Multi-source heterogeneous data processing method

Abstract

According to the multi-source heterogeneous data processing method provided by the invention, equipment data standardization is realized through an extensible protocol conversion module, a hybrid processing pipeline is constructed based on an Flink flow engine and a WarpStream component, state consistency is guaranteed through a Checkpoint mechanism, a dynamic parallelism degree adjustment strategy is adopted to cope with data flow fluctuation, and a multi-stage data fusion strategy is used in a hybrid manner, so that the multi-source heterogeneous data processing efficiency is improved. And deep mining and efficient processing of production data are realized, so that the data fusion error rate is reduced, and the abnormal detection omission ratio is reduced.

Description

Technical Field

[0001] The present invention relates to the technical field of industrial Internet, and in particular to a method for processing multi-source heterogeneous data. Background Art

[0002] As the photovoltaic industry transitions toward intelligent manufacturing, the core role of manufacturing execution systems (MES) in cell production is becoming increasingly prominent. Photovoltaic cell production involves multiple processes, including texturing, diffusion, etching, and coating. These processes require real-time collection of heterogeneous data from multiple sources, including equipment sensor data (such as temperature, pressure, and current), image detection data (such as EL defect images), and process parameters (such as slurry ratio). However, existing MES systems face the following technical bottlenecks in data processing:

[0003] Difficulty accessing multi-protocol devices: Production equipment comes from different manufacturers and uses multiple communication protocols such as Modbus, OPC UA, and MQTT. This leads to complex protocol conversion at the data acquisition layer, making it difficult for edge gateways to achieve unified access and dynamic resource scheduling.

[0004] Stream and batch processing separation: In traditional architectures, real-time stream data (such as sensor time series data) and batch data (such as historical process records) use separate processing frameworks (such as Storm + Hadoop), resulting in fragmented data processing processes and difficulty supporting dynamic optimization of process parameters.

[0005] Inefficient multimodal data fusion: Structured data (equipment logs) and unstructured data (defect images) differ significantly in time scale and feature dimension. Existing methods use a single fusion strategy without distinguishing between data lifecycle stages, resulting in insufficient feature association accuracy (measured error rate reaches 12%-18%).

[0006] Insufficient dynamic adaptability: The existing system lacks a compensation mechanism for device status fluctuations (such as clock offset and network jitter), and the data fusion weight is fixed and cannot be dynamically adjusted according to data quality, affecting the accuracy of anomaly detection (missed detection rate > 8%).

[0007] To address these issues, existing multi-protocol industrial gateways have partially addressed access issues but fall short of addressing the need for integrated batch and stream processing. Existing multi-source data fusion methods propose a layered fusion framework but fail to differentiate between data lifecycle stages. Therefore, a full-process data processing solution for photovoltaic cell production is urgently needed. Summary of the Invention

[0008] In response to the shortcomings of the existing technology, the purpose of the present invention is to propose a multi-source heterogeneous data processing method, which realizes efficient processing and deep mining of production data through edge layer protocol conversion, stream batch processing integration architecture and three-stage data fusion mechanism.

[0009] On the one hand, a method for processing multi-source heterogeneous data is provided, comprising the following steps:

[0010] (1) At the edge computing layer, the protocol conversion module is used to standardize the multi-source heterogeneous data of photovoltaic cell production equipment, enabling unified access of devices with different communication protocols to the edge gateway;

[0011] (2) Use the Flink stream computing engine to process real-time data in a streaming manner, and integrate batch processing and stream processing frameworks through the WarpStream component to build a unified data processing process;

[0012] (3) Data fusion of structured data and unstructured data in three stages: early, middle and late stages:

[0013] Early fusion stage: Feature extraction and correlation analysis of real-time data through the streaming computing engine;

[0014] Mid-term fusion stage: The intermediate results of stream processing are stored in a distributed message queue and aligned with the historical batch data in time series;

[0015] Late fusion stage: Combine knowledge graphs and machine learning models to perform quality assessment and optimized output on the fused data.

[0016] Preferably, the protocol conversion module in step (1) supports dynamic adaptation of Modbus, OPCUA, and MQTT protocols, dynamically allocates computing tasks based on the hardware resources of the edge gateway, and enables preemptive transmission channels for high-priority device data.

[0017] Preferably, the integration of the WarpStream component in step (2) includes:

[0018] The FlinkCheckpoint mechanism is used to implement stream processing state persistence.

