Heterogeneous information fusion system and method in complex industrial scene

Abstract

The invention relates to the technical field of industrial communication and Internet of Things, in particular to a heterogeneous information fusion system and method in a complex industrial scene. Comprising a deterministic transmission module, a distributed identification analysis module and a dynamic load balancing module which work cooperatively, and the three modules cooperate to realize global information real-time communication from an industrial site to an industrial chain application layer. Microsecond-level deterministic transmission of industrial field key data is realized through deep coupling of the OPC UA and the TSN and real-time optimization of an embedded system, experimental results show that in a network environment with background traffic interference, the end-to-end transmission delay of the OPC UA key service data is averagely reduced by 72.3%, time synchronization jitter is controlled within a range of + / -20 ns, and the time synchronization jitter is controlled within a range of + / -20 ns. And the time sequence consistency of the industrial control system is effectively ensured.

Description

Technical Field

[0001] This invention relates to the field of industrial internet and intelligent manufacturing technology, and in particular to a heterogeneous information fusion system and method for complex industrial scenarios. Background Technology

[0002] With the deepening of the intelligent transformation of global manufacturing, the Industrial Internet has become a key infrastructure for improving industrial production efficiency and intelligence. In complex industrial scenarios, achieving deep integration and real-time communication of heterogeneous information among equipment, systems, and supply chain levels is a core prerequisite for supporting advanced applications such as intelligent manufacturing and digital twins. However, existing technologies still have significant shortcomings in dealing with the transmission, identification, and scheduling of multi-source heterogeneous data in industrial settings.

[0003] Firstly, at the industrial field information transmission level, while the Open Platform Unified Communication Architecture (OPC UA) can effectively solve the semantic interoperability problem of heterogeneous systems, its communication architecture based on traditional Ethernet is difficult to meet the stringent requirements for deterministic data transmission in scenarios such as motion control and real-time monitoring. Although Time-Sensitive Networking (TSN) technology provides deterministic transmission capabilities at the data link layer, current research on the integration of OPC UA and TSN is mostly limited to point-to-point direct connection test frameworks, failing to fully consider the collaborative optimization of diverse device types in industrial fields and TSN switch scheduling mechanisms. Existing solutions generally lack efficient mapping mechanisms from OPC UA Pub / Sub messages to TSN frames and have not deeply optimized for the real-time performance of embedded devices, resulting in significant fluctuations in transmission latency of critical business data under background traffic interference, making it difficult to achieve microsecond-level timing guarantees.

[0004] Secondly, at the level of supply chain information traceability, the existing industrial internet identifier resolution system suffers from severe fragmentation. Different identifier systems such as Handle, OID, and Ecode are incompatible with each other, especially lacking support for the identification of virtual entity data. Although traditional centralized resolution systems based on DNS architecture are simple in structure, they have the risk of single point of failure, and their resolution efficiency drops sharply with the increase in request volume. Although some studies have attempted to build decentralized resolution networks using distributed hash table (DHT) technology, most solutions have not been optimized for the spatiotemporal constraints of industrial scenarios, resulting in data retrieval latency between nodes exceeding the expected threshold and long average query path lengths, failing to meet the real-time requirements of low-latency data queries in complex industrial scenarios.

[0005] Finally, at the high-concurrency processing level of identifier resolution services, traditional load balancing algorithms such as round-robin scheduling, weighted round-robin, and minimum connection count mostly use static weights or single performance indicators, making it difficult to adapt to the short-cycle, high-density access characteristics of industrial identifier resolution request traffic. Although existing dynamic load balancing research has introduced predictive algorithms to optimize request scheduling, it lacks sufficient mining of the temporal correlation characteristics of identifier request traffic, and the accuracy of the predictive models is limited. This leads to increased system response latency and decreased throughput under sudden traffic scenarios, seriously affecting the availability and reliability of identifier resolution services.

[0006] Therefore, this application proposes a heterogeneous information fusion system and method for complex industrial scenarios. Summary of the Invention

[0007] To overcome the shortcomings of existing technologies and solve the technical problems existing in the background technology, this invention proposes a heterogeneous information fusion system and method for complex industrial scenarios.

[0008] This invention is achieved through the following technical solution: A heterogeneous information fusion system for complex industrial scenarios includes a deterministic transmission module, a distributed identifier resolution module, and a dynamic load balancing module that work together to achieve real-time information connectivity across the entire industrial field and the application layer of the industrial chain. The deterministic transmission module achieves low-latency deterministic transmission of heterogeneous information in industrial settings by integrating OPC UA with time-sensitive networking. The distributed identifier resolution module constructs an identifier resolution system using a distributed hash table and an improved Kademlia protocol to support unified identifier and efficient resolution of multi-source data. The dynamic load balancing module achieves balanced scheduling of identifier resolution requests based on traffic prediction and the improved Cheetah Optimization Algorithm (OBL-CO).

[0009] Preferably, the deterministic transmission module includes an OPC UA embedded server deployed in an embedded industrial gateway, a mapping mechanism between OPC UA Pub / Sub messages and TSN frames, and a TSN network configuration scheme based on the IEEE 802.1Qcc protocol; The OPC UA embedded server is used for semantic modeling and integration of multi-source heterogeneous data, and the mapping mechanism encapsulates application layer messages into TSN data frames and optimizes transmission timing. The TSN network configuration scheme ensures deterministic transmission of critical service flows through centralized scheduling; Preferably, the OPC UA embedded server includes an OPC UA information model building unit, an embedded deployment unit, and a real-time optimization unit; The information model building unit converts industrial field equipment data into standardized node objects by defining a set of model objects, instantiating models, and mapping models into encoding. The embedded deployment unit is based on the RK3568 processor platform and achieves embedded deployment of the server and client by cross-compiling the Open62541 open source library; The real-time optimization unit improves the Linux kernel by integrating the PREEMPT_RT patch, reducing the average peak latency to within a preset threshold.

