Full-link digital twin-driven industrial beat management system and method

Abstract

The invention relates to the field of intelligent manufacturing, in particular to a full-link digital twin-driven industrial beat management system, in the system, an intelligent hardware layer realizes multi-dimensional data acquisition and flexible production, and a digital twin layer completes multi-source data space-time alignment and virtual production line construction; the intelligent application layer realizes dynamic beat management, abnormal diagnosis scheduling and quality-beat symbiotic optimization, and the edge computing node and the cloud server realize hierarchical intelligent decision. The system is adopted for industrial beat management, specifically, multi-target beat modeling, self-calibration process balancing, multi-modal fusion monitoring diagnosis, intelligent evolution dynamic scheduling and quality-beat symbiotic optimization are included, and the problems that all links of traditional industrial beat management are independent, and full-link cooperation is lacked are solved; deep cooperation of the production rhythm, customer requirements, quality efficiency and equipment health is achieved, and production efficiency, quality stability and flexible production capacity can be remarkably improved.

Description

Technical Field

[0001] This invention relates to the field of intelligent manufacturing, specifically to an industrial cycle management system and method driven by end-to-end digital twins. Background Technology

[0002] Industrial cycle time management, as a core component of manufacturing systems, directly impacts production efficiency, quality, and cost. With the integration of smart manufacturing and the Industrial Internet, the amount of data generated in manufacturing workshops during production is exploding, making traditional production management methods ill-suited to the complex and ever-changing production environment. For example, manufacturing enterprises use traditional methods such as test data acquisition systems, quality control systems, and production scheduling systems to support test process management, but many problems still exist in practical applications: 1. Traditional cycle time calculation only considers output and time, relies on a fixed formula (Takt Time = Available Time / Demand) and static process balance, and lacks dynamic adaptation to multi-variety mixed production and real-time quality fluctuations; 2. Traditional data acquisition and monitoring focuses on the deployment of single-device sensors (such as RFID, PLC status) and has not formed a deep integration of multi-modal data of "equipment-quality-logistics" (such as only realizing equipment networking). 3. Traditional scheduling control relies heavily on manually setting constraints. This method, based on rule engines or classic algorithms, can only schedule limited capacity and is difficult to handle real-time decision-making in complex scenarios such as sudden orders and equipment failures.

[0003] In summary, traditional industrial cycle time management methods all suffer from the problem of failing to incorporate implicit constraints such as equipment health and energy consumption into cycle time calculations, making them unsuitable for dynamic production environments. Each link (scheduling, quality, logistics) operates independently, lacking end-to-end collaboration; anomaly handling relies on manual intervention, resulting in low response efficiency. During product changeovers, process balancing and equipment parameter adjustments depend on manual experience, leading to long cycle time stabilization after changeovers and failing to meet the demands of flexible production.

[0004] Currently, to address the challenges of improving production line balance, enhancing quality, reducing energy consumption, and increasing flexible production capabilities, many experts have proposed technical solutions that combine artificial intelligence with industrial cycle time management. However, the results have been less than satisfactory. For example: (1) Patent document CN114418177A discloses a method for predicting material distribution for new products in a digital twin workshop based on generative adversarial networks (GANs). This method constructs a GAN model by acquiring sample data of uncertain factors affecting the production process and sample data of production cycle time in the digital twin workshop. The model is then trained and optimized based on historical sample data of similar products and real-time monitoring sample data of new products to obtain the working cycle time of the production unit in the digital twin workshop. However, this method still has shortcomings in the optimization of the structure and parameters of the GAN, affecting the accuracy and adaptability of the prediction model.

[0005] (2) Patent document CN114218763A proposes a method for dynamic virtual reconfiguration of production lines based on digital twins. This method involves real-time dynamic data acquisition of production element information from the physical production line, preprocessing and standardizing the acquired data to construct a digital twin model of the production line, and unifying and encapsulating the physical production line data within the digital twin model to establish a dynamic virtual reconfiguration system. The simulation operation status of the established digital twin model is then evaluated. However, this method still needs improvement in terms of accuracy and efficiency in dynamic virtual reconfiguration of production lines, and it is difficult to meet the ever-changing production demands and technological developments.

[0006] (3) Patent document CN115097788A proposes an intelligent control platform based on a digital twin factory. This platform includes industrial equipment providing underlying hardware support and data sources, an edge computing layer responsible for data calculation and storage, an industrial PaaS layer providing operation services and data analysis components, a digital twin factory containing multiple digital twins, and industrial applications based on digital twin technology. Through collaborative data processing at each level and the construction of physical digital models, it achieves functions such as accurate prediction and control of the production process and full lifecycle management of equipment, achieving accurate mapping, virtual-real interaction, and intelligent intervention. However, the platform's cycle time management is static, failing to incorporate implicit constraints such as equipment health and energy consumption into cycle time calculation. The depth of multimodal data fusion is insufficient, and information silos exist. Scheduling decisions rely on preset rules, and there is a lack of a symbiotic optimization mechanism between quality and cycle time. The flexible production capacity is insufficient, making it difficult to adapt to the dynamic needs of multi-variety mixed-line production.

[0007] (4) Patent document CN116449781A proposes a smart factory MES system based on digital twins. This system consists of a physical workshop, a twin workshop, a management system, twin data, and a Kanban system. The physical workshop completes production and processing tasks and provides real-time data feedback. The twin workshop guides production through 3D modeling and simulation optimization. The twin data integrates data from multiple systems to form a unified data source. The management system provides digital control services, and the Kanban system displays real-time information from the workshop, thus addressing the lack of autonomous decision-making and simulation optimization mechanisms in traditional MES systems. However, this system has a single optimization dimension for cycle time calculation, does not consider equipment load balancing, energy consumption, and carbon emissions, fails to achieve precise spatiotemporal alignment of multi-source data, has weak autonomy in scheduling decisions and relies on manual error handling, lacks a dynamic correlation mechanism between quality and cycle time, and lacks rapid changeover adaptation capabilities, making it difficult to meet the needs of flexible production and dynamic production environments.

[0008] Therefore, there is an urgent need for a full-link adaptive cycle management system and method based on digital twins and data intelligence to solve problems such as fragmented optimization of single links, static rule dependence, and data silos of multiple systems in existing technologies, and to achieve deep collaboration between production rhythm and customer needs, quality efficiency, and equipment health. Summary of the Invention

[0009] The purpose of this invention is to propose an industrial cycle time management system and method driven by a full-link digital twin, in order to solve the problems in the existing industrial cycle time management system where each link operates independently and lacks full-link collaboration, resulting in low production efficiency, unstable quality and difficulty in adapting to dynamic production environments.

[0010] The objective of this invention is achieved through the following approach: An end-to-end digital twin-driven industrial cycle management system includes a smart hardware layer, a digital twin layer, and a smart application layer. The smart hardware layer comprises several equipment data acquisition units and several flexible production devices. The smart hardware layer utilizes the equipment data acquisition units to collect raw heterogeneous data on equipment status, operating actions, and material trajectories in real time, and sends it to the digital twin layer. This data is then transformed into regularized and correlated high-quality data, providing a multi-dimensional data foundation for the smart application layer. This enables the smart application layer to accurately optimize cycle plans according to production requirements and generate global optimization instructions for real-time cycle adjustments, thereby adjusting the parameters of the flexible production devices in the smart hardware layer.

[0011] Preferably, in the smart hardware layer, several equipment data acquisition units and several flexible production equipment are distributed and matched according to workstations / equipment groups. The equipment data acquisition units in each workstation / equipment group are deployed close to the corresponding flexible production equipment to form a local "data acquisition-action control" closed-loop architecture.

