A cloud-edge collaboration-based intelligent control method and system for an injection molding machine

Abstract

The application belongs to the technical field of injection molding machine control, and discloses an intelligent control method and system of an injection molding machine based on cloud-edge collaboration, constructs a three-dimensional hierarchical system of process characteristics, response requirements and priorities, classifies full-process control tasks of the injection molding machine, fixes deployment of high real-time tasks of core process control and equipment safety protection on the edge, ensures stable and reliable control of key processes, deploys non-real-time optimization tasks on the cloud, realizes global process optimization and production scheduling through cloud computing power, and dynamically adjusts the task deployment boundary based on network status and production conditions. A lightweight fusion control model is deployed on the edge, special control strategies are customized for each core process of injection molding, and a redundant backup mechanism is matched to avoid production interruption caused by equipment failure. Meanwhile, a cloud-edge collaborative federal incremental learning framework is built, the edge combines the local optimization model of individualized conditions of a single device, and the cloud completes multi-device model parameter aggregation and global optimization.

Description

Technical Field

[0001] This invention belongs to the field of injection molding machine control technology, specifically a cloud-edge collaborative intelligent control method and system for injection molding machines. Background Technology

[0002] Injection molding machines are the core equipment for injection molding. With the rapid advancement of intelligent manufacturing, traditional injection molding machine control modes can no longer meet the production needs of multi-variety, small-batch, high-precision, and low-energy-consumption production. Existing injection molding machine control technologies are mainly divided into two categories, both of which have unavoidable technical problems:

[0003] Pure cloud-based centralized control mode: All operating data of the injection molding machine is uploaded to the cloud, and process optimization and control command issuance are achieved through big data analysis and AI models in the cloud. However, the core processes of injection molding, such as injection and mold closing, have extremely high requirements for real-time control. Cloud transmission has inherent problems such as network latency, jitter, and packet loss, which cannot meet the requirements of hard real-time closed-loop control and can easily lead to control failure, equipment failure, or even safety accidents.

[0004] Currently, some cloud-edge collaborative injection molding machine control solutions have emerged in the industry, but most of them do not take into account the characteristics of the injection molding process for in-depth task decomposition and collaborative optimization. The core defects are: the task is not dynamically divided based on the real-time requirements of the injection molding process; the boundary between hard real-time control tasks and non-real-time optimization tasks is blurred, and it is impossible to balance control real-time performance and global optimization; the edge device only undertakes data acquisition and command execution functions, fails to achieve local intelligent closed-loop control, relies too much on the cloud, and is prone to control interruption when the network is abnormal.

[0005] The model training adopts a centralized cloud training and full distribution mode, which does not take into account the individual working conditions of a single injection molding machine, resulting in poor model adaptability; it lacks a sound disaster recovery backup and security management mechanism, and network interruption or single point of failure can easily lead to production stoppage, and the security protection capability of core data and models is insufficient. Summary of the Invention

[0006] The purpose of this invention is to provide an intelligent control method and system for injection molding machines based on cloud-edge collaboration, so as to solve one or more problems mentioned in the background art.

[0007] To achieve the above objectives, the present invention provides the following technical solution: an intelligent control method for injection molding machines based on cloud-edge collaboration, comprising the following specific steps:

[0008] Furthermore, the task deployment phase controls the entire process of the injection molding machine. Combining the characteristics of the injection molding process, response time requirements, and process importance, a three-dimensional hierarchical system of process, real-time performance, and priority is constructed. Hard real-time tasks have a response time ≤10ms and the highest priority. They include core closed-loop control such as mold closing position control, injection pressure control, injection speed control, and holding pressure control, as well as equipment emergency shutdown protection functions. The hard real-time tasks are directly related to product accuracy and equipment safety, and are bound to core process parameter thresholds and trigger protection mechanisms.

[0009] The soft real-time task response time is between 10ms and 1s, including adaptive adjustment of process parameters, real-time detection of product defects, online monitoring of equipment status and pre-diagnosis of minor faults. The system dynamically adjusts the soft real-time task response time threshold based on raw material characteristics and mold wear status.

[0010] Non-real-time task response time ≥1s, including global process parameter optimization, multi-machine collaborative production scheduling, mold wear prediction and defect root cause tracing;

[0011] A three-layer collaborative architecture is established, consisting of the edge layer, the terminal layer, and the cloud layer. The terminal layer comprises the injection molding machine body, various sensors, and actuators; the edge layer consists of the local edge controller of the injection molding machine and the workshop edge gateway; and the cloud layer is the industrial cloud platform. A fixed deployment and dynamic migration strategy is adopted. Hard real-time tasks are fixedly deployed on the edge controller, soft real-time tasks are deployed on the workshop edge gateway by default, and non-real-time tasks are fixedly deployed on the cloud platform. The system monitors network quality and process conditions in real time and dynamically adjusts the task deployment boundaries.

[0012] The system has a built-in three-level task priority preemption scheduling mechanism. Hard real-time tasks have the highest execution authority and computing resource allocation rights. When an execution request is initiated, they can immediately preempt system resources from soft real-time tasks and non-real-time tasks, suspending the computing power occupation of low-priority tasks to meet the hard real-time control requirement of ≤10ms. Soft real-time tasks have the next highest priority and are allocated computing resources only during the execution intervals of hard real-time tasks. They are responsible for tasks with medium time sensitivity, such as adaptive adjustment of process parameters and real-time defect detection. Non-real-time tasks have the lowest priority and are executed only when the system's computing power and network resources are completely idle. They are used for non-time-sensitive tasks such as global process optimization, multi-machine scheduling, and defect tracing.

[0013] Furthermore, the data preprocessing stage, based on the collaborative architecture and task deployment logic determined in the task deployment stage, synchronously collects multi-source heterogeneous data from the entire injection molding process through the edge controller of the edge layer and the workshop edge gateway. This includes injection molding machine operation data, process production data, online detection data, and environmental data. The injection molding machine operation data includes core parameters such as mold closing position, mold opening position, screw speed, screw position, injection pressure, holding pressure, and barrel temperature. The process production data includes raw material batches and mold parameters. The environmental data includes workshop temperature and humidity, and dust concentration.

[0014] Data standardization and preprocessing are performed locally at the edge. Hardware clock synchronization technology is used to add microsecond-level timestamps to all data. Abnormal interference data is removed through a triple noise reduction mechanism of moving average filtering, wavelet denoising and process anomaly threshold removal. The injection molding process-specific feature engineering module extracts key feature quantities that are strongly correlated with product quality and equipment status. Data dimensionality is reduced through feature importance ranking algorithm.

[0015] The preprocessed feature data is sorted according to real-time performance and process priority. Hard real-time control data is directly input into the local hard real-time control model of the edge controller, while soft real-time and cloud task data are uploaded to the cloud platform according to a preset cycle. At the same time, the data is cached in layers at the edge, and the cached data is encrypted and compressed at the edge.

[0016] Furthermore, the edge closed-loop control stage is based on the real-time key feature data of the edge obtained in the data preprocessing stage. An injection molding process adaptive control model that has been pre-trained in the cloud and optimized in a lightweight manner is deployed in the edge controller. The model adopts a control architecture that integrates a lightweight temporal Transformer network, a fuzzy PID algorithm and an injection molding process mechanism model to achieve dual control of data-driven and mechanism-constrained control. The key feature quantities collected in real time at the edge are used as inputs and the control parameters of each actuator of the injection molding machine are used as outputs.