[0010] Preferably, the mapping mechanism between OPC UA Pub / Sub messages and TSN frames includes a message encapsulation unit, a priority mapping unit, and a frame structure reconstruction unit; The message encapsulation unit encapsulates the published dataset into a dataset message by the dataset writer, and the writer group aggregates it into a complete network message; the priority mapping unit maps the priority of OPC UA messages to the 8-level queue of TSN, and allocates high-priority queues for high real-time messages. The frame structure reconstruction unit extends TSN features by inserting a 4-byte VLAN tag, where the priority code field carries traffic priority information.

[0011] Preferably, the TSN network configuration scheme includes a time synchronization unit, a traffic shaping unit, and a network configuration unit; the time synchronization unit is based on the IEEE 802.1AS protocol and implements hardware-level timestamp marking through an Intel I210 Gigabit Ethernet controller, so that the time synchronization jitter of the master and slave devices is controlled within ±20ns; the traffic shaping unit is based on the IEEE 802.1Qbv gating scheduling mechanism and implements precise control over the data stream transmission timing through a pre-configured periodic gating list; the network configuration unit adopts a fully centralized configuration model, collects the service characteristics of terminal devices through a centralized user configurator, and generates a global scheduling table by a centralized network configurator.

[0012] Preferably, the distributed identifier resolution module includes a multi-adaptive identifier encoding scheme, a distributed enterprise node network, and a small-world network optimization mechanism; the multi-adaptive identifier encoding scheme integrates the advantages of the Handle system and Ecode encoding to assign globally unique identifiers to physical entities and virtual data; The distributed enterprise node network is constructed based on the improved Kademlia protocol to build a two-level topology of "industry domain - enterprise node", which supports the registration and resolution of identifiers; the small-world network optimization mechanism reconstructs the node topology through a probabilistic connection model to reduce the average query path length.

[0013] Preferably, the multi-adaptive identifier encoding scheme includes a composite identifier object type definition unit, a dual mapping unit, and a metadata extension unit; the composite identifier object type definition unit defines HybridIdentifierType in the OPC UA information model, integrating the dual attributes of Handle and Ecode; the dual mapping unit binds the Handle encoding as the primary key for cross-domain interaction to the national identifier node, and uses the Ecode encoding for fast intra-domain retrieval; the metadata extension unit stores the metadata returned by the identifier resolution system through the definition of UAVariableType, including attributes such as manufacturer and maintenance date.

[0014] Preferably, the distributed enterprise node network includes a node identity generation unit, a routing table initialization unit, and a node dynamic maintenance unit; the node identity generation unit encrypts the node's IP address, port number, and timestamp using a hash algorithm to generate a 160-bit Node ID; the routing table initialization unit establishes an initial connection through a bootstrap node, constructs a dynamic routing table containing 160 K buckets, and maintains information on k active nodes using an LRU replacement strategy; the node dynamic maintenance unit implements dynamic election of the bootstrap node using the RAFT consensus algorithm and handles the joining and leaving transactions of enterprise nodes.

[0015] Preferably, the small-world network optimization mechanism includes a probabilistic permutation model, a network training unit, and a performance degradation detection unit; the probabilistic permutation model calculates the node permutation probability based on a preset formula, as follows: ; Here, a, b, and c represent three different nodes in the network, with the distances between the nodes denoted as |ab| and |ac|, respectively. γ and β are adjustment parameters greater than 0. The larger the value of γ, the higher the probability that node a will replace node b with node c when node c is closer to node a; the larger the value of β, the higher the probability that this replacement will occur when node b is farther from node a. The network training unit collects node information through network-wide broadcasting and performs progressive node replacement operations according to the replacement probability. The performance degradation detection unit is equipped with a performance degradation detection mechanism, which triggers a re-optimization process when the average number of route hops deviates from the baseline value by more than 10%.

[0016] Preferably, the dynamic load balancing module includes a Transformer-based traffic prediction model, an improved Cheetah Optimization Algorithm (OBL-CO), and a load balancing system framework. The traffic prediction model extracts the temporal correlation features of the identified request traffic through a multi-head attention mechanism to predict server load. The OBL-CO algorithm combines Oppositional Learning (OBL) and stochastic parameter nonlinear optimization to dynamically calculate server weights and allocate requests. The load balancing system framework is based on a reverse proxy architecture and integrates data collection, load assessment, traffic scheduling, and cluster management functions.

[0017] Preferably, the Transformer-based traffic prediction model includes an input mapping unit, a time feature extraction unit, and a prediction output unit; the input mapping unit compresses the grid-shaped identifier request data into a list format, and reduces the input dimension from R to R through an embedding layer. T×H×W Extend to R T×N×1 The time feature extraction unit calculates dynamic global temporal correlation through a multi-head attention mechanism, as shown in the following formula: ; The prediction output unit converts features into prediction information through a fully connected layer and outputs the number of server cluster connection requests at the next time step.