[0019] Use distributed snapshot technology to ensure consistency in batch and stream mixed processing;

[0020] Enable back pressure control and automatic retry mechanism for abnormal data flow, and dynamically adjust the parallelism of computing nodes based on workload.

[0021] Preferably, the early fusion stage of step (3) further comprises:

[0022] Defect feature extraction is performed on unstructured image data using a lightweight CNN model, and TensorRT is used on the edge to accelerate the inference process.

[0023] Through the time window mechanism, structured sensor data and image feature data are associated with production batches to build a full-chain production view. The window length is dynamically adjusted to 5-30 seconds according to the process rhythm.

[0024] Preferably, during the mid-fusion phase of step (3):

[0025] The distributed message queue adopts a hybrid architecture of Kafka and Redis, and stores data in a hierarchical manner according to its timeliness, with a retention period of ≤1 hour for real-time data and ≥30 days for historical data.

[0026] The timing alignment process introduces the Dynamic Time Warping (DTW) algorithm to compensate for device clock deviation and calibrate the timestamp offset based on the device heartbeat signal.

[0027] Preferably, the late fusion stage of step (3) specifically includes:

[0028] Build a knowledge graph of photovoltaic cell production processes, including equipment parameters, material properties, and process standard entities, and trigger dynamic updates of the graph based on process change events;

[0029] An LSTM-GAN hybrid model is used to perform anomaly detection and data enhancement on the fused data. A transfer learning mechanism is introduced during model training to reuse cross-production line features.

[0030] The output data includes multi-dimensional information such as quality scores, equipment health indicators, and process optimization suggestions, and is connected to the factory ERP system through RESTful API. When the scores of three consecutive batches are less than 90 points, process parameter adjustment instructions are automatically triggered.

[0031] Preferably, it also includes:

[0032] A lightweight data preprocessing module is deployed on the edge gateway to perform data denoising, null value filling, and unit normalization operations. Denoising uses a hybrid algorithm of wavelet transform and Kalman filtering;

[0033] Set a dynamic weight strategy to adjust the weight distribution in the fusion process based on the reliability and timeliness of the data source. Reliability assessment is based on the quantitative calculation of the historical failure rate of the equipment.

[0034] Preferably, the dynamically adjusting the parallelism of computing nodes includes:

[0035] Monitor CPU / GPU resource utilization in real time and automatically scale out compute container instances when utilization remains above 85% for three minutes.

[0036] Dedicated resource pools are allocated for delay-sensitive data processing tasks, and QoS levels are set to guarantee computing bandwidth.

[0037] Preferably, the dynamic update of the process knowledge graph includes:

[0038] Capture device parameter changes through OPCUA event subscription mechanism;

[0039] When a process standard version update is detected, the graph verification process is automatically triggered and the nodes for manual review are marked;

[0040] Establish a graph version history library to support version rollback and difference comparison.

[0041] Preferably, the implementation of the dynamic weight strategy includes:

[0042] Construct a weight coefficient matrix W = [w1, w2, ..., wn], where wi = α·Ri + β·Ti, wi = α·Ri + β·Ti, Ri is the reliability factor, Ti is the timeliness factor, α+β = 1 and α∈[0.6,0.8];

[0043] The weight coefficient is recalculated every 8 hours, and the coefficient mutation is smoothed by the sliding average algorithm;

[0044] A device health check work order is triggered for data sources with a weight change rate greater than 15% for three consecutive times.

[0045] On the other hand, a multi-source heterogeneous data processing system is provided, which includes a memory, a processor, and a computer program stored in the memory and runnable on the processor. When the processor executes the computer program, it implements the multi-source heterogeneous data processing method of the photovoltaic cell MES system as described above.

[0046] The present invention provides a multi-source heterogeneous data processing method that achieves device data standardization through an extensible protocol conversion module, builds a hybrid processing pipeline based on the Flink streaming engine and WarpStream components, ensures state consistency through a Checkpoint mechanism, and adopts a dynamic parallelism adjustment strategy to cope with data flow fluctuations. A hybrid multi-stage data fusion strategy is used to achieve deep mining and efficient processing of production data, thereby reducing the data fusion error rate and the anomaly detection missed rate. BRIEF DESCRIPTION OF THE DRAWINGS

[0047] Figure 1 A diagram of the architecture of a multi-source data processing system provided by the present invention;

[0048] Figure 2 A multi-source heterogeneous data processing process provided by the present invention;

[0049] Figure 3 A schematic diagram of a multi-source heterogeneous data system provided by the present invention. DETAILED DESCRIPTION

[0050] The following will clearly and completely describe the technical solutions in the embodiments of the present invention in conjunction with the accompanying drawings. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative efforts are within the scope of protection of the present invention.

