An IT-OT fusion-based full-chain open-source intelligent manufacturing test bed system construction method

Abstract

The application discloses an IT-OT fusion-based full-chain open-source intelligent manufacturing test bed system construction method, relates to the technical field of intelligent manufacturing, and comprises the following steps: constructing a multi-site manufacturing production line system, obtaining original sensing data by acquiring multi-dimensional signals in the equipment operation process through a sensor group on each site; deploying a cloud-edge-end collaborative system architecture, reading the original sensing data through the edge layer, pushing the data to the cloud after analysis and conversion, and constructing a data resource pool in the cloud; constructing an intelligent analysis model after processing the data in the resource pool to obtain a manufacturing process dataset, identifying the equipment operation state to generate an optimized control instruction, and then executing the instruction; constructing a fusion computing model of each site based on the dataset, correcting and predicting the state of each fusion computing model through execution result data, associating and coupling each fusion computing model, and forming a production line-level digital twin. The method has good expansibility and system reuse capability.

Description

Technical Field

[0001] This invention relates to the field of intelligent manufacturing technology, specifically to a method for constructing a full-chain open-source intelligent manufacturing testbed system based on IT-OT convergence. Background Technology

[0002] With the deep integration of new-generation information technology and manufacturing, intelligent manufacturing has become an important development direction for promoting the transformation and upgrading of the manufacturing industry. Enterprises and research institutions are constantly exploring system architectures and application models for manufacturing sites, focusing on key technologies such as the Industrial Internet, industrial big data, artificial intelligence, and cyber-physical systems, and are gradually carrying out research and development and application practices of related technologies.

[0003] In the research and development, system verification, and engineering promotion of key technologies in intelligent manufacturing, it is usually necessary to build a test platform that can simulate the real manufacturing environment to achieve technology verification, system integration testing, and application effect evaluation. However, most existing intelligent manufacturing experimental platforms or test systems are built around single equipment, single processes, or partial automation units. For example, they conduct experimental verification for independent application scenarios such as equipment condition monitoring systems, flexible assembly units, or automated logistics systems, lacking a system-level test environment that can cover the entire manufacturing process. Although some test platforms have achieved some automation functions, they still have shortcomings overall. First, existing test platforms lack a full-process manufacturing production line, making it difficult to conduct system-level technology verification for the entire manufacturing process under real processing conditions. Second, independent barriers exist between the information technology (IT) layer and the equipment control (OT) layer, and there is a lack of a unified data interaction mechanism and collaborative operation mode between systems, making it difficult to achieve data connectivity and collaborative control between the equipment layer, control layer, and management layer.

[0004] In addition, existing testing platforms typically rely on proprietary equipment and commercial software for construction, resulting in a relatively closed system architecture, inconsistent interface standards, and insufficient scalability and reusability. This hinders the access of multiple types of devices, multi-level platform collaboration, and verification across multiple application scenarios, and makes it difficult to support continuous iteration and upgrades for scientific research, teaching, and industry-academia-research collaborative innovation. Summary of the Invention

[0005] In order to solve the above-mentioned technical problems, this application proposes the following technical solution: This application provides a method for constructing a full-chain open-source intelligent manufacturing testbed system based on IT-OT convergence, including: Construct a multi-site hardware manufacturing production line system, and obtain raw sensing data by acquiring multi-dimensional physical signals during equipment operation through sensor groups deployed at each site; Based on the hardware manufacturing production line system, a cloud-edge-device collaborative system architecture is deployed. The original sensor data of the terminal layer device is read through the edge layer, parsed and converted into a format and then pushed to the cloud. A manufacturing process data resource pool is built in the cloud. The data in the manufacturing process data resource pool is processed to obtain a structured manufacturing process dataset; An artificial intelligence analysis model is built based on the structured manufacturing process dataset to identify and predict equipment operating status and generate optimized control commands. The optimized control command is sent to the equipment control system to execute the corresponding control action, and the execution result data is fed back to the manufacturing process data resource pool. Based on the structured manufacturing process dataset, a fusion computing model is constructed for each site. A real-time mapping relationship between each fusion computing model and physical equipment is established. After correcting and predicting the status of each fusion computing model through the execution result data, the fusion computing models are associated and coupled to form a production line-level digital twin.

[0006] In one possible implementation, a multi-site hardware manufacturing production line system is constructed. Sensor arrays deployed at each site acquire multi-dimensional physical signals during equipment operation, obtaining raw sensor data, including: Construct a multi-site hardware manufacturing production line system that includes CNC machining stations, additive manufacturing stations, quality inspection stations, assembly stations, packaging stations, and warehousing and logistics stations; Sensor arrays are deployed at various sites to collect multi-dimensional physical signals during equipment operation in real time. The multi-dimensional physical signals collected at the same site are then aligned by timestamps and stitched together to form a unified equipment state vector. The unified device state vector of each station is used as the original sensor data of each station.

