Intelligent injection molding monitoring method based on edge computing

Abstract

This invention discloses an intelligent injection molding monitoring method based on edge computing, relating to the field of edge computing technology. The edge gateway aligns and normalizes the cavity pressure, screw displacement, and melt temperature signals, then inputs a causal invariant feature vector obtained by feature extraction branch fusion trained with environmental invariance constraints. Based on the Mahalanobis distance between this vector and the local safety boundary map, the V / P switching critical point is determined. The compensation control quantity is calculated using a proportional-integral-derivative (PID) method and sent to the injection molding machine controller within 10 to 50 milliseconds via an end-to-end collaborative bus to complete closed-loop correction. Edge nodes can also asynchronously upload features to the cloud for prototype aggregation and updates based on information entropy, and activate local safety boundary map degradation control when the network deteriorates. This invention can achieve millisecond-level closed-loop control at the edge, improving cross-production line generalization capability and control continuity under network anomalies.

Description

Technical Field

[0001] This invention relates to the field of edge computing technology, specifically to an intelligent injection molding monitoring method based on edge computing. Background Technology

[0002] Injection molding injects molten plastic into the mold cavity through a screw, and completes the product molding through stages such as mold clamping, injection, pressure holding, cooling, and mold opening. The entire production cycle is usually from a few seconds to tens of seconds. Among them, the V / P switching point when switching from speed control mode to pressure control mode and the peak pressure of holding pressure are the core control variables that determine the dimensional accuracy, internal stress distribution and appearance defects of the product. This critical window period is usually only a few tens of milliseconds to tens of milliseconds.

[0003] Existing industrial IoT monitoring solutions generally adopt a layered architecture of sensor acquisition, fieldbus uploading, cloud server centralized analysis, and optimized parameter distribution. Field sensors are responsible for collecting multi-physical field signals such as mold cavity pressure, screw displacement, and melt temperature, and uploading them to the cloud data center via industrial Ethernet. The cloud server performs offline or near real-time quality prediction and anomaly identification on the uploaded data based on a pre-trained deep learning model, and distributes the identification results or optimized process parameter recommendations to the factory manufacturing execution system or human-machine interface for operators to adjust parameters before the start of the next production cycle. The structure of this solution includes a field sensor layer, an industrial network transmission layer, a cloud data processing and model inference layer, and a parameter distribution and human-machine interaction layer. These four layers are connected in sequence through a network to form a unidirectional serial data flow path.

[0004] In the above-mentioned scheme, data needs to undergo multi-level transmission and queuing processing through fieldbus, wide area network and cloud server. The total latency of communication and inference usually reaches hundreds of milliseconds or even seconds, while the effective decision window for V / P switching is only tens of milliseconds. This makes it impossible for the scheme to complete the closed-loop control of anomaly detection, compensation calculation, instruction issuance and execution correction within the current production cycle. It can only be used for post-event analysis or parameter optimization for the next cycle. Abnormal products have already been produced as scrap in the current cycle. At the same time, the existing edge monitoring model learns the correlation between features and quality labels based on black-box statistical fitting, without distinguishing between environmentally related random noise and components that truly reflect the causal relationship of quality. Once the mold, raw material batch or injection molding machine is changed, the model accuracy will drop significantly, making it difficult to migrate and reuse across multiple production lines. In addition, the edge gateway has limited computing power and labeled sample resources, and the industrial field network is subject to jitter and packet loss risks. The existing system is highly dependent on cloud connection. When the network is abnormal, the monitoring capability is completely lost. There is a lack of deterministic transmission and local security degradation guarantee mechanism. Therefore, we propose an intelligent injection molding monitoring method based on edge computing. Summary of the Invention

[0005] The purpose of this invention is to provide an intelligent injection molding monitoring method based on edge computing.