[0017] Dedicated control strategies are implemented for each injection molding process. During the mold closing stage, the mold closing speed curve and clamping force are adjusted based on the mold closing position and clamping force data, combined with the mold material, and a segmented buffering strategy is adopted. During the injection stage, the melt flow state is predicted based on data including screw position and injection pressure. During the holding pressure stage, melt cooling and shrinkage are compensated. During the plasticizing and cooling stage, parameters are optimized based on the melting characteristics of the raw material.

[0018] The edge control redundancy backup module monitors the equipment status in real time. When it detects that the parameters exceed the threshold, it triggers emergency shutdown protection and starts the backup module. When the system detects a fault in the edge controller, it automatically switches to the temporary control mode of the workshop edge gateway.

[0019] Furthermore, in the model iteration stage, a federated incremental learning framework for edge-cloud collaboration is constructed based on the lightweight control model deployed in the edge-end closed-loop control stage and the real-time operation data of the edge. This enables the adaptive iterative update of the process model. The workshop edge gateway incrementally fine-tunes the locally deployed lightweight control model and defect detection model based on local real-time production data and the working conditions of a single injection molding machine, using a fixed production cycle as a unit. Only the high-level feature network parameters of the model are updated, and a working condition adaptation factor is introduced.

[0020] The edge uploads the incremental parameters of the model and the effect evaluation indicators to the cloud platform. The cloud platform uses a weighted federated average algorithm to aggregate the parameters. The weights are allocated according to the production scale and representativeness of each equipment. Combined with the cloud injection molding process knowledge base, the global model is optimized. The incremental parameters are then sent to the edge and a second fine-tuning deployment is completed.

[0021] The contribution of equipment is comprehensively judged based on four core indicators: production module, operational stability, working condition diversity, and quality compliance rate. Equipment with a larger production module, more stable operation, more comprehensive coverage of working conditions, and higher quality data has a higher contribution and a greater parameter aggregation weight; the weight is dynamically allocated in the cloud according to the contribution.

[0022] The aggregation strategy is adjusted through closed-loop verification until the model performance reaches the preset target, and an iteration log is established to record parameter changes and operating condition adaptation.

[0023] After the local model and the global model in the cloud complete parameter aggregation, a small-batch trial production verification will be performed on the edge to compare the control accuracy, product defect rate, energy consumption data and other core indicators before and after the model iteration. Only when the verification results meet the standards will the new model be officially used. If the verification fails, the system will automatically roll back to the previous stable model and upload the reason for the verification failure to the cloud, and readjust the parameter aggregation strategy and local fine-tuning scheme.

[0024] Furthermore, the multi-machine collaborative optimization stage is based on the global process model optimized in the model iteration stage and the desensitized data uploaded from the edge. The cloud platform constructs a global digital twin model of the injection molding workshop to simulate the operating status of the injection molding machines, changes in process parameters, and the product molding process. It also obtains the peak and valley periods of electricity consumption in the workshop, electricity price information, and the production tasks and equipment status of each injection molding machine in real time. With the three-dimensional optimization objectives of the lowest overall energy consumption, the shortest production delivery cycle, and the lowest product defect rate, the production plan and process parameters of each injection molding machine are dynamically adjusted.

[0025] The digital twin model and the physical injection molding machine adopt a millisecond-level synchronous mapping mechanism. The equipment status, process parameters, and action commands collected by the end-layer sensors are transmitted to the digital twin space in real time. The digital mirror synchronously replicates all the operating states and action processes of the physical equipment. When the physical equipment undergoes parameter adjustments, fault alarms, or state switching, the digital twin model will complete synchronous updates within 10ms to ensure that the digital space is completely consistent with the physical world.

[0026] During off-peak hours, high-energy-consuming, high-volume production tasks are prioritized, while process parameters are optimized to improve production efficiency. During peak hours, process parameters are optimized and adjusted to reduce equipment operating power, while balancing the production load of each injection molding machine. The mold sharing scheduling module dynamically allocates mold resources based on the production tasks and mold status of each injection molding machine to reduce replacement time.

[0027] Based on the full-cycle operation data of all equipment in the workshop, the wear status of molds and the remaining service life of core equipment components are predicted by integrating machine learning models with injection molding equipment mechanism models. Maintenance plans are generated in advance and distributed to management personnel. The equipment operation data and product quality data after maintenance are then transmitted back to the cloud platform.

[0028] Furthermore, the quality closed-loop control stage is based on the digital twin simulation capability of the multi-machine collaborative optimization stage and the molding data collected in real time at the edge. The workshop edge gateway collects melt filling status and product molding data in real time through in-mold vision inspection equipment and in-mold pressure sensors. The local lightweight defect detection model embeds a special injection molding defect detection algorithm to identify various molding abnormalities such as insufficient melt filling, flash, bubbles, shrinkage marks, and cracks in real time. Abnormal signals are synchronously input into the adaptive control model of the edge controller to adjust process parameters in real time.

[0029] When a molding defect is detected in a product at the edge, the defect type, defect severity, and the corresponding full-process parameters, equipment operation data, and raw material environment data before and after the defect are uploaded to the cloud platform. Based on the defect tracing knowledge base and the pre-trained root cause analysis model, the cloud platform uses big data correlation analysis and process mechanism reasoning to locate the root cause of the defect and generates corresponding process parameter correction schemes and equipment maintenance suggestions based on the root cause.

[0030] The defect data adopts an anomaly-triggered real-time upload mechanism. Once the edge defect detection model identifies a molding anomaly, it immediately packages the complete defect data and transmits it with priority, without waiting for the regular data upload cycle, thus minimizing the response time of cloud-based root cause analysis. The uploaded defect data package fully contains continuous working condition data from the 5 molds before the defect occurs to the 3 molds after the defect occurs, covering the complete cycle of defect incubation, generation, and molding. It also includes related information such as raw material batch, mold status, and environmental parameters to avoid root cause localization errors caused by data loss or gaps.

[0031] After digital twin simulation verification, the data is sent to the edge device. The edge device receives the correction plan, adjusts the process parameters in real time, and continuously collects product quality data in subsequent production cycles to verify the correction effect. The verification results are then sent back to the cloud to update the defect traceability knowledge base and analysis model.

[0032] Furthermore, the disaster recovery and security management phase constructs a multi-layered, end-to-end disaster recovery and security management system encompassing the edge, cloud, and terminal. Edge devices monitor the network communication quality with the cloud in real time, employing a triple architecture of main link, backup link, and emergency link. The main link is an industrial dedicated line, the backup link is a 5G / 4G redundant link, and the emergency link is a local area network backup. When network interruption, latency, or jitter exceeds a preset threshold, the edge device automatically switches to offline operation mode, continuing to complete closed-loop control of the entire process based on locally cached models, process parameters, and historical data. Simultaneously, all production data is encrypted and cached locally.

[0033] After the network is restored to normal, the edge device automatically re-uploads the data cached during the offline period to the cloud and synchronizes the model update and optimization scheme in the cloud. It switches back to online collaboration mode. Data and model security adopts a multi-layer protection strategy. The cloud physically isolates and controls the data of different enterprises and workshops, and adopts a three-level authorization mechanism of role, permission and data. Only authorized personnel can access the corresponding data. The edge cloud realizes dual backup of data and model and sets up off-site backup nodes.

[0034] Firewalls and intrusion detection systems are deployed at both the edge and the cloud. Security vulnerability scans and system upgrades are performed regularly. Triple identity verification is set up for model distribution and updates, including device verification, personnel authorization verification, and encryption verification. Data transmission is encrypted using national cryptographic algorithms. At the same time, a security early warning mechanism monitors security anomalies in data transmission, model updates, and device operation in real time, triggering early warnings and taking emergency response measures in a timely manner.