[0007] In one possible implementation, a cloud-edge-device collaborative system architecture is deployed based on the aforementioned hardware manufacturing production line system. The edge layer reads the raw sensor data from the terminal layer devices, parses and converts the data, and then pushes it to the cloud. A manufacturing process data resource pool is then built in the cloud, including: Based on the aforementioned hardware production line system, a cloud-edge-device three-layer collaborative system architecture is deployed. The terminal layer includes VPLCs, execution devices, logistics and warehousing equipment, and quality inspection equipment at each site. The edge layer reads the device operation status data, process parameter data, and production management data of the terminal layer device through the Modbus protocol and OPC UA protocol; The edge layer deploys an industrial IoT platform and a SCADA system. It connects to serial devices via serial gateways and RS-485 / RS-232, and integrates OPC UA and Modbus-TCP through the SCADA system to achieve unified access and data acquisition for multi-source heterogeneous devices. Edge nodes perform protocol parsing, format conversion, and timestamp alignment on the collected device operation status data, and perform lightweight mapping on high-frequency signals in the device operation status data to generate standardized data; The standardized data is pushed to the cloud via MQTT / HTTP / RESTful protocols; The cloud platform employs a hybrid storage architecture combining time-series databases, relational databases, and object storage to construct a data resource pool for the manufacturing process.

[0008] In one possible implementation, the edge node performs protocol parsing, format conversion, and timestamp alignment on the collected device operating status data, and applies a lightweight mapping formula to the high-frequency signals in the device operating status data: in, For downsampling operation, These represent the mean, variance, and peak value within the sliding window, respectively. For vector concatenation, for The time-series signal vector of the device's high-frequency operating status obtained from the original acquisition at the edge node at any given time. for The fused feature vector is output after the edge lightweight mapping process at each time step.

[0009] In one possible implementation, the data in the manufacturing process data resource pool is processed to obtain a structured manufacturing process dataset, including: Null value detection and outlier removal are performed on each data stream in the manufacturing process data resource pool. Low-rank matrix factorization is used to impute discrete missing data to complete data cleaning. The formula is as follows: in, The set of locations of the observed data. , It is a low-rank matrix factorization factor. The regularization coefficient is . This represents the optimal imputation value for the discrete missing data after iterative optimization using low-rank matrix factorization. For the first in the manufacturing process data resource pool The equipment sensor, raw production condition data, and the first Observations of dimensional data features The latent feature vectors of the data sample dimensions; Time-domain features, frequency-domain features, and statistical features are extracted from the cleaned data. Multi-source sensor data from the same device are aligned by timestamps and spliced ​​into a unified device state vector. The health of the device at each site is calculated based on the time cumulative effect of device load and temperature. Production management data is obtained from the manufacturing process data resource pool. The equipment status vectors and equipment health of each station are associated and fused with the production management data according to the equipment code and timestamp to form a structured manufacturing process dataset.

[0010] In one possible implementation, the data in the manufacturing process data resource pool is processed to obtain a structured manufacturing process dataset, including: Null value detection and outlier removal are performed on each data stream in the manufacturing process data resource pool. Low-rank matrix factorization is used to impute discrete missing data to complete data cleaning. The formula is as follows: in, The set of locations of the observed data. , It is a low-rank matrix factorization factor. The regularization coefficient is used. Time-domain features, frequency-domain features, and statistical features are extracted from the cleaned data. Multi-source sensor data from the same device are aligned by timestamps and spliced ​​into a unified device state vector. The health of the device at each site is calculated based on the time cumulative effect of device load and temperature. Production management data is obtained from the manufacturing process data resource pool. The equipment status vectors and equipment health of each station are associated and fused with the production management data according to the equipment code and timestamp to form a structured manufacturing process dataset.

[0011] In one possible implementation, the formula for calculating the equipment health of each site based on the time-cumulative effect of equipment load and temperature is as follows: in, For equipment health, For rated load, The fatigue strength coefficient, For load index, For temperature sensitivity coefficient, For the integration time Next Real-time operating load of the equipment For the integration time Next Real-time operating temperature of the equipment.

[0012] In one possible implementation, an artificial intelligence analysis model is constructed based on the structured manufacturing process dataset to identify and predict equipment operating status anomalies, and generate optimized control commands, including: A state-space model is constructed based on the structured manufacturing process dataset, and the formula is: in, The device is in hidden state. To control the input, These are the observation features extracted from the structured manufacturing process dataset. , It is a nonlinear mapping function. for The system process noise at any given time represents the state error caused by unmodelable random operating condition disturbances and environmental interference during equipment operation. for The system observation noise represents the random measurement errors generated during sensor acquisition, data transmission, and data preprocessing. for The device's predicted hidden state vector updated at each time step. It follows a zero-mean Gaussian distribution. The process noise covariance matrix is... It follows a zero-mean Gaussian distribution. To observe the noise covariance matrix; The hidden states of the device are recursively estimated using the state-space model, and the Mahalanobis distance anomaly score is calculated using the following formula: in, Scoring for Mahalanobis distance anomalies, This is a predicted value for the current moment based on historical data. The residual covariance matrix; After judging the abnormal operating state of the equipment based on the Mahalanobis distance anomaly score, the optimal control command is obtained by optimizing the solution based on the Mahalanobis distance anomaly score.

[0013] In one possible implementation, the formula for obtaining the optimal control command based on the Mahalanobis distance anomaly score is as follows: in, This is the optimal control command. For the target state, This is an abnormal threshold. This represents the risk penalty coefficient.

[0014] In one possible implementation, a fusion computing model for each site is constructed based on the structured manufacturing process dataset. A real-time mapping relationship is established between each fusion computing model and the physical equipment. After correcting and predicting the state of each fusion computing model using the execution result data, the fusion computing models are coupled together to form a production line-level digital twin, including: Mechanism models based on physical laws were constructed for each site; The IT-side neural network is trained based on the structured manufacturing process dataset. The output of the neural network is then used as a residual correction term and superimposed onto the mechanism model to form a mechanism-data fusion computing model. A real-time mapping relationship is established between each fusion computing model and the physical device. The formula for the mechanism-data fusion computing model is: In the formula, is the predicted final state output of the device from the mechanism-data fusion calculation model at time t, and is the high-precision device operating state characterization value after fusing physical mechanism constraints and data residual corrections. For the physical mechanism model function of the equipment, for The standardized equipment operating condition input feature vector is input to the model at each time step. A nonlinear residual correction function constructed for the IT-side neural network. This is the complete set of trainable parameters for the IT-side neural network. The execution result data is used to correct each fusion computing model, and state prediction is performed based on the corrected fusion computing models. By associating and coupling the various fusion computing models according to the production line topology, a production line-level digital twin is formed.