[0051] It should be noted that the terms used herein are only for describing specific embodiments and are not intended to limit the exemplary embodiments according to the present application. As used herein, unless the context clearly indicates otherwise, the singular form is also intended to include the plural form. In addition, it should be understood that when the terms "comprise" and / or "include" are used in this specification, they indicate the presence of features, steps, operations, devices, components and / or combinations thereof.

[0052] Example 1

[0053] like Figure 1 The figure shows a multi-source data processing architecture provided by the present invention. The data architecture is based on multiple core databases and modules. These databases and modules are interconnected and jointly support the management and monitoring of the entire production process.

[0054] The master data management module includes material management, product management, factory model, user management, work calendar, BOM management, process model, quality master data, and production master data. Among them, material management is used to store and manage all material information related to production, including material code, name, specifications, inventory, etc. Product management is used to record detailed information about products, such as product model, specifications, BOM (bill of materials), etc. The factory model is used to describe the physical layout of the factory, production line configuration, equipment information, etc. User management is used to manage the permissions, roles and identity information of system users. The work calendar is used to record the factory's working hours, holidays, shifts and other information. BOM management is used to record the material composition and quantity of the product in detail. The process model is used to define the production process, parameters and steps of the product. Quality master data is used to store standards, specifications and testing methods related to quality management. Production master data includes production plans, work orders, batches and other information directly related to production.

[0055] The production planning module includes ERP work order planning and work order management. ERP work order planning receives production plans from the ERP system, including product model, quantity, and delivery date. Work order management records each work order's status, quantity, completion time, and other information in detail, allowing for tracking production progress.

[0056] The production execution module includes production batches, feeding and clearing, and output and storage. The production batch module records information for each production batch, including batch number, production time, and feeding quantity. The feeding and clearing module records the feeding and clearing process, including feeding quantity, clearing signals, and the number of debris. The output and storage module records the output and storage quantity of each batch, as well as the relevant machine, storage location, and time information.

[0057] The material collection and return management module is mainly a material issuance and return module, which records the material issuance process from the warehouse to the production line, as well as the material return process after production is completed.

[0058] The quality management module includes sampling, inspection, and processing. Sampling and inspection records information such as sampling quantity, inspection batch, and inspection results to monitor product quality. The processing and feedback module records the handling process and results of quality issues, as well as related information such as slurry consumption.

[0059] Data interaction and integration include automated collection, system integration, and configuration. Automated collection uses sensors, barcode scanners, and other devices to automatically capture data from the production process. System integration connects data with other systems, such as ERP and WMS (Warehouse Management System), enabling information sharing and collaboration. The configuration module offers flexible configuration options to meet the customized needs of different businesses and supports system expansion and upgrades.

[0060] This MES system data architecture diagram illustrates a highly integrated and meticulously managed production environment. Through modules such as master data management, production planning, production execution, material receipt and return management, and quality management, it enables comprehensive monitoring and optimization of production processes. Furthermore, through automated data collection, system integration, and flexible configuration, the system's adaptability and scalability are enhanced, providing enterprises with powerful production management and decision-making support capabilities.

[0061] like Figure 2 As shown, this embodiment provides a multi-source heterogeneous data processing method, which includes:

[0062] (1) At the edge computing layer, the protocol conversion module is used to standardize the multi-source heterogeneous data of photovoltaic cell production equipment, enabling unified access of devices with different communication protocols to the edge gateway;

[0063] (2) Use the Flink stream computing engine to process real-time data in a streaming manner, and integrate batch processing and stream processing frameworks through the WarpStream component to build a unified data processing process;

[0064] (3) Data fusion of structured data and unstructured data in three stages: early, middle and late stages:

[0065] Early fusion stage: Feature extraction and correlation analysis of real-time data through the streaming computing engine;

[0066] Mid-term fusion stage: The intermediate results of stream processing are stored in a distributed message queue and aligned with the historical batch data in time series;

[0067] Late fusion stage: Combine knowledge graphs and machine learning models to perform quality assessment and optimized output on the fused data.

[0068] The protocol conversion module described in step (1) supports dynamic adaptation of Modbus, OPCUA, and MQTT protocols, dynamically allocates computing tasks based on the hardware resources of the edge gateway, and enables preemptive transmission channels for high-priority device data.