[0035] This invention also provides an intelligent control system for injection molding machines based on cloud-edge collaboration. Based on the above method, it adopts a three-layer architecture of end layer, edge layer, and cloud layer, with the following specific components:

[0036] The end layer is a sensing and execution module, which consists of the injection molding machine body, various sensors and execution mechanisms. The sensors include detection units such as mold closing position detection sensors, injection pressure detection sensors, and barrel temperature detection sensors, which collect real-time data on the entire injection molding process, including operation, process and environmental data. The execution mechanism receives control commands from the side layer and completes actions such as mold closing, injection, and pressure holding.

[0037] The edge layer is a local control and data processing module, which includes an edge controller and a workshop edge gateway. The edge controller deploys a lightweight adaptive control model and a redundant backup module to perform hard real-time closed-loop control of the entire injection molding process, monitor equipment operating parameters in real time, and trigger hard real-time emergency shutdown protection. The workshop edge gateway is used for data acquisition, standardized preprocessing, noise reduction, and feature extraction, completes data diversion and local caching, and simultaneously realizes local model incremental fine-tuning and data interaction with the cloud.

[0038] The cloud layer is a global optimization and security management module, consisting of an industrial cloud platform and supporting software. It deploys a global process model, a digital twin model, a federated incremental learning framework, and a disaster recovery and security management system. It receives incremental parameters and data uploaded from the edge layer, completes global model aggregation optimization, workshop multi-machine collaborative optimization, defect root cause analysis, and maintenance plan generation, while realizing data and model security protection, off-site backup, and emergency response to network anomalies.

[0039] The beneficial effects of this invention are as follows:

[0040] 1. This invention constructs a three-dimensional hierarchical system of process characteristics, response requirements, and priorities to classify the control tasks of the entire injection molding machine process. The core process control and equipment safety protection tasks with high real-time requirements are fixedly deployed at the edge to ensure stable and reliable control of key processes; non-real-time optimization tasks are deployed in the cloud, and global process optimization and production scheduling are achieved through cloud computing power; the system can dynamically adjust the task deployment boundaries based on network status and production conditions.

[0041] 2. This invention deploys a lightweight fusion control model at the edge, customizes dedicated control strategies for each core process of injection molding, and incorporates a redundancy backup mechanism to avoid production interruptions caused by equipment failure. At the same time, it builds an edge-cloud collaborative federated incremental learning framework, combining the local optimization model of individual equipment operating conditions at the edge with the cloud to complete the aggregation and global optimization of multi-equipment model parameters. The local real-time closed-loop control at the edge and the continuous model iteration in the cloud work together to improve the accuracy of process parameter control, reduce product defects, extend the service life of equipment and molds, and improve overall production quality.

[0042] 3. This invention enables collaborative scheduling of multiple devices in the workshop through cloud-based digital twins, dynamically optimizes production arrangements based on energy usage patterns, and coordinates production with energy consumption, efficiency, and yield as objectives, effectively reducing the overall production cost of the workshop; quality control forms a closed loop, with the edge end identifying molding anomalies in real time and making rapid adjustments, and the cloud end completing defect root cause analysis and solution optimization, continuously improving product quality; at the same time, it constructs a multi-layered disaster recovery and security protection system for the edge end, which can operate independently and stably when the network is abnormal, coupled with data backup and security verification mechanisms, to avoid production interruption and data security risks. Attached Figure Description

[0043] Figure 1 This is the overall flowchart of the cloud-edge collaborative intelligent control system for injection molding machines according to the present invention;

[0044] Figure 2 This is a flowchart of the edge closed-loop control sub-process of the present invention. Detailed Implementation

[0045] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0046] like Figures 1 to 2 As shown, this embodiment of the invention provides an intelligent control method for injection molding machines based on cloud-edge collaboration, including the following specific steps:

[0047] In this embodiment of the invention, the task deployment phase controls the entire process of the injection molding machine. Combining the characteristics of the injection molding process, response time requirements, and process importance, a three-dimensional hierarchical system of process, real-time performance, and priority is constructed. Hard real-time tasks have a response time ≤10ms and the highest priority. They include core closed-loop control such as mold closing position control, injection pressure control, injection speed control, and holding pressure control, as well as equipment emergency shutdown protection functions. The hard real-time tasks are directly related to product accuracy and equipment safety, and are bound to core process parameter thresholds and trigger protection mechanisms.

[0048] The soft real-time task response time is between 10ms and 1s, including adaptive adjustment of process parameters, real-time detection of product defects, online monitoring of equipment status and pre-diagnosis of minor faults. The system dynamically adjusts the soft real-time task response time threshold based on raw material characteristics and mold wear status.

[0049] Non-real-time task response time ≥1s, including global process parameter optimization, multi-machine collaborative production scheduling, mold wear prediction and defect root cause tracing;

[0050] A three-layer collaborative architecture is constructed, consisting of the edge layer, the terminal layer, and the cloud layer. The terminal layer comprises the injection molding machine body, various sensors, and actuators; the edge layer consists of the local edge controller of the injection molding machine and the workshop edge gateway; and the cloud layer is the industrial cloud platform. A fixed deployment and dynamic migration strategy is adopted. The hard real-time tasks are fixedly deployed on the edge controller to perform hard real-time closed-loop control and emergency protection. The soft real-time tasks are deployed on the workshop edge gateway by default, and the non-real-time tasks are fixedly deployed on the cloud platform. The system monitors network quality and process conditions in real time and dynamically adjusts the task deployment boundaries.

[0051] The three-dimensional grading system has fixed judgment criteria and an adaptive adjustment mechanism. The process characteristic dimension is divided into levels according to the core molding process, auxiliary operation process, and production management process. The real-time dimension is divided into levels according to the control command execution delay and data transmission timeliness. The priority dimension is divided into levels according to equipment safety assurance, product precision control, and production efficiency improvement. The grading results of the three types of tasks are automatically updated according to the actual production conditions such as raw material model change, mold type switch, and product specification adjustment. The system has a built-in injection molding process knowledge base, which can automatically match the grading thresholds corresponding to different products and raw materials.

[0052] The dynamic migration of tasks has clear triggering conditions and an uninterrupted execution process. When network latency is consistently high and data packet loss rate increases, the system automatically pushes some soft real-time tasks down to the edge gateway for execution. When network communication returns to stability and transmission quality meets the standards, the tasks are automatically migrated back to their original deployment location. The migration process first completes model loading, parameter synchronization and status calibration on the target node, and then closes the execution permissions of the original node.

[0053] The end-layer sensors and actuators are directly connected to the edge controller via an industrial bus to upload and collect data and receive control commands in real time. The edge controller and the workshop edge gateway communicate via a high-speed local area network to complete data forwarding, model interaction and task collaboration. The workshop edge gateway and the cloud-layer industrial cloud platform communicate via the industrial internet to complete data uploading, model distribution and scheduling command reception.

[0054] In this embodiment of the invention, the data preprocessing stage is based on the collaborative architecture and task deployment logic determined in the task deployment stage. It synchronously collects multi-source heterogeneous data of the entire injection molding process through the edge controller of the edge layer and the workshop edge gateway. This includes injection molding machine operation data, process production data, online detection data, and environmental data. The injection molding machine operation data includes core parameters such as mold closing position, mold opening position, screw speed, screw position, injection pressure, holding pressure, and barrel temperature. The process production data includes raw material batches and mold parameters. The environmental data includes workshop temperature and humidity and dust concentration.

[0055] The edge controller and workshop edge gateway have built-in multi-protocol adaptive unified conversion modules, which can automatically be compatible with mainstream industrial communication protocols in injection molding workshops such as Modbus, Profinet, EtherCAT, and CANopen. They can identify the differentiated data formats and sampling frequencies of different sensors, actuators, and injection molding machines, and automatically convert heterogeneous data such as voltage, current, temperature, pressure, and position into a unified standard data format, while also aligning data dimensions and unifying units.