[0015] In one possible implementation, the execution result data is used to correct each fusion computing model, and state prediction is performed based on the corrected fusion computing models, including: An extended Kalman filter algorithm is used to recursively correct each fusion calculation model; The prior prediction and prior error covariance are calculated based on the posterior estimate from the previous time step, using the following formula: in, for Extended Kalman filter prior predictions of the hidden state of the device at any given time; This refers to the nonlinear fusion state transition function corresponding to the aforementioned mechanism-data fusion computation model. for The posterior optimal estimate of the hidden state of the device at time step; for Process control inputs for timing equipment; for The prior error covariance matrix of the state at any given time. for Time-Nonlinear Fusion State Transition Function For the Jacobian matrix of the state variables, for The transpose of the time-matrix Jacobian matrix; for The posterior error covariance matrix at time step [time]. The process noise covariance matrix; The Kalman gain is calculated based on the prior error covariance and observation noise covariance matrices, using the following formula: in, For Kalman gain, The observation matrix; To observe the noise covariance matrix; The prior prediction value is corrected using the measured values ​​of the OT-side sensors in the execution result data and the Kalman gain, as shown in the formula: in, for The posterior optimal estimate of the hidden state of the equipment, obtained after observation correction by extended Kalman filter at each time step, represents the high-precision, low-error final state output of the equipment after fusing model predictions and OT measured data. Prior predictions for IT virtual models based on OT mechanisms; These are measured values ​​from the OT-side sensors. The prior error covariance is corrected using the Kalman gain, as shown in the following formula: in, for The updated posterior error covariance matrix is ​​corrected at each step. For the corresponding dimension, for The prior error covariance matrix of the filtering iteration at each time step; Based on the corrected fusion computing models, the state estimates of each device at the next moment are predicted.

[0016] Compared with the prior art, the beneficial effects of this application are as follows: This application achieves high-fidelity real-time status awareness throughout the entire manufacturing process through a multi-site modular production line system and adaptive data acquisition. At the same time, it establishes a cloud-edge-device collaborative system architecture to realize data connectivity between the OT-side equipment control system and the IT-side business management system, improve the collaborative operation capability of multiple devices and systems, reduce response latency, and ensure that the data is collectable, visible, and open-source throughout the entire chain.

[0017] This application establishes a real-time mapping relationship between physical entities and virtual models within a cyber-physical system framework at the virtual-physical integration level, constructing equipment-level and production line-level digital twin models. Driven by both data and mechanisms, the digital twin models enable real-time monitoring and optimal simulation of the manufacturing system's operational status. Attached Figure Description

[0018] Figure 1 A flowchart illustrating a method for constructing a full-chain open-source intelligent manufacturing testbed system based on IT-OT convergence, provided for embodiments of this application; Figure 2 The cloud-edge-device architecture diagram of the intelligent manufacturing testbed provided in the embodiments of this application. Detailed Implementation

[0019] The present solution will now be described in conjunction with the accompanying drawings and specific embodiments.

[0020] Figure 1 A flowchart illustrating a method for constructing a full-chain open-source intelligent manufacturing testbed system based on IT-OT convergence, as provided in this application embodiment, is shown below. Figure 1 This embodiment describes a method for constructing a full-chain open-source intelligent manufacturing testbed system based on IT-OT convergence, comprising: S101 constructs a multi-site hardware manufacturing production line system, and obtains raw sensing data by acquiring multi-dimensional physical signals during equipment operation through sensor groups deployed at each site.

[0021] In this embodiment, a typical manufacturing process is functionally decoupled and modularized to construct a multi-site hardware manufacturing production line system, including CNC machining stations, additive manufacturing stations, quality inspection stations, assembly stations, packaging stations, and warehousing and logistics stations. Each station is connected via a physical conveyor system to form a complete manufacturing chain. The CNC machining station includes CNC lathes, CNC milling machines, or mill-turn machining centers to perform the machining of typical mechanical parts. The additive manufacturing station includes metal additive manufacturing equipment or polymer additive manufacturing equipment. The quality inspection station includes vision inspection equipment, coordinate measuring machines, or laser measuring equipment.

[0022] Sensor arrays are deployed at each site to collect multidimensional physical signals during equipment operation in real time. The multidimensional physical signals collected at the same site are aligned by timestamp and spliced ​​into a unified device state vector. The unified device state vector of each site is used as the raw sensor data of each site.

[0023] Taking a CNC machining station as an example, its sensor configuration includes a cutting force sensor, an acceleration sensor, and a thermocouple temperature sensor. The cutting force sensor is installed between the spindle and the tool post, with a range of 0~10kN and an accuracy of ±0.5%FS, used for real-time monitoring of cutting force / load. The acceleration sensor is installed at the spindle bearing housing, with a range of ±50g and a frequency response of 0.5Hz~10kHz, used to pick up vibration amplitude. The thermocouple temperature sensor is installed near the cutting area, with an accuracy of ±0.5℃, used to collect machining temperature. The sampling frequency of each sensor is set according to the signal type: the sampling frequency for cutting force and vibration signals is greater than or equal to 1kHz, and the sampling frequency for temperature signals is greater than or equal to 10Hz. The multi-station hardware manufacturing production line system provides the physical objects and equipment sources for data acquisition in subsequent steps; the raw physical signals collected by the sensors at each station constitute the source of subsequent data stream processing.