[0069] A lightweight data preprocessing module is deployed on the edge gateway to perform data denoising, null value filling, and unit normalization operations. Denoising uses a hybrid algorithm of wavelet transform and Kalman filtering. A dynamic weight strategy is set to adjust the weight distribution in the fusion process according to the reliability and timeliness of the data source. Reliability assessment is based on the quantitative calculation of the historical failure rate of the equipment.

[0070] The implementation of the dynamic weight strategy includes: constructing a weight coefficient matrix W = [w1, w2, ..., wn], where wi = α·Ri + β·Ti, wi = α·Ri + β·Ti, Ri is the reliability factor, Ti is the timeliness factor, α+β = 1 and α∈[0.6, 0.8]; recalculating the weight coefficient every 8 hours and smoothing coefficient mutations using a sliding average algorithm; triggering a device health check work order for data sources with a weight change rate greater than 15% for three consecutive times.

[0071] The integration of the WarpStream component in step (2) includes: implementing stream processing state persistence through the FlinkCheckpoint mechanism, specifically, the Checkpoint regulator in WarpStream periodically sends CheckpointBarrier to all source operators, and when the data source reaches a specific offset, the current position is immediately written to the HDFS persistent storage; using distributed snapshot technology to ensure the consistency of stream-batch mixed processing, specifically, inserting a logical barrier in the streaming data channel to trigger all parallel processing nodes to asynchronously generate local state snapshots, and aligning batch tasks with stream processing snapshots through metadata version numbers to ensure that batch jobs are started based on the latest snapshot generated by stream processing; enabling backpressure control and automatic retry mechanism for abnormal data streams, and notifying upstream operators to temporarily slow down and reduce the amplitude through the TCP backpressure protocol. When the proportion of abnormal data continues to exceed the standard, data intake is temporarily suspended. The automatic retry mechanism can be implemented through shard retry, rescheduling abnormal data shards to healthy nodes for execution, and dynamically adjusting the parallelism of computing nodes based on workload. The dynamic adjustment of computing node parallelism includes: real-time monitoring of CPU / GPU resource utilization, automatically expanding computing container instances when utilization remains ≥85% for three minutes; allocating exclusive resource pools to delay-sensitive data processing tasks, and setting QoS levels to guarantee computing bandwidth.

[0072] The early fusion stage of step (3) also includes: extracting defect features from unstructured image data through a lightweight CNN model, and using TensorRT on the edge to accelerate the inference process; using a time window mechanism to associate structured sensor data with image feature data according to production batches to build a full-link production view, and the window length is dynamically adjusted to 5-30 seconds according to the process cycle.

[0073] In the early stages, solutions need to focus on information integrity, real-time performance, and deep interaction between modalities. In terms of information integrity, structured data and unstructured data are directly integrated at the data input stage. Structured data includes process parameters and equipment status, and unstructured data includes EL detection images and sensor timing. Feature splicing or mapping alignment is used to achieve full-dimensional information integration and improve real-time monitoring capabilities. In terms of information real-time performance, it is suitable for scenarios that require rapid response (such as defect detection). Joint modeling is used to reduce information transmission delays and improve anomaly recognition accuracy. In terms of deep interaction between modalities, EL images (unstructured) and production batch parameters (structured) are mapped to the same feature space through a unified encoder (such as Transformer), enhancing the correlation between process parameters and defect features. Early fusion improves production monitoring accuracy through real-time performance and deep interaction.

[0074] In the mid-term fusion stage of step (3): the distributed message queue adopts a hybrid architecture of Kafka and Redis, and stores data in a hierarchical manner according to the timeliness of the data, where the real-time data retention period is ≤1 hour and the historical data retention period is ≥30 days; the timing alignment process introduces the dynamic time warping algorithm (DTW) to compensate for the device clock deviation, and calibrates the timestamp offset based on the device heartbeat signal.

[0075] An attention mechanism is introduced into the model's middle layer to dynamically weight structured data (such as temperature and current) and unstructured data (such as voiceprint signals) to screen key features and optimize process control logic. Multimodal feature cross-validation is supported (such as correlation analysis between sensor timing and device logs) to reduce false alarm rates. Two types of data are processed in stages (such as independently extracting EL defect features before fusing them with structured parameters) to reduce the impact of unstructured data noise on the global model. Mid-term fusion balances efficiency and robustness through dynamic feature optimization to adapt to complex process scenarios.