[0012] Preferably, within each workstation / equipment group, one equipment data acquisition unit corresponds to at least multiple flexible production equipment in the corresponding area. The equipment data acquisition unit works collaboratively through multiple sensors, simultaneously covering the multimodal data acquisition needs of the corresponding flexible production equipment.

[0013] Preferably, the digital twin layer includes a spatiotemporal calibration system and a 3D modeling module. The spatiotemporal calibration system uses the ST-CNN algorithm to achieve spatiotemporal alignment of multi-source data and constructs a causal chain traceability model about "process-quality-logistics". This model is used to output the causal correlation analysis results of the entire production chain to the intelligent application layer in real time during production. The 3D modeling module constructs a 1:1 virtual production line using a 3D scanner, integrating equipment PLC simulation models and physical parameters to achieve real-time state mirroring.

[0014] Preferably, the intelligent application layer includes a dynamic beat management platform, an anomaly diagnosis and scheduling engine, and a quality-beat symbiotic optimization module; The dynamic cycle management platform uses reinforcement learning algorithms to solve for the optimal theoretical cycle in real time, outputs the optimal cycle scheme, and drives the dynamic adjustment of equipment parameters to ensure that the production line balance rate, flexibility and adaptability and production efficiency are optimal. The anomaly diagnosis and scheduling engine supports a three-level self-healing mechanism. Through multimodal data fusion diagnosis and digital thread-driven scheduling, it forms a closed-loop technical link of "anomaly identification - root cause location - autonomous handling - scheduling optimization" to support the system's full-link collaboration and flexible production needs. The quality-cycle symbiotic optimization module establishes a Pareto optimization model of process capability index and cycle deviation to achieve dynamic coupling optimization of quality and cycle.

[0015] Preferably, the dynamic beat management platform includes a multi-objective beat calculation engine and a self-calibrating process balancing system: The multi-objective cycle time calculation engine uses reinforcement learning to solve the optimal theoretical cycle time and the dynamic cycle time compensation coefficient of each workstation in real time based on a four-dimensional optimization model that includes customer needs, quality loss costs, equipment load balancing, energy consumption and carbon emissions. This provides a target basis for the self-calibrating process balancing system, drives the adaptive adjustment of flexible production equipment parameters, and works in conjunction with the anomaly diagnosis and scheduling engine and the quality-cycle time symbiotic optimization module to stabilize the production line balance rate. The self-calibrating process balancing system works in collaboration with station-level machine vision, LSTM time prediction network and genetic algorithm. It first calculates the operation time of the station in real time and predicts the process time fluctuation, then generates a dynamic recombination scheme that includes station sequence, equipment parameters and operation allocation, and outputs equipment parameter adjustment instructions to calibrate the process balancing of the production line in real time.

[0016] Preferably, the anomaly diagnosis and scheduling engine integrates the Transformer anomaly diagnosis model and the digital thread-driven scheduling engine: The Transformer anomaly diagnosis model is based on the Transformer architecture. It integrates the multi-dimensional frequency domain features of vibration sensors, the pixel spatial features of 4K vision cameras, and the NLP word embedding vectors of operation logs through a multi-head attention mechanism. It calculates real-time classification results of anomaly types, conclusions on the location of anomaly roots, and results of impact path analysis. This provides core decision-making basis for the three-level self-healing mechanism (equipment level / workstation level / system level) and the digital thread-driven scheduling engine, enabling early prediction and location of anomaly roots. The digital thread-driven scheduling engine utilizes a three-level algorithm that combines Monte Carlo simulation, deep reinforcement learning, and digital twin pre-validation. Based on digital threads that span the entire "order receiving - process planning - equipment control" chain, it calculates a rescheduling scheme adapted to dynamic production scenarios. Through full lifecycle data flow fusion and virtual environment pre-validation, it coordinates with a dynamic cycle time management platform and anomaly diagnosis model to ensure continuous and stable operation of the production line and improve production flexibility and response efficiency.

[0017] Preferably, several distributed local intelligent data processing execution terminals are deployed on the production site as edge computing nodes. After acquiring the original heterogeneous data of the intelligent hardware layer, key information is filtered and uploaded to the digital twin layer. At the same time, based on the real-time data collected locally, the parameters of the flexible production equipment in the intelligent hardware layer are adjusted at the workstation level according to the actual local status and combined with the global optimization instructions of the intelligent application layer, so that the real-time production cycle of each flexible production equipment meets the requirements.

[0018] Preferably, it also includes a cloud server, which is deployed with a reinforcement learning beat optimization model. Combined with the high-quality data output by the digital twin layer, it generates global optimization instructions covering the entire production line to guide the beat optimization scheme of the intelligent application layer. Preferably, the input data of the reinforcement learning beat optimization model includes customer order information, equipment health data and real-time energy consumption data, wherein the equipment health data is PHM predicted remaining lifetime data, and the output data is a dynamic beat compensation coefficient, wherein the dynamic beat compensation coefficient has a range of ±15%.

[0019] The method for industrial cycle time management using the above system includes the following steps: 1) Utilize real-time data collected by the smart hardware layer and customer order information, equipment health data, and real-time energy consumption data obtained from the cloud server to construct a four-dimensional optimization model that includes customer needs, quality loss costs, equipment load balancing, and energy consumption and carbon emissions. Solve the optimal theoretical beat in real time through the reinforcement learning beat optimization model deployed on the cloud server and output the dynamic beat compensation coefficient. 2) The edge computing node operation action recognition model outputs the workstation operation time in real time, and combines the LSTM time prediction network to predict the process time fluctuation. The genetic algorithm generates a dynamic recombination scheme containing workstation sequence, equipment parameters and operation allocation. It receives the dynamic cycle compensation coefficient of the cloud server and the global optimization instructions of the intelligent application layer, and drives the flexible production equipment in the intelligent hardware layer to perform parameter adaptive adjustment. 3) The equipment data acquisition unit of the intelligent hardware layer transmits the original heterogeneous data of equipment status, operation actions and material trajectories to the digital twin layer. The spatiotemporal calibration system realizes spatiotemporal alignment based on the ST-CNN algorithm and builds a causal chain traceability model about "process-quality-logistics". Through the Transformer anomaly diagnosis model, multimodal features are integrated to output anomaly type classification results, root cause location conclusions and impact path analysis in real time. 4) Based on the 3D modeling module of the digital twin layer, the 1:1 virtual production line is constructed. The digital thread-driven scheduling engine uses Monte Carlo simulation, deep reinforcement learning and digital twin pre-validation to generate rescheduling schemes that are adapted to dynamic scenarios such as emergency order insertion and equipment failure through three-level collaboration and combined with the correlation analysis results output by the causal chain tracing model. The schemes are then synchronized to the intelligent application layer and edge computing nodes. 5) The quality-cycle time symbiosis optimization module establishes a Pareto optimization model of process capability index and cycle time deviation. Combining the anomaly diagnosis results and dynamic cycle time compensation coefficient, it uses a multi-objective particle swarm optimization algorithm to solve the optimal working point in real time. When the process capability index drops to a preset threshold, it automatically increases the sampling frequency and triggers a cycle time extension strategy. 6) Each edge computing node uploads the parameter adjustment effect and production cycle time compliance status to the cloud server. The cloud server summarizes the data of the entire system every day and updates the anomaly diagnosis model and reinforcement learning cycle time optimization model through federated learning to continuously improve the system's adaptive capability and cycle time optimization accuracy.