[0018] Preferably, the improved Cheetah Optimization Algorithm (OBL-CO) includes an opposition learning optimization unit, a search strategy optimization unit, an attack strategy optimization unit, and a parameter nonlinearity improvement unit; the opposition learning optimization unit enhances the search space exploration capability by generating the opposite solution of the current solution, and the formula for generating the opposite solution is: ; The search strategy optimization unit improves the search strategy based on opposition learning, and the position update formula is: ; The attack strategy optimization unit improves the attack strategy based on adversarial learning; the position update formula is: ; The parameter nonlinear improvement unit performs nonlinear improvement on the random step size and interaction coefficients. The nonlinear improvement of the random step size is expressed by the following formula: ; The nonlinear improvement of the interaction coefficient is achieved through the following formula: ; in, Indicates from A value randomly selected from the range. Represents the boundary values ​​of individual algorithms. Indicates the initial position of an individual. , , They represent the first In the next iteration, the leader, the cheetah Fitness value with neighboring cheetahs. , , They represent the first In the next iteration, the leader, the cheetah The relative positions of the cheetahs and their neighbors.

[0019] Parameters can be dynamically adjusted using preset formulas.

[0020] Preferably, the load balancing system framework includes: The data acquisition and evaluation unit collects server performance metrics, including CPU performance, memory, disk I / O, network bandwidth, and requests per second, through the Linux system's proc file or command line. The load balancing calculation unit calculates the composite load weight based on the server's initial performance evaluation metrics and dynamic performance evaluation metrics, using the following formula: ; in As the initial performance evaluation metric, As an indicator for evaluating residual capacity, To request traffic evaluation metrics.

[0021] The traffic scheduling unit, based on the reverse proxy architecture, performs traffic allocation according to the optimal node information output by the OBL-CO algorithm; The cluster management unit manages the cluster servers and provides an information collection interface to the data acquisition module.

[0022] A method for heterogeneous information fusion in complex industrial scenarios includes the following steps: S1: Collect heterogeneous data from the industrial field through an OPC UA embedded server, construct an OPC UA information model and complete model instantiation and encoding mapping, design a mapping mechanism from OPC UA Pub / Sub messages to TSN frames, and combine IEEE 802.1AS time synchronization and IEEE 802.1Qbv gating scheduling to achieve deterministic data transmission based on the TSN network; S2: Adopt a multi-adaptive identification coding scheme that integrates the advantages of Handle and Ecode to uniformly identify data, register the identification to a distributed node network built on the improved Kademlia protocol, iteratively query identification information based on the Kademlia protocol, and apply the small-world theory to optimize the network topology to achieve efficient identification query; S3: Collect recursive server performance metrics, calculate initial and dynamic performance evaluation metrics for the server, predict identifier request traffic based on the Transformer model, input the prediction results into the improved Cheetah Optimization Algorithm (OBL-CO) to calculate the server dynamic weights, and dynamically schedule requests to the recursive parsing server according to the weights to achieve load balancing.

[0023] Preferably, in S1, the OPC UA information model realizes the digital expression of equipment status, process and control information by defining a recursive nested structure of component sets, components, attribute sets and attribute elements; the model instantiation injects the physical parameters of the industrial site into the predefined model framework, and realizes parameter reuse through inheritance; the mapping of OPC UA Pub / Sub messages to TSN frames distinguishes traffic priorities through the PCP field of the VLAN tag.

[0024] Preferably, in S2, the prefix of the multi-adaptive identifier encoding follows a management structure of "top-level organization / sub-organization", and the suffix is ​​defined by the local naming authority. When registering the identifier, the enterprise node submits authentication information to the industry secondary node through the guiding node, and obtains the identifier prefix after verification. The iterative query based on the Kademlia protocol establishes a logical space mapping relationship through the XOR distance calculation rule of the 160-bit node ID.

[0025] Preferably, the S3 body comprises: Collect recursive server performance metrics, including CPU utilization ( ), memory utilization ( ), disk I / O utilization ) and network bandwidth utilization ( ); Calculate the initial performance evaluation metrics of the server: ; Calculate dynamic performance evaluation metrics for servers: ; The Transformer model is used to predict the number of request connections in the next time step, and the OBL-CO algorithm is used to calculate the dynamic weight of the server. Requests are allocated based on weights to minimize response time and maximize load balancing.

[0026] According to the method of claim 18, the execution of the OBL-CO algorithm includes: Initialize the cheetah population by randomly generating the initial positions of N cheetahs in D-dimensional space; For each individual, an opposite solution is generated, and the opposite solution is calculated using the following formula; ; Choose a search, wait, or attack strategy based on random variables; Apply a nonlinear improved parameter update strategy to non-opposing individuals; Compare the fitness of the current solution with that of the complementary solution, and retain the better solution; Repeat the iteration until the termination condition is met, and output the optimal server allocation scheme.

[0027] The fitness function is defined as follows: ; in Total runtime latency, For server load balancing. , which is a weighting factor used to measure the priority of the two optimization objectives, and .

[0028] The beneficial effects of this invention are: 1. This invention achieves microsecond-level deterministic transmission of critical data in industrial settings through deep coupling of OPC UA and TSN, and real-time optimization of embedded systems. Experimental results show that in network environments with background traffic interference, the end-to-end transmission latency of critical OPC UA business data is reduced by an average of 72.3%, and the time synchronization jitter is controlled within ±20 ns, effectively ensuring the timing consistency of industrial control systems.