[0056] Data standardization and preprocessing are completed locally at the edge. Hardware clock synchronization technology is used to add microsecond-level timestamps to all data to achieve accurate time synchronization of multi-source data. Abnormal interference data is removed through a triple noise reduction mechanism of moving average filtering, wavelet noise reduction and process abnormal threshold removal. The injection molding process-specific feature engineering module extracts key feature quantities that are strongly correlated with product quality and equipment status. Data dimensionality is reduced through feature importance ranking algorithm.

[0057] Data preprocessing is performed independently and locally at the edge. First, hardware clock synchronization technology is used to align the time of all collected data. Then, moving average filtering is used to remove random interference noise, wavelet denoising is used to remove high-frequency abnormal signals, and process abnormality thresholds are used to filter out invalid data that exceeds the range. The feature extraction stage focuses on core dimensions that are strongly related to the quality of injection molded products and the operating status of equipment, including five categories of features: position features, pressure features, temperature features, speed features, and time features. After extraction, effective core data is retained while reducing the amount of transmission and storage. Data distribution follows real-time and process priority rules. Hard real-time control data is transmitted through a dedicated channel and sent directly to the edge controller for immediate closed-loop control. Soft real-time data is transmitted in batches at set times. Non-real-time data is packaged and transmitted during network idle periods. Different types of data are transmitted through independent channels and with independent bandwidth.

[0058] The preprocessed feature data is sorted according to real-time performance and process priority. Hard real-time control data is directly input into the local hard real-time control model of the edge controller to meet the ≤10ms response requirement. Soft real-time and cloud task data are uploaded to the cloud platform according to a preset cycle. At the same time, the data is cached in layers at the edge and the cached data is encrypted and compressed at the edge.

[0059] Edge data encryption uses a lightweight symmetric encryption algorithm, which is adapted to the computing power of edge hardware. Compression adopts a lossless data compression method to preserve data characteristics to the greatest extent and reduce storage usage. The tiered cache is divided into a first-level real-time cache, a second-level periodic cache, and a third-level historical cache. The first-level cache stores the core real-time data of the current production cycle and supports millisecond-level reading. The second-level cache stores single-batch production data and synchronizes it to the cloud on a regular basis. The third-level cache stores historical production data and archives and backs it up regularly. Different levels of cache are managed independently and expired data is automatically cleaned up.

[0060] In this embodiment of the invention, the edge closed-loop control stage is based on the real-time key feature data of the edge obtained in the data preprocessing stage. An injection molding process adaptive control model that has been pre-trained and lightweight optimized in the cloud is deployed in the edge controller. The model adopts a control architecture that integrates a lightweight temporal Transformer network, a fuzzy PID algorithm and an injection molding process mechanism model to achieve dual control of data-driven and mechanism-constrained control. The key feature quantities collected in real time at the edge are used as inputs and the control parameters of each actuator of the injection molding machine are used as outputs.

[0061] The lightweight temporal Transformer network is a customized lightweight structure for injection molding processes. It consists of four interconnected modules: a temporal feature embedding layer, a single-head temporal attention layer, a point-by-point feedforward network layer, and a process constraint output layer. The embedding layer performs a 64-dimensional vector mapping on the input temporal features. The single-head attention layer focuses only on the temporal dependencies between injection molding processes. The hidden layer of the feedforward network has a dimension of 128, and the activation function is ReLU. The network output and the output of the fuzzy PID algorithm are weighted and fused with a weight ratio of 0.6:0.4. The fused result is input into the injection molding process mechanism model for physical constraint correction, and finally outputs the control quantities for each process: mold closing, injection, holding pressure, and plasticizing.

[0062] The model was pre-trained in the cloud using a full-condition injection molding dataset with a batch size of 32, an initial learning rate of 1e-4, and a cosine annealing learning rate scheduling strategy. The training rounds were 50, with the mean square error (MSE) < 0.001 and the response time ≤ 10ms as the dual convergence conditions. After pre-training, the model volume was compressed through quantization pruning.

[0063] The lightweight optimization of the model adopts a combined implementation process of weight quantization and channel pruning. First, the floating-point parameters of the model are converted into fixed-point parameters through 8-bit weight quantization to reduce the model's storage and computational overhead. Then, through a process feature-oriented channel pruning strategy, redundant neurons and invalid feature channels that do not contribute to the control effect of injection molding process are automatically removed, and only key network structures that are strongly correlated with core control parameters such as position, pressure, and temperature are retained. After pruning optimization, the model size is reduced, the inference speed is improved, and the memory usage and computing power consumption are adapted to the hardware configuration of the edge controller.

[0064] Dedicated control strategies are implemented for each injection molding process. During the mold closing stage, the mold closing speed curve and clamping force are adjusted based on the mold closing position and clamping force data, combined with the mold material. A segmented buffering strategy is adopted to avoid mold closing impact and mold damage, and to extend mold life. During the injection stage, the melt flow state is predicted based on data such as screw position and injection pressure to ensure uniform and stable melt filling. During the holding pressure stage, melt cooling and shrinkage are compensated to avoid defects such as shrinkage marks, flash, and insufficient glue in the product. During the plasticizing and cooling stage, parameters are optimized based on the melting characteristics of the raw materials to improve efficiency while ensuring plasticizing quality.

[0065] The mold closing process automatically adjusts the mold closing speed according to the mold size and material, with slow mold closing for positioning, fast mold closing for molding, and low-pressure mold protection to prevent mold impact damage; the injection process controls the injection speed in stages according to the screw position, with slow filling in the first stage, rapid molding in the middle stage, and stable pressure to prevent backflow in the last stage; the pressure holding process adjusts the pressure in stages, with high pressure compensation in the early stage and low pressure for shape maintenance in the later stage, adapting to the cooling and shrinkage rate of different raw materials; the plasticizing and cooling process adjusts the barrel temperature and screw speed according to the melting point of the raw material to ensure that the raw material is fully melted and does not degrade, and the cooling time is automatically adapted to the thickness of the product.

[0066] The edge control redundancy backup module monitors the equipment status in real time. When it detects that the parameters exceed the threshold, it triggers emergency shutdown protection and starts the backup module. When the system detects a fault in the edge controller, it automatically switches to the temporary control mode of the workshop edge gateway.

[0067] The redundancy backup mechanism adopts a dual-unit synchronous operation architecture of the main control unit and the backup control unit. The two units synchronize process parameters, equipment status and control commands in real time. When a fault such as parameter exceeding the threshold, module crash, or communication interruption is detected, the backup unit is immediately triggered to start. The switching time is controlled at the millisecond level, which does not affect the core control process.

[0068] The handover of control power between the primary and backup control units adopts a smooth and seamless switching process. Before the switchover, the backup unit has completed full parameter synchronization and status warm-up. During the switchover, the control command output is locked first, the data verification is completed, the control power is transferred instantly, and finally the command output is unlocked. After the control power is transferred, the backup unit immediately takes over the entire process control, and the faulty primary unit automatically enters the maintenance state.

[0069] As a secondary redundant backup, the workshop edge gateway only takes over hard real-time control permissions when the edge controller fails completely and cannot be recovered. After taking over, it automatically restricts the operation of non-essential auxiliary functions and retains only the core process control functions such as mold closing, injection, pressure holding, and plasticizing, so as to maximize equipment safety and production continuity. After the fault is cleared, the system automatically completes edge data synchronization, parameter calibration, and control permission revert.