[0024] S102 is based on a cloud-edge-device collaborative system architecture deployed in the hardware manufacturing production line system. It reads the raw sensor data of the terminal layer devices through the edge layer, and pushes it to the cloud after parsing and format conversion, and builds a manufacturing process data resource pool in the cloud.

[0025] In this embodiment, a three-layer collaborative system architecture of cloud-edge-device is deployed based on the hardware production line system. The terminal layer is located at the bottom of the architecture and is the source of manufacturing data and the final execution end of control commands. It includes VPLCs, execution devices, logistics and warehousing equipment, and quality inspection equipment at each site. Among them, the VPLC, as the core control unit of the production line, interacts with the lower-level physical devices and issues commands through the Modbus / OPC UA industrial standard protocol. The execution devices include 3D printing equipment, multiple processing units, ground-rail robots, cleaning units, carving units, assembly units, and packaging units, which undertake specific processing, handling, and assembly tasks. The logistics and warehousing equipment includes storage units, automated guided vehicles (AGVs), conveying units, and finished product warehouses, realizing the automatic flow of materials and inventory management. The quality inspection equipment includes inspection units, which are responsible for online or offline quality inspection and generate quality result data.

[0026] As the intermediate hub connecting the terminal layer and the cloud, the edge layer undertakes four core functions: multi-protocol access, data standardization and conversion, real-time processing, and lightweight AI inference. The industrial IoT platform, as the core software platform of the edge layer, connects to legacy serial port devices via serial gateways and RS-485 / RS-232 communication lines, and integrates mainstream industrial Ethernet protocols such as OPCUA, Modbus-TCP, and S7 through the SCADA system, enabling unified access and data acquisition from heterogeneous devices. The protocol conversion and data aggregation module performs unified parsing, format conversion, and timestamp alignment of raw data from different devices and protocol formats, and performs lightweight mapping of high-frequency signals to generate standardized JSON / ProtoBuf data formats.

[0027] The formula for lightweight mapping of high-frequency signals is: in, For downsampling operation, These represent the mean, variance, and peak value within the sliding window, respectively. For vector concatenation, for The time-series signal vector of the device's high-frequency operating status obtained from the original acquisition at the edge node at any given time. for The fused feature vector is output after lightweight mapping processing at the edge. The high-frequency raw signal from the OT side is compressed into statistical features and then uploaded to the cloud IT resource pool to alleviate bandwidth pressure.

[0028] The edge data uplink channel pushes standardized data to the cloud via MQTT / HTTP / RESTful protocols, with MQTT serving as the primary transmission protocol to ensure real-time and reliable data transmission. The edge AI inference engine is deployed in a 5G gateway and KubeEdge containerized runtime environment, supporting local real-time inference of lightweight AI models with millisecond-level response times. It also features local data caching and breakpoint resumption capabilities to ensure data integrity under network fluctuation scenarios.

[0029] The cloud layer serves as the physical carrier of the manufacturing process data resource pool and the core engine for intelligent analysis. Specifically, it comprises the manufacturing process data resource pool, an AI analysis model cluster, and a digital twin simulation engine. The manufacturing process data resource pool uniformly receives all standardized data reported from the edge layer, constructing a thematic data lake covering equipment operation data, process parameter data, and production management data. Internally, the data resource pool employs a hybrid storage architecture: a time-series database stores high-frequency equipment status data, a relational database stores structured business data such as production orders and quality records, and a distributed object storage system stores unstructured data such as equipment logs and inspection images. Data quality is ensured through a data governance process. The AI ​​analysis model cluster is deployed in a high-performance computing environment in the cloud, utilizing massive amounts of historical data from the data resource pool for offline training of complex models, including quality prediction models, equipment remaining life prediction models, and intelligent scheduling optimization models. After training, model parameters or compressed inference models are distributed to the edge AI inference engine via the MQTT protocol, enabling continuous iterative updates of the models.

[0030] The digital twin simulation engine builds a virtual image of the physical production line based on real-time data streams, enabling visualized mapping of the production process, anomaly simulation and deduction, and what-if analysis, which are used for process parameter optimization, capacity bottleneck diagnosis, and root cause location of failures.

[0031] This architecture adopts a two-way collaborative working mechanism of uplink data acquisition and downlink intelligent control. The data uplink path is as follows: each device at the terminal layer uploads real-time data to the edge layer IoT platform or SCADA system via the Modbus / OPC UA protocol. After protocol conversion and standardization, the data is pushed to the cloud manufacturing process data resource pool via the MQTT / HTTP protocol to complete the persistent storage and unified management of the data.

[0032] Intelligent downlink path: After the cloud-based AI analysis model cluster completes model training and optimization based on the data resource pool, it distributes the inference model to the edge layer KubeEdge container environment via the MQTT protocol. The edge AI inference engine performs inference calculations in real time and transmits the decision results / control parameters back to the terminal layer VPLC and execution device via the Modbus / OPC UA protocol to achieve closed-loop intelligent control.

[0033] S103 processes the data in the manufacturing process data resource pool to obtain a structured manufacturing process dataset.