[0020] The beneficial effects of this invention are as follows: 1. By using multi-objective cycle time modeling, a multi-dimensional optimization model is constructed that includes customer needs, quality loss costs, equipment load balancing, energy consumption and carbon emissions. This enables the dynamic solution of theoretical cycle time, overcomes the limitations of traditional solutions that rely on fixed formulas and static process balancing, improves the production line balancing rate (from 85% to over 95%), and solves the problem that the cycle time calculation in existing technologies does not incorporate implicit constraints such as equipment health and energy consumption. 2. By adopting multimodal fusion real-time monitoring technology, through the collaborative work of various sensors such as vibration sensors, 4K vision cameras, and UWB positioning modules, a unified spatiotemporal mapping of equipment, quality, and logistics data is achieved, which solves the problem in existing technologies that data acquisition and monitoring focus on the deployment of single device sensors and the independent operation of each link without full-link collaboration. 3. Based on a digital thread-driven scheduling engine and reinforcement learning algorithm, intelligent three-level scheduling management is realized, which significantly improves the flexibility and response speed of production and overcomes the limitations of existing technologies that rely on rule engines or classical algorithms for limited capacity scheduling. 4. By using a dynamic coupling optimization model of process capability index and cycle time deviation, the optimization of quality and cycle time is achieved (quality-related cycle time anomalies are reduced by 60%, error prevention and omission detection rate is <0.5%, and the Cpk compliance rate of key processes is increased to 98%), avoiding batch rework caused by quality problems in traditional solutions and effectively reducing production costs. 5. Supports multi-product mixed-line production. Through intelligent process reorganization and equipment parameter adjustment, it can quickly adapt to the production needs of different products, overcoming the shortcomings of existing technologies that rely on manual experience for process balancing and equipment parameter adjustment. This reduces multi-product changeover time by 50%, emergency order response time by ≤5 minutes, unplanned equipment downtime by 40%, unit product energy consumption by 15%-20%, and new product introduction cycle by 30%. It supports mixed-line production of ≥20 products, with adaptive cycle time adjustment response time ≤100ms, adapting to personalized production needs. Attached Figure Description

[0021] Figure 1 This is a schematic diagram illustrating the specific principle of the industrial cycle management system described in this invention. Detailed Implementation

[0022] like Figure 1 As shown, an end-to-end digital twin-driven industrial cycle management system includes a smart hardware layer, a digital twin layer, and a smart application layer. The smart hardware layer includes several equipment data acquisition units and several flexible production devices. The smart hardware layer uses the equipment data acquisition units to collect raw heterogeneous data on equipment status, operation actions, and material trajectories in real time, and sends it to the digital twin layer. This data is then transformed into regularized and correlated high-quality data, providing a multi-dimensional data foundation for the smart application layer. This enables the smart application layer to accurately optimize cycle plans according to production requirements and generate global optimization instructions for real-time cycle adjustments, thereby adjusting the parameters of the flexible production devices in the smart hardware layer.

[0023] In the smart hardware layer, several equipment data acquisition units and several flexible production equipment are distributed and matched according to workstations / equipment groups. The equipment data acquisition units in each workstation / equipment group are deployed nearby with the corresponding flexible production equipment, forming a local "data acquisition-action control" closed-loop architecture.

[0024] Within each workstation / equipment group, one equipment data acquisition unit corresponds to at least one (e.g., one to three) flexible production equipment in the area. The equipment data acquisition unit works collaboratively through multiple sensors to simultaneously cover the multimodal data acquisition needs of the corresponding flexible production equipment.

[0025] The digital twin layer includes a spatiotemporal calibration system and a 3D modeling module. The spatiotemporal calibration system uses the ST-CNN algorithm to achieve spatiotemporal alignment of multi-source data and constructs a causal chain traceability model about "process-quality-logistics". This model is used to output the causal correlation analysis results of the entire production chain to the intelligent application layer in real time during production. This provides causal constraints for the intelligent application layer to deal with dynamic scenarios such as mixed production lines of multiple varieties, emergency order insertion, and equipment failure, ensuring that the decision-making meets the real-time adjustment requirements. The 3D modeling module constructs a 1:1 virtual production line using a 3D scanner, integrates equipment PLC simulation models and physical parameters, and realizes real-time state mirroring. It provides a virtual simulation environment for intelligent application layer's cycle time adjustment, scheduling schemes, and changeover optimization, allowing for early verification of decision feasibility and avoiding cycle time imbalance, quality risks, and cost waste caused by trial and error in actual production.

[0026] The intelligent application layer includes a dynamic beat management platform, an anomaly diagnosis and scheduling engine, and a quality-beat symbiotic optimization module. The dynamic cycle management platform uses reinforcement learning algorithms to solve for the optimal theoretical cycle in real time, outputs the optimal cycle scheme, and drives the dynamic adjustment of equipment parameters. It uses multi-objective algorithms and real-time process self-calibration to ensure that the production line balance rate, flexibility and adaptability and production efficiency are optimal. The anomaly diagnosis and scheduling engine supports a three-level self-healing mechanism. Through multimodal data fusion diagnosis and digital thread-driven scheduling, it forms a closed-loop technical link of "anomaly identification - root cause location - autonomous handling - scheduling optimization" to support the system's full-link collaboration and flexible production needs. The quality-cycle symbiotic optimization module establishes a Pareto optimization model of process capability index and cycle deviation to achieve dynamic coupling optimization of quality and cycle.

[0027] The dynamic beat management platform includes a multi-objective beat calculation engine and a self-calibrating process balancing system. The multi-objective cycle time calculation engine uses reinforcement learning (PPO algorithm) based on a four-dimensional optimization model that includes customer demand, quality loss cost, equipment load balancing, energy consumption and carbon emissions. It solves the optimal theoretical cycle time and the dynamic cycle time compensation coefficient of each workstation in real time, providing a target basis for the self-calibrating process balancing system and driving the adaptive adjustment of flexible production equipment parameters. At the same time, it works in conjunction with the anomaly diagnosis and scheduling engine and the quality-cycle time symbiotic optimization module to stabilize the production line balance rate. Its core purpose is to break through the limitations of traditional cycle time calculation that relies on fixed formulas and single-objective optimization, adapt to dynamic production environments such as multi-variety mixed lines and order fluctuations, and achieve deep collaboration between production rhythm and multiple objectives. The self-calibrating process balancing system works collaboratively with station-level machine vision (YOLOv8 model), LSTM time prediction network, and genetic algorithm. It first calculates the operation time of each station in real time (accuracy ±200ms) and predicts process time fluctuations (accuracy ±2%). Then, it generates a dynamic reorganization scheme that includes station sequences, equipment parameters, and operation allocations, and outputs equipment parameter adjustment instructions such as servo motor speed compensation (±5%). The core purpose is to overcome the limitations of traditional process balancing that relies on human experience and has a lag in adjustment. It adapts to dynamic production scenarios such as multi-product mixed lines and fluctuations in operation time, calibrates the process balancing of the production line in real time, avoids cycle imbalance caused by efficiency differences of a single station, ensures that the production line balance rate is stable at over 95%, supports rapid changeover and flexible production, and improves overall production efficiency.