[0029] 2. This invention solves the interoperability problem of the identification system through a multi-adaptive coding scheme. The distributed parsing network optimized based on the small-world theory reduces the average broadcast hops of information queries by 16.7%-25.8% while maintaining the high reliability of the decentralized architecture, significantly improving the retrieval speed of massive identification data and the overall scalability of the system.

[0030] 3. This invention combines Transformer traffic prediction with an improved OBL-CO algorithm for dynamic load balancing. It can accurately capture the temporal patterns of request traffic and quickly converge to the optimal scheduling scheme. Compared with traditional algorithms such as minimum connections and IRLB, under 4500 concurrent requests, the average response time is reduced by 8.7%-13.6%, the load balancing degree is improved by 2.1%-7.9%, and the system throughput is improved by 29.8%-113.4%, which greatly enhances the processing capacity and stability of the identifier resolution service in high-concurrency scenarios.

[0031] 4. This invention organically integrates three key technologies: deterministic transmission, distributed identifier resolution, and dynamic load balancing, and constructs a collaborative information fusion framework covering "cloud-edge-device". It completely breaks down the information barriers between industrial sites and industrial chain levels, and provides a solid technical foundation for intelligent collaborative optimization and full life cycle management in complex industrial scenarios. Attached Figure Description

[0032] Figure 1 A schematic diagram of the overall architecture of a heterogeneous information fusion system in a complex industrial setting; Figure 2 : A schematic diagram of the protocol stack mapping of the OPC UA and TSN converged network model; Figure 3 Architecture diagram of a distributed identifier resolution system; Figure 4 Functional block diagram of a load balancing system; Figure 5 : Schematic diagram of the Transformer-based traffic prediction algorithm framework. Detailed Implementation

[0033] The present invention will be further illustrated below with reference to specific embodiments. It should be understood that these embodiments are for illustrative purposes only and are not intended to limit the scope of the invention. Experimental methods in the following embodiments, unless otherwise specified, are generally performed under conventional conditions or as recommended by the manufacturer.

[0034] Unless otherwise defined, all technical and scientific terms used herein have the same meaning as commonly understood by one of skill in the art. The reagents and raw materials used in this invention are readily available through conventional means, and unless otherwise specified, they shall be used in accordance with conventional methods in the art or as per the product instructions. Furthermore, any methods and materials similar to or equivalent to those described herein may be applied to the methods of this invention. The invention will now be further described with reference to the accompanying drawings and specific embodiments. The preferred embodiments and materials described herein are for illustrative purposes only.

[0035] (1) System composition and module functions Deterministic Transmission Module: This module is the foundation for the system to achieve reliable acquisition and real-time transmission of industrial field data. Its core lies in the deep integration of OPC UA and TSN technologies.

[0036] The OPC UA embedded server utilizes the open62541 open-source library and is deployed on an industrial gateway based on the RK3568 processor. This server is responsible for semantic modeling of multi-source heterogeneous data (such as loom speed, spindle vibration, and fabric tension) from industrial sites like textile production lines. It constructs a unified information model that includes basic equipment attributes, dynamic production process data, control component feature data, and production line configuration metadata, and supports a publish / subscribe (Pub / Sub) communication mechanism.

[0037] The TSN Network Configuration and Scheduling Unit: Based on the IEEE 802.1Qcc standard, it adopts a fully centralized network configuration model (CUC+CNC). The core functions of this unit include: Time synchronization: Based on the IEEE 802.1AS protocol, and utilizing the hardware timestamp function of the Intel I210 network card, nanosecond-level clock synchronization between master and slave devices is achieved on a Linux kernel equipped with the PREEMPT_RT real-time patch. In the experiment, the synchronization deviation was controlled within ±100 ns.

[0038] Traffic scheduling and shaping: Based on the IEEE 802.1Qbv gating mechanism, a periodic gating list (GCL) is designed. For example, the 500-microsecond transmission period is divided into multiple time slots. Exclusive transmission windows (e.g., 0-50 microseconds) are allocated to high-priority OPC UA critical service data (such as control commands, PCP=7), while the remaining time slots (e.g., 300-500 microseconds) are allocated to low-priority background traffic (BE flow, PCP=0), thereby achieving traffic isolation and deterministic latency guarantees.

[0039] Message mapping mechanism: Design mapping rules from OPC UA Pub / Sub messages to TSN data frames, mapping parameters such as publication period and priority in OPC UA messages to the PCP field in the VLAN tag of the TSN frame, so that the TSN switch can schedule according to priority.

[0040] (2) Distributed identifier resolution module This module provides globally unique identifier management and efficient resolution services for industrial entities and virtual data. Its innovation lies in the optimization of the encoding scheme and network structure.

[0041] Multi-adaptive Identifier Encoding Scheme: This scheme integrates the advantages of the Handle system and the Ecode encoding system to design a composite encoding structure. For example, the prefix uses Handle's "top-level organization / sub-organization" format to ensure global uniqueness, while the suffix uses Ecode's version number, NSI, and primary key structure to enhance readability and compatibility within the domain. This encoding is bound to the real-time data model through the HybridIdentifierType object type in the OPC UA information model.

[0042] A distributed network based on the improved Kademlia protocol: A two-tier resolution architecture of "industry domain - enterprise node" is constructed. The enterprise node layer builds a distributed hash table (DHT) network based on the Kademlia protocol, with each node maintaining a routing table containing 160 K buckets. Identification data is distributed and stored on the k logically nearest nodes in the network according to its hash value.

