based on the evolved manufacturing knowledge graph, it calls the hybrid inference engine to generate an optimized decision instruction set for the current manufacturing condition; Update module: used to execute the optimized decision instruction set and synchronously collect the actual running results after execution as feedback data; associate and bind the feedback data with the corresponding original decision context to form a feedback sample set, and use the feedback sample set to trigger the next round of iterative updates of knowledge evolution and data governance strategies.

[0016] The beneficial effects of this invention are as follows: This invention is successful. Attached Figure Description

[0017] Figure 1 This is a schematic diagram of a data management method for intelligent manufacturing based on big data processing provided in an embodiment of the present invention; Figure 2 This is a schematic diagram of the intelligent manufacturing data management process provided in an embodiment of the present invention; Figure 3 This is a schematic diagram of a smart manufacturing data management system module based on big data processing, provided in an embodiment of the present invention. Detailed Implementation

[0018] 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.

[0019] At least one embodiment of the present invention discloses a data management method and system for intelligent manufacturing based on big data processing, comprising: like Figures 1-2 As shown, a data management method for intelligent manufacturing based on big data processing includes the following steps: Step 1: Collect multi-source heterogeneous data streams from the manufacturing site to generate an initial dataset; Step 2: Perform dynamic data governance operations based on value density assessment on each data item in the initial dataset to obtain a high-value dataset; Step 3: Transfer the high-value dataset to the knowledge graph construction layer to generate an initial manufacturing knowledge graph; dynamically add, delete, adjust weights, and modify the topology of nodes and edges in the initial manufacturing knowledge graph to form an evolved manufacturing knowledge graph; Step 4: Based on the evolved manufacturing knowledge graph, call the hybrid inference engine to generate an optimized decision instruction set for the current manufacturing condition; Step 5: Execute the optimized decision instruction set and simultaneously collect the actual running results after execution as feedback data; associate and bind the feedback data with the corresponding original decision context to form a feedback sample set, and use the feedback sample set to trigger the next round of knowledge evolution and data governance strategy iteration update.

[0020] The system architecture of this invention includes a data acquisition layer, a dynamic data governance layer, a knowledge graph construction and evolution layer, a hybrid reasoning decision-making layer, an execution feedback acquisition layer, and a dual-path iterative update layer. These layers are loosely coupled through standardized interfaces and message middleware, and end-to-end consistency is ensured by a unified data model and time synchronization mechanism.

[0021] At the data acquisition layer, sensor networks, programmable logic controllers, manufacturing optimization execution systems, data acquisition and monitoring devices, visual inspection equipment, and manual operation terminals deployed on the manufacturing site constitute a multi-source heterogeneous data input end. These devices push raw data streams, such as equipment operating status, process parameters, quality inspection results, energy consumption data, image features, and operation logs, to edge computing nodes in real time through unified architecture protocols for manufacturing control, industrial fieldbus communication protocols, message queue telemetry transmission protocols, or RESTful API interfaces. The edge computing nodes' built-in timestamp alignment uses PTP (Precise Time Protocol) or NTP (Network Time Protocol) to synchronize data packets from different sources at the nanosecond or millisecond level, ensuring that all data items have a unified time base. Subsequently, format normalization converts heterogeneous formats such as key-value pair text format, Extensible Markup Language format, comma-separated value format, and binary stream into a unified Remote Procedure Call serialization framework format or columnar storage format, and adds semantic annotation information. Semantic annotation is driven by a predefined manufacturing ontology library, for example, mapping temperature sensor 001 to <equipment entity: injection molding machine 03> — [monitoring] — > <attribute: melt temperature>, thereby generating an initial dataset with structured semantic labels, which is then streamed to the dynamic data governance layer.