[0006] The technical problem this invention aims to solve is that existing cloud-based centralized injection molding monitoring solutions suffer from high overall communication and inference latency due to the need for multi-level transmission and queuing of data through fieldbus, wide area network, and cloud server. This makes it difficult to complete closed-loop control of anomaly detection, compensation calculation, instruction issuance, and execution correction within the critical decision window of tens of milliseconds during V / P switching. Consequently, abnormal products are often produced as scrap within the current production cycle. Furthermore, existing end-side monitoring models fail to distinguish between environmental noise components and components that truly reflect the causal relationship of quality in multiphysics signals, resulting in poor generalization ability of the model after mold, raw material batch, or machine change. The existing system is insufficient and difficult to migrate and reuse across multiple production lines. In addition, the existing system is highly dependent on cloud connectivity. Once the network jitters or is interrupted, the monitoring and control capabilities are completely lost. It lacks deterministic transmission and local security degradation guarantee mechanisms. Existing edge-side monitoring solutions are generally limited by edge computing power and labeled sample resources. It has long been believed in the field that edge-side models cannot simultaneously achieve millisecond-level real-time performance and stable generalization ability across molds / batches / machines. This invention breaks through this cognitive limitation by training the feature extraction branch through environmental invariance constraints and combining it with hard decision degradation of local security boundary maps. It simultaneously achieves millisecond-level closed-loop control and stable extraction of causal features across production lines at the edge.

[0007] To address the aforementioned technical problems, this invention provides an intelligent injection molding monitoring method based on edge computing, comprising: acquiring cavity pressure signals, screw displacement signals, and melt temperature signals respectively collected by a cavity pressure sensor, a screw displacement sensor, and a melt temperature sensor deployed at the injection molding machine site; aligning the timestamps of the three signals using an edge gateway, and normalizing them according to the historical mean and standard deviation of each signal channel to obtain an aligned multimodal input sequence; inputting the multimodal input sequence into a first feature extraction branch, a second feature extraction branch, and a third feature extraction branch respectively to obtain corresponding first local temporal features, second local temporal features, and third local temporal features, wherein the three feature extraction branches are based on historical samples of multiple environmental domains divided according to differences in mold, raw material batch, and machine, and are adjusted by applying environmental invariance constraints. The training yields a feature vector that filters out pseudo-correlation features related to the environment. The three local time-series features are concatenated along the channel dimension and processed by a fully connected fusion layer to obtain a causal invariant feature vector. Based on the mean vector and covariance matrix of the historical qualified sample features in the locally cached safety boundary map, the Mahalanobis distance between the causal invariant feature vector and the mean vector is calculated. When the Mahalanobis distance is greater than the preset judgment threshold, the abnormal critical point of V / P switching is determined at the current moment. Based on the error between the actual feature and the center of the target safety interval, the compensation control quantity is calculated using proportional, integral, and derivative methods. The compensation control quantity is encapsulated into a control command and sent to the injection molding machine controller within a time window of 10 to 50 milliseconds through the end-to-end collaborative bus to adjust at least one parameter in the holding pressure setpoint and the V / P switching time point, thus completing the closed-loop correction within a single cycle.

[0008] Based on this, the present invention also provides a cross-production line small sample knowledge co-evolution mechanism: edge nodes calculate the information entropy of locally cached causal feature samples according to category distribution, and determine the feature vectors with information entropy greater than a preset threshold as high information entropy feature vectors and asynchronously upload them to the cloud server. The cloud server calculates the aggregation weight based on the number of local samples uploaded by each edge node and the average information entropy, and performs weighted fusion on the local prototypes of each edge node to obtain the updated global process prototype, and distributes it to each edge node to complete the hot update of local model parameters in the background thread mode, so as to continuously absorb long-tail defect knowledge without interrupting the real-time control loop. At the same time, the present invention also provides a flexible degradation mechanism for network quality awareness: the edge gateway monitors the latency and packet loss rate of the edge cloud transmission link according to a preset monitoring cycle, and removes the dependence on cloud online verification when the link quality deteriorates. It calls the locally cached security boundary map to independently complete the anomaly judgment and compensation control quantity calculation, and uses time-sensitive network technology to mark the deterministic sending timestamp of the downlink control command. When the command has not arrived after the set delivery time limit, the injection molding machine controller automatically activates the locally preset safety pressure holding fallback curve to maintain the pressure holding parameters within the safe range.

[0009] Compared with the prior art, the beneficial effects of the present invention by adopting the above technical solution are as follows:

[0010] 1. This invention pushes the computing power of causal feature extraction, critical point detection, and compensation control down to the edge gateway, avoiding the high latency caused by data transmission and queuing through multiple levels of fieldbus, wide area network, and cloud server. It can complete the entire closed loop from feature extraction and anomaly judgment to compensation command issuance within a time window of 10 to 50 milliseconds, advance the anomaly interception to before the waste is generated, and realize real-time correction within a single cycle.