[0043] Small-world network optimization: To address the issue of long query paths in traditional Kad networks, Kleinberg's small-world network theory is introduced to optimize the node topology. By dynamically adjusting the node routing table using the following probabilistic permutation formula and adding a small number of remote connections, the average query path length is significantly shortened. Experiments show that the optimized network reduces the average broadcast hops by 16.7% to 25.8% across different network sizes (128-2048 nodes).

[0044] ; Dynamic load balancing module: This module is responsible for handling high-concurrency requests to the identifier resolution service and achieving balanced distribution of server cluster load through prediction and optimization.

[0045] Transformer traffic prediction model: This model uses the request traffic matrix of a recursive server cluster over a historical time period. As input, the data dimensions are transformed and embedded into a high-dimensional space through the input mapping module. Then, the temporal correlation features of request traffic are extracted using the following multi-head self-attention mechanism formula, and finally, the predicted number of request connections for each server at the next time step is output. The model training uses 40% Dropout and the Adam optimizer. Experiments show that its predicted MAE is 19.595 and RMSE is 37.442.

[0046] ; Improved Cheetah Optimization Algorithm (OBL-CO): The load balancing problem is modeled as an optimization problem with the objectives of minimizing total response time and maximizing load balancing, and the fitness function is: ; To address the premature convergence problem of the basic Co algorithm, this invention incorporates an Opposites Learning (OBL) strategy to generate alternative solutions, thereby expanding the search range, and adjusts the search step size. and interaction coefficient Nonlinear improvements are made to enhance convergence speed and global optimization capabilities.

[0047] Load assessment unit: Real-time collection of performance indicators such as CPU utilization, memory usage, disk I / O and network bandwidth of server nodes, combined with initial performance weights and predicted number of request connections, dynamically calculates the comprehensive load weight of each server using the following formula, providing a decision basis for the OBL-CO algorithm.

[0048] ; (3) System workflow Data Acquisition and Identification: The OPC UA embedded server on the industrial gateway collects data from field devices and builds an information model. Subsequently, it assigns global identifiers conforming to a multi-adaptive coding scheme to data objects and registers them with a distributed identifier resolution system.

[0049] Deterministic transmission: OPC UA Pub / Sub data is mapped to different priority queues in the TSN network based on service criticality. The TSN switch schedules data according to a pre-configured gating list to ensure low-latency, deterministic transmission of critical data (such as ST streams) even in the presence of background traffic (BE streams).

[0050] Request Scheduling and Resolution: When an application layer (such as a digital twin platform) initiates an identifier resolution request, the load balancing module starts. First, the Transformer model predicts future traffic; then, the OBL-CO algorithm calculates the optimal request allocation scheme based on the prediction results and the real-time server load status; finally, the request is scheduled to the selected recursive resolution server.

[0051] Information traceability and fusion: The recursive server queries the target identifier in the distributed identifier resolution network, obtains the associated metadata (such as equipment information, production batch, maintenance records), and returns the results to the application layer, which is then integrated with the OPC UA real-time data stream to support full lifecycle management and intelligent decision-making.

[0052] (4) Overall system implementation like Figure 1 As shown, the implementation of this invention system covers the complete information chain from the industrial field equipment layer to the industrial chain application layer. In a typical implementation case, we take the intelligent upgrade of a textile production line as the application scenario. The system first collects the operating data of the Picanol OMNIplus 800 air-jet loom through an RK3568 industrial gateway (CPU with quad-core A55 architecture, main frequency 2.0GHz, memory 2GB) deployed in the workshop, including spindle speed, fabric tension, warp yarn balance, and equipment start / stop status. The gateway encapsulates the raw data from the loom PLC into a semantic information model conforming to the OPCUA standard through a Modbus to OPC UA protocol conversion module.

[0053] (5) Implementation of the deterministic transmission module like Figure 2 As shown, the implementation of this module is crucial to ensuring data real-time performance.

[0054] Building an OPC UA embedded server: A cross-compilation environment was set up on the RK3568 gateway. The host machine used Ubuntu 20.04 LTS operating system and installed the aarch64-linux-gnu-gcc 9.4.0 cross-compilation toolchain. The v1.3.8 version source code of the Open62541 open-source library was obtained from the official repository, and single-file compilation mode (-DUA_ENABLE_AMALGAMATION=ON) and MBEDTLS encryption support were enabled during configuration. The compiled libopen62541.a static library (approximately 1.7MB) and header files were transferred to the / usr / local / lib path of the gateway via SCP. The UaModeler tool was used to build a textile production line information model, defining the LoomMachineType object type, including variable nodes such as WeavingSpeed ​​(vehicle speed) and WarpTension (warp tension), and exporting it as an XML file. The server program loaded this XML file, called the UA_Server_addObjectNode() function to dynamically create the address space, and started the service through the UA_Server_run() function, with the subscription period set to 1ms.

[0055] TSN network configuration and optimization: The network hardware uses a Sennheiser HP6E series TSN switch, with Intel I210 network interface cards (NICs) for the terminals. On the RK3568 gateway, the PREEMPT_RT real-time patch is applied to the Linux kernel (version 5.10), and the kernel option CONFIG_PREEMPT_RT_FULL=y is configured. Time synchronization is achieved using the linuxPTP package. The ptp4l process runs on the gateway and is configured as the master clock. The path delay between the master and slave clocks is precisely measured and compensated using the Pdelay_Req / Pdelay_Resp message group to ensure that the master-slave clock deviation is stable within ±100 nanoseconds.