[0076] The late fusion stage of step (3) specifically includes: constructing a knowledge graph of photovoltaic cell production process, including equipment parameters, material properties and process standard entities, and triggering dynamic updates of the graph based on process change events; using an LSTM-GAN hybrid model to detect anomalies and enhance data on the fused data, and introducing a transfer learning mechanism to reuse cross-production line features during model training; the output data includes multi-dimensional information on quality scores, equipment health indicators and process optimization suggestions, and is connected to the factory ERP system through a RESTful API. When the scores of three consecutive batches are less than 90 points, process parameter adjustment instructions are automatically triggered. The dynamic update of the process knowledge graph includes: capturing equipment parameter changes through the OPCUA event subscription mechanism; automatically triggering the graph verification process and marking nodes for manual review when a process standard version update is detected; and establishing a graph version history library to support version rollback and difference comparison.

[0077] At the decision-making level, structured results (such as yield statistics) are integrated with unstructured results (such as EL defect classification) to achieve more fault-tolerant quality judgments through weighted voting or confidence matching. Dynamic adjustment of fusion strategies (such as prioritizing high-precision sensor data) is supported to adapt to the needs of different production scenarios. Separate processing of two types of data (such as structured data processed by the MES system and unstructured data analyzed by edge computing nodes) reduces the system's computing load and is suitable for distributed MES architectures. Late fusion achieves efficient decision-making through fault tolerance and resource separation, making it suitable for distributed system architectures.

[0078] The present invention provides a multi-source heterogeneous data processing method that achieves device data standardization through an extensible protocol conversion module, builds a hybrid processing pipeline based on the Flink streaming engine and WarpStream components, ensures state consistency through a Checkpoint mechanism, and adopts a dynamic parallelism adjustment strategy to cope with data flow fluctuations. A hybrid multi-stage data fusion strategy is used to achieve deep mining and efficient processing of production data, thereby reducing the data fusion error rate and the anomaly detection missed rate.

[0079] Example 2

[0080] In summary, the present invention provides an optical multi-source heterogeneous data processing system, including a processor 201, a memory 202, a communication module 303, and a computer program stored in the memory. When the processor 201 executes the computer program, the multi-source heterogeneous data processing method is implemented.

[0081] Those skilled in the art will appreciate that the embodiments of the present application can be provided as methods, systems, or computer program products. Therefore, the present application can adopt the form of a complete hardware embodiment, a complete software embodiment, or an embodiment in combination with software and hardware. Moreover, the present application can adopt the form of a computer program product implemented on one or more computer-usable storage media (including but not limited to magnetic disk storage, CD-ROM, optical storage, etc.) that contain computer-usable program code.

[0082] Those skilled in the art will appreciate that the embodiments of the present application can be provided as methods, systems, or computer program products. Therefore, the present application can adopt the form of a complete hardware embodiment, a complete software embodiment, or an embodiment in combination with software and hardware. Moreover, the present application can adopt the form of a computer program product implemented on one or more computer-usable storage media (including but not limited to magnetic disk storage, CD-ROM, optical storage, etc.) that contain computer-usable program code.

[0083] These computer program instructions may also be stored in a computer readable memory that can direct a computer or other programmable data processing device to operate in a specific manner as specified in the instructions, so that the instructions stored in the computer readable memory produce an article of manufacture including an instruction device, which implements the process Figure 1 a process or multiple processes and / or boxes Figure 1 The function specified in one or more boxes.

[0084] These computer program instructions may also be stored in a computer readable memory that can direct a computer or other programmable data processing device to operate in a specific manner as specified in the instructions, so that the instructions stored in the computer readable memory produce an article of manufacture including an instruction device, which implements the process Figure 1 a process or multiple processes and / or boxes Figure 1 The function specified in one or more boxes.

[0085] It should be noted that the technical features in the above embodiments can be combined in any way, and the technical solutions formed by the combination all fall within the scope of protection of this application. In this article, terms such as "including", "comprising" or any other variants thereof are intended to cover non-exclusive inclusion, so that the process, method, article or device including a series of elements includes not only those elements, but also includes other elements that are not explicitly listed, or also includes elements inherent to such process, method, article or device. In the absence of further restrictions, the elements defined by the sentence "including a..." do not exclude the presence of other identical elements in the process, method, article or device including the elements.

[0086] While embodiments of the present invention have been shown and described, it will be appreciated by those skilled in the art that various changes, modifications, substitutions, and variations may be made to these embodiments without departing from the principles and spirit of the invention, and that the scope of the invention is defined by the appended claims and their equivalents.