[0011] 2. This invention trains the feature extraction branch by applying environmental invariance constraints based on multi-environment domain samples, which can filter out pseudo-correlation noise caused by differences in molds, raw material batches and machines, and retain only causal invariant features that are stable across environments. This allows the end-side model to maintain monitoring accuracy without retraining when changing production lines or machines, and improves the model's cross-production line generalization ability.

[0012] 3. This invention filters and uploads only high information entropy feature vectors rather than original time-series curves or image data at edge nodes. This reduces the bandwidth consumption of edge cloud transmission and the storage pressure on the cloud, while enabling the cloud to continuously update the global process prototype based on scarce long-tail defect samples. Furthermore, it achieves continuous evolution of model knowledge through hot updates without interrupting the real-time inference service at the edge.

[0013] 4. This invention enables the system to perform hard-decision degradation control based on the locally cached security boundary map when the edge cloud link experiences jitter or interruption, by using a network quality-aware flexible degradation protocol and a time-sensitive network deterministic timestamp mechanism. It also enables the use of a local security pressure-holding fallback curve when instruction transmission times out, thereby avoiding mold damage or batch scrap due to network anomalies and improving the robustness and reliability of the system. Attached Figure Description

[0014] Figure 1 This is a schematic diagram of the entire process of the overall system edge-cloud in this embodiment of the invention;

[0015] Figure 2 This is a schematic diagram of the closed-loop process of local real-time monitoring of the edge gateway in an embodiment of the present invention;

[0016] Figure 3 This is a schematic diagram of the network adaptive degradation and end-to-cloud collaborative update process in an embodiment of the present invention. Detailed Implementation

[0017] The following is in conjunction with the appendix Figure 1-3 The specific embodiments of the present invention will be further described below. It should be noted that the description of these embodiments is for the purpose of helping to understand the present invention, but does not constitute a limitation of the present invention.

[0018] Furthermore, the technical features involved in the various embodiments of the present invention described below can be combined with each other as long as they do not conflict with each other.

[0019] In terms of overall system architecture, the edge computing-based intelligent injection molding monitoring method provided by this invention is jointly executed by a multi-physics sensor group deployed at the injection molding machine site, an edge gateway, an end-to-end collaborative bus, an edge cloud transmission link, and a cloud server. The multi-physics sensor group includes a cavity pressure sensor installed at the injection molding machine cavity, a screw displacement sensor installed at the screw position, and a melt temperature sensor installed at the barrel position. All three types of sensors are connected to the edge gateway via high-speed data lines. The edge gateway carries a multi-modal feature fusion module, a critical point detection and compensation control module, and a local safety boundary map caching module. The three modules are connected by an internal data bus to jointly complete causal invariant feature extraction and V / P switching. The edge gateway performs real-time processing functions such as critical point determination and compensation control quantity calculation. It is bidirectionally connected to the injection molding machine controller through the end-to-end collaborative bus to issue closed-loop compensation commands within a time window of 10 to 50 milliseconds. The edge gateway is also bidirectionally connected to the cloud server through the edge cloud transmission link to asynchronously upload high information entropy feature vectors and receive global prototype parameters issued by the cloud. The cloud server carries a prototype federation aggregation module and a global process prototype map module to perform weighted fusion of local prototypes uploaded by multiple edge nodes and update the global process prototype map. The above three-layer structure of field perception, edge real-time closed loop and cloud-edge asynchronous collaboration forms a causal calling relationship based on the order of data generation, processing and knowledge feedback.