[0034] In this embodiment, null value detection and outlier removal are performed on each data stream in the manufacturing process data resource pool. Low-rank matrix decomposition is used to impute discrete missing data, completing data cleaning. The formula is as follows: in, The set of locations of the observed data. , It is a low-rank matrix factorization factor. The regularization coefficient is . This represents the optimal imputation value for the discrete missing data after iterative optimization using low-rank matrix factorization. For the first in the manufacturing process data resource pool The equipment sensor, raw production condition data, and the first Observations of dimensional data features This represents the latent feature vectors of the data sample dimensions.

[0035] Temporal, frequency, and statistical features are extracted from the cleaned data. Multi-source sensor data from the same device are aligned by timestamps and concatenated into a unified device state vector. The device health of each site is calculated based on the time-cumulative effect of device load and temperature, using the following formula: in, For equipment health, For rated load, The fatigue strength coefficient, For load index, For temperature sensitivity coefficient, For the integration time Next Real-time operating load of the equipment For the integration time Next Real-time operating temperature of the equipment.

[0036] Production management data is obtained from the manufacturing process data resource pool. The equipment status vectors and equipment health of each station are associated and fused with the production management data according to the equipment code and timestamp to form a structured manufacturing process dataset, which can be called by subsequent artificial intelligence analysis models.

[0037] S104, based on a structured manufacturing process dataset, constructs an artificial intelligence analysis model to identify and predict equipment operating status anomalies, and generates optimized control commands.

[0038] In this embodiment, a state-space model is constructed based on a structured manufacturing process dataset, and the formula is as follows: in, The device is in hidden state. To control the input, These are the observation features extracted from the structured manufacturing process dataset. , It is a nonlinear mapping function. for The system process noise at any given time represents the state error caused by unmodelable random operating condition disturbances and environmental interference during equipment operation. for The system observation noise represents the random measurement errors generated during sensor acquisition, data transmission, and data preprocessing. for The device's predicted hidden state vector updated at each time step. It follows a zero-mean Gaussian distribution. This is the process noise covariance matrix, used to quantify the fluctuation range of random disturbances in equipment operating conditions. It follows a zero-mean Gaussian distribution. The observation noise covariance matrix is ​​used to quantify the fluctuation range of data observation errors.

[0039] The hidden states of the device are recursively estimated using a state-space model, and the Mahalanobis distance anomaly score is calculated using the following formula: in, Scoring for Mahalanobis distance anomalies, This is a predicted value for the current moment based on historical data. The residual covariance matrix; After judging the abnormal operating state of the equipment based on the Mahalanobis distance anomaly score, the optimal control command is obtained by optimization based on the Mahalanobis distance anomaly score. The formula is: in, This is the optimal control command. For the target state, This is an abnormal threshold. This represents the risk penalty coefficient.

[0040] S105 will send the optimized control command to the equipment control system to execute the corresponding control action, and feed back the execution result data to the manufacturing process data resource pool.

[0041] In this embodiment, control commands are sent to the OT-side equipment control system through edge computing nodes, causing the manufacturing equipment to perform corresponding control actions, including feed rate adjustment, spindle speed adjustment, cooling start / stop, and tool replacement, etc. The execution results, such as actual control effect, equipment response status, and control deviation, are fed back to the manufacturing process data resource pool, forming a closed-loop operation mechanism of data acquisition, intelligent analysis, control feedback, and execution verification.

[0042] S106: Based on the structured manufacturing process dataset, construct the fusion computing model of each site, establish the real-time mapping relationship between each fusion computing model and physical equipment, and after correcting and predicting the status of each fusion computing model through the execution result data, associate and couple the fusion computing models to form a production line-level digital twin.

[0043] In this embodiment, a mechanistic model based on physical laws is constructed for each site, with different structural compositions for different equipment types. Taking the CNC machining unit as an example, its mechanistic model consists of three modules connected in series: a cutting force sub-model, a temperature rise sub-model, and a vibration sub-model. The cutting force sub-model... Taking the instantaneous cutting area and cutting speed as inputs, the output is the cutting force, where, The instantaneous cutting force at time t. The cutting force coefficient of the workpiece material. Let be the instantaneous cross-sectional area of ​​the cut at time t. Let be the instantaneous cutting linear velocity of the tool at time t. This is the cutting loss correction factor. Temperature rise sub-model. Taking cutting force as input, the output is the temperature rise in the cutting zone, where, Let be the instantaneous temperature of the cutting region at time t. This is the initial reference temperature for the cutting area. Here, m is the cutting heat distribution coefficient, and ρ is the equivalent mass of the cutting zone. Given the specific heat capacity of the workpiece material under constant pressure, k is the thermal time constant, t is the processing time, and the vibratory sub-model... The vibration response is output using the inherent characteristics of the system as parameters. Among them, Let be the instantaneous vibration response amplitude of the system at time t. The initial amplitude of the excited vibration. The natural frequency of the structure, The initial phase of the vibration, For the system damping ratio, Let ω be the system's natural angular frequency, and t be the processing time. The three sub-models are connected in series, with the output of the previous stage serving as the input of the next stage, together forming a complete description of the processing mechanism. Mechanism models for other equipment types are constructed according to the same principle.