[0028] The anomaly diagnosis and scheduling engine integrates the Transformer anomaly diagnosis model and a digital thread-driven scheduling engine: The Transformer anomaly diagnosis model is based on the Transformer architecture. It integrates multi-dimensional frequency domain features (e.g., 128-dimensional frequency domain features) from vibration sensors, pixel spatial features (e.g., 224×224 pixels) from 4K vision cameras, and NLP word embedding vectors from operation logs through a multi-head attention mechanism. It calculates real-time classification results of anomaly types (equipment failure / quality fluctuation / logistics blockage), anomaly root cause location conclusions, and impact path analysis results. The warning delay is ≤200ms and the error prevention and omission rate is <0.5%. The core purpose is to overcome the limitations of traditional single-sensor monitoring, delayed anomaly identification, and reliance on manual investigation, and achieve accurate and rapid diagnosis of production anomalies. Its role is to provide core decision-making basis for the three-level self-healing mechanism (equipment level / workstation level / system level) and digital thread-driven scheduling engine, predict and locate the root cause of anomalies in advance, avoid the spread of anomalies affecting the overall cycle time, ensure the continuous and stable operation of the production line, and improve the responsiveness to dynamic production scenarios in conjunction with the scheduling system. The digital thread-driven scheduling engine utilizes a three-tiered algorithm system—Monte Carlo simulation (strategic layer), deep reinforcement learning (tactical layer), and digital twin pre-validation (execution layer)—to work collaboratively. Based on digital threads (dynamic directed acyclic graphs) spanning the entire production chain from order receipt to process planning to equipment control, it calculates rescheduling schemes adapted to dynamic production scenarios (including process sequence adjustments, backup resource activation, and process route optimization). In the event of an emergency order, the scheduling time is ≤3 minutes and the cycle time compliance rate is ≥98%. The core objective is to overcome the limitations of traditional rule engines or classic algorithms in terms of limited capacity scheduling, address complex scenarios such as sudden orders and equipment failures, and achieve dynamic collaborative scheduling across the entire production chain. Its function is to avoid scheduling scheme execution conflicts through full lifecycle data flow fusion and virtual environment pre-validation, quickly respond to demands such as multi-product mixed lines and anomaly handling, and coordinate with the dynamic cycle time management platform and anomaly diagnosis model to ensure continuous and stable production line operation, improving production flexibility and response efficiency.

[0029] In this invention, the system deploys several distributed local intelligent data processing execution terminals at the production site (near workstations / equipment groups) as edge computing nodes. After acquiring the raw heterogeneous data of the intelligent hardware layer, it filters key information and uploads it to the digital twin layer / cloud server. At the same time, based on the real-time data (equipment status, operation actions) collected locally, and according to the actual local status, combined with the global optimization instructions of the intelligent application layer, it adjusts the parameters of the flexible production equipment in the intelligent hardware layer at the workstation level so that the real-time production cycle of each flexible production equipment meets the requirements.

[0030] The system also includes a cloud server, which is equipped with a reinforcement learning beat optimization model. Combined with high-quality data output from the digital twin layer, it generates global optimization instructions covering the entire production line to guide the beat optimization scheme of the intelligent application layer and the parameter adjustment of the flexible production equipment by the edge computing nodes. Preferably, the input data of the reinforcement learning beat optimization model includes customer order information, equipment health data and real-time energy consumption data, wherein the equipment health data is PHM predicted remaining lifetime data, and the output data is a dynamic beat compensation coefficient, wherein the dynamic beat compensation coefficient has a range of ±15%.

[0031] The method for industrial cycle time management using the above system includes the following steps: 1) Multi-objective beat modeling: Utilizing real-time data collected from the intelligent hardware layer and customer order information, equipment health (PHM predicted remaining life) data, and real-time energy consumption data obtained from the cloud server, a four-dimensional optimization model is constructed, which includes customer demand (Td), quality loss cost (Qloss), equipment load balancing (Eload), and energy consumption carbon emissions (Ccarbon). The optimal theoretical beat is solved in real time through the reinforcement learning beat optimization model (PPO algorithm) deployed on the cloud server, and the dynamic beat compensation coefficient is output. 2) Self-calibrating process balancing: The edge computing node runs the YOLOv8 action recognition model (training dataset of 500,000+ samples) to output the real-time operation time of the workstation (accuracy ±200ms). Combined with the LSTM time prediction network, it predicts the process time fluctuation (accuracy ±2%). The genetic algorithm generates a dynamic recombination scheme that includes workstation sequence, equipment parameters and operation allocation. It receives the dynamic cycle compensation coefficient from the cloud server and the global optimization instructions from the intelligent application layer, driving the flexible production equipment in the intelligent hardware layer to perform parameter adaptive adjustment (such as servo motor speed compensation ±5%). 3) Multimodal Fusion Monitoring and Anomaly Diagnosis: The equipment data acquisition unit of the intelligent hardware layer transmits the original heterogeneous data of equipment status, operation actions, and material trajectories to the digital twin layer. The spatiotemporal calibration system achieves spatiotemporal alignment (nanosecond-level time synchronization, millimeter-level spatial mapping) based on the ST-CNN algorithm and constructs a causal chain traceability model about "process-quality-logistics". Through the Transformer anomaly diagnosis model, multimodal features are fused to output anomaly type classification results, root cause location conclusions, and impact path analysis in real time (early warning delay ≤200ms, error prevention and omission rate <0.5%). 4) Intelligent Evolutionary Dynamic Scheduling: Based on the 1:1 virtual production line built by the digital twin layer 3D modeling module, the digital thread-driven scheduling engine uses three levels of collaboration: Monte Carlo simulation (strategic layer), deep reinforcement learning (tactical layer), and digital twin pre-verification (execution layer). Combined with the correlation analysis results output by the causal chain tracing model, it generates a rescheduling scheme adapted to dynamic scenarios such as emergency order insertion and equipment failure (scheduling time ≤ 3 minutes, cycle time compliance rate ≥ 98%), and synchronizes it to the intelligent application layer and edge computing nodes. 5) Quality-Beat Cycle Co-existence Optimization: The quality-beat cycle co-existence optimization module establishes a Pareto optimization model of process capability index (Cpk≥1.33) and cycle deviation (T_d≤5%). Combining the anomaly diagnosis results and dynamic cycle compensation coefficient, the multi-objective particle swarm optimization algorithm is used to solve the optimal working point in real time. When the process capability index drops to a preset threshold (e.g., 1.5), the sampling frequency is automatically increased and the cycle extension strategy is triggered to achieve dynamic coupling optimization of quality and cycle. 6) Data closed-loop iteration: Each edge computing node uploads data such as the effect of parameter adjustment and the production cycle time target to the cloud server. The cloud server summarizes the data of the entire system every day and updates the anomaly diagnosis model (weight update magnitude ≤ 5%) and the reinforcement learning cycle time optimization model through federated learning to continuously improve the system's adaptive ability and cycle time optimization accuracy.

[0032] Based on the above method, the following is an example: An end-to-end digital twin-driven industrial cycle management system, comprising a smart hardware layer, a digital twin layer, and a smart application layer.

[0033] 1) Smart Hardware Layer Includes equipment data acquisition units and flexible production equipment: 1-1) Equipment Data Acquisition Unit Deploy vibration sensors (PCB 352C03), 4K vision cameras (Basler ace 2), and UWB positioning modules to collect real-time data on equipment status, operational actions, and material trajectory.

[0034] Among them, the vibration sensor has a sampling rate of 1kHz, the 4K vision camera has a frame rate of 60fps, and the UWB positioning module has an accuracy of ±10cm.

[0035] For example, a vibration sensor (sampling rate 1kHz) is added to the stator winding machine, a three-coordinate measuring instrument (accuracy ±5μm) is integrated into the rotor dynamic balancing machine, and the AGV is equipped with UWB positioning (update frequency 10Hz).

[0036] 1-2) Flexible production equipment Features include servo-driven adjustable speed conveyor lines (speed range 0.5-10m / min, accuracy ±0.1m / min), collaborative robots (UR10e), force control sensors (ATI Nano17), and even vision guidance systems (recognition speed ≤1 second / part) supporting dynamic adjustment of workstation-level cycle time.

[0037] 2) Digital Twin Layer Includes a spatiotemporal calibration system and a 3D modeling module: This embodiment uses Siemens Tecnomatix to build a digital twin of the production line, synchronizes the PLC control logic with the physical parameters of the equipment (such as the torque-speed curve of the winding machine), and develops a spatiotemporal calibration system to achieve nanosecond-level time alignment (PTP protocol) and millimeter-level spatial mapping (coordinate transformation matrix) of sensor data.