[0020] In terms of causal invariant feature extraction, the edge gateway first aligns the cavity pressure signal, screw displacement signal, and melt temperature signal with a unified clock reference, and then normalizes them according to the historical mean and standard deviation of each signal channel to obtain the aligned multimodal input sequence. Subsequently, the aligned cavity pressure signal, screw displacement signal, and melt temperature signal are input into the first feature extraction branch, the second feature extraction branch, and the third feature extraction branch, respectively. Each branch is composed of a lightweight convolutional network to extract the local temporal feature representation of the corresponding signal channel, resulting in the first local temporal feature, the second local temporal feature, and the third local temporal feature. After concatenating the three local temporal features along the channel dimension, they are input into a fully connected fusion layer for processing to obtain the fused feature representation, which is denoted as the feature extraction network output. ;

[0021] The network parameters of the three feature extraction branches mentioned above are determined during the training phase in the following way: historical samples from different molds, different raw material batches, and different machines are labeled as different environment domains according to their source, forming an environment domain set. For each environment domain Calculate the feature extraction network in this domain With classifier Experience risk By adding invariance constraints and minimizing the following total loss function, we obtain the feature extraction network parameters with cross-environment stability:

[0022] ;

[0023] in, This represents the set of environment domains constructed during training. Representation of feature extraction network With classifier In the environmental domain Experience and risk This represents the gradient of the empirical risk with respect to the classifier weights when the classifier weights are fixed at 1.0. It is used to measure whether the classifier has reached its optimum in this environment domain. The closer the gradient is to zero, the closer the classifier is to its optimum in this environment domain. The weight coefficients of the invariance constraint term are used to balance the two optimization objectives of minimizing empirical risk and cross-environment invariance constraint. By jointly constraining the feature extraction network in multiple environmental domains, the features learned by the network after training retain only the stable causal invariant components across molds, batches, and machines, and eliminate the pseudo-correlation components that drift with environmental changes. The parameters of the trained feature extraction network are deployed to the edge gateway and used for forward inference on the real-time multi-physics data stream during the inference phase. The causal invariant feature vector is directly output for use by the subsequent critical point detection module. The inference process no longer requires the participation of environmental discrimination labels.

[0024] In terms of millisecond-level critical point detection and closed-loop compensation control, the edge gateway continuously receives the above-mentioned causal invariant feature vector sequence, and the feature vector at the current moment is denoted as... The edge gateway calculates the Mahalanobis distance between the current feature vector and the mean of the security boundary based on the statistical parameters of the locally cached security boundary map, and uses this distance as the critical point criterion.

[0025] ;

[0026] in, express The causal invariant eigenvector at time step, This represents the mean vector of the characteristics of historical qualified samples in the safety boundary map. The covariance matrix representing the characteristics of historical qualified samples. Its inverse matrix, express The distance between the moment feature and the center of the safety boundary is measured. The safety boundary map is composed of a set of safety threshold intervals for multi-physics parameters, statistically derived from historical qualified process window data by the edge gateway, and the mean vector. With covariance matrix All data are statistically derived from the characteristics of historical qualified samples in the historical qualified process window data and cached locally on the edge gateway for local fallback decision-making in case of network anomalies.

[0027] Set the judgment threshold ,when If the current time has reached the abnormal critical point of V / P switching, it is determined that the current time has reached the critical point. Otherwise, the data of the next time moment is monitored. After the critical point is determined to have been reached, the error between the current actual characteristics and the center of the target safe interval is used as the basis for the determination. The compensation control quantity is calculated using proportional, integral, and derivative methods:

[0028] ;

[0029] in, express The compensation control quantity output at any time is used to correct the holding pressure of the injection molding machine or the switching timing parameters. express The error between the actual characteristics at any given moment and the center of the target's safe zone. , , These are the proportional gain coefficient, integral gain coefficient, and derivative gain coefficient, all of which are preset fixed values. This represents the time interval between adjacent sampling moments; the edge gateway will compensate for the control quantity. Encapsulated as control commands, they are directly sent to the injection molding machine controller within a 10 to 50 millisecond time window via the end-to-end collaborative bus. The controller adjusts the holding pressure setpoint or V / P switching point in real time accordingly to complete the adaptive correction within the current cycle. In parallel with the command sending process, the edge gateway temporarily stores the original causal feature vectors involved in the current cycle in the local cache queue for asynchronous calls by the subsequent end-to-cloud collaborative update module. This temporary storage process does not affect the real-time performance of the current control loop.

