Intelligent control method and system for an extrusion apparatus

Abstract

The application discloses an intelligent control method and system of an extrusion device, comprising: acquiring an adjacency matrix of a physical topology graph constructed based on mechanical assembly distance and thermodynamic conduction coupling strength of each component of the extrusion device, and mapping multi-modal sensor data into graph node features; using a graph neural network to aggregate the graph node features in a spatial dimension to obtain a spatial coupling feature sequence; inputting the spatial coupling feature sequence into a structured time sequence state evolution model with linear calculation complexity to extract time sequence evolution features; based on the time sequence evolution features, synchronously outputting a multi-motor frequency fine-tuning vector and a temperature zone convergence prediction time through a joint decision network, dynamically triggering a start-up phase jump according to the convergence prediction time, and realizing collaborative control according to the frequency fine-tuning vector. The application explicitly encodes the spatial coupling relationship between device components by using a physical topology graph, and realizes high-precision space-time joint reasoning and control under low computing power consumption.

Description

Technical Field

[0001] This invention relates to the fields of industrial automation control and industrial artificial intelligence application technology, specifically to an intelligent control method and system for extrusion equipment. Background Technology

[0002] Currently, the production control of large industrial equipment such as plastic extruders mainly relies on traditional PLC hard-coded control or conventional AI control methods. Traditional PLC hard-coded control typically relies on fixed empirical delay parameters, such as a fixed 60-second wait after the spindle starts, without considering the actual temperature and humidity of the equipment. Furthermore, the independent linear acceleration of multiple motors makes them highly susceptible to transient overpressure or mechanical shock due to physical coupling. In addition, conventional AI control schemes often simply concatenate all sensor data into a one-dimensional array and input it into a neural network. This approach severs the topological distances and thermodynamic conduction relationships between the various temperature zones of the extruder, and between the screw and the die in real physical space. Simultaneously, the preheating cycle of extrusion equipment often lasts for tens of minutes. Traditional Long Short-Term Memory (LSTM) networks suffer from a severe "forgetting" effect when processing such ultra-long time series, and the computational complexity at the edge increases quadratically, making it difficult to meet the lightweight requirements of millisecond-level inference in industrial control computers. Therefore, existing technologies suffer from technical problems such as fragmented data representation, high computational requirements for long-time series modeling, and susceptibility to forgetting. Summary of the Invention

[0003] The purpose of this invention is to provide an intelligent control method and system for extrusion equipment. This invention utilizes a physical topology diagram to explicitly encode the spatial coupling relationships between equipment components, achieving high-precision spatiotemporal joint reasoning and control with low computational power consumption.

[0004] The technical solution of the present invention: an intelligent control method for an extrusion equipment, comprising the following steps: Step S1: Obtain the adjacency matrix of the physical topology graph. The adjacency matrix is ​​constructed based on the mechanical assembly distance and thermodynamic conduction coupling strength of each component of the extrusion equipment, and the multimodal sensor data is mapped into graph node features. Step S2: Use a graph neural network to aggregate the graph node features in the spatial dimension to obtain a spatially coupled feature sequence; Step S3: Input the spatial coupling feature sequence into a structured temporal state evolution model with linear computational complexity to extract temporal evolution features; Step S4: Based on the time-series evolution characteristics, the multi-motor frequency fine-tuning vector and temperature zone convergence prediction time are synchronously output through the joint decision network. The start-up stage jump is dynamically triggered according to the convergence prediction time, and coordinated control is achieved according to the frequency fine-tuning vector.

[0005] The edge weights of the adjacency matrix described above are constructed based on the following relationship: ; in, For nodes With nodes The mechanical assembly distance between them For nodes With nodes Thermodynamic conduction coupling strength between them and Configurable scaling factor; The edge weights of the adjacency matrix are symmetrically normalized to obtain a normalized adjacency matrix, which is used for spatial dimension aggregation in graph neural networks.

[0006] In the aforementioned method, the structured temporal state evolution model is a selective state space model. The discretization step size, discretization input matrix, and output projection matrix of the selective state space model are all adaptively generated based on the current time step input through linear projection and nonlinear activation, thereby achieving content-dependent memory selection. The selective state-space model completes state recursion and output projection with linear computational complexity through a hardware-aware parallel scanning algorithm.