[0056] Traffic scheduling employs centralized configuration. The CUC (Central User Configurator) collects service flow characteristics from the OPC UA publisher, including: ST traffic 1: period 20ms, frame length 1500 bytes, priority 7 (corresponding to PCP 6).

[0057] ST traffic 2: period 50ms, frame length 1500 bytes, priority 5 (corresponding to PCP 3).

[0058] BE background traffic: priority 0.

[0059] Based on these requirements, the CNC (Central Network Configurator) calculates and distributes the Gated List (GCL) to the TSN switch. The GCL is divided in a 500-microsecond time interval. Queue 1 (corresponding to ST traffic 1): time slot 0-50 microseconds, gate status is "open".

[0060] Queues 2 and 3 (corresponding to ST traffic 2): time slots 50-300 microseconds, gate status is "open".

[0061] Queue 4 (corresponding to BE traffic): time slot 300-500 microseconds, gate status is "open".

[0062] With this scheduling, the transmission time of a single frame for ST traffic 1 and 2 is approximately 12 microseconds at a bandwidth of 1Gbps. The allocated time slot width is sufficient to complete the transmission, thus avoiding interference from non-critical traffic. Experiments show that under this configuration, the average end-to-end transmission latency of ST traffic 1 is 24.88 microseconds, a reduction of 72.3% compared to 123.28 microseconds in a traditional Ethernet environment.

[0063] (6) Implementation of the distributed identifier resolution module like Figure 3 As shown, this module enables unified management and efficient parsing of identifiers.

[0064] Implementation of multi-adaptive identifier coding: The following is an example of a unique identifier assigned to a weaving machine: Prefix (following the Handle system): 86.G17.33.01.05.TX2023 86: China's country code G17: Textile Industry Classification Code 33.01.05: Administrative division code of Xiaoshan District, Hangzhou City, Zhejiang Province TX2023: Unique Code for Textile Enterprises Suffix (following Ecode structure): V1.A1.348.1.4.0200022.OMNIplus800_001 V1: Version number A1: Coding System Identifier (GS1) 348: Code of the Ministry of Textile Industry 1: Equipment type (loom) 4: Management permissions (workshop level) 0200022: Equipment Subcategory (Air-jet Looms) OMNIplus800_001: Device Serial Number In the OPC UA server, by defining a custom object type called HybridIdentifierType, the Handle encoding and Ecode encoding mentioned above are stored in two variable nodes, Loom_Handle and Loom_Ecode, respectively, and a reference relationship is established with the Loom_Machine_001 object node representing the physical loom.

[0065] Building a distributed enterprise node network: Four Alibaba Cloud Elastic Compute Service (ECS) instances are used to simulate enterprise nodes, with configurations shown in Table 4.3 (e.g., the Beijing region instance is ecs.r6e.xlarge, 2-core CPU, 4GB memory, CentOS 7 operating system). A P2P network based on the Kademlia protocol is built using the Golang language and the libp2p-go dependency library. Upon startup, each node generates a 160-bit Node ID using the SHA-1 hash algorithm based on its IP address, port number, and timestamp. Each node maintains 160 K-buckets, with the K-bucket capacity k value set according to the network size; in a 128-node network, it is set to 4. When a node joins the network, it interacts with the pre-defined bootstrap node and gradually builds its routing table through FIND_NODE requests.

[0066] Implementation of small-world network optimization: After network initialization is complete, the optimization process begins. The optimization algorithm is also implemented in Golang. A probability permutation formula is defined. The parameters are: probability dimension γ=2, and the scaling factor β=3 for route hop count and XOR distance. The optimization process includes: Network-wide broadcast: Initiate a flood broadcast to collect information from all nodes.

[0067] Probability Calculation and Replacement: For each node a, iterate through nodes b in its K buckets and randomly selects node c in the network, calculating the probability P of replacing b with c. If the random number is less than P, perform the replacement operation and store the information of node c in the expanded storage space.

[0068] Iterative convergence: Repeat the above process 50 times until the average number of broadcast hops in the network tends to stabilize.

[0069] After optimization, a test was conducted on a 128-node network, and the average number of broadcast hops decreased from about 5.2 hops before optimization to about 4.2 hops, resulting in a performance improvement of about 19.2%.

[0070] (7) Implementation of dynamic load balancing module like Figure 4 and Figure 5As shown, this module is the core for handling high-concurrency requests.

[0071] Training and deployment of the Transformer traffic prediction model: Data Preparation: 3240 samples of identifier resolution requests (108 hours, 2-minute intervals) were extracted from the Nginx logs of an actual textile production line project. These samples were then normalized to their maximum and minimum values. Using the request sequence of the previous 30 time steps (i.e., the past hour) as input, the number of requests in the next time step was predicted. The training and test sets were divided in a 7:3 ratio.

[0072] Model building and training: Implement the Transformer model using Python's PyTorch framework. Key parameters are set as follows: Input dimension: (30, N, 1), where N is the number of servers, which is 10 in this embodiment.

[0073] Embedding layer dimension (D): 32 Multi-head attention count (I): 2 Feedforward network hidden layer dimension: 64 Dropout rate: 0.4 Optimizer: Adam, initial learning rate 0.01, which decreases to 0.001 and 0.0001 at epochs 125 and 187, respectively.