[0030] In the cross-production line small sample long-tail defect edge-cloud co-evolution, edge nodes calculate the information entropy of locally cached causal feature samples according to their category distribution, which serves as a measure of the amount of long-tail defect information contained in the sample:

[0031] ;

[0032] in, Representing feature samples Information entropy corresponding to the category distribution This indicates that the sample was determined to be the first. The probability of class defects, This represents the total number of defect categories. A higher information entropy value indicates greater uncertainty in the sample category, making it more likely to be a long-tail defect sample and thus more valuable for uploading. Edge nodes filter out high information entropy feature vectors according to a preset information entropy threshold. Only this feature vector itself is asynchronously uploaded to the cloud server; the original stress curve or image data is not uploaded to reduce bandwidth consumption for edge data transmission. Each edge node calculates the mean vector of its local defect category features, which serves as its local prototype. And count the number of samples participating in the aggregation locally. Average information entropy of uploaded samples ;

[0033] The cloud server calculates the aggregation weight based on both data volume and information entropy for the local prototypes uploaded from each edge node.

[0034] ;

[0035] in, Indicates the first The weight of each edge node in this round of aggregation Indicates the first The number of samples participating in the aggregation from each edge node. Indicates the first The average information entropy of samples uploaded by each edge node This represents the total number of edge nodes participating in this round of aggregation. As shown in the expression, the aggregation weight is directly proportional to the product of the number of local samples and the average information entropy of the corresponding edge node. Furthermore, the sum of the aggregation weights of all edge nodes participating in this round of aggregation is always 1. This ensures that edge nodes with a larger number of samples and higher information entropy—that is, those with a larger amount of long-tail defect information—receive higher weights during the aggregation process. The cloud server then performs weighted fusion of the local prototypes of each edge node according to the aggregation weight, updating the global process prototype map.

[0036] ;

[0037] in, This represents the updated global process prototype. Indicates the first The edge node at the ... The local prototype of the wheel is updated. The cloud server distributes the updated global prototype parameters to each edge node. The edge node completes the hot update of local model parameters in the background thread without interrupting the current real-time inference process, thereby decoupling the model accuracy improvement and real-time inference service. Since the edge nodes do not communicate directly with each other, but only share knowledge indirectly through the cloud, the direct exchange of raw data between edge nodes is avoided, further reducing the risk of data transmission and privacy leakage.

[0038] In terms of flexible degradation and deterministic transmission under network jitter, the edge gateway periodically collects round-trip latency and packet loss rate metrics of the edge cloud transmission link and compares them with preset thresholds. When either latency or packet loss rate exceeds the preset threshold, it is determined that the network is in a state of transient congestion. The edge gateway automatically detaches from the dependence on cloud-based online verification and switches to local autonomous control mode. It calls the latest version of the security boundary map in the local cache and, combined with the aforementioned causal invariant feature vector, independently completes anomaly detection and compensation control quantity calculation without waiting for cloud feedback. For each control command sent to the injection molding machine controller, the edge gateway adopts... Time-Sensitive Networking (TSN) technology marks deterministic transmission timestamps and sets a timeout limit for instruction delivery. The injection molding machine controller detects the arrival time of the instruction. If the instruction arrives within the timeout limit, the compensation action is executed normally according to the instruction. If no valid instruction is received after the timeout limit, the injection molding machine's locally preset safety pressure holding fallback curve is automatically activated to maintain the current pressure holding parameters within the preset safety range, avoiding mold damage or batch scrap due to missing instructions. The edge gateway continuously monitors the link quality. When the link quality indicators recover to within the threshold and remain within the set period, it automatically exits the local autonomous control mode and re-activates the end-to-cloud collaborative working mode.

[0039] As an alternative implementation, in the causal invariant feature extraction stage, the aforementioned invariant risk minimization loss function can be omitted. Instead, a method based on nonlinear independent component analysis can be used to perform blind source separation on the cavity pressure signal, screw displacement signal, and melt temperature signal. The signal is decomposed into environmentally related interference source components and process-related causal source components. Only the causal source components are retained as input features for the subsequent critical point detection module. This can also achieve the technical effect of filtering out pseudo-correlation features and improving the generalization ability across production lines.

[0040] As an alternative implementation method, in the edge-cloud collaborative architecture, instead of using the aforementioned prototype federated aggregation method to update the global knowledge base, a gradient federated averaging method can be used. That is, each edge node uploads the gradient update amount calculated by the local model for high information entropy samples. The cloud performs a weighted average of the gradients uploaded by multiple edge nodes and updates the global model parameters uniformly. Then, the updated global model parameters are distributed to each edge node to replace the local model. This can also achieve the technical goal of cross-production line small sample knowledge collaborative evolution.