[0022] After receiving the initial dataset, the dynamic data governance layer initiates a real-time scoring mechanism based on value density assessment. The core of this mechanism is the data value assessment function. R(d) is calculated by querying a historical decision log database (such as an Elasticsearch index) to determine the number of times the data item d has been accessed in the past 30 days; F(d) is returned by the feature contribution value or feature ranking importance value from the currently deployed gradient boosting decision tree prediction model or lightweight gradient prediction model; U(d) is calculated by the frequency of the data item being read or subscribed to within a sliding time window (such as the last 5 minutes). The weighting coefficients α, β, and γ are initially set to 0.4, 0.4, and 0.2, respectively, but will be dynamically adjusted based on the contribution of the feedback sample set. The data value score calculates V(d) independently for each data item and routes it to one of three parallel processing pipelines based on the result: When V(d) ≥ T1 (e.g., T1 = 0.85), the data item enters the full-precision processing pipeline, retains the original sampling rate (e.g., 1kHz), and performs feature enhancement via GPU acceleration; when T2 ≤ V(d) < T1 (e.g., T2 = 0.6), it enters the downsampling processing pipeline, using the LTTB (Maximum Triangle Three-Bug Downsampling) algorithm to compress the sampling rate to 1 / 10 of the original rate, and allocates medium CPU resources; when V(d) < T2, only metadata (including data source ID, time range, data type, and semantic tags) is extracted, the original payload is discarded, and it enters the metadata extraction pipeline. The outputs of the three pipelines are encapsulated into Avro (serialized) format messages by KafkaProducer (distributed message queue producer) and published to a Kafka (message queue) topic named high_value_data_topic for consumption by the knowledge graph construction layer.

[0023] The knowledge graph construction layer subscribes to the aforementioned topics through KafkaConsumer (a distributed message queue consumer) to receive high-value datasets. This layer includes entity recognition, relation extraction, and attribute fusion. Entity recognition uses a BiLSTM-CRF model (a bidirectional long short-term memory network combined with a conditional random field model) to extract equipment entities (e.g., CNC machine tool 07), process entities (e.g., milling process), and quality indicator entities (e.g., surface roughness Ra) from text logs or structured fields. Relationship extraction utilizes a pre-trained BERT model (a bidirectional encoder representation transformation model) combined with rule templates to identify causal or correlational relationships such as equipment-execution-process and process-impact-quality. Attribute fusion performs conflict detection and fusion on attributes of the same entity from different data sources (e.g., temperature, vibration, current), generating unified attribute values ​​using a weighted average or confidence-first strategy. The results of these processes are written into the initial manufacturing knowledge graph, which is stored in the JanusGraph database as RDF (Resource Description Framework) triples. Nodes represent entities, edges represent relations, and both nodes and edges carry weight and timestamp attributes.

[0024] Subsequently, the system initiates the knowledge graph evolution mechanism. This mechanism employs a dual-buffer architecture: the MasterGraph resides in a memory cache (such as RedisGraph) to support millisecond-level inference requests; the ShadowGraph serves as an offline update area, receiving new observation data and feedback information. Local subgraph reconstruction is performed in the ShadowGraph: first, entities, relations, and attributes in the new data are parsed; then, the existing node set of the ShadowGraph is traversed. If a new entity does not exist, a new node is created and its weight is initialized to 0.5; if it already exists, the attribute values ​​are updated, and the node weight adjustment value is calculated based on the Exponentially Weighted Moving Average (EWMA) algorithm. For relation edges, if it is a newly added relation, an edge is created and its initial weight is set to 0.6; if it is an existing relation, the edge weight is adjusted according to the causal strength of the new observation (such as the Granger causality test p-value) or the correlation coefficient. Simultaneously, the system periodically checks the graph topology. If a node has no in-degree or out-degree for 7 consecutive days, it is marked as an isolated node and deleted; if two edges express the same semantics and their weight difference is less than 0.05, they are merged into one edge. All operations are recorded with timestamps and operation types to form a version snapshot.

[0025] After reconstruction, the system calculates the difference metric for the primary and secondary graphs on the critical path. The critical path is defined as the shortest causal chain from the equipment node through the process node to the quality node, such as injection molding machine 03 → injection molding → product dimensional deviation. The difference metric is calculated using weighted Euclidean distance, with the weights determined by the importance of each node in historical fault diagnosis. If the difference metric is below the tolerance threshold (e.g., 0.15), an atomic switch is performed: all data of the secondary graph is written to the memory area of ​​the primary graph, and the old primary graph memory is released; if it is above the threshold, conflict resolution is initiated: the most recent 30 decision chains are traced back to locate the node with the largest difference (e.g., the injection pressure node), and its weight in the primary and secondary graphs (e.g., 0.78 for the primary graph and 0.92 for the secondary graph) and attribute distribution are compared to generate a correction suggestion (e.g., adjusting the weight to 0.85). This suggestion is submitted to the automatic verification step, which verifies the inference accuracy of the corrected graph in historical scenarios through Monte Carlo simulation; if the accuracy improves by more than 2%, the suggestion is adopted and the secondary graph is reconstructed; otherwise, it is rolled back and the conflict log is recorded.