[0070] In this embodiment of the invention, the model iteration stage is based on the lightweight control model deployed in the edge-end closed-loop control stage and the real-time operation data of the edge end to construct a federated incremental learning framework for edge-cloud collaboration, so as to realize the adaptive iterative update of the process model. The workshop edge gateway takes a fixed production cycle as the unit, and the production cycle can be dynamically adjusted according to the batch of products and the change of raw materials. Based on the local real-time production data and the working conditions of a single injection molding machine, the locally deployed lightweight control model and defect detection model are incrementally fine-tuned. Only the high-level feature network parameters of the model are updated, and the working condition adaptation factor is introduced to ensure that the model fits the characteristics of a single machine.

[0071] The operating condition adaptation factor is a parameter adjustment coefficient customized for the individual characteristics of a single injection molding machine. It comprehensively considers personalized operating condition information such as the equipment's operating years, mechanical wear, differences in installation environment, raw material usage habits, and mold replacement frequency. It is automatically calculated and generated by the edge based on the equipment's real-time operating data and historical operating condition records. During the local model incremental fine-tuning process, the operating condition adaptation factor can dynamically adjust the weight ratio of the model network parameters, so that the local control model and defect detection model fit the actual operating state of a single piece of equipment.

[0072] The edge device uploads incremental model parameters and performance evaluation indicators to the cloud platform. These performance evaluation indicators include product defect rate, control accuracy, and energy consumption level. The cloud platform uses a weighted federated average algorithm to aggregate parameters, with weights allocated according to the production scale and representativeness of each device's operating conditions. It then optimizes the global model by combining the cloud-based injection molding process knowledge base, which contains optimal solutions for different raw materials, molds, and products. The incremental parameters are then distributed to the edge device for secondary fine-tuning and deployment. The aggregation strategy is adjusted through closed-loop verification until the model performance reaches the preset target. An iteration log is established to record parameter changes and operating condition adaptation.

[0073] The formula for calculating the weights in the weighted federated average algorithm is as follows:

[0074]

[0075] Indicates the first Weighting coefficients of injection molding machines in cloud parameter aggregation;

[0076] Indicates the first The actual number of modules produced by the injection molding machine on that day;

[0077] Indicates all within the workshop Total daily production modulus of injection molding machines in Taiwan;

[0078] This indicates the serial number of the injection molding machine in the workshop. =1, 2, 3, ... ;

[0079] This indicates the total number of injection molding machines participating in federated learning within the workshop.

[0080] Local incremental fine-tuning only updates the parameters of the top fully connected layer and attention layer of the model, freezes the bottom feature extraction layer to retain general process features, fixes the local fine-tuning learning rate at 1e-5, and the fine-tuning cycle is one complete production batch. After each batch is completed, the incremental parameters and three evaluation indicators, namely defect rate, control accuracy and energy consumption, are uploaded.

[0081] The cloud aggregation cycle is 24 hours. After aggregation, it is combined with the cloud injection molding process knowledge base for global optimization. Then, the incremental parameters are sent to the edge to perform 100 steps of secondary fine-tuning, and the model deployment and effect verification are completed simultaneously. The system automatically records the model iteration log, including parameter changes, working condition adaptability, and performance improvement.

[0082] The cloud-based injection molding process knowledge base includes a standard process parameter library, a historical best solution library, and an anomaly handling case library corresponding to different raw material types, mold specifications, and product structures. The system automatically matches the corresponding process data based on the current raw material model, mold number, and product size. The knowledge base supports automatic updates, continuously recording high-quality parameters and production-verified process solutions after each model iteration.

[0083] Edge-cloud collaborative federated incremental learning fully protects the data privacy of individual devices. The edge only uploads incremental model parameters and performance evaluation indicators, without uploading original production data or device privacy information. When aggregating parameters in the cloud, priority is given to device parameters with large production data volume, strong representativeness of operating conditions, and high operational stability. The iterated model can quickly adapt to the personalized operating conditions of different equipment, raw materials, and molds, improving control accuracy and defect identification capabilities. The iteration log automatically generated by the system fully records the parameter changes, operating condition adaptability, and performance improvement.

[0084] In this embodiment of the invention, the multi-machine collaborative optimization stage is based on the global process model optimized in the model iteration stage and the desensitized data uploaded from the edge. The cloud platform constructs a global digital twin model of the injection molding workshop to realize the real-time mapping between physical equipment and digital twins, simulate the operating status of injection molding machines, changes in process parameters, and product molding process, and obtain the peak and valley periods of electricity consumption in the workshop, electricity price information, and the production tasks and equipment status of each injection molding machine in real time. With the three-dimensional optimization objectives of the lowest overall energy consumption, the shortest production delivery cycle, and the lowest product defect rate, the production plan and process parameters of each injection molding machine are dynamically adjusted.

[0085] The workshop-wide digital twin model comprises four modeling dimensions: equipment operation layer, process control layer, production scheduling layer, and energy consumption management layer. It can collect real-time data from all dimensions of physical equipment to complete digital mirror mapping. The model can fully simulate the entire injection molding process, dynamic changes in process parameters, product molding effects, and energy consumption. All scheduling decisions are first simulated and verified in the digital space to ensure there are no process conflicts, no abnormal equipment loads, and no quality risks before being sent to physical equipment for execution. At the same time, production tasks are flexibly allocated based on the workshop's peak and off-peak electricity consumption periods and electricity price information. High-energy-consuming, high-volume production is prioritized during off-peak periods, while parameters are optimized during peak periods to reduce energy consumption, balance the production load of each piece of equipment, and simultaneously optimize mold resource scheduling to reduce mold changeover time.

[0086] The three-dimensional optimization objectives adopt an adaptive dynamic weight allocation mechanism. The system automatically adjusts the weight ratio of each objective according to real-time production needs: when production order delivery is tight, the weight of production delivery cycle is increased to prioritize production efficiency and order delivery; when it is during peak power grid consumption, the weight of energy consumption control is increased to prioritize reducing the overall energy consumption of the workshop; when product quality fluctuates greatly, the weight of product yield rate is increased to ensure molding quality; the system autonomously decides the weight allocation by combining multi-dimensional information such as production plan, electricity price information, equipment status, and quality data, and outputs the globally optimal scheduling scheme that best fits the current production scenario.

[0087] During off-peak hours, high-energy-consuming, high-volume production tasks are prioritized, while process parameters are optimized to improve production efficiency. During peak hours, process parameters are optimized and adjusted to reduce equipment operating power, while avoiding equipment idleness and overload operation. The production load of each injection molding machine is balanced. The mold sharing scheduling module dynamically allocates mold resources based on the production tasks and mold status of each injection molding machine to reduce replacement time.

[0088] Based on the full-cycle operation data of all equipment in the workshop, the wear status of molds and the remaining service life of core equipment components are predicted by integrating machine learning models with injection molding equipment mechanism models. Maintenance plans are generated in advance and distributed to management personnel. The equipment operation data and product quality data after maintenance are transmitted back to the cloud platform to continuously optimize the prediction model.

[0089] The equipment wear prediction model is a fusion structure of the LightGBM regression model and the equipment mechanism model. The input features are filtered by injection molding process-specific feature engineering and include core features such as total number of mold closing times, peak and fluctuation values ​​of injection pressure, mean and deviation of barrel temperature, running time, and material wear coefficient. The model output includes three results: mold wear percentage, remaining service life of the three core components (connector, screw, and heating coil), and wear warning level.

[0090] The model training uses full-cycle operation data of all equipment in the workshop. Before training, outlier removal and feature standardization are completed. The LightGBM model tree depth is set to 6, the learning rate is 0.1, and the subsample ratio is 0.8. The convergence conditions are prediction mean absolute error (MAE) < 5% and life prediction deviation ≤ 8 hours. When the mold wear is > 85% and the remaining life of the component is < 72 hours, the model automatically triggers a high-level maintenance warning and generates a maintenance plan and parts replacement suggestions.