[0007] The aforementioned method further includes: Binary masking tags are injected into multimodal sensor data. When a sensor disconnection is detected, a preset physical default value is filled in and the corresponding masking tag is set to invalid. Nodes with invalid masking tags are skipped from the normalization process. During the message passing process of the graph neural network, for nodes whose masking is invalid, the aggregation features of neighboring nodes are used for inference compensation to maintain the hidden state reasoning of that node.

[0008] The aforementioned method, wherein the preset physical default value includes: For the temperature sensor, the ambient temperature is set to 25.0℃, corresponding to the initial state assumption of the chiller; For the speed sensor, set it to zero speed 0.0 Hz, which corresponds to the default value in the shutdown state; For the barometric pressure sensor, set it to standard atmospheric pressure of 0.101 MPa.

[0009] The aforementioned method, which dynamically triggers the startup phase jump based on the convergence prediction time, includes: When the convergence prediction time is less than the preset trigger threshold, the fixed delay of the current stage is skipped and the process enters the next stage. At the same time, a minimum protection time for the stage is set to prevent premature jumps caused by false triggers.

[0010] In the aforementioned method, the preset trigger threshold is 5 seconds, and the minimum protection time for the stage is 10 seconds. The multi-motor frequency fine-tuning vector is compressed to [-1, 1] by the Tanh activation function and then multiplied by the physical limiting coefficient, so that the single-cycle change is limited to ±0.5 Hz / s.

[0011] The aforementioned method further includes: predicting the air pressure or temperature trend after a preset time period based on the time-series evolution characteristics; and outputting a smooth deceleration command in advance when the predicted value approaches the alarm threshold to achieve predictive error prevention intervention; the preset time period is 15 seconds.

[0012] In the aforementioned method, the adjacency matrix is ​​pre-generated by the cloud server based on the mechanical assembly distance and thermodynamic conduction coupling strength of each component of the extrusion equipment and then distributed to the edge; the graph neural network and the structured temporal state evolution model are trained by the cloud using historical operating data and then deployed to the edge.

[0013] An intelligent control system for an extrusion equipment includes: The topology construction module is used to obtain the adjacency matrix of the physical topology graph. The adjacency matrix is ​​constructed based on the mechanical assembly distance and thermodynamic conduction coupling strength of each component of the extrusion equipment, and maps multimodal sensor data into graph node features. The spatial feature aggregation module is used to aggregate the graph node features in the spatial dimension using a graph neural network to obtain a spatially coupled feature sequence. The temporal feature extraction module is used to input the spatially coupled feature sequence into a structured temporal state evolution model with linear computational complexity to extract temporal evolution features; The control decision module is used to synchronously output the multi-motor frequency fine-tuning vector and temperature zone convergence prediction time through a joint decision network based on the time-series evolution characteristics, dynamically trigger the start-up stage jump according to the convergence prediction time, and realize coordinated control according to the frequency fine-tuning vector.

[0014] Compared with the prior art, the present invention has the following beneficial effects: 1. This invention reconstructs a spatiotemporal diagram that conforms to the actual physical assembly structure of the equipment by constructing a physical topology diagram based on mechanical assembly distance and thermodynamic conduction coupling strength, thereby enabling the model to have physical interpretability in the prior art and solving the problem of data representation being fragmented in physical coupling relationships.

[0015] 2. This invention introduces a structured state-space model with linear computational complexity to replace the quadratic complexity of traditional LSTM calculations. It achieves lossless context memory for power-on cycles of more than 30 minutes on edge industrial control computers, meeting the lightweight requirement of millisecond-level inference latency.

[0016] 3. Based on spatiotemporal joint inference on the physical topology graph, this invention synchronously outputs multi-motor frequency fine-tuning vectors, eliminating air pressure spikes and mechanical impacts caused by independent acceleration, and improving product yield.