[0026] The hybrid reasoning engine generates an optimized decision instruction set based on an evolved manufacturing knowledge graph. This engine incorporates two reasoning methods: deterministic rule-based reasoning and graph neural network reasoning. Rule-based reasoning loads the Drools (Business Rule Management System) rule base and uses the Rete (pattern matching) algorithm to match node states (e.g., equipment vibration > 5 mm / s) with rule conditions (e.g., IF high vibration AND normal temperature THEN, recommend checking bearings), outputting a deterministic conclusion. The graph neural network inputs the graph into a three-layer GAT (Graph Attention Network), with each layer performing neighborhood aggregation: for each equipment node, it collects features from its neighboring process nodes, calculates weights through an attention mechanism, performs a weighted sum, and then activates the result using ReLU (Revised Linear Unit) to generate a hidden state vector. Finally, the classification head performs a Softmax (Normalized Exponential Function) operation on the hidden state vector, outputting an action probability distribution, such as parameter adjustment (probability 0.6), process rearrangement (0.3), and equipment switching (0.1). The system simultaneously calculates the uncertainty metric of the current operating condition (such as the prediction entropy based on a Bayesian neural network). If the uncertainty is high (>0.7), the weight of rule-based reasoning is reduced (e.g., 0.3), and the weight of GAT output is increased (0.7); conversely, the opposite adjustment is made. The highest probability action after fusion is converted into a specific instruction, such as adjusting the injection pressure from 80MPa to 85MPa.

[0027] The optimized decision-making instruction set is issued to the shop floor control system via a RESTful API (Concrete State Transfer Application Programming Interface) conforming to the ISA-95 Level 3 (Manufacturing Operations Management Level 3) standard. The API gateway verifies the signature and permissions of the instructions before forwarding them to the corresponding PLC (Programmable Logic Controller) or HMI (Human-Machine Interface) terminal. During execution, the data acquisition layer simultaneously initiates feedback data capture: equipment operating parameters are transmitted back in real time by sensors, product quality is assessed by the online visual inspection system outputting OK / NG labels and defect coordinates, energy consumption is measured by smart meters, and operator confirmation is triggered by HMI buttons. All feedback data is appended with the same UUID (Universally Unique Identifier) ​​as the original instruction and records the execution start and end times (accurate to milliseconds), ambient temperature and humidity, batch number, and other contextual information, forming a structured data stream written to the InfluxDB time-series database.

[0028] The feedback data then proceeds to the association and binding step. This step first aligns the multidimensional indicators (such as temperature, pressure, and roughness) according to the decision execution time window (e.g., 0-10 minutes after the instruction is issued), and fills in missing values ​​using linear interpolation. Next, the indicators of different dimensions are Min-Max normalized and mapped to the [0,1] interval. The system calculates the weighted Euclidean distance between the normalized feedback data and the expected target (e.g., roughness ≤ 0.8μm). If the distance is < 0.1, it is marked as positive reinforcement; > 0.3 as negative correction; and in between as neutral neglect. This label is bound to the normalized data to form a feedback sample, which is stored in the Redis (in-memory database) feedback memory pool as key-value pairs. The key is the instruction UUID, and the value is a JSON (structured data) object containing the execution context, expected target, actual result, and deviation value.

[0029] Finally, the dual-path parallel update mechanism is triggered. In the knowledge graph evolution path, the system traverses the graph nodes involved in the feedback samples (such as injection pressure and surface roughness), and calculates the weight update value based on the deviation feedback label: positive reinforcement increases the node weight by 0.05, negative correction decreases it by 0.1, and neutral neglect remains unchanged; edge weights are corrected based on the joint deviation of the two end nodes. The topology adjustment parameters are driven by the isolated node detection results. In the data governance path, the system statistically analyzes the actual contribution of data item d in each feedback sample (such as whether it appears in high-impact decisions), and calculates Δα, Δβ, and Δγ accordingly: if d appears frequently in positive samples, then Δα = +0.02; if its feature importance decreases significantly in negative samples, then Δβ = -0.03; the heat increment Δγ is calculated based on the change in access frequency within the feedback window. The updated α, β, and γ immediately take effect in the next round of data value assessment.

[0030] Through the above closed-loop process, autonomous iteration from raw data collection to knowledge evolution and then to strategy optimization is achieved. The components work closely together through middleware such as Kafka (distributed message queue), Redis (in-memory database), InfluxDB (time series database), and JanusGraph (distributed graph database). Data flow and control flow are strictly synchronized, ensuring the real-time, accuracy and adaptability of manufacturing data management.

[0031] To enable those skilled in the art to fully understand and implement this invention, the specific implementation principles of this invention are further supplemented below with a specific application scenario.