[0044] The IT-side neural network is trained based on a structured manufacturing process dataset. The output of the neural network is then superimposed onto the mechanism model as a residual correction term, forming a mechanism-data fusion computational model. A real-time mapping relationship between each fusion computational model and the physical equipment is established. The formula for the mechanism-data fusion computational model is as follows: In the formula, is the predicted final state output of the device from the mechanism-data fusion calculation model at time t, and is the high-precision device operating state characterization value after fusing physical mechanism constraints and data residual corrections. This is the physical mechanism model function for the equipment, corresponding to the aforementioned mechanism deduction model constructed based on equipment dynamics and operating condition state space. It is used to output basic theoretical state prediction results based on the inherent physical laws, process constraints, and operating mechanisms of intelligent manufacturing equipment. for The standardized equipment operating condition input feature vectors input to the model at each moment are taken from a structured manufacturing process dataset and include multi-dimensional input features such as equipment control parameters, operating loads, operating temperatures, vibration timing, and production process parameters. A nonlinear residual correction function is constructed for the IT-side neural network to fit system nonlinear deviations, implicit performance degradation of equipment, and random disturbance errors of operating conditions that are difficult to model using pure mechanistic models. This is the complete set of trainable parameters for the IT-side neural network, including network weights and bias terms. The optimal residual correction mapping rule is determined through iterative training and optimization using a structured manufacturing process dataset.

[0045] The extended Kalman filter algorithm is used to recursively correct each fusion calculation model. The prior prediction and prior error covariance at the current time are calculated based on the posterior estimate from the previous time step, using the following formula: in, for The extended Kalman filter prior prediction of the hidden state of the device at time 1 is the device state prediction obtained by extrapolating the optimal state of the previous time 1 before the current observation data is introduced. The nonlinear fusion state transition function corresponding to the aforementioned mechanism-data fusion calculation model integrates the physical mechanism constraints of the equipment and the residual correction characteristics of the neural network, and is used to realize the iterative deduction of the timing state of the equipment. for The posterior optimal estimate of the hidden state of the device at time step 1 is the accurate device state value after observation correction and iterative optimization at the previous time step 2. for The process control inputs of the equipment at all times include actual control parameters such as equipment speed, feed rate, and working pressure; for The prior error covariance matrix of the current time state is used to characterize the overall prediction error fluctuation of the prior prediction state at the current time. for Time-Nonlinear Fusion State Transition Function The Jacobian matrix of the state variables is used to perform local linearization of the nonlinear state transition equation, adapting it to the extended Kalman filter iterative logic. for The transpose of the time-matrix Jacobian matrix; for The posterior error covariance matrix at time step 1 represents the error convergence accuracy of the optimal estimated state at the previous time step 2. This is the system process noise covariance matrix, consistent with the previous definition of state-space model parameters, used to quantify the state prediction error caused by random operating conditions and environmental disturbances of the equipment.

[0046] The Kalman gain is calculated based on the prior error covariance and the observation noise covariance matrix, using the following formula: in, For Kalman gain, The observation matrix; To observe the noise covariance matrix.

[0047] The prior predictions are corrected using the measured values ​​of the OT-side sensors and the Kalman gain from the execution results data, using the following formula: in, for The posterior optimal estimate of the hidden state of the equipment, obtained after observation correction by extended Kalman filter at each time step, represents the high-precision, low-error final state output of the equipment after fusing model predictions and OT measured data. Prior predictions for IT virtual models based on OT mechanisms; These are measured values ​​from the OT-side sensors. The prior error covariance is corrected using Kalman gain, as shown in the following formula: in, for The updated posterior error covariance matrix after time-time correction characterizes the error convergence accuracy and uncertainty range of the optimal estimated state of the device at the current time, and is used for filtering iteration calculation at the next time time. This is the identity matrix for the corresponding dimension, used to ensure dimension matching and linear constraint stability in matrix operations; for The prior error covariance matrix of the filtering iteration at each time step.

[0048] Based on the corrected fusion computing models, the state estimates of each device at the next moment are predicted, and the anomaly scores of each site are calculated.

[0049] The overall health index of the production line is calculated based on the anomaly scores of each site. The formula is as follows: in, For the overall health index, For the first Abnormal ratings for each site, For the first The upper limit of the abnormal score set for the normal operation status of equipment at each site.

[0050] By linking and coupling the various fusion computing models, a production line-level digital twin is formed, thereby completing the construction of a full-chain intelligent manufacturing testbed system.

[0051] Figure 2 This application provides an embodiment of an open-source, IT-OT-integrated smart manufacturing testbed system architecture, which is divided into a cloud layer, an edge layer, and a terminal layer from top to bottom. The cloud layer includes a manufacturing operations management platform, a digital twin simulation engine, a manufacturing process data resource pool, and an AI analysis model cluster. The cloud and edge layer communicate bidirectionally via MQTT / HTTP / RESTful protocols. The edge layer includes an industrial IoT platform, SCADA (OPC UA, MODBUS-TCP, S7, etc.), a serial gateway, 485 / 232 communication lines, and a 5G gateway + KubeEdge + edge AI inference engine. The industrial IoT platform connects to SCADA and the 5G gateway + KubeEdge + edge AI inference engine via MQTT protocol. SCADA connects to devices via the serial gateway and 485 / 232 communication lines. The edge layer and terminal layer communicate bidirectionally via Modbus / OPC UA protocol. The terminal layer includes 1#VPLC, 2#VPLC, and 3#VPLC. 1#VPLC, 2#VPLC, and 3#VPLC are interconnected via the PROFINET protocol. 1#VPLC is connected to the 3D printing unit, storage unit, AGV, and finished product warehouse via Modbus and PROFINET protocols. 2#VPLC is connected to the processing unit, ground-rail robot, and cleaning unit via Modbus and PROFINET protocols. 3#VPLC is connected to the engraving unit, inspection unit, conveyor line, assembly unit, and packaging unit via Modbus and PROFINET protocols.