[0017] 4. This invention, through feature masking and graph structure spatial neighbor compensation, can maintain high confidence control output even when some non-core sensors are disconnected, thus eliminating erroneous shutdowns and improving production line continuity. Attached Figure Description

[0018] Figure 1 This is a flowchart of the overall system architecture and data flywheel closed loop of the present invention; Figure 2 This is the network topology diagram of the core architecture of the algorithm based on the spatiotemporal graph topology (GNN) and state space model (SSM) of this invention; Figure 3 This is a schematic diagram of the sensor data masking and fault-tolerant preprocessing flow of the present invention; Figure 4 This is a schematic diagram comparing the timing of the dynamic jump start-up phase of this invention with that of traditional PLC fixed delay control; Figure 5 This is a flowchart of the online hot update process for the shadow mode model. Detailed Implementation

[0019] The present invention will be further described below with reference to the accompanying drawings and embodiments, but this should not be construed as limiting the present invention.

[0020] Example 1: This example provides an intelligent control method for an extrusion equipment. This method aims to solve the problems in existing technologies, such as the lack of physical coupling due to isolated processing of sensor data, and the computational bottleneck and forgetting effect in long-term time-series modeling. By constructing a physical topology graph to explicitly encode spatial relationships and introducing a state-space model with linear complexity to process temporal features, it achieves high-precision spatiotemporal joint reasoning and control with low computational consumption.

[0021] Specifically, the method in this embodiment includes the following steps: Step S1: Obtain the adjacency matrix of the physical topology graph. The adjacency matrix is ​​constructed based on the mechanical assembly distance and thermodynamic conduction coupling strength of each component of the extrusion equipment, and the multimodal sensor data is mapped into graph node features.

[0022] In this step, the physical topology graph is an abstract representation of the physical structure of the extrusion equipment. Unlike existing technologies that simply concatenate sensor data into a one-dimensional array, this embodiment defines each component of the extrusion equipment as a node of the graph, and the physical relationships between components as edges of the graph. In this embodiment, the extrusion equipment includes 11 temperature zones (T1-T11), a spindle motor, an air pump motor, and a die head motor, totaling 14 physical nodes. Each node corresponds strictly one-to-one with the physical entities and sensors: nodes 0-10 correspond to the heating sections of temperature zones T1-T11, mapping temperature sensor data; node 11 corresponds to the spindle motor and transmission mechanism (Spindle), mapping speed and current sensor data; node 12 corresponds to the air pump motor and air pressure pipeline (AirPump), mapping speed and air pressure sensor data; and node 13 corresponds to the die head extrusion mechanism (DieHead), mapping speed and temperature sensor data. This "physical topology graph" not only includes the mechanical connections between components, but more importantly, it integrates implicit coupling relationships such as thermodynamic conduction. For example, although two adjacent temperature zones may be close in mechanical distance, their heat conduction intensity may be weaker due to the presence of insulation material; conversely, although the spindle motor and the die head are far apart in mechanical distance, they are strongly coupled through material transfer.

[0023] Preferably, the edge weights of the adjacency matrix are constructed according to the following relationship: ; in, For nodes With nodes The mechanical assembly distance between them For nodes With nodes Thermodynamic conduction coupling strength between them and This is a configurable scaling factor; the default value is... =1.0、 =0.8. For example, temperature zones T1 and T2 are adjacent in mechanical assembly, with a distance of d(T1,T2)=0.3 meters, and there is a strong thermal conduction relationship κ(T1,T2)=0.9. Then w(T1,T2)=1.0 / 0.3+0.8×0.9≈3.33+0.72=4.05, which is approximately 2.89 after normalization.

[0024] The edge weights of the adjacency matrix are symmetrically normalized to obtain a normalized adjacency matrix, which is used for spatial dimension aggregation in the graph neural network. The symmetric normalization process employs... ,in For degree matrix, This represents the original non-normalized adjacency matrix. This normalization ensures the stability of the graph convolution values ​​and eliminates degree distribution bias. It should be understood that the specific values ​​of the 14 nodes and the adjacency matrix mentioned above are only illustrative examples. In actual applications, they can be adjusted according to the actual component configuration and physical parameters of the device. This embodiment does not impose any limitations on this.