[0032] In an automotive parts injection molding workshop, 20 injection molding machines, a matching mold temperature control system, online visual inspection equipment, and a Manufacturing Execution System (MES) are deployed. Upon system startup, the data acquisition layer first obtains real-time melt temperature, injection pressure, holding time, and other process parameters from the PLCs of each injection molding machine via the OPCUA (Unified Architecture Protocol for Manufacturing Control) protocol; it receives cooling water inlet and outlet temperature and flow data from the mold temperature controller via the MQTT protocol; and it retrieves the current production batch number, mold number, and planned cycle time from the MES system via RESTful API. Simultaneously, industrial cameras deployed at the end of the production line capture product appearance images at a frequency of 5 frames per second, and edge AI extracts surface defect coordinates and type labels. All of the above raw data streams are pushed to edge computing nodes deployed on the workshop edge server. These nodes have built-in PTP time synchronization, aligning data packets from different devices to a unified nanosecond-level timestamp according to the IEEE 1588 (Precision Clock Synchronization International) standard, ensuring that the peak injection pressure and the product shrinkage mark location correspond precisely on the timeline. Subsequently, format normalization was performed on pressure sensor A3 based on a predefined manufacturing ontology library (such as an ISO 13374-4 compatible equipment semantic model). The mapping is <Equipment Entity: Injection Molding Machine 12> — [Monitoring] — > <Attribute: Injection Pressure>. PLC (Programmable Logic Controller) data in JSON (key-value text) format, binary image features, and MES (Manufacturing Execution System) work orders in CSV (comma-separated values) format are uniformly converted into Apache Avro (Remote Procedure Call Serialization Framework) format. After adding structured semantic tags, an initial dataset is formed and continuously output to the dynamic data governance layer.

[0033] After receiving the initial dataset, the dynamic data governance layer immediately calls the data value assessment function for each data item d. For example, regarding the data item of melt temperature of injection molding machine 12, the system queries the historical decision log in Elasticsearch and finds that it has been called 42 times in the past 30 days, so R(d) = 42 / Max(R) = 0.84; at the same time, the LightGBM (Lightweight Gradient Boosting) model currently deployed online for predicting product warpage returns a SHAP value of 0.31 for this feature, ranking 5th among all 128 features, and after normalization, F(d) = 0.92; in addition, in the last 5 minutes, this data item has been subscribed to 18 times by the knowledge graph construction layer, and the maximum subscription frequency of the sliding window is 30 times, so U(d) = 18 / 30 = 0.6. The initial weighting coefficients are set to α = 0.4, β = 0.4, and γ = 0.2, then V(d) = 0.4 × 0.84 + 0.4 × 0.92 + 0.2 × 0.6 = 0.824. Since T2 = 0.6 ≤ V(d) = 0.824 < T1 = 0.85, this data item is routed to the downsampling processing pipeline. The LTTB algorithm compresses its original 1kHz sampling rate to 100Hz, preserving key waveform inflection points, while allocating moderate CPU resources for moving average filtering. The processed data, along with semantic tags, is encapsulated into Avro (serialized format) messages by the KafkaProducer (distributed message queue producer) and published to high-value data topics.

[0034] The knowledge graph construction layer consumes messages on this topic through KafkaConsumer (a distributed message queue consumer). Entity recognition uses a BiLSTM-CRF (Bidirectional Long Short-Term Memory Network combined with Conditional Random Field) model to extract three entities from the operation log of injection molding machine 12-20240515-batch A7: injection molding machine 12, injection molding process, and product size deviation. Relationship extraction is based on a fine-tuned bidirectional encoder representation transformation model, combined with the rule template [equipment] affecting [quality indicators] in [process], identifying two relationships: injection molding machine 12—[execution]→injection molding process and injection molding process—[cause]→product size deviation. Attribute fusion performs confidence-weighted fusion of three-source temperature data from PLC (melt temperature 85℃), infrared thermal imager (83℃), and MES process card (set value 86℃). Because the PLC sensor has a shorter calibration cycle, it is given a weight of 0.6, ultimately generating a unified attribute value of 85.2℃. The aforementioned triples are written into the JanusGraph (distributed graph) database to form the initial manufacturing knowledge graph.