[0052] In this embodiment, "multiple" refers to two or more. "And / or" describes the relationship between related objects, indicating that three relationships can exist. For example, A and / or B can represent the existence of A alone, the simultaneous existence of A and B, or the existence of B alone. A and B can be singular or plural. The character " / " generally indicates that the preceding and following related objects have an "or" relationship.

[0053] The above description is merely a specific embodiment of this application. Any variations or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in this application should be included within the protection scope of this application.

Claims

1. A method for constructing a full-chain open-source intelligent manufacturing testbed system based on IT-OT convergence, characterized in that, include: Construct a multi-site hardware manufacturing production line system, and obtain raw sensing data by acquiring multi-dimensional physical signals during equipment operation through sensor groups deployed at each site; Based on the hardware manufacturing production line system, a cloud-edge-device collaborative system architecture is deployed. The original sensor data of the terminal layer device is read through the edge layer, parsed and converted into a format and then pushed to the cloud. A manufacturing process data resource pool is built in the cloud. The data in the manufacturing process data resource pool is processed to obtain a structured manufacturing process dataset; An artificial intelligence analysis model is built based on the structured manufacturing process dataset to identify and predict equipment operating status and generate optimized control commands. The optimized control command is sent to the equipment control system to execute the corresponding control action, and the execution result data is fed back to the manufacturing process data resource pool. Based on the structured manufacturing process dataset, a fusion computing model is constructed for each site. A real-time mapping relationship between each fusion computing model and physical equipment is established. After correcting and predicting the status of each fusion computing model through the execution result data, the fusion computing models are associated and coupled to form a production line-level digital twin.

2. The method for constructing a full-chain open-source intelligent manufacturing testbed system based on IT-OT convergence according to claim 1, characterized in that, A multi-site hardware manufacturing production line system is constructed, and multi-dimensional physical signals during equipment operation are acquired through sensor groups deployed at each site to obtain raw sensor data, including: Construct a multi-site hardware manufacturing production line system that includes CNC machining stations, additive manufacturing stations, quality inspection stations, assembly stations, packaging stations, and warehousing and logistics stations; Sensor arrays are deployed at various sites to collect multi-dimensional physical signals during equipment operation in real time. The multi-dimensional physical signals collected at the same site are then aligned by timestamps and stitched together to form a unified equipment state vector. The unified device state vector of each station is used as the original sensor data of each station.

3. The method for constructing a full-chain open-source intelligent manufacturing testbed system based on IT-OT convergence according to claim 1, characterized in that, Based on the aforementioned hardware manufacturing production line system, a cloud-edge-device collaborative system architecture is deployed. The edge layer reads the raw sensor data from the terminal layer devices, parses and converts it before pushing it to the cloud. A manufacturing process data resource pool is then built in the cloud, including: Based on the aforementioned hardware production line system, a cloud-edge-device three-layer collaborative system architecture is deployed. The terminal layer includes VPLCs, execution devices, logistics and warehousing equipment, and quality inspection equipment at each site. The edge layer reads the device operation status data, process parameter data, and production management data of the terminal layer device through the Modbus protocol and OPC UA protocol; The edge layer deploys an industrial IoT platform and a SCADA system. It connects to serial devices via serial gateways and RS-485 / RS-232, and integrates OPC UA and Modbus-TCP through the SCADA system to achieve unified access and data acquisition for multi-source heterogeneous devices. Edge nodes perform protocol parsing, format conversion, and timestamp alignment on the collected device operation status data, and perform lightweight mapping on high-frequency signals in the device operation status data to generate standardized data; The standardized data is pushed to the cloud via MQTT / HTTP / RESTful protocols; The cloud platform employs a hybrid storage architecture combining time-series databases, relational databases, and object storage to construct a data resource pool for the manufacturing process.

4. The method for constructing a full-chain open-source intelligent manufacturing testbed system based on IT-OT convergence according to claim 3, characterized in that, The edge node performs protocol parsing, format conversion, and timestamp alignment on the collected device operation status data, and applies a lightweight mapping formula to the high-frequency signals in the device operation status data: in, For downsampling operation, These represent the mean, variance, and peak value within the sliding window, respectively. For vector concatenation, for The time-series signal vector of the device's high-frequency operating status obtained from the original acquisition at the edge nodes at any given time. for The fused feature vector is output after the edge lightweight mapping process at each time step.

5. The method for constructing a full-chain open-source intelligent manufacturing testbed system based on IT-OT convergence according to claim 1, characterized in that, The data in the manufacturing process data resource pool is processed to obtain a structured manufacturing process dataset, including: Null value detection and outlier removal are performed on each data stream in the manufacturing process data resource pool. Low-rank matrix factorization is used to impute discrete missing data to complete data cleaning. The formula is as follows: in, The set of locations of the observed data. , It is a low-rank matrix factorization factor. The regularization coefficient is . This represents the optimal imputation value for the discrete missing data after iterative optimization using low-rank matrix factorization. For the first in the manufacturing process data resource pool The equipment sensor, raw production condition data, and the first Observations of dimensional data features The latent feature vectors of the data sample dimensions; Time-domain features, frequency-domain features, and statistical features are extracted from the cleaned data. Multi-source sensor data from the same device are aligned by timestamps and spliced ​​into a unified device state vector. The health of the device at each site is calculated based on the time cumulative effect of device load and temperature. Production management data is obtained from the manufacturing process data resource pool. The equipment status vectors and equipment health of each station are associated and fused with the production management data according to the equipment code and timestamp to form a structured manufacturing process dataset.