[0025] By constructing an adjacency matrix, these complex physical relationships can be quantified into mathematical expressions, enabling subsequent neural network models to learn based on real-world physical prior knowledge, thereby improving the model's generalization ability and interpretability. Subsequently, the system acquires real-time data from various sensors via a fieldbus (such as Modbus RTU), including temperatures in each temperature zone, motor speeds and currents, and air pump pressure, and maps this multimodal sensor data into graph node features. At this point, because the equipment is in a cold start-up state, the temperature sensor readings in each temperature zone are close to the ambient temperature of 25.0℃, and the speed sensor reading is 0.0 Hz.

[0026] Step S2: Use a graph neural network to aggregate the graph node features in the spatial dimension to obtain a spatially coupled feature sequence.

[0027] After constructing the physical topology and mapping the node features, this step utilizes a Graph Neural Network (GNN) to aggregate features in the spatial dimension. Through a message-passing mechanism, the GNN enables each node to aggregate feature information from its neighboring nodes. For example, for a specific temperature zone node on an extruder, its features not only include its own temperature sensor readings but also aggregate features from neighboring temperature zones, the spindle motor, and other neighboring nodes via the GNN, thus enabling it to perceive changes in the thermodynamic state of the surrounding environment. This spatial aggregation allows the model to capture the local coupling relationships between various components of the equipment, such as the physical phenomenon of a decrease in temperature in adjacent temperature zones due to an increase in spindle speed. Ultimately, the GNN outputs a spatial coupling feature sequence, where each feature vector incorporates physical coupling information within the local space of the equipment.

[0028] Step S3: Input the spatial coupling feature sequence into a structured temporal state evolution model with linear computational complexity to extract temporal evolution features.

[0029] Traditional recurrent neural networks (RNNs) or long short-term memory networks (LSTMs) typically exhibit computational complexity that increases quadratically with sequence length when processing long-term temporal data, and they also suffer from long-range forgetting problems. Extrusion equipment often has a warm-up cycle lasting tens of minutes, generating sensor data across thousands of time steps, making it difficult for traditional models to achieve real-time inference on edge industrial control computers. This embodiment introduces a structured temporal state evolution model with linear computational complexity, whose complexity is only linearly related to sequence length, significantly reducing computational resource consumption. Furthermore, by introducing a structured state matrix, this model can effectively memorize long-term historical information, avoiding the forgetting effect of traditional models on long sequences.

[0030] Figure 2 This is the network topology diagram of the core architecture of the algorithm based on the spatiotemporal graph topology (GNN) and state space model (SSM) of this invention. The network is divided into four layers from bottom to top: input layer, spatiotemporal graph GNN layer, SSM state space temporal layer, and joint decision output layer.

[0031] Input Layer: Accesses multimodal sensor timing data, including three types of raw timing curves: temperature zone, spindle operation, and air pump operation, corresponding to the measured sensor data of each physical node of the extruder; Spatio-Temporal Graph GNN Layer: Constructs a graph network based on the device's physical topology adjacency matrix, performs spatial coupling and aggregation on the features of each node, integrates the mechanical assembly and thermodynamic coupling relationships between components, and outputs spatial features; SSM (State Space Model Layer): A serial selective state space model is used to extract time-series features with linear complexity, replacing the traditional LSTM loop structure. This avoids long-term forgetting and secondary computational overhead, and transforms spatial features into spatiotemporal fusion time-series features. Joint Decision Output Layer: Through a multi-task joint decision network, the frequency fine-tuning of the three types of motors—spindle, air pump, and mold head—is simultaneously decoupled and output, thereby achieving multi-motor coordinated speed control.

[0032] Preferably, the structured temporal state evolution model is a selective state space model. The discretization step size, discretization input matrix, and output projection matrix of the selective state space model are all adaptively generated based on the current time step input through linear projection and nonlinear activation, realizing content-dependent memory selection. The selective state space model completes state recursion and output projection with linear computational complexity through a hardware-aware parallel scanning algorithm.