[0035] Subsequently, the knowledge graph evolution mechanism was activated. Upon receiving new observation data, the sub-graph found that the product size deviation node already existed, but its associated holding time attribute was updated from 3.2s to 3.5s. The system used the EWMA algorithm to calculate the node weight correction value with λ=0.3: new weight = 0.3×1 + 0.7×0.78 = 0.846 (original weight 0.78). For the newly added mold temperature controller 12—[control]→melt temperature relationship, the system established a new edge and initialized the weight to 0.6. Simultaneously, topology detection revealed that the old mold temperature sensor 09 node had no data inflow for 7 consecutive days, marking it as an isolated node and deleting it. After reconstruction, the system calculated the differences between the main and sub-graphs on the critical path injection molding machine 12→injection molding→product size deviation: in the main graph, the injection pressure weight was 0.78, and the holding time was 0.72; in the sub-graph, they were 0.81 and 0.85 respectively. The weighted Euclidean distance was calculated using the importance of each node in the historical fault diagnosis as a weight (injection pressure 0.6, holding time 0.4). Since the value is below the tolerance threshold of 0.15, an atomic switch is performed to fully load the subgraph into the main graph memory area of ​​RedisGraph (memory graph database).

[0036] The hybrid inference engine then generates decisions based on the updated graph. Rule-based inference matches the Drools rule: if the melt temperature > 85℃ and the product size deviation > 0.1mm, then it suggests lowering the melt temperature. The current melt temperature is 85.2℃ and the size deviation is 0.12mm, triggering this rule. The graph neural network inputs the graph into a three-layer GAT (Graph Attention Network): the first layer aggregates the features of the mold temperature controller 12 and the injection molding process in the neighborhood of injection molding machine 12; the second layer aggregates the product size deviation and surface shrinkage features in the neighborhood of the injection molding process; the third layer generates the hidden state vector of the equipment nodes. The classification head outputs the action probability distribution: parameter adjustment 0.65, process rearrangement 0.25, equipment switching 0.10. The system simultaneously calculates the Bayesian GAT prediction entropy, which is 0.68 (< 0.7), indicating low uncertainty in the judgment. Therefore, the rule-based inference weight is set to 0.7, and the GAT weight to 0.3. After fusion, lowering the melt temperature yields the highest overall score, generating an instruction to adjust the melt temperature setpoint of injection molding machine 12 from 86℃ to 84℃.

[0037] After being signed and verified by the API (Application Programming Interface) gateway, the instruction is sent to the HMI (Human-Machine Interface) terminal of injection molding machine 12 via the RESTful API (Realistic State Transfer Application Programming Interface). During execution, the data acquisition layer synchronously captures feedback: the PLC reports that the new melt temperature is stable at 84.1℃; the vision inspection system outputs that the product size deviation has decreased to 0.08mm and is marked as OK; the smart meter records a decrease in energy consumption of 2.3kWh; and the operator clicks to confirm the effect on the HMI. All feedback data is appended with the original instruction's universally unique identifier: UUIDcmd-20240515-12345, and the ambient temperature of 25℃, humidity of 60%, batch A7, and other context information are recorded and written to the time-series database.

[0038] The associated binding aligns the feedback data within a 0–10 minute window after command execution, and missing vibration data is filled in using linear interpolation. After Min-Max normalization, a size deviation of 0.08 is mapped to 0.2 (target ≤ 0.1 corresponds to a normalized value of 0.0), and the energy consumption reduction is mapped to 0.7. The system calculates the weighted Euclidean distance. (Expected energy consumption baseline normalized value 0.5), between 0.1 and 0.3, marked as neutral and ignored. This feedback sample is stored in Redis (in-memory database) with the key cmd-20240515-12345.

[0039] The dual-path update mechanism is triggered as follows: In the knowledge graph path, the melt temperature node maintains its weight of 0.846 due to its neutral result; the weight of the product size deviation node also remains unchanged. In the data governance path, the system analysis shows that the instruction did not significantly improve quality, but the melt temperature data item was frequently used in this decision-making process, resulting in a U(d) boost of 15, hence Δγ = +0.015; while its SHAP value in the LightGBM (Lightweight Gradient Boosting) model decreased from 0.31 to 0.28, hence Δβ = -0.02. After the update, β = 0.38 and γ = 0.215, which are immediately used for the next round of data scoring. Thus, the system completes a full closed loop from data acquisition to strategy optimization to model self-evolution, achieving adaptive response and continuous optimization to complex manufacturing conditions.