6. The method for constructing a full-chain open-source intelligent manufacturing testbed system based on IT-OT convergence according to claim 5, characterized in that, The formula for calculating the equipment health of each site based on the time-cumulative effect of equipment load and temperature is as follows: in, For equipment health, For rated load, The fatigue strength coefficient, For load index, For temperature sensitivity coefficient, For the integration time Next Real-time operating load of the equipment For the integration time Next Real-time operating temperature of the equipment.

7. The method for constructing a full-chain open-source intelligent manufacturing testbed system based on IT-OT convergence according to claim 1, characterized in that, An artificial intelligence analysis model is constructed based on the structured manufacturing process dataset to identify and predict equipment operating status anomalies, and generate optimized control commands, including: A state-space model is constructed based on the structured manufacturing process dataset, and the formula is: in, The device is in hidden state. To control the input, These are the observation features extracted from the structured manufacturing process dataset. , It is a nonlinear mapping function. for The system process noise at any given moment represents the state error caused by unmodelable random operating condition disturbances and environmental interference during equipment operation. for The system observation noise represents the random measurement errors generated during sensor acquisition, data transmission, and data preprocessing. for The device's predicted hidden state vector updated at each time step. It follows a zero-mean Gaussian distribution. The process noise covariance matrix is... It follows a zero-mean Gaussian distribution. To observe the noise covariance matrix; The hidden states of the device are recursively estimated using the state-space model, and the Mahalanobis distance anomaly score is calculated using the following formula: in, Scoring for Mahalanobis distance anomalies, This is a predicted value for the current moment based on historical data. The residual covariance matrix; After judging the abnormal operating state of the equipment based on the Mahalanobis distance anomaly score, the optimal control command is obtained by optimizing the solution based on the Mahalanobis distance anomaly score.

8. The method for constructing a full-chain open-source intelligent manufacturing testbed system based on IT-OT convergence according to claim 7, characterized in that, The formula for obtaining the optimal control command based on the Mahalanobis distance anomaly score is as follows: in, This is the optimal control command. For the target state, This is an abnormal threshold. This represents the risk penalty coefficient.

9. The method for constructing a full-chain open-source intelligent manufacturing testbed system based on IT-OT convergence according to claim 1, characterized in that, Based on the structured manufacturing process dataset, a fusion computing model is constructed for each site. A real-time mapping relationship is established between each fusion computing model and the physical equipment. After correcting and predicting the state of each fusion computing model using the execution result data, the fusion computing models are coupled together to form a production line-level digital twin, including: Mechanism models based on physical laws were constructed for each site; The IT-side neural network is trained based on the structured manufacturing process dataset. The output of the neural network is then used as a residual correction term and superimposed onto the mechanism model to form a mechanism-data fusion computing model. A real-time mapping relationship is established between each fusion computing model and the physical device. The formula for the mechanism-data fusion computing model is: In the formula, is the predicted final state output of the device from the mechanism-data fusion calculation model at time t, and is the high-precision device operating state characterization value after fusing physical mechanism constraints and data residual corrections. For the physical mechanism model function of the equipment, for The standardized equipment operating condition input feature vector is input to the model at each time step. A nonlinear residual correction function constructed for the IT-side neural network. This is the complete set of trainable parameters for the IT-side neural network. The execution result data is used to correct each fusion computing model, and state prediction is performed based on the corrected fusion computing models. By associating and coupling the various fusion computing models according to the production line topology, a production line-level digital twin is formed.

10. The method for constructing a full-chain open-source intelligent manufacturing testbed system based on IT-OT convergence according to claim 9, characterized in that, The execution result data is used to correct each fusion computing model, and state prediction is performed based on the corrected fusion computing models, including: An extended Kalman filter algorithm is used to recursively correct each fusion calculation model; The prior prediction and prior error covariance at the current time are calculated based on the posterior estimate from the previous time step, using the following formula: in, for Extended Kalman filter prior predictions of the hidden state of the device at any given time; This refers to the nonlinear fusion state transition function corresponding to the aforementioned mechanism-data fusion computation model. for The posterior optimal estimate of the hidden state of the device at time step; for Process control inputs for timing equipment; for The prior error covariance matrix of the state at any given time. for Time-Nonlinear Fusion State Transition Function For the Jacobian matrix of the state variables, for The transpose of the time-matrix Jacobian matrix; for The posterior error covariance matrix at time step [time]. The process noise covariance matrix; The Kalman gain is calculated based on the prior error covariance and observation noise covariance matrices, using the following formula: in, For Kalman gain, The observation matrix; To observe the noise covariance matrix; The prior prediction value is corrected using the measured values ​​of the OT-side sensors in the execution result data and the Kalman gain, as shown in the formula: in, for The posterior optimal estimate of the hidden state of the equipment, obtained after observation correction by extended Kalman filter at each time step, represents the high-precision, low-error final state output of the equipment after fusing model predictions and OT measured data. Prior predictions for IT virtual models based on OT mechanisms; These are measured values ​​from the OT-side sensors. The prior error covariance is corrected using the Kalman gain, as shown in the following formula: in, for The updated posterior error covariance matrix is ​​corrected at each step. The identity matrix is ​​the one corresponding to the dimension. for The prior error covariance matrix of the filtering iteration at each time step; Based on the corrected fusion computing models, the state estimates of each device at the next moment are predicted.

Images