Claims

1. A method for processing multi-source heterogeneous data, characterized in that: The following steps are involved: Step 1: At the edge computing layer, the protocol conversion module performs protocol standardization on the multi-source heterogeneous data of photovoltaic cell production equipment, enabling unified access of devices with different communication protocols to the edge gateway. Step 2: Use the Flink stream computing engine to process real-time data in a streaming manner. Integrate the batch processing and stream processing frameworks through the WarpStream component to build a unified data processing process. Step 3: Fusion of structured data and unstructured data in three stages: early, middle, and late: Early fusion stage: Feature extraction and correlation analysis of real-time data through the streaming computing engine; Mid-term fusion stage: The intermediate results of stream processing are stored in a distributed message queue and aligned with the historical batch data in time series; Late fusion stage: Combine knowledge graphs and machine learning models to perform quality assessment and optimized output on the fused data.

2. The method according to claim 1, characterized in that The protocol conversion module described in step 1 supports dynamic adaptation of Modbus, OPCUA, and MQTT protocols, dynamically allocates computing tasks based on the hardware resources of the edge gateway, and enables preemptive transmission channels for high-priority device data.

3. The method according to claim 1, characterized in that The integration of the WarpStream component described in step 2 includes: implementing stream processing state persistence through the Flink Checkpoint mechanism; Use distributed snapshot technology to ensure consistency in batch and stream mixed processing; Enable back pressure control and automatic retry mechanism for abnormal data flow, and dynamically adjust the parallelism of computing nodes based on workload.

4. The method according to claim 1, wherein The early fusion phase of step 3 also includes: Defect feature extraction is performed on unstructured image data using a lightweight CNN model, and TensorRT is used on the edge to accelerate the inference process. Through the time window mechanism, structured sensor data and image feature data are associated with production batches to build a full-chain production view. The window length is dynamically adjusted to 5-30 seconds according to the process rhythm.

5. The method according to claim 1, wherein During the mid-stage fusion phase of step 3: The distributed message queue adopts a hybrid architecture of Kafka and Redis, and stores data in a hierarchical manner according to its timeliness, with a retention period of ≤1 hour for real-time data and ≥30 days for historical data. The timing alignment process introduces a dynamic time warping algorithm to compensate for device clock deviation and calibrates the timestamp offset based on the device heartbeat signal.

6. The method according to claim 1, characterized in that The late fusion phase of step 3 specifically includes: Build a knowledge graph of photovoltaic cell production processes, including equipment parameters, material properties, and process standard entities, and trigger dynamic updates of the graph based on process change events; An LSTM-GAN hybrid model is used to perform anomaly detection and data enhancement on the fused data. A transfer learning mechanism is introduced during model training to reuse cross-production line features. The output data includes multi-dimensional information such as quality scores, equipment health indicators, and process optimization suggestions, and is connected to the factory ERP system through RESTful API. When the scores of three consecutive batches are less than 90 points, process parameter adjustment instructions are automatically triggered.

7. The method according to any one of claims 1 to 6, characterized in that Also includes: A lightweight data preprocessing module is deployed on the edge gateway to perform data denoising, null value filling, and unit normalization operations. Denoising uses a hybrid algorithm of wavelet transform and Kalman filtering; Set a dynamic weight strategy to adjust the weight distribution in the fusion process based on the reliability and timeliness of the data source. Reliability assessment is based on the quantitative calculation of the historical failure rate of the equipment.

8. The method according to claim 3, characterized in that The dynamic adjustment of the parallelism of computing nodes includes: Monitor CPU / GPU resource utilization in real time and automatically scale out compute container instances when utilization remains above 85% for three minutes. Dedicated resource pools are allocated for delay-sensitive data processing tasks, and QoS levels are set to guarantee computing bandwidth.

9. The method according to claim 6, characterized in that The dynamic update of the process knowledge graph includes: Capture device parameter changes through OPC UA event subscription mechanism; When a process standard version update is detected, the graph verification process is automatically triggered and the nodes for manual review are marked; Establish a graph version history library to support version rollback and difference comparison.

10. The method according to claim 7, characterized in that The implementation of the dynamic weight strategy includes: Construct a weight coefficient matrix W = [w1, w2, ..., wn], where wi = α·Ri + β·Ti, wi = α·Ri + β·Ti, Ri is the reliability factor, Ti is the timeliness factor, α+β = 1 and α∈[0.6,0.8]; The weight coefficient is recalculated every 8 hours, and the coefficient mutation is smoothed by the sliding average algorithm; A device health check work order is triggered for data sources with a weight change rate greater than 15% for three consecutive times.

Images