[0040] like Figure 2 As shown, a smart manufacturing data management system based on big data processing includes: Data acquisition module: used to collect multi-source heterogeneous data streams from the manufacturing site and generate an initial dataset; Data processing module: used to perform dynamic data governance operations based on value density assessment on each data item in the initial dataset to obtain a high-value dataset; The graph construction module is used to transmit the high-value dataset to the knowledge graph construction layer to generate an initial manufacturing knowledge graph; and to dynamically add, delete, correct weights, and adjust the topology of nodes and edges in the initial manufacturing knowledge graph to form an evolved manufacturing knowledge graph. Instruction generation module: Based on the evolved manufacturing knowledge graph, it calls the hybrid inference engine to generate an optimized decision instruction set for the current manufacturing condition; Update module: used to execute the optimized decision instruction set and synchronously collect the actual running results after execution as feedback data; associate and bind the feedback data with the corresponding original decision context to form a feedback sample set, and use the feedback sample set to trigger the next round of iterative updates of knowledge evolution and data governance strategies.

[0041] The embodiments of the present invention have been described above. However, 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 equivalent embodiments under the guidance of the present embodiments, and all of them are within the protection scope of the present embodiments.

Claims

1. A data management method for intelligent manufacturing based on big data processing, characterized in that, include: Collect multi-source heterogeneous data streams from the manufacturing site to generate an initial dataset; Perform dynamic data governance operations based on value density assessment on each data item in the initial dataset to obtain a high-value dataset; The high-value dataset is transferred to the knowledge graph construction layer to generate an initial manufacturing knowledge graph. The nodes and edges in the initial manufacturing knowledge graph are dynamically added, deleted, weighted, and their topology is adjusted to form the evolved manufacturing knowledge graph. Based on the evolved manufacturing knowledge graph, the hybrid inference engine is invoked to generate an optimized decision instruction set for the current manufacturing conditions; The optimized decision instruction set is executed, and the actual running results after execution are collected synchronously as feedback data. The feedback data is associated and bound with the corresponding original decision context to form a feedback sample set, which is then used to trigger the next round of knowledge evolution and iterative updates of data governance strategies.

2. The intelligent manufacturing data management method based on big data processing according to claim 1, characterized in that, Perform dynamic data governance operations based on value density assessment on each data item in the initial dataset to obtain a high-value dataset, including: Through data value assessment function Each data item in the initial dataset is scored in real time; where V(d) represents the value score of data item d, R(d) represents the reference frequency of the data item in the historical decision log, F(d) represents the feature importance score of the feature associated with the data item in the currently deployed prediction model, U(d) represents the usage popularity index of the data item in the current sliding time window, and α, β, and γ are preset weighting coefficients that satisfy α+β+γ=1. The weighting coefficients are dynamically adjusted according to the actual contribution of the feedback sample set. When V(d) is higher than the first threshold T1, the original precision is preserved and all computing resources are allocated to enter the full-precision processing pipeline. When V(d) is between the first threshold T1 and the second threshold T2, perform compressed sampling and allocate moderate computing resources to enter the downsampling pipeline; When V(d) is below the second threshold T2, only metadata is extracted and the original payload is discarded, and the metadata extraction pipeline is entered. Three processing pipelines are run in parallel, and the data items selected after value density assessment are aggregated to form a high-value dataset; The high-value dataset contains hierarchical data that has undergone full-precision processing, downsampling, and metadata extraction. The high-value dataset is sent to the knowledge graph construction layer via a data transmission interface as high-quality input data for entity recognition, relation extraction, and graph generation.

3. The intelligent manufacturing data management method based on big data processing according to claim 1, characterized in that, The initial manufacturing knowledge graph is dynamically updated by adding, deleting, adjusting weights, and refining topology to form an evolved manufacturing knowledge graph, including: Read the equipment nodes, process nodes, quality nodes, causal relationship edges, and visual attribute nodes from the initial manufacturing knowledge graph; Based on the initial manufacturing knowledge graph, a dual-buffered knowledge graph architecture is constructed. The main graph is used to support the current real-time inference task, and the secondary graph serves as an update buffer to receive new observation data and feedback information. Perform local subgraph reconstruction in the subgraph, including dynamic node addition and deletion operations, edge weight correction operations, and topology adjustment operations, to obtain the reconstruction result; Based on the reconstruction results, the logical output differences of the main and secondary graphs on the critical path from the equipment node through the process node to the quality node are obtained, and the difference measurement value is obtained. Based on the comparison result between the difference metric and the preset tolerance threshold, perform a sub-map switching operation or a conflict resolution operation to obtain the operation result; Based on the operation results, an evolved manufacturing knowledge graph is output, which includes node weights, edge relationships, and topological structure updated with feedback information.