[0074] Training cycles: 250 Batch size: 16 After training, the model is deployed to a load balancer server to receive request traffic sequences in real time and output predictions of the number of request connections to each server at future times. .

[0075] Improved implementation of the Cheetah Optimization Algorithm (OBL-CO): The load balancing problem is modeled as finding the optimal request allocation matrix K to minimize the fitness function. , where η is set to 0.5.

[0076] The OBL-CO algorithm is implemented using MATLAB, and its core steps are as follows: Initialization: Set the cheetah population size to 300 and the maximum number of iterations to 200. Randomly initialize the cheetah positions, with each position representing a possible request allocation scheme.

[0077] Opposites Learning (OBL): For each individual in the current population According to the formula Generate its opposite individual.

[0078] Strategy Selection and Updates: Search strategy: For the current individual and its counterpart, an improved search strategy is adopted, with the step size update formula as follows: ; in, Indicates from A value randomly selected from the range. Represents the boundary values ​​of individual algorithms. Indicates the initial position of an individual. , , They represent the first The fitness values ​​of the leader, cheetah, and neighboring cheetahs at the next iteration. , , They represent the first The relative positions of the leader, cheetah, and neighboring cheetahs in the next iteration.

[0079] Attack Strategy: An improved attack strategy is adopted, and its interaction coefficients are updated as follows: ; Waiting strategy: follow the formula Keep the position unchanged.

[0080] Elite selection: Compare the fitness of current individuals with that of their counterparts and retain the better ones.

[0081] Return strategy: If the optimal solution does not improve after multiple iterations, then follow the formula. Reset some individuals to be near the current optimal solution.

[0082] In a scenario where 1000 parsing requests are distributed across 10 servers, this algorithm typically converges to a satisfactory solution within 80 generations.

[0083] Load assessment and scheduling: The load balancer (using an Nginx reverse proxy server) periodically collects performance metrics from 10 recursive DNS servers via a custom script: CPU utilization (%), memory usage (%), disk I / O rate (MB / s), and network bandwidth utilization (Mbps). The initial performance weight for each server is then calculated using a formula. and dynamic weights The scaling factor is set to... Combined with the predictions from the Transformer Based on the optimal allocation scheme calculated by OBL-CO, Nginx dynamically distributes newly arriving identifier resolution requests to each recursive server through the upstream module and weight configuration.

[0084] (8) System Workflow Example Data Acquisition and Identifier Registration: The OPC UA server on the RK3568 gateway continuously reads the rotational speed data (e.g., 650 RPM) of the loom OMNIplus800_001. Simultaneously, this data point is bound to the identifier 86.G17.33.01.05.TX2023 / V1.A1.348.1.4.0200022.OMNIplus800_001.WeavingSpeed, and registration is completed by calling the RESTful API provided by the identifier resolution system.

[0085] Deterministic transmission: The rotation speed data, as critical process data, is encapsulated by the OPC UA Pub / Sub mechanism and mapped to the PCP 3 field of the TSN frame according to its priority (5). The TSN switch forwards it to subscribers (such as edge computing nodes) without conflict within a time slot of 50-300 microseconds according to the GCL.

[0086] Resolution Request and Load Balancing: The digital twin platform needs to obtain the historical maintenance records of the loom, so it sends a resolution request for the loom's identifier to the recursive node. After receiving this request, the load balancer initiates the traffic prediction and OBL-CO optimization process described above, and finally schedules the request to the Shanghai region ECS server (IP: 192.168.1.103) with the highest current overall load weight for processing.

[0087] Information Traceability and Fusion: After receiving a request, the Shanghai server queries the identifier in its local DHT storage. If no match is found, the Kademlia iterative query algorithm is used. After an average of about 4.2 hops, the metadata associated with the identifier (including manufacturer, last maintenance date, etc.) is found on the Beijing node. This metadata is returned to the digital twin platform and fused with real-time rotation speed data to jointly drive the operation and fault prediction analysis of the virtual loom model.

[0088] The above specific implementation methods fully demonstrate that the technical solution of the present invention has clear feasibility and can effectively solve the problem of heterogeneous information fusion in complex industrial scenarios.

[0089] Any aspects not covered in this invention are applicable to existing technologies.

Claims

1. A heterogeneous information fusion system for complex industrial scenarios, characterized in that, It includes a deterministic transmission module, a distributed identifier resolution module, and a dynamic load balancing module that work together to achieve real-time information connectivity across the entire industrial field and the application layer of the industrial chain. The deterministic transmission module achieves low-latency deterministic transmission of heterogeneous information in industrial settings by integrating OPC UA with time-sensitive networking. The distributed identifier resolution module constructs an identifier resolution system using a distributed hash table and an improved Kademlia protocol to support unified identifier and efficient resolution of multi-source data. The dynamic load balancing module achieves balanced scheduling of identifier resolution requests based on traffic prediction and the improved Cheetah Optimization Algorithm (OBL-CO).

2. The heterogeneous information fusion system for complex industrial scenarios according to claim 1, characterized in that, The deterministic transmission module includes an OPC UA embedded server deployed in an embedded industrial gateway, a mapping mechanism between OPC UA Pub / Sub messages and TSN frames, and a TSN network configuration scheme based on the IEEE 802.1Qcc protocol. The OPC UA embedded server is used for semantic modeling and integration of multi-source heterogeneous data, and the mapping mechanism encapsulates application layer messages into TSN data frames and optimizes transmission timing. The TSN network configuration scheme ensures deterministic transmission of critical service flows through centralized scheduling.