[0038] 2-1) Spatiotemporal calibration system Based on the ST-CNN algorithm, spatiotemporal alignment of multi-source data is achieved, and a causal chain traceability model of "process-quality-logistics" is constructed.

[0039] 2-1-1) ST-CNN algorithm technical path: It uses three-dimensional convolutional kernels (such as 5×5×3) to extract features in both spatial and temporal dimensions.

[0040] Spatial convolution kernels extract spatial features such as device position and orientation, while temporal convolution kernels capture temporal patterns such as action sequences and material flow.

[0041] The timing alignment strategy employs a cross-camera timing synchronization algorithm based on feature point matching. It calculates the timing offset by identifying the time difference of the appearance of common feature points in multiple cameras, achieving nanosecond-level synchronization accuracy.

[0042] Traditional PTP protocols only achieve clock synchronization between network devices, failing to address the substantial data asynchrony issues caused by inconsistent data acquisition times from different sensors and differences in data transmission latency, and even more so, the problem of spatial coordinate system uniformity. ST-CNN, on the other hand, directly aligns data content through deep learning algorithms, achieving dual alignment at both the nanosecond temporal and millimeter spatial levels.

[0043] 2-1-2) Causal chain tracing model: This model is actually a graph-based multi-relationship network. Nodes represent various entities in the production process (processes, equipment, materials, quality inspection points), and edges represent causal relationships between entities (such as the impact of "changes in processing parameters of a certain process" on "quality indicators of subsequent processes"). By analyzing the propagation of real-time data in this network, end-to-end traceability is achieved.

[0044] 2-2) 3D Modeling Module A 1:1 virtual production line is constructed using a 3D scanner, integrating equipment PLC simulation models and physical parameters to achieve real-time state mirroring.

[0045] The equipment PLC simulation model is not built directly in a single off-the-shelf software, but is a control logic model created based on industrial automation standards (such as IEC 61131-3) using PLC programming software such as Siemens TIA Portal and Rockwell Studio 5000.

[0046] These models were then imported into professional digital twin platforms or simulation software (such as NX Mechatronics Concept Designer, Dassault Systèmes' 3DEXPERIENCE platform, Unity Industrial Edition, or NVIDIA Omniverse).

[0047] In a digital twin platform, a real-time data connection is established with the physical PLC via industrial communication protocols such as OPC UA (Unified Architecture) and MQTT. The PLC's real-time status data (such as I / O signals and register values) is mapped and drives the corresponding simulation actuators and sensors in the virtual model. Simultaneously, the spatial state of the 3D model (such as robot joint angles and conveyor belt positions) is updated based on this data. The final result is a virtual mirror image that is highly synchronized with the physical production line, containing its geometry, motion state, and control system logic. This virtual production line can be used for real-time monitoring, advanced simulation, and debugging to demonstrate the high degree of consistency between the posture and movements of a device (such as a robotic arm) in the virtual environment and the real-time video stream or status data of the physical device.

[0048] 3) Intelligent Application Layer It includes a dynamic beat management platform, an anomaly diagnosis and scheduling engine, and a quality-beat symbiosis optimization module.

[0049] 3-1) Dynamic beat management platform It includes a multi-objective cycle time calculation engine and a self-calibrating process balancing system, which solves for the optimal theoretical cycle time in real time through a reinforcement learning algorithm (PPO algorithm).

[0050] 3-1-1) The core of the multi-objective cycle time calculation engine is a multi-objective optimization solver. The objective functions typically include maximizing production efficiency (output per unit time), minimizing energy consumption, maximizing overall equipment efficiency (OEE), balancing production line load, and meeting order delivery deadlines. Constraints include equipment capacity limits, logical relationships between processes, and material availability.

[0051] In this embodiment, the multi-objective takt calculation engine continuously receives real-time data (equipment status, order information, energy consumption) from the digital twin layer, as well as external inputs (such as order changes and emergency order insertions). The PPO algorithm learns and outputs the optimal global takt setting and the takt compensation coefficient for each workstation in the current state by exploring different takt strategies in a simulated environment (digital twin model) and evaluating their comprehensive performance on these multi-objectives.

[0052] 3-1-2) Self-calibrating process balancing system: The self-calibrating process balancing system includes a time database (storing historical data of standard and actual working hours for each workstation), a balance calculation module, and a feedback control loop.

[0053] The self-calibration process balancing system uses a balance calculation module to calculate the production line balance rate in real time using the following formula: Production line balance rate = Total working hours of all workstations / (Number of workstations * Bottleneck working hours) When the actual working hours of a certain workstation continuously deviate from the standard working hours due to reasons such as operator proficiency or slight differences in materials, resulting in a decrease in the balance rate, the system will automatically calibrate the balance between processes by fine-tuning the speed of the conveyor sections before and after the workstation (i.e., dynamically allocating buffer time) or suggesting the reallocation of work tasks, so as to ensure that the production line operates at the highest efficiency.

[0054] 3-2) Anomaly Diagnosis and Scheduling Engine 3-2-1) Transformer Anomaly Diagnosis Model: This is not a completely off-the-shelf model, but a deep learning model specifically trained on industrial time-series data (vibration signals, current sequences, etc.) based on the Transformer architecture. Its advantage lies in using a self-attention mechanism to capture long-range dependencies and complex patterns between multi-sensor data, thereby accurately diagnosing the type and severity of equipment anomalies (such as tool wear, bearing failure).

[0055] 3-2-2) Digital Thread-Driven Scheduling Engine: This engine relies on digital threads throughout the entire design, production, and operation and maintenance process. When an anomaly occurs, the engine can automatically assess the scope of impact (such as which work-in-process is affected and which subsequent processes will be halted) based on the relationships between product models, process models, and equipment models, and dynamically generate rescheduling solutions (such as skipping faulty equipment, activating backup resources, and adjusting process routes) based on the current status (equipment health, order priority, and alternative process paths).

[0056] 3-2-3) Three-level self-healing mechanism: 3-2-3-1) Equipment-level self-healing: PLC or edge computing nodes automatically execute preset recovery programs (such as robot returning to home position, clearing jammed materials).

[0057] 3-2-3-2) Workstation / Unit-level Self-Healing: Edge computing nodes coordinate the actions of multiple devices within a workstation, or adjust cycle time and process parameters to compensate for or isolate anomalies.

[0058] 3-2-3-3) System-level self-healing: The cloud or central scheduling system starts up and performs global optimization operations such as order rearrangement, production line reconstruction, and resource rescheduling.

[0059] 3-3) Quality-Temperature Symbiosis Optimization Module A Pareto optimization model is established for the process capability index (Cpk) and cycle time deviation (T_d) to achieve dynamic coupling optimization of quality and cycle time. The establishment of the Pareto optimization model is as follows: 3-3-1) Define decision variables: usually the cycle time setting value or cycle time compensation coefficient of each workstation.

[0060] 3-3-2) Constructing the objective function: Objective 1: Maximize the overall quality level of the production line (which can be expressed as the weighted average or minimum value of the Cpk of key processes). Objective 2: Minimize the production line cycle time deviation (T_d, which can be defined as the sum of squares of the deviations between the actual cycle time and the theoretical optimal cycle time, or the waiting time of the bottleneck process).

[0061] 3-3-3) Determine the constraints: including order delivery time, equipment capacity limitations, and the sequence of processes.

[0062] 3-3-4) Solving the Pareto front: Using a multi-objective evolutionary algorithm (such as NSGA-II) or an optimization method based on a surrogate model, simulations are performed on historical data and real-time digital twin models to obtain a set of non-dominated solutions (Pareto optimal solution set), which is the set of solutions where improving any one objective will lead to a worsening of the other objective.