[0091] The maintenance plan includes the maintenance equipment number, maintenance component name, maintenance content, optimal maintenance time, and parts replacement recommendations. After being sent to the operation and maintenance terminal, the execution status is recorded synchronously. After the maintenance is completed, the management personnel enter the maintenance results. After the equipment is put back into operation, the system automatically collects operation data and quality data and feeds them back to the cloud to optimize the equipment wear prediction model.

[0092] In this embodiment of the invention, the quality closed-loop control stage is based on the digital twin simulation capability of the multi-machine collaborative optimization stage and the molding data collected in real time at the edge. The workshop edge gateway collects melt filling status and product molding data in real time through in-mold visual inspection equipment, in-mold pressure sensors, etc. The local lightweight defect detection model embeds a special injection molding defect detection algorithm to identify various molding abnormalities such as insufficient melt filling, flash, bubbles, shrinkage marks, and cracks in real time. Abnormal signals are synchronously input into the adaptive control model of the edge controller to adjust the process parameters of injection, holding pressure and other stages in real time.

[0093] The defect detection model is a lightweight YOLOv5n improved structure adapted to in-mold inspection scenarios, including an image preprocessing module, a C2f lightweight backbone layer, a PAN neck fusion layer, and a defect detection output layer. The input data consists of three types of fused data: in-mold visual 640×640 RGB images, in-mold pressure time-series curves, and temperature time-series features. After normalization and channel concatenation, the data is input into the model. The output terminal simultaneously outputs four results: defect category, defect pixel coordinates, defect confidence, and defect area ratio, covering six major categories of common injection molding defects: missing glue, flash, bubbles, shrinkage marks, cracks, and deformation.

[0094] The model training uses a labeled injection molding defect sample set, which includes defect data under different raw materials, molds, and process parameters. The batch size is 16, the initial learning rate is 5e-4, an early stopping strategy is adopted to avoid overfitting, and the training rounds are 30. The convergence criteria are defect recognition accuracy ≥95%, recall ≥93%, and single frame inference time ≤50ms. The model inference results are directly mapped to process parameter adjustment instructions.

[0095] The real-time edge defect correction adopts a targeted adjustment strategy, which adjusts the corresponding process parameters for different defect types to avoid large-scale parameter fluctuations affecting production stability. For missing glue defects, the injection pressure and injection speed are increased; for flash defects, the holding pressure and clamping force are reduced; and for bubble defects, the barrel temperature and cooling time are optimized.

[0096] The process modification plan generated in the cloud needs to be simulated and executed multiple times through digital twins to verify the quality improvement effect, production stability and energy consumption fluctuation. Only after confirming that the simulation defect rate is lower than the preset standard and the energy consumption fluctuation is within a reasonable range can it be sent to the edge for execution, so as to ensure that the product quality can be improved quickly after the plan is implemented.

[0097] After the corrective solution is implemented, the system continuously collects product quality data, equipment operation data, and process parameter data for subsequent continuous production cycles, compares various indicators before and after the defect correction, and automatically calculates the improvement rate and effectiveness of the solution. Valid corrective solutions are automatically included in the cloud-based defect traceability knowledge base. For solutions with poor verification results, the system will re-trigger root cause analysis and solution optimization, and continue iterating until the defect is resolved.

[0098] When a molding defect is detected in a product at the edge, the defect type, defect severity, and the corresponding full-process parameters, equipment operation data, and raw material environment data before and after the defect are uploaded to the cloud platform. Based on the defect tracing knowledge base and the pre-trained root cause analysis model, the cloud platform uses big data correlation analysis and process mechanism reasoning to locate the root cause of the defect and generates corresponding process parameter correction schemes and equipment maintenance suggestions based on the root cause.

[0099] The defect root cause analysis model is a fusion structure of decision tree classification and Bayesian probabilistic inference. The input includes six types of data: defect type, defect severity, process parameter sequences from the previous and next 10 models, equipment status data, raw material type, and environmental temperature and humidity. The model first performs initial root cause screening using decision trees, then calculates the probability of each potential root cause using Bayesian inference, outputting a root cause probability ranking and confidence level. The model's inference confidence threshold is set to 0.8; root causes with a probability > 0.8 are directly identified as core root causes, while those with a probability between 0.5 and 0.8 are classified as secondary root causes. Simultaneously, a corresponding correction plan is matched using a defect tracing knowledge base. Root cause types cover four main categories: process parameter deviation, equipment failure, raw material anomaly, and environmental fluctuation. Correction plans include process parameter adjustments, equipment inspection points, and raw material replacement suggestions. After verification of feasibility through digital twin simulation, the plans are deployed to the edge for execution.

[0100] After digital twin simulation verification, the data is sent to the edge device. The edge device receives the correction plan, adjusts the process parameters in real time, and continuously collects product quality data in subsequent production cycles to verify the correction effect. The verification results are then sent back to the cloud to update the defect traceability knowledge base and analysis model.

[0101] The defect tracing knowledge base stores defects categorized by type. Each type of defect corresponds to a complete list of root causes, corrective measures, and verification results. Defect cases uploaded from the edge are automatically categorized and entered into the knowledge base after being verified as valid in the cloud. When the system calls the knowledge base, it accurately matches similar cases based on defect characteristics and operating condition data, quickly outputs the optimal solution, and records the implementation effect of the solution to continuously optimize the matching accuracy of the knowledge base.

[0102] In this embodiment of the invention, the disaster recovery and security management phase constructs a multi-layered, end-to-end disaster recovery and security management system encompassing the edge, cloud, and terminal. Edge devices monitor the network communication quality with the cloud in real time. The communication quality indicators include latency, jitter, and packet loss rate. A triple architecture of main link, backup link, and emergency link is adopted. The main link is an industrial leased line, the backup link is a 5G / 4G redundant link, and the emergency link is a local area network backup. When a network interruption, latency, or jitter exceeds a preset threshold, the edge device automatically switches to offline operation mode. Based on locally cached models, process parameters, and historical data, it continues to complete the closed-loop control of the entire process flow. At the same time, all production data is encrypted and cached locally.

[0103] In the offline operation mode, the system automatically activates the core parameter protection mechanism to lock key process parameters such as mold closing position, injection pressure, and holding time to prevent arbitrary modification, and executes the control process according to preset standard parameters; at the same time, it completely caches all local control commands, equipment operating status, and product quality data in millisecond-level time series, prohibiting unauthorized external commands from accessing and data tampering; after the network is restored, the system can completely trace back the entire production process in the offline stage.

[0104] After the network is restored to normal, the edge device automatically re-uploads the data cached during the offline period to the cloud and synchronizes the model update and optimization scheme in the cloud. It switches back to online collaboration mode. Data and model security adopts a multi-layer protection strategy. The cloud physically isolates and controls the data of different enterprises and workshops, and adopts a three-level authorization mechanism of role, permission and data. Only authorized personnel can access the corresponding data. The edge cloud realizes dual backup of data and model and sets up off-site backup nodes.

[0105] The three communication links automatically switch according to a fixed priority. The main industrial leased line link is used first. When the main link experiences delays, packet loss, or other abnormalities, it automatically switches to the 5G / 4G backup link. When the backup link is abnormal, it switches to the local area network emergency link.

[0106] The system monitors the core quality indicators of the three communication links in real time, such as latency, packet loss rate, and jitter. It automatically determines the availability of the links according to preset thresholds. When the main link meets the standards, it exclusively occupies communication resources. When the main link fails to meet the standards, it immediately switches to the backup link. If the backup link is still abnormal, it quickly switches to the emergency local area network.