4. The intelligent manufacturing data management method based on big data processing according to claim 3, characterized in that, Perform local subgraph reconstruction in the subgraph, including dynamic node addition and deletion operations, edge weight correction operations, and topology adjustment operations, including: Analyze new observation data to extract entity information, relationship information, and attribute information; Traverse the existing node set in the sub-map to determine whether the entities in the new observation data already exist in the map; For entities that do not exist in the subgraph, add corresponding entity nodes in the subgraph and initialize node attribute values ​​and node weights; For entities that already exist in the sub-map, update the node attribute values ​​based on the new observation data, and calculate the corrected values ​​of the node weights based on the node's historical behavior data and the current observation data; Traverse the existing edge set in the subgraph and determine whether the relationship in the new observation data already exists in the graph; For relations that do not exist in the subgraph, create new relation edges in the subgraph and initialize the edge weights; For existing relationships in the subgraph, the edge weights are adjusted based on the causal strength or correlation strength in the new observation data. Detect changes in node degree distribution and connectivity in the subgraph, and perform node deletion or edge deletion operations when isolated nodes or redundant edges are detected. Record the timestamps and operation types of all add, delete, and modify operations, maintain the time version information of the subgraph, and form the subgraph after the local subgraph reconstruction is completed.

5. The intelligent manufacturing data management method based on big data processing according to claim 3, characterized in that, Based on the comparison results between the difference metric and the preset tolerance threshold, perform sub-map switching operations or conflict resolution operations, including: Obtain the difference metric value between the primary and secondary maps, which is defined as the weighted Euclidean distance between the output results of the primary and secondary maps on the critical path; Determine whether the difference metric is lower than a preset tolerance threshold; If the difference metric is lower than the preset tolerance threshold, a subgraph switching operation is performed, atomically replacing all nodes, edges and attribute data of the subgraph with the main graph, and releasing the memory space occupied by the original main graph. If the difference metric value is higher than the preset tolerance threshold, the conflict resolution operation is initiated, including: tracing back the historical decision chain to locate the key node or key edge that caused the abnormal difference metric value. Analyze the attribute and weight differences of key nodes or key edges in the main graph and subgraph; generate conflict resolution suggestions, including node weight correction schemes or edge weight correction schemes. Submit conflict resolution suggestions to the manual verification interface or the automatic verification system for verification; If the verification passes, the subgraph is corrected according to the conflict resolution recommendations, and the local subgraph reconstruction is re-executed; if the verification fails, the subgraph is rolled back to its pre-reconstruction state, and the conflict event log is recorded.

6. The intelligent manufacturing data management method based on big data processing according to claim 1, characterized in that, Based on the evolved manufacturing knowledge graph, a hybrid inference engine is invoked to generate a set of optimization decision instructions for the current manufacturing conditions, including: Read the node weights, edge relationships, and topology of the evolved manufacturing knowledge graph after updating with feedback information; Based on the node weights and edge relationships, the Rete algorithm is used to perform pattern matching between the node states and rule conditions in the knowledge graph, execute deterministic rule reasoning, and output the rule reasoning results. The evolved manufacturing knowledge graph is input into a three-layer graph attention network. Based on the topology, neighborhood aggregation, attention weighting calculation and ReLU nonlinear transformation are performed in each layer. The hidden state vectors of equipment nodes and process nodes are obtained through multi-layer message passing. The hidden state vectors of the equipment nodes and process nodes are classified and calculated, and the action probability distribution for different optimization objectives is output. The action probability distribution includes the probability values ​​of candidate solutions for parameter adjustment, process reordering or equipment switching. Obtain the uncertainty metric value of the current manufacturing condition, and dynamically adjust the fusion weight of the rule inference result and the action probability distribution based on the uncertainty metric value; Based on the fusion weight, the probability values ​​of the candidate solutions for parameter adjustment, process reordering or equipment switching are weighted and fused with the rule reasoning results to obtain the candidate solution with the highest probability value after fusion. The candidate solutions are generated into corresponding optimization decision instruction sets, which include parameter adjustment instructions, process reordering instructions, or equipment switching instructions. The optimized decision instruction set is output, which includes the fusion result of deterministic decision based on rule reasoning and optimized decision based on action probability distribution.