3. The heterogeneous information fusion system for complex industrial scenarios according to claim 2, characterized in that, The OPC UA embedded server includes an OPC UA information model building unit, an embedded deployment unit, and a real-time optimization unit; The information model building unit converts industrial field equipment data into standardized node objects by defining a set of model objects, instantiating models, and mapping models into encoding. The embedded deployment unit is based on the RK3568 processor platform and achieves embedded deployment of the server and client by cross-compiling the Open62541 open source library; The real-time optimization unit improves the Linux kernel by integrating the PREEMPT_RT patch, reducing the average peak latency to within a preset threshold.

4. The heterogeneous information fusion system for complex industrial scenarios according to claim 2, characterized in that, The mapping mechanism between OPC UA Pub / Sub messages and TSN frames includes a message encapsulation unit, a priority mapping unit, and a frame structure reconstruction unit. The message encapsulation unit encapsulates the published dataset into a dataset message by the dataset writer, and the writer group aggregates it into a complete network message. The priority mapping unit maps the priority of OPC UA messages to the 8-level queues of TSN, and allocates high-priority queues for high-real-time messages; The frame structure reconstruction unit extends TSN features by inserting a 4-byte VLAN tag, where the priority code field carries traffic priority information.

5. A heterogeneous information fusion system for complex industrial scenarios according to claim 2, characterized in that, The TSN network configuration scheme includes a time synchronization unit, a traffic shaping unit, and a network configuration unit; The time synchronization unit is based on the IEEE 802.1AS protocol and implements hardware-level timestamp marking through the Intel I210 Gigabit Ethernet controller, so that the time synchronization jitter of the master and slave devices is controlled within ±20ns. The flow shaping unit is based on the IEEE 802.1Qbv gating scheduling mechanism, which uses a pre-configured periodic gating list to precisely control the timing of data stream transmission. The network configuration unit adopts a fully centralized configuration model, which collects the service characteristics of terminal devices through a centralized user configurator and generates a global scheduling table through a centralized network configurator.

6. The heterogeneous information fusion system for complex industrial scenarios according to claim 1, characterized in that, The distributed identifier resolution module includes a multi-adaptive identifier encoding scheme, a distributed enterprise node network, and a small-world network optimization mechanism; the multi-adaptive identifier encoding scheme integrates the advantages of the Handle system and Ecode encoding to assign globally unique identifiers to physical entities and virtual data; The distributed enterprise node network is constructed based on the improved Kademlia protocol to build a two-level topology of "industry domain - enterprise node", which supports the registration and resolution of identifiers; The small-world network optimization mechanism reconstructs the node topology through a probabilistic connectivity model, thereby reducing the average query path length.

7. A method applicable to heterogeneous information fusion systems in complex industrial scenarios according to any one of claims 1-6, characterized in that, Includes the following steps: S1: Collect heterogeneous data from industrial sites through an OPC UA embedded server, construct an OPC UA information model and complete model instantiation and encoding mapping, design a mapping mechanism from OPC UA Pub / Sub messages to TSN frames, combine IEEE 802.1AS time synchronization and IEEE 802.1Qbv gating scheduling, and realize deterministic data transmission based on the TSN network; S2: Adopting a multi-adaptive identifier encoding scheme that integrates the advantages of Handle and Ecode to uniformly identify data, registering the identifiers to a distributed node network built on the improved Kademlia protocol, iteratively querying identifier information based on the Kademlia protocol, and applying the small-world theory to optimize the network topology to achieve efficient identifier querying; S3: Collect recursive server performance metrics, calculate initial and dynamic performance evaluation metrics for the server, predict identifier request traffic based on the Transformer model, input the prediction results into the improved Cheetah optimization algorithm to calculate the server dynamic weight, and dynamically schedule requests to the recursive parsing server according to the weight to achieve load balancing.

8. The method for heterogeneous information fusion system in complex industrial scenarios according to claim 7, characterized in that, In S1, the OPC UA information model realizes the digital expression of device status, process and control information by defining a recursive nested structure of component sets, components, attribute sets and attribute elements; The model instantiation injects the physical parameters of the industrial site into a predefined model framework and achieves parameter reuse through inheritance; the mapping of OPC UA Pub / Sub messages to TSN frames distinguishes traffic priorities through the PCP field of the VLAN tag.

9. The method for heterogeneous information fusion system in complex industrial scenarios according to claim 8, characterized in that, In S2, the prefix of the multi-adaptive identifier encoding follows the management structure of "top-level organization / sub-organization", and the suffix is ​​defined by the local naming authority. When registering the identifier, the enterprise node submits authentication information to the industry secondary node through the guiding node, and obtains the identifier prefix after verification. The iterative query based on the Kademlia protocol establishes a logical space mapping relationship through the XOR distance calculation rule of the 160-bit node ID.

10. The method for heterogeneous information fusion system in complex industrial scenarios according to claim 9, characterized in that, In S3, the recursive server performance metrics include CPU utilization, memory utilization, disk I / O utilization, and network bandwidth utilization; the initial server performance evaluation metrics and dynamic performance evaluation metrics are calculated using preset formulas; the OBL-CO algorithm outputs the optimal server allocation scheme by initializing the cheetah population, generating complementary solutions, selecting strategies, updating parameters, retaining optimal solutions, and iteratively converging.

Images