[0063] 3-3-5) Decision Making and Mapping: The system or operator selects an optimal solution from the Pareto front based on current preferences (quality or efficiency) and sends the corresponding cycle time settings to the production line. This model continuously updates the Pareto front by learning from data.

[0064] It is worth noting that this industrial cycle time management system also includes edge computing nodes and cloud servers. The edge computing nodes run a YOLOv8 action recognition model (training dataset of 500,000+ samples), outputting the real-time operation time of each workstation (accuracy ±200ms). The cloud server deploys a reinforcement learning cycle time optimization model, taking customer orders, equipment health (PHM predicts remaining lifespan), and real-time energy consumption data as input, and outputting dynamic cycle time compensation coefficients (range ±15%).

[0065] Specifically, the industrial cycle time management method of the above system includes the following steps: Step 1: Multi-objective cycle time modeling and optimization algorithm By integrating customer needs, quality loss costs, equipment load balancing, and energy consumption and carbon emissions into a unified optimization framework, and using reinforcement learning (PPO algorithm), an optimization model with four-dimensional constraints is constructed to solve for the optimal theoretical cycle time in real time. The optimization objective function is as follows:

[0066] The parameters of the above formula are defined as follows: Topt: Optimal theoretical cycle time (unit: seconds / piece), the overall production rhythm recommended by the system; Td: Customer demand takt time (unit: seconds / piece), calculated based on order delivery cycle and daily production target. The calculation formula is: Td = Daily demand / Daily working hours. Qloss: Quality loss cost (unit: yuan / piece), including scrap cost, rework cost and quality penalty cost, which is calculated through real-time quality inspection data; Eload: Equipment load balance index (dimensionless), calculated as the standard deviation of equipment utilization rate at each workstation, reflecting the balance of the production line; Ccarbon: Energy consumption carbon emissions (unit: kgCO2 / unit), calculated by collecting real-time equipment energy consumption data (such as electricity and gas consumption) and multiplying it by the carbon emission coefficient; α, β, γ, δ: Dynamic weighting coefficients (dimensionless), with initial values ​​set based on historical data and updated online via the PPO algorithm, satisfying the constraint α+β+γ+δ=1.

[0067] Compared with traditional reinforcement learning algorithms (such as DQN), the PPO algorithm in this embodiment effectively solves the model convergence problem under production data fluctuations through trust domain optimization and shearing mechanism: ① Trust domain constraint: Limit the policy update range (within 20%) to prevent policy collapse due to data fluctuations; ② Adaptive weight adjustment: The dynamic weight coefficients are adaptively adjusted through real-time data training. Based on gradient backpropagation and loss function optimization, the gradient contribution of each objective is calculated and the weight ratio is dynamically allocated. ③ Multi-objective collaboration: By shearing the objective function to handle multiple competing objectives, over-optimization of a single objective is avoided, and Pareto optimality is achieved.

[0068] Table 1. Comparison of the improved PPO algorithm and the traditional algorithm in multi-objective cycle time optimization in this embodiment.

[0069] Step 2: Self-calibration process balancing By utilizing real-time identification of operation time at the workstation level machine vision, combined with LSTM network prediction of process time fluctuations, and using a genetic algorithm to generate dynamic recombination schemes, the workstation equipment parameters are adaptively adjusted, as detailed below: 2-1) The workstation-level machine vision system uses the YOLOv8 model (2 million pixels resolution) to identify operation actions and material flow status in real time. It extracts image features through convolutional neural networks, outputs the start and end time points and type classification of operation actions (such as picking, assembling, and placing), and then calculates the time consumption of workstation operations.

[0070] 2-2) The system uses an LSTM network to predict process time fluctuations. Its inputs include historical time consumption data, equipment status parameters and material characteristics, and the output is the predicted process time within the future time window (accuracy ±2%).

[0071] 2-3) Dynamic recombination using genetic algorithms: The process by which a genetic algorithm generates an optimal solution through dynamic recombination is as follows: 2-3-1) Coding Design: Chromosome coding includes workstation sequences, equipment parameters, and operation allocation schemes; 2-3-2) Fitness function: The optimization objectives are cycle time stability, equipment utilization, and balance rate; 2-3-3) Selection, Crossover, and Mutation: New solutions are generated through roulette wheel selection, single-point crossover, and uniform mutation operations; 2-3-4) Iterative optimization: Iterate up to 100 times to output the optimal workstation reorganization scheme; The system drives the station equipment (such as servo motors) to adaptively adjust parameters to achieve speed compensation (±5%) and ensure process balance.

[0072] Step 3: Multimodal fusion real-time monitoring and anomaly diagnosis 3-1) Spatiotemporal alignment data acquisition The deployed vibration sensors (accuracy ±0.1g), 4K vision camera (frame rate 60fps), and UWB positioning module (accuracy ±10cm) achieve nanosecond-level time synchronization through the PTP protocol, establish a coordinate system transformation matrix (deviation ≤0.5mm), and realize the spatiotemporal unified mapping of equipment, quality, and logistics data.

[0073] 3-2) Diagnosis of Self-Learning Anomalies 3-2-1) Multimodal feature fusion method: The system integrates three types of heterogeneous data through the Transformer architecture: 3-2-1-1) Vibration signal: Collect triaxial acceleration data (128-dimensional feature vector) of the equipment, extract frequency domain features through Fourier transform, and characterize the mechanical state of the equipment; 3-2-1-2) Visual images: 4K camera captures workstation operation scenes (224×224 pixel RGB images), spatial features are extracted through convolutional neural network to identify material positions and abnormal operation actions; 3-2-1-3) Operation Log: Natural Language Processing (NLP) technology extracts text keywords (such as "emergency stop" and "material shortage") and converts them into word embedding vectors.

[0074] 3-2-2) Attention Mechanism and Latent Association Recognition: Transformer's multi-head attention mechanism calculates the correlation weights between data from different modalities to uncover implicit causal chains. 3-2-2-1) Correlation between equipment vibration and quality fluctuation: When the vibration signal shows high-frequency abnormalities, the attention weight is automatically allocated to the quality inspection data, triggering causal analysis (such as bearing wear leading to assembly deviation). 3-2-2-2) Operation logs are linked to logistics congestion: When the frequency of "waiting for materials" in the logs increases, the system links UWB positioning data to identify logistics congestion points.

[0075] 3-2-3) Implementation of exception classification: The anomaly classification model uses an encoder-decoder structure: 3-2-3-1) Encoder: Maps multimodal data into a unified feature vector; 3-2-3-2) Decoder: Outputs the probability of anomaly type (equipment failure / quality fluctuation / logistics blockage) through the Softmax layer; 3-2-3-3) Real-time classification: Early warning delay ≤200ms, meeting the real-time requirements of production.

[0076] Table 2 Dimensions and characteristics of multimodal data fusion

[0077] Step 4: Intelligent Evolutionary Dynamic Scheduling and Flexible Control 4-1) Digital Thread-Driven Scheduling 4-1-1) Methods and Forms for Constructing Digital Threads: The digital thread in this embodiment is a data flow framework that spans the entire product lifecycle. The thread constructed by this system covers the entire link of "order receipt - process planning - equipment control", specifically including: 4-1-1-1) Data Layer: Integrates ERP (order data), PLM (process data), MES (execution data), and equipment control system (real-time status data), and achieves information fusion through a unified data model (such as the OPC UA standard); 4-1-1-2) Model layer: Based on digital twin technology, a virtual production line model is constructed to map the state of the physical entity in real time; 4-1-1-3) Service Layer: Provides a scheduling algorithm library (Monte Carlo simulation, deep reinforcement learning, etc.) and supports three-level scheduling decisions; The digital thread is represented as a dynamically updated directed acyclic graph (DAG), where nodes represent orders, process segments, or equipment, and edges represent data flow and constraints.