[0033] The selective state-space model is an improvement on the traditional state-space model. Traditional state-space models typically have fixed parameters, making it difficult to dynamically adjust the memory strategy based on different input content. The "selectivity" mechanism introduced in this embodiment allows the model's parameters to adaptively change according to the current input data. Specifically, the discretization step size... Discretized input matrix and output projection matrix It is no longer a fixed constant, but rather determined by the current input features. It is generated in real time through a linear projection layer and a nonlinear activation function (preferably SiLU, i.e., the Swish function). This design achieves "content-dependent memory selection," meaning the model can autonomously decide which information to remember and which to forget based on the importance of the input data at any given moment. For example, during the startup process of an extrusion device, when sensor data fluctuates drastically or reaches a critical turning point, the model can adaptively adjust the step size and input matrix to more effectively encode this key information into the hidden state; while during periods of stable data, the model can update the state more smoothly, thereby improving the responsiveness to critical events while maintaining long-term memory capabilities.

[0034] Furthermore, the selective state-space model completes state recursion and output projection with linear computational complexity through a hardware-aware parallel scanning algorithm. This contrasts sharply with traditional recurrent neural networks or long short-term memory networks. Long short-term memory networks, when processing long sequences, must perform recursive calculations step-by-step, making parallelization impossible. Their computational complexity increases quadratically with sequence length, and when the sequence length reaches thousands of steps (such as the 30-minute boot cycle of an extrusion device), gradient vanishing or exploding is highly likely, leading to severe "long-range forgetting." While the selective state-space model used in this embodiment still logically relies on time-series dependencies in its state recursion, by transforming the recursion process into matrix operations and utilizing the parallel computing capabilities of modern GPUs, combined with a hardware-aware parallel scanning algorithm, it can achieve linear computational complexity. The computation of the entire sequence is completed in linear time. This means that no matter how long the input time series is, the computational load is only proportional to the sequence length, greatly reducing the computing burden on edge industrial control computers and making it possible to process ultra-long time series data in real time at resource-constrained industrial edges. In terms of parameter quantity, under the same expressive power, the number of parameters in Long Short-Term Memory networks increases quadratically with the hidden layer dimension, while the number of parameters in Selective State Space models shows a better growth relationship with the key dimensions, resulting in lower storage and computational burden at the edge.

[0035] Specifically, the discretized state transition equations of the selective state-space model are as follows: First, input based on the current time step Adaptive generation of discretization parameters: Discretization step size ,in The value range is configured. and Parameters are constrained for optimal results. =0.001、 =0.1, ensuring that the step size is dynamically adjusted within a physically reasonable range; Discretized input matrix ,in From input Generates through linear projection; discretizes the state transition matrix. ,in The initialization method for the state matrix initialized for HiPPO-like applications is as follows: , Initialize to The repetition of this ensures the decay characteristics of long-range memory, and it is usually initialized to a negative value to ensure stable decay of the state; Output projection matrix Similarly, it is input Generated by linear projection.

[0036] Subsequently, the hidden state and output are recursively calculated using a hardware-aware parallel scanning algorithm: Hidden state update: ; Output projection: , in, The learnable skip connection parameter is initialized to 1.0.

[0037] Although the above recursive process is similar in form to that of a recurrent neural network, it is because... , , All are input The system is dynamically generated, and the computation process is optimized into parallel scanning operations. Therefore, while maintaining linear complexity, it achieves dynamic perception and selective memorization of the input content.

[0038] As a preferred option, the structural hyperparameters of the selective state-space model are set as follows: Input projection dimension The feature dimension equals the flattened output of the graph neural network, which is the number of nodes multiplied by the hidden layer dimension of the GNN (e.g., 14 nodes × 32 dimensions = 448); internal expansion dimension. ( =2), used to enhance internal expressive power; local causal convolution kernel size =4, used to capture local temporal patterns; state space dimension =N=64, this dimension setting ensures long-term memory capacity while keeping the computational load of single-step inference within the tolerance range of the edge control computer; the discretization method adopts the zero-order hold (ZOH) method, discretizing the continuous state equation into a recursive form to ensure numerical stability. This embodiment, by introducing this selective state-space model, not only solves the computational bottleneck and forgetting problem of traditional models in long-term modeling, but also improves the model's ability to capture key state changes through a content-dependent mechanism, providing an efficient and accurate means of extracting temporal features for the intelligent control of extrusion equipment.