[0041] It should be noted that the sensor types, network structures, control algorithms, and numerical ranges in the above embodiments are merely examples provided to facilitate understanding of the technical solutions of this invention. Those skilled in the art can make appropriate adjustments to the specific parameters and implementation methods according to the actual production line configuration without departing from the technical concept of this invention, and all such adjustments should fall within the protection scope of this invention.

[0042] Although the present invention has been disclosed above with reference to preferred embodiments, it is not intended to limit the present invention. Any person skilled in the art can make possible changes and modifications without departing from the spirit and scope of the present invention. Therefore, any modifications, equivalent changes and alterations made to the above embodiments based on the technical essence of the present invention without departing from the content of the technical solution of the present invention shall fall within the protection scope defined by the claims of the present invention.

Claims

1. An intelligent injection molding monitoring method based on edge computing, characterized in that, include: Acquire the cavity pressure signal, screw displacement signal, and melt temperature signal respectively collected by the cavity pressure sensor, screw displacement sensor, and melt temperature sensor deployed on the injection molding machine site; The edge gateway timestamps the cavity pressure signal, screw displacement signal, and melt temperature signal, and normalizes them according to the historical mean and standard deviation of each signal channel to obtain the aligned multimodal input sequence. The multimodal input sequence is input into the first feature extraction branch, the second feature extraction branch, and the third feature extraction branch, respectively, to obtain the first local temporal feature, the second local temporal feature, and the third local temporal feature corresponding to the cavity pressure signal, the screw displacement signal, and the melt temperature signal, respectively. The first feature extraction branch, the second feature extraction branch, and the third feature extraction branch are based on historical samples from multiple environmental domains and are trained by calculating the empirical risk for each environmental domain and superimposing an invariant constraint term composed of the gradient of the classifier weight fixed value, and minimizing the total loss function. The first local temporal feature, the second local temporal feature, and the third local temporal feature are concatenated along the channel dimension and processed through a fully connected fusion layer to obtain a causal invariant feature vector. The Mahalanobis distance between the causal invariant feature vector and the mean vector is calculated based on the mean vector and covariance matrix of the historical qualified sample features in the locally cached security boundary map. When the Mahalanobis distance is greater than the preset judgment threshold, it is determined that the current time has reached the abnormal critical point of V / P switching. Based on the error between the actual feature corresponding to the causal invariant feature vector and the center of the target safe interval, the compensation control quantity is calculated using proportional, integral, and differential methods. The compensation control quantity is encapsulated as a control command and sent to the injection molding machine controller within a time window of 10 to 50 milliseconds via the end-to-end collaborative bus. This allows the injection molding machine controller to adjust at least one parameter, either the holding pressure setpoint or the V / P switching timing, according to the control command, thereby completing the closed-loop correction within a single cycle.

2. The intelligent injection molding monitoring method based on edge computing according to claim 1, characterized in that: The training process of the first feature extraction branch, the second feature extraction branch, and the third feature extraction branch includes: calculating the empirical risk of the feature extraction network and the classifier in each environment domain, and superimposing an invariant constraint term; obtaining the network parameters of the first feature extraction branch, the second feature extraction branch, and the third feature extraction branch by minimizing the total loss function formed by the sum of the empirical risk and the invariant constraint term; wherein, the invariant constraint term is determined according to the gradient of the empirical risk with respect to the classifier weight when the classifier weight is fixed to a set value, and is used to constrain the classifier learned by the feature extraction network in the multiple environment domains to remain optimal.

3. The intelligent injection molding monitoring method based on edge computing according to claim 1, characterized in that: The safety boundary map is composed of a set of safety threshold intervals for multi-physics parameters obtained by the edge gateway based on historical qualified process window data. The mean vector is calculated based on the characteristics of historical qualified samples, and the covariance matrix is ​​obtained based on the distribution statistics of the characteristics of the historical qualified samples.

4. The intelligent injection molding monitoring method based on edge computing according to claim 1, characterized in that: The calculation of the compensation control quantity using proportional, integral, and derivative methods includes: multiplying the proportional gain coefficient by the error as the proportional term; multiplying the integral gain coefficient by the cumulative amount of the error at multiple sampling times as the integral term; multiplying the derivative gain coefficient by the rate of change of the error between adjacent sampling times as the derivative term; and adding the proportional, integral, and derivative terms to obtain the compensation control quantity.