[0078] ① Basic structure: It consists of “nodes” and “directed edges”. Nodes correspond to entities in the entire production chain (orders, processes, equipment, materials, etc.), and directed edges represent logical dependencies between entities (such as “order → process route” and “process → equipment”).

[0079] The nodes carry content that includes all elements such as order information (customer requirements, delivery time limits), process data (process sequence, parameters), equipment status (health, availability), and material information (flow trajectory).

[0080] Directed edge function: Clarifies the constraint relationship between entities (such as "a certain process needs to be performed on a designated device" or "materials need to complete the preceding processing before they can be transferred").

[0081] ② Core constraint: "Acyclic" means that there are no circular dependencies (such as process A → process B → process A), ensuring that the production process is logically smooth and executable.

[0082] ③ Dynamic characteristics: Nodes can be added or deleted or edge relationships can be adjusted according to real-time production changes (emergency order insertion, equipment failure), adapting to dynamic production scenarios.

[0083] Directed acyclic graphs can provide data support for three-level scheduling: the strategic layer simulates order delivery scenarios, the tactical layer optimizes resource allocation, and the execution layer verifies that the scheduling scheme is conflict-free, ensuring that a feasible solution is generated within 3 minutes for emergency order insertion.

[0084] 4-1-2) Three-level scheduling mechanism: 4-1-2-1) Strategic Layer (Monte Carlo Simulation): Simulate various possible scenarios for order delivery, assess risks, and develop strategic plans; 4-1-2-2) Tactical Layer (Deep Reinforcement Learning): Dynamically adjust production scheduling and optimize resource allocation based on real-time data; 4-1-2-3) Execution Layer (Digital Twin Pre-verification): Verify the feasibility of the scheduling scheme in a virtual environment to avoid actual execution conflicts; When an emergency order is inserted, the dispatch time is ≤3 minutes and the cycle time compliance rate is ≥98%.

[0085] 4-2) Self-optimizing rapid changeover 4-2-1) Modular quick-change tooling design: The independently designed modular quick-change tooling has the following structural features: 4-2-1-1) Positioning mechanism: A hydraulic clamping device and a tapered pin are used for positioning to ensure a repeatability accuracy of ±0.05mm; 4-2-1-2) Interface Standard: Unified pneumatic and electrical interfaces, supporting quick plugging and unplugging; 4-2-1-3) Driving method: Servo motor driven switching action.

[0086] 4-2-2) Optimization of reinforcement learning actions: Application of reinforcement learning in transformation sequence optimization: 4-2-2-1) State space: includes tooling status, equipment availability, and order priority; 4-2-2-2) Action Space: Defines operations such as disassembly, installation, and adjustment; 4-2-2-3) Reward function: The optimization objectives are minimizing the changeover time and preserving accuracy; 4-2-2-4) PPO Algorithm Training: Output the optimal transformation action sequence to reduce invalid actions; In this embodiment, the system automatically generates the optimal changeover sequence, reducing the changeover time from 12 minutes to 5 minutes, without the need for manual updates to the process documents.

[0087] Step 5: Quality-Temperature Coexistence Optimization (Dynamic Coupling Optimization) 5-1) Pareto optimization model: Establish a Pareto optimization model for process capability index (Cpk) and cycle time deviation (T_d): Minimize F(Cpk,Td)=[1 / Cpk,Td] Constraints: Cpk≥1.33, Td≤5%.

[0088] 5-2) Solution process of Multi-Objective Particle Swarm Optimization (MOPSO): 5-2-1) Particle initialization: Each particle represents a potential solution (a combination of Cpk and T_d); 5-2-2) Non-dominated ordering: Evaluating the Pareto dominance of particles in the target space; 5-2-3) Leader Selection: Select a global leader from the set of non-dominated solutions; 5-2-4) Speed ​​Update: Adjust flight direction based on the particle's own experience and the leader's experience; 5-2-5) Iterate until convergence: Output the Pareto front, from which decision-makers select the optimal operating point.

[0089] 5-3) Adaptive quality control strategy: When Cpk drops to 1.5, the system automatically triggers the following dynamic adjustment strategy: 5-3-1) Cycle time increased by 2%: Reduce production line pace, extend operation time of key processes, and reduce quality risks; 5-3-2) Increase the frequency of spot checks: Increase the proportion of quality spot checks from 5% to 10%, and conduct 100% full inspection when necessary; 5-3-3) Feedback control: Monitor Cpk changes in real time. If it remains below 1.5 for three consecutive cycles, extend the cycle time further (up to a maximum of 5% in total).

[0090] The specific steps for industrial cycle time management using the system constructed in this embodiment include: (1) Deploy sensors and flexible devices according to the hardware list, and connect edge computing nodes and cloud platforms; (2) Import the 3D model of the production line and configure the equipment PLC simulation logic and time-space calibration parameters; (3) Train the YOLOv8 action recognition model and the reinforcement learning beat optimization model (the dataset must include historical operation data, quality inspection data, and device status data). (4) Integrate with the enterprise ERP / MES system and configure interfaces for order receipt and production plan synchronization; (5) Verify the system's adaptive capabilities by simulating scenarios such as order fluctuations, equipment failures, and quality anomalies; ⑹ Finally, each edge computing node analyzes local data in real time and adjusts the conveyor speed autonomously (compensating ±0.5m / min). The cloud server summarizes the data daily and updates the anomaly diagnosis model through federated learning (weight update magnitude ≤5%) to continuously improve the system's adaptability.

[0091] The system constructed in this embodiment was then used on a production line for typical scenario testing, for example: Emergency order insertion (15% surge in orders): The scheduling engine generates a new process sequence within 3 minutes, the robot's motion trajectory is automatically adjusted, and the cycle time compliance rate is 98.5%.

[0092] Equipment malfunction (winding machine vibration exceeds standard): The malfunction diagnosis engine issues a warning 1.5 hours in advance and automatically switches to backup equipment. The switching time is 2 minutes and does not affect the overall cycle time.

[0093] Quality fluctuation (critical rotor imbalance): The system automatically extends the cycle time by 5 seconds, increases the sampling rate from 5% to 20%, avoids batch rework, and reduces quality loss costs by 35%.

[0094] Ultimately, it was found that, compared with the traditional solution, the various indicators of the present invention are far superior to the cycle time management of the traditional solution, as detailed in Table 3: Table 3 .

[0095] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications made to the present invention by those skilled in the art without departing from the spirit of the present invention shall fall within the protection scope of the present invention.

Claims

1. A full-link digital twin-driven industrial cycle management system, characterized in that, It includes a smart hardware layer, a digital twin layer, and a smart application layer. The smart hardware layer includes several equipment data acquisition units and several flexible production devices. The smart hardware layer uses the equipment data acquisition units to collect raw heterogeneous data about equipment status, operation actions, and material trajectories in real time, and sends it to the digital twin layer. The data is then transformed into regularized and correlated high-quality data, providing a multi-dimensional data foundation for the smart application layer. This enables the smart application layer to accurately optimize the cycle time scheme according to production requirements and generate global optimization instructions for real-time cycle time adjustment, thereby adjusting the parameters of the flexible production devices in the smart hardware layer.

2. The industrial cycle management system according to claim 1, characterized in that, In the smart hardware layer, several equipment data acquisition units and several flexible production equipment are distributed and matched according to workstations / equipment groups. The equipment data acquisition units in each workstation / equipment group are deployed nearby with the corresponding flexible production equipment, forming a local "data acquisition-action control" closed-loop architecture.