7. The intelligent manufacturing data management method based on big data processing according to claim 4, characterized in that, Execute the optimization decision instruction set and synchronously collect the actual running results after execution as feedback data, including: Read parameter adjustment instructions, process reordering instructions, or equipment switching instructions generated based on the fusion of rule reasoning and action probability distribution from the optimization decision instruction set; The parameter adjustment command, process rescheduling command, or equipment switching command is sent to the PLC controller or HMI terminal. During the execution of parameter adjustment instructions, process rescheduling instructions, or equipment switching instructions, multi-dimensional execution data, including actual equipment operating parameters, product quality inspection results, energy consumption metering values, and operator confirmation feedback, are collected in real time. The multidimensional execution data is appended with a unique tag identical to the original decision instruction, and the execution start and end times, environmental conditions, and operation context are recorded to form a structured feedback data stream and write it into the InfluxDB time series database. The feedback data stream is output, which includes the actual operating parameters of the equipment, product quality test results, energy consumption measurement values ​​and their corresponding original decision contexts, corresponding to the parameter adjustment instruction, process rescheduling instruction or equipment switching instruction.

8. The intelligent manufacturing data management method based on big data processing according to claim 5, characterized in that, The feedback data is associated and bound with the corresponding original decision context to form a feedback sample set, which is then used to trigger the next round of knowledge evolution and data governance strategy iteration updates, including: A time alignment operation is performed on the actual operating parameters of the equipment, the product quality test results, and the energy consumption metering values ​​to obtain the alignment operation results. Based on the alignment operation results, the indicators scattered in different collection periods are unified into the decision execution time window, and the missing values ​​are filled in by linear interpolation method to generate time-aligned feedback data. The indicators of different dimensions in the time-aligned feedback data are normalized and converted into dimensionless values ​​in the range [0,1] to generate normalized feedback data. The deviation between the normalized feedback data and the expected target is obtained based on the weighted Euclidean distance. A deviation feedback label is generated based on the deviation value. The deviation feedback label includes positive reinforcement, negative correction or neutral ignore types. The normalized feedback data is associated and bound with the deviation feedback label to form a feedback sample set; The feedback sample set is stored in the Redis feedback memory pool in the form of key-value pairs, where the key is a unique identifier for the decision instruction and the value is a composite object containing the execution context, expected target, actual result and the deviation value. The updated value set of the relevant node weights and edge relationships in the knowledge graph is calculated based on the deviation feedback labels in the feedback sample set. The updated value set includes node weight increments, edge weight correction coefficients, and topology adjustment parameters. The adjusted set of weighting coefficients in the data value assessment function is calculated based on the actual contribution of the normalized feedback data items in the feedback sample set. This adjusted set includes the increment of the citation frequency weight. Increment of feature importance weights and the increment of the heat weight. ; Based on the updated value set and the adjusted value set, a dual-path parallel update operation is performed.

9. A data management method for intelligent manufacturing based on big data processing according to claim 8, characterized in that, Based on the updated value set and the adjusted value set, a dual-path parallel update operation is performed, including: In the knowledge graph evolution path, the node weight increment is accumulated to the current node weight value to obtain the updated node weight value; the edge weight correction coefficient is multiplied by the current relation edge weight value to obtain the corrected edge weight value; the topology adjustment parameter is applied to the node addition / deletion and edge connection relationship adjustment of the graph structure to complete the update of the knowledge graph from the current state to the evolved state; In the data governance path, the incremental reference frequency weight is added to the current reference frequency weight to obtain the updated reference frequency weight; the incremental feature importance weight is added to the current feature importance weight to obtain the updated feature importance weight; the incremental usage popularity weight is added to the current usage popularity weight to obtain the updated usage popularity weight; and the updated data value assessment function is recalculated using the three updated weight values, so that the knowledge graph evolution and data governance strategy are updated in synergy.

10. A smart manufacturing data management system based on big data processing, used to execute the smart manufacturing data management method based on big data processing as described in any one of claims 1-9, characterized in that, include: Data acquisition module: used to collect multi-source heterogeneous data streams from the manufacturing site and generate an initial dataset; Data processing module: used to perform dynamic data governance operations based on value density assessment on each data item in the initial dataset to obtain a high-value dataset; The graph construction module is used to transmit the high-value dataset to the knowledge graph construction layer to generate an initial manufacturing knowledge graph; and to dynamically add, delete, correct weights, and adjust the topology of nodes and edges in the initial manufacturing knowledge graph to form an evolved manufacturing knowledge graph. Instruction generation module: Based on the evolved manufacturing knowledge graph, it calls the hybrid inference engine to generate an optimized decision instruction set for the current manufacturing condition; Update module: used to execute the optimized decision instruction set and synchronously collect the actual running results after execution as feedback data; associate and bind the feedback data with the corresponding original decision context to form a feedback sample set, and use the feedback sample set to trigger the next round of iterative updates of knowledge evolution and data governance strategies.

Images