5. The intelligent injection molding monitoring method based on edge computing according to claim 1, characterized in that: Also includes: The edge gateway calculates the information entropy of locally cached causal feature samples according to the category distribution, and identifies causal invariant feature vectors with information entropy greater than a preset information entropy threshold as high information entropy feature vectors. The high information entropy feature vector is asynchronously uploaded to the cloud server, and the number of local samples and the average information entropy of the uploaded high information entropy feature vector are counted. The cloud server calculates the aggregation weight based on the number of local samples uploaded by each edge node and the average information entropy, and then performs weighted fusion on the local prototypes uploaded by each edge node according to the aggregation weight to obtain the updated global process prototype. The cloud server distributes the updated global process prototype to each edge node, and the edge nodes complete the hot update of local model parameters in the background thread mode.

6. The intelligent injection molding monitoring method based on edge computing according to claim 5, characterized in that: The aggregation weight is proportional to the product of the number of local samples and the average information entropy of the corresponding edge node, and the sum of the aggregation weights of each edge node participating in this round of aggregation is 1.

7. The intelligent injection molding monitoring method based on edge computing according to claim 1, characterized in that: Also includes: The edge gateway monitors the latency and packet loss rate of the edge cloud transmission link according to a preset monitoring cycle; When at least one of the latency and packet loss rate exceeds a preset threshold, the edge gateway detaches from the dependence on cloud-based online verification, calls the latest version of the security boundary map cached locally, and independently completes the anomaly determination and compensation control quantity calculation. Time-Sensitive Networking (TSN) technology is used to mark deterministic transmission timestamps for control commands sent to the injection molding machine controller, and a timeout limit for command delivery is set.

8. The intelligent injection molding monitoring method based on edge computing according to claim 7, characterized in that, Also includes: When the injection molding machine controller detects that the control command has not arrived within the timeout period, it activates the injection molding machine's locally preset safety pressure holding fallback curve to maintain the pressure holding parameters within the preset safety range. After the edge gateway detects that the latency and packet loss rate of the edge cloud transmission link have recovered to within the preset threshold and maintained the preset recovery confirmation period, it exits the local autonomous control mode and re-enables the end-cloud collaborative working mode.

9. An intelligent injection molding monitoring system based on edge computing, characterized in that, This includes multiphysics sensor arrays, edge gateways, and cloud servers; The multiphysics sensor group is deployed on the injection molding machine site and includes a cavity pressure sensor, a screw displacement sensor, and a melt temperature sensor to collect cavity pressure signals, screw displacement signals, and melt temperature signals. The edge gateway is connected to the multiphysics sensor group and includes a multimodal feature fusion module, a critical point detection and compensation control module, and a local safety boundary map caching module. The multimodal feature fusion module is used to perform timestamp alignment and normalization processing on the cavity pressure signal, screw displacement signal, and melt temperature signal, and then input the pre-trained feature extraction branch to obtain a causal invariant feature vector. The critical point detection and compensation control module is used to determine the critical point based on the Mahalanobis distance between the causal invariant feature vector and the safety boundary map cached in the local safety boundary map caching module, and to calculate the compensation control amount. The edge gateway is connected to the injection molding machine controller via an end-to-end collaborative bus, and is used to send the encapsulated compensation control quantity to the injection molding machine controller within a time window of 10 to 50 milliseconds. The edge gateway is also connected to the cloud server via an edge cloud transmission link. The cloud server includes a prototype federation aggregation module and a global process prototype map module, which are used to receive high information entropy feature vectors asynchronously uploaded by the edge gateway and send the aggregated and updated global process prototype parameters to the edge gateway.

10. The intelligent injection molding monitoring system based on edge computing according to claim 9, characterized in that: The edge gateway is also used to periodically monitor the latency and packet loss rate of the edge cloud transmission link. When at least one of the latency and packet loss rate exceeds a preset threshold, it removes the online verification dependency on the cloud server and calls the latest version of the security boundary map in the local security boundary map cache module to independently complete the anomaly judgment and compensation control quantity calculation. The injection molding machine controller is used to enable the locally preset safety pressure holding fallback curve when the downlink control command does not arrive within the set timeout limit.

Images