3. The industrial cycle management system according to claim 2, characterized in that, Within each workstation / equipment group, one equipment data acquisition unit corresponds to at least multiple flexible production equipment in the area. The equipment data acquisition unit works collaboratively through multiple sensors to simultaneously cover the multimodal data acquisition needs of the corresponding flexible production equipment.

4. The industrial cycle management system according to claim 1, characterized in that, The digital twin layer includes a spatiotemporal calibration system and a 3D modeling module. The spatiotemporal calibration system uses the ST-CNN algorithm to achieve spatiotemporal alignment of multi-source data and constructs a causal chain traceability model about "process-quality-logistics". This model is used to output the causal correlation analysis results of the entire production chain to the intelligent application layer in real time during production. The 3D modeling module constructs a 1:1 virtual production line using a 3D scanner, integrating equipment PLC simulation models and physical parameters to achieve real-time state mirroring.

5. The industrial cycle management system according to claim 1, characterized in that, The intelligent application layer includes a dynamic beat management platform, an anomaly diagnosis and scheduling engine, and a quality-beat symbiotic optimization module. The dynamic cycle management platform uses reinforcement learning algorithms to solve for the optimal theoretical cycle in real time, outputs the optimal cycle scheme, and drives the dynamic adjustment of equipment parameters to ensure that the production line balance rate, flexibility and adaptability and production efficiency are optimal. The anomaly diagnosis and scheduling engine supports a three-level self-healing mechanism. Through multimodal data fusion diagnosis and digital thread-driven scheduling, it forms a closed-loop technical link of "anomaly identification - root cause location - autonomous handling - scheduling optimization" to support the system's full-link collaboration and flexible production needs. The quality-cycle symbiotic optimization module establishes a Pareto optimization model of process capability index and cycle deviation to achieve dynamic coupling optimization of quality and cycle.

6. The industrial cycle management system according to claim 5, characterized in that, The dynamic beat management platform includes a multi-objective beat calculation engine and a self-calibrating process balancing system. The multi-objective cycle time calculation engine uses reinforcement learning to solve the optimal theoretical cycle time and the dynamic cycle time compensation coefficient of each workstation in real time based on a four-dimensional optimization model that includes customer needs, quality loss costs, equipment load balancing, energy consumption and carbon emissions. This provides a target basis for the self-calibrating process balancing system, drives the adaptive adjustment of flexible production equipment parameters, and works in conjunction with the anomaly diagnosis and scheduling engine and the quality-cycle time symbiotic optimization module to stabilize the production line balance rate. The self-calibrating process balancing system works in collaboration with workstation-level machine vision, LSTM time prediction network and genetic algorithm. It first calculates the operation time of workstations in real time and predicts process time fluctuations, then generates a dynamic recombination scheme that includes workstation sequence, equipment parameters and operation allocation, and outputs equipment parameter adjustment instructions to calibrate the process balancing of the production line in real time.

7. The industrial cycle management system according to claim 5, characterized in that, The anomaly diagnosis and scheduling engine integrates the Transformer anomaly diagnosis model and a digital thread-driven scheduling engine: The Transformer anomaly diagnosis model is based on the Transformer architecture. It integrates the multi-dimensional frequency domain features of vibration sensors, the pixel spatial features of 4K vision cameras, and the NLP word embedding vectors of operation logs through a multi-head attention mechanism. It calculates real-time classification results of anomaly types, conclusions on the location of anomaly roots, and results of impact path analysis. This provides core decision-making basis for the three-level self-healing mechanism (equipment level / workstation level / system level) and the digital thread-driven scheduling engine, enabling early prediction and location of anomaly roots. The digital thread-driven scheduling engine utilizes a three-level algorithm that combines Monte Carlo simulation, deep reinforcement learning, and digital twin pre-validation. Based on digital threads that span the entire "order receiving - process planning - equipment control" chain, it calculates a rescheduling scheme adapted to dynamic production scenarios. Through full lifecycle data flow fusion and virtual environment pre-validation, it coordinates with a dynamic cycle time management platform and anomaly diagnosis model to ensure continuous and stable operation of the production line and improve production flexibility and response efficiency.

8. The industrial cycle management system according to claim 1, characterized in that, Several distributed local intelligent data processing execution terminals are deployed on the production site as edge computing nodes. After acquiring the raw heterogeneous data from the intelligent hardware layer, key information is filtered and uploaded to the digital twin layer. At the same time, based on the real-time data collected locally, the parameters of the flexible production equipment in the intelligent hardware layer are adjusted at the workstation level according to the actual local status and combined with the global optimization instructions of the intelligent application layer, so that the real-time production cycle of each flexible production equipment meets the requirements.

9. The industrial cycle management system according to claim 1, characterized in that, It also includes a cloud server, which is equipped with a reinforcement learning beat optimization model. Combined with high-quality data output from the digital twin layer, it generates global optimization instructions covering the entire production line to guide the beat optimization scheme of the intelligent application layer. Preferably, the input data of the reinforcement learning beat optimization model includes customer order information, equipment health data and real-time energy consumption data, wherein the equipment health data is PHM predicted remaining lifetime data, and the output data is a dynamic beat compensation coefficient, wherein the dynamic beat compensation coefficient is within the range of ±15%.

10. An industrial cycle time management method based on the system of claim 9, characterized in that, Includes the following steps: 1) Utilize real-time data collected by the smart hardware layer and customer order information, equipment health data, and real-time energy consumption data obtained from the cloud server to construct a four-dimensional optimization model that includes customer needs, quality loss costs, equipment load balancing, and energy consumption and carbon emissions. Solve the optimal theoretical beat in real time through the reinforcement learning beat optimization model deployed on the cloud server and output the dynamic beat compensation coefficient. 2) The edge computing node operation action recognition model outputs the workstation operation time in real time, and combines the LSTM time prediction network to predict the process time fluctuation. The genetic algorithm generates a dynamic recombination scheme containing workstation sequence, equipment parameters and operation allocation. It receives the dynamic cycle compensation coefficient of the cloud server and the global optimization instructions of the intelligent application layer, and drives the flexible production equipment in the intelligent hardware layer to perform parameter adaptive adjustment. 3) The equipment data acquisition unit of the intelligent hardware layer transmits the original heterogeneous data of equipment status, operation actions and material trajectories to the digital twin layer. The spatiotemporal calibration system realizes spatiotemporal alignment based on the ST-CNN algorithm and builds a causal chain traceability model about "process-quality-logistics". Through the Transformer anomaly diagnosis model, multimodal features are integrated to output anomaly type classification results, root cause location conclusions and impact path analysis in real time. 4) Based on the 3D modeling module of the digital twin layer, the 1:1 virtual production line is constructed. The digital thread-driven scheduling engine uses Monte Carlo simulation, deep reinforcement learning and digital twin pre-validation to generate rescheduling schemes that are adapted to dynamic scenarios such as emergency order insertion and equipment failure through three-level collaboration and combined with the correlation analysis results output by the causal chain tracing model. The schemes are then synchronized to the intelligent application layer and edge computing nodes. 5) The quality-cycle time symbiosis optimization module establishes a Pareto optimization model of process capability index and cycle time deviation. Combining the anomaly diagnosis results and dynamic cycle time compensation coefficient, it uses a multi-objective particle swarm optimization algorithm to solve the optimal working point in real time. When the process capability index drops to a preset threshold, it automatically increases the sampling frequency and triggers a cycle time extension strategy. 6) Each edge computing node uploads the parameter adjustment effect and production cycle time compliance status to the cloud server. The cloud server summarizes the data of the entire system every day and updates the anomaly diagnosis model and reinforcement learning cycle time optimization model through federated learning to continuously improve the system's adaptive capability and cycle time optimization accuracy.

Images