[0107] After a network outage, the edge device automatically enters an offline autonomous operation mode, retaining all hardware real-time control functions, software real-time monitoring functions, and local data caching functions. Non-real-time data uploads and cloud command reception are suspended. The local cache adopts a cyclic overwrite strategy, prioritizing the storage of core process data, equipment operation data, and product quality data. After the network is restored, the edge device re-uploads all data from the offline period in chronological order, first synchronizing model parameters and process optimization schemes, and then uploading production records and quality data.

[0108] The off-site backup node is physically isolated from the local node and automatically synchronizes core data, global model, process knowledge base, and production history to the cloud on a regular basis. The backup frequency is set according to the importance of the data. The core model and parameters are backed up in real time, and production data is backed up on a scheduled basis. The backup data has an integrity verification mechanism, and in the event of data loss or damage, it can be quickly restored directly from the off-site backup node.

[0109] Firewalls and intrusion detection systems are deployed at both the edge and the cloud. Security vulnerability scans and system upgrades are performed regularly. Triple identity verification is set up for model distribution and updates, including device verification, personnel authorization verification, and encryption verification. Data transmission is encrypted using national cryptographic algorithms. At the same time, a security early warning mechanism monitors security anomalies in data transmission, model updates, and device operation in real time, triggering early warnings and taking emergency response measures in a timely manner.

[0110] The system performs triple identity verification for critical operations such as model distribution, command execution, and data access. This includes verification of the device's unique hardware identifier, verification of the operator's permission level, and verification of the data encryption key. Operations can only be executed if all three verifications pass. Security alerts cover four major scenarios: network attacks, data tampering, unauthorized access, and device anomalies. The alert levels are divided into three levels: prompt, warning, and emergency. Prompt-level alerts only record system logs, warning-level alerts automatically push reminder information to the operation and maintenance terminal, and emergency-level alerts immediately lock operation permissions, disconnect abnormal communication links, and start the backup system.

[0111] Security vulnerability scanning adopts a fully automatic scheduled scanning mode, which is executed during off-peak production periods without affecting normal production operations. The scan covers all dimensions of edge hardware, edge software, cloud platform, and communication links. System upgrades are divided into emergency security upgrades and routine functional upgrades. Emergency upgrades are executed immediately, while routine upgrades are scheduled to be completed during production breaks. The current system version is automatically backed up before the upgrade, and a quick rollback is possible if the upgrade fails.

[0112] This invention also provides an intelligent control system for injection molding machines based on cloud-edge collaboration. Based on the above method, it adopts a three-layer architecture consisting of an end layer, an edge layer, and a cloud layer, with the following specific components:

[0113] The end layer is a sensing and execution module, which consists of the injection molding machine body, various sensors and execution mechanisms. The sensors include detection units such as mold closing position detection sensors, injection pressure detection sensors, and barrel temperature detection sensors, which collect real-time data on the entire injection molding process, including operation, process and environmental data. The execution mechanism receives control commands from the side layer and completes actions such as mold closing, injection, and pressure holding.

[0114] The process involves data acquisition and uploading at the edge layer, inference and computation at the edge layer, transmission of control commands, action response of the actuator, and feedback of operational results. Each step is equipped with timing standards and latency thresholds. After completing an action, the actuator immediately sends back status feedback data, and the edge model quickly verifies the execution effect.

[0115] The edge layer is a local control and data processing module, which includes an edge controller and a workshop edge gateway. The edge controller deploys a lightweight adaptive control model and a redundant backup module to perform hard real-time closed-loop control of the entire injection molding process, monitor equipment operating parameters in real time, and trigger hard real-time emergency shutdown protection. The workshop edge gateway is used for data acquisition, standardized preprocessing, noise reduction, and feature extraction, completes data diversion and local caching, and simultaneously realizes local model incremental fine-tuning and data interaction with the cloud.

[0116] The cloud layer is a global optimization and security management module, consisting of an industrial cloud platform and supporting software. It deploys a global process model, a digital twin model, a federated incremental learning framework, and a disaster recovery and security management system. It receives incremental parameters and data uploaded from the edge layer, completes global model aggregation optimization, workshop multi-machine collaborative optimization, defect root cause analysis, and maintenance plan generation, while realizing data and model security protection, off-site backup, and emergency response to network anomalies.

[0117] It should be noted that, in this document, relational terms such as "first" and "second" are used only to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such process, method, article, or apparatus.

[0118] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. A cloud-edge collaborative intelligent control method for injection molding machines, characterized in that, Includes the following steps: During the task deployment phase, for the entire process control task of the injection molding machine, a three-dimensional hierarchical system is constructed based on the process requirements, response time requirements and process importance of each process. Different priority control task types are divided, and a three-layer collaborative architecture of end layer, edge layer and cloud layer is built. A strategy combining fixed deployment and dynamic migration is adopted, and the task deployment boundary is dynamically adjusted according to network quality and process conditions. In the data preprocessing stage, based on the aforementioned collaborative architecture and task deployment logic, multi-source heterogeneous data from the entire injection molding process are collected. Standardized preprocessing, time synchronization, noise reduction, and feature extraction are completed at the edge. The data is then distributed and transmitted according to real-time requirements and process priorities. In the edge closed-loop control stage, based on the real-time key feature data after edge preprocessing, a lightweight adaptive control model with multi-algorithm fusion is deployed on the edge controller. Dedicated control strategies are configured for the process characteristics of each injection molding process, and a redundant backup mechanism is provided. During the model iteration phase, an edge-cloud collaborative federated incremental learning framework is constructed based on the edge control model and real-time operation data. The local model is incrementally fine-tuned in combination with the personalized working conditions of a single injection molding machine. The global model is optimized through cloud parameter aggregation and iterative updates are completed through closed-loop verification. In the multi-machine collaborative optimization stage, a global digital twin model of the workshop is constructed based on the global optimization model and edge-desensitized data. Combined with peak and valley electricity consumption and equipment status, production is dynamically scheduled to predict equipment loss and generate maintenance plans. During the closed-loop quality control phase, based on digital twin simulation capabilities and edge forming data, edge detection models are used to identify forming anomalies and correct them in real time. Defect-related data are uploaded to the cloud to complete root cause analysis, generate correction plans, and verify and optimize them. During the disaster recovery and security management phase, a multi-layered disaster recovery and security system is constructed, encompassing the edge, cloud, and endpoints. This system employs a multi-link architecture and data backup strategy, with security protection and anomaly warning mechanisms at both the edge and cloud ends.

2. The intelligent control method for injection molding machines based on cloud-edge collaboration according to claim 1, characterized in that, During the task deployment phase, the three-dimensional hierarchical system is a three-dimensional system of process, real-time performance, and priority. Combining process characteristics, real-time requirements, and priority dimensions, control tasks are divided into three categories—hard real-time, soft real-time, and non-real-time—based on response time and process importance. Hard real-time tasks have a response time ≤10ms and the highest priority, corresponding to closed-loop control of core injection molding processes and emergency equipment shutdown protection functions. Soft real-time tasks have a response time of 10ms~1s, corresponding to adaptive adjustment of process parameters, real-time detection of product defects, online monitoring of equipment status, and pre-diagnosis of minor faults. Non-real-time tasks have a response time ≥1s, corresponding to global process parameter optimization, multi-machine collaborative production scheduling, mold wear prediction, and defect root cause tracing. The three-layer collaborative architecture consists of the injection molding machine body, various sensors and actuators at the end layer, the local edge controller of the injection molding machine and the workshop edge gateway at the edge layer, and the industrial cloud platform at the cloud layer. The fixed deployment and dynamic migration strategy is as follows: hard real-time tasks are fixedly deployed on the edge controller, soft real-time tasks are deployed on the workshop edge gateway by default, and non-real-time tasks are fixedly deployed on the cloud platform. The system monitors network quality and process conditions in real time and dynamically adjusts the task deployment boundaries.

3. The intelligent control method for injection molding machines based on cloud-edge collaboration according to claim 2, characterized in that, In the data preprocessing stage, multi-source heterogeneous data, including injection molding machine operation data, process production data, online detection data, and environmental data, are synchronously collected through the edge controller of the edge layer and the workshop edge gateway. The data preprocessing is completed locally at the edge, including adding microsecond-level timestamps to all data using hardware clock synchronization technology, removing abnormal interference data through a triple noise reduction mechanism of moving average filtering, wavelet noise reduction, and process anomaly threshold removal, and extracting key feature quantities and completing data dimensionality reduction through injection molding process-specific feature engineering. The preprocessed feature data is distributed according to real-time performance and process priority. Hard real-time control data is directly input into the lightweight adaptive control model of the edge controller, while the remaining data is uploaded to the cloud platform according to a preset cycle. At the same time, the edge device completes data encryption, compression and hierarchical caching.

4. The intelligent control method for injection molding machines based on cloud-edge collaboration according to claim 3, characterized in that, In the edge closed-loop control stage, the lightweight adaptive control model is a control model pre-trained and lightweight optimized in the cloud. It adopts a control architecture that integrates a lightweight temporal Transformer network, a fuzzy PID algorithm, and an injection molding process mechanism model to achieve dual control driven by data and constrained by mechanism. The model takes key feature quantities collected in real time at the edge as input and control parameters of each actuator of the injection molding machine as output. The dedicated control strategy matches the corresponding control logic to the process characteristics of each process in the entire injection molding process. The redundancy backup mechanism monitors the equipment status in real time. When the parameter exceeds the threshold, it triggers emergency shutdown protection and starts the backup control module. When the system detects an edge controller failure, it automatically switches to the temporary takeover control mode of the workshop edge gateway.

5. The intelligent control method for injection molding machines based on cloud-edge collaboration according to claim 4, characterized in that, During the model iteration phase, the edge-cloud collaborative federated incremental learning framework, with a fixed production cycle as the unit, allows the workshop edge gateway to incrementally fine-tune the locally deployed lightweight control model and defect detection model based on local real-time production data and the working conditions of a single injection molding machine. The edge device uploads incremental model parameters and performance evaluation indicators to the cloud platform. The cloud platform uses a weighted federated average algorithm to aggregate the parameters. The weights are allocated according to the production scale and representativeness of each device's operating conditions according to preset rules. Combined with the cloud-based injection molding process knowledge base, the global model is optimized. The incremental parameters are then distributed to the edge device for secondary fine-tuning and deployment. Through production closed-loop effect verification, the parameter aggregation strategy is dynamically adjusted until the model performance reaches the preset target. An iteration log is synchronously established in the cloud to record parameter changes and operating condition adaptation.

6. The intelligent control method for injection molding machines based on cloud-edge collaboration according to claim 5, characterized in that, In the multi-machine collaborative optimization stage, the workshop global digital twin model can simulate the injection molding machine's operating status, process parameter changes, and product molding process. The cloud platform uses three-dimensional optimization goals: lowest overall energy consumption in the workshop, shortest production delivery cycle, and lowest product defect rate. It combines real-time data on peak and off-peak electricity consumption, electricity prices, and the production tasks and equipment status of each injection molding machine to dynamically adjust the production plans and process parameters of each injection molding machine and simultaneously optimize mold resource scheduling. Based on the full-cycle operation data of all equipment in the workshop, it predicts the wear status of molds and the remaining service life of core equipment components by fusing machine learning models with injection molding equipment mechanism models. It generates maintenance plans in advance and sends them to the operation and maintenance terminal. After maintenance, the equipment operation data and product quality data are transmitted back to the cloud platform.

7. The intelligent control method for injection molding machines based on cloud-edge collaboration according to claim 6, characterized in that, During the closed-loop quality control phase, the workshop edge gateway collects melt filling status and product molding data in real time through in-mold detection equipment. It identifies injection molding abnormalities in real time through a local lightweight defect detection model. Abnormal signals are synchronously input into the lightweight adaptive control model of the edge controller to adjust process parameters in real time. When the edge detects a product molding defect, it uploads the defect-related full-process data to the cloud platform. Based on the defect tracing knowledge base and the pre-trained root cause analysis model, the cloud platform locates the root cause of the defect through big data correlation analysis and process mechanism reasoning, and generates corresponding process parameter correction schemes and equipment maintenance suggestions. After the corrected solution is verified by digital twin simulation, it is sent to the edge. The edge controller receives the data and adjusts the process parameters in real time. The edge continuously collects product quality data for subsequent production cycles to verify the effect of the corrected solution and sends the verification results back to the cloud to update the defect tracing knowledge base and analysis model.

8. The intelligent control method for injection molding machines based on cloud-edge collaboration according to claim 7, characterized in that, During the disaster recovery and security management phase, the edge-cloud multi-layered disaster recovery and security system adopts a triple communication architecture of main link, backup link, and emergency link. Edge devices monitor the network communication quality with the cloud in real time. When a network interruption, latency, or jitter exceeds a preset threshold, the edge device automatically switches to offline operation mode and continues to complete the closed-loop control of the entire process based on locally cached models, process parameters, and historical data. At the same time, all production data is encrypted and cached locally. After the network is restored, the edge device automatically retransmits the offline cached data, synchronizes the cloud model updates and optimization schemes, and switches back to online collaborative mode. Data and model security employ a multi-layered protection strategy. The cloud enables physical isolation and access control of data, while the edge cloud provides local and off-site dual backups of data and models. Security protection modules are deployed on both the edge and the cloud. The model distribution and updates are subject to triple identity verification of devices, personnel, and encryption. Data transmission is encrypted throughout the entire process, and a security early warning mechanism monitors security anomalies across the entire chain in real time and triggers emergency response procedures.

9. A cloud-edge collaborative intelligent control system for injection molding machines, based on the method described in any one of claims 1-8, characterized in that, The system adopts a three-layer collaborative architecture consisting of the endpoint layer, edge layer, and cloud layer, specifically including: The sensing and execution module, deployed at the edge layer, consists of the injection molding machine body, various sensors, and actuators. The sensors collect real-time data on the entire injection molding process, including operation, process, and environment. The actuators receive control commands from the edge layer and complete the actions of the entire injection molding process. The local control and data processing module, deployed at the edge layer, includes an edge controller and a workshop edge gateway. The edge controller deploys a lightweight adaptive control model and a redundant backup module to perform hard real-time closed-loop control of the entire injection molding process, monitor equipment operating parameters in real time, and trigger hard real-time emergency shutdown protection. The workshop edge gateway is used for data acquisition, standardized preprocessing, feature extraction, data splitting, and local caching, and completes local model incremental fine-tuning and data interaction with the cloud. The global optimization and security management module is deployed in the cloud and consists of an industrial cloud platform and supporting software units. It deploys a global process model, a digital twin model, a federated incremental learning framework, and a disaster recovery and security management system. It receives incremental model parameters and de-identified production data uploaded from the edge layer, and completes global model aggregation optimization, multi-machine collaborative optimization in the workshop, defect root cause analysis, and maintenance plan generation. The cloud synchronously realizes data and model security protection, off-site backup, and emergency handling of network anomalies.

Images