[0039] Specifically, the model receives the spatially coupled feature sequence output in step S2, evolves and extracts the features in the time dimension, captures the dynamic law of equipment status changes over time, such as the temperature rise curve of the temperature zone, the fluctuation trend of air pressure, etc., and finally outputs the temporal evolution features containing spatiotemporal information.

[0040] Preferably, the method also includes feature masking and fault tolerance mechanisms to address situations such as sensor disconnection or data anomalies. Specifically, binary masking tags are injected into the multimodal sensor data. When sensor disconnection is detected, a preset physical default value is filled in and the corresponding masking tag is set to invalid. Nodes with invalid masking tags skip the normalization process. During message passing in the graph neural network, for nodes with invalid masking tags, inference compensation is performed using the aggregated features of neighboring nodes to maintain the hidden state inference of that node. Figure 3 This is a schematic diagram of the sensor data masking and fault-tolerant preprocessing process of the present invention, used to implement data preprocessing in scenarios of sensor disconnection / data anomaly. After receiving data from multiple sensors, the system first determines whether the data is abnormal or the sensor is disconnected. For the normal data branch: the true sampled value of the sensor is retained, and the corresponding node's Mask tag is assigned a value of 1. For the abnormal disconnection branch: abnormal data is discarded, and a preset physical default value is filled in, and the corresponding node's Mask tag is assigned a value of 0. The two data are then synthesized into a spatiotemporal graph feature tensor carrying a masking tag. The Z-Score normalization of the failed nodes with Mask=0 is skipped in the preprocessing stage. Finally, the preprocessed feature data is sent to the ST-GNN+SSM inference model. Subsequently, the GNN network uses the features of neighboring nodes to perform hidden state compensation for the failed nodes.

[0041] In industrial environments, it is common for sensors to experience disconnections or data anomalies due to aging wiring, loose connectors, or environmental interference. Traditional control schemes often trigger shutdown protection due to the loss of a single sensor, leading to production line interruptions. This embodiment introduces a feature masking mechanism, enabling the system to maintain high-confidence inference and control even when some sensors fail, greatly improving the system's robustness.

[0042] Specifically, binary masking labels are auxiliary input features used to identify the validity of data. For each sensor node, the system adds a dimension to the input data vector to indicate whether the sensor data of that node is valid. For example, when the masking label is 1, it means that the sensor data of that node is normal, and the model can use the data for inference normally; when the masking label is 0, it means that the sensor data of that node is missing or abnormal, and the model will ignore the original data of that node and rely on other information for inference instead.

[0043] When the edge detects a sensor disconnection or data exceeding the physically reasonable range, the system automatically executes a fault-tolerant processing procedure. First, it fills in preset physical default values ​​to replace the missing sensor readings. These preset physical default values ​​are not arbitrarily set, but are based on assumptions about the physical state of the equipment. For example, for the temperature zone sensor, it is set to an ambient temperature of 25.0℃, corresponding to the initial cold-running state assumption; for the spindle speed sensor, it is set to zero speed (0.0 Hz), corresponding to the default value in the shutdown state; for the spindle current sensor, it is set to zero current (0.0A), corresponding to the default value in the shutdown state; for the air pump speed sensor, it is set to zero speed (0.0 Hz), corresponding to the default value in the shutdown state; for the air pump pressure sensor, it is set to standard atmospheric pressure (0.101 MPa); for the die head speed sensor, it is set to zero speed (0.0 Hz), corresponding to the default value in the shutdown state; for the die head temperature sensor, it is set to an ambient temperature of 25.0℃; for the ambient temperature sensor, it is set to the local annual average workshop temperature of 25.0℃; for the cooling water temperature sensor, it is set to the industrial cooling water standard temperature of 20.0℃. The choice of these default values ​​is intended to provide the model with a relatively safe and physically reasonable initial state input, avoiding drastic fluctuations in the model output caused by all-zero inputs or random noise.

[0044] Secondly, the system invalidates the masking flag of the corresponding node (e.g., sets it to 0). This operation is crucial, as it informs subsequent neural network modules that the input data for this node is not the actual observation value, but rather a default value that has been filled in. Therefore, the reliance on the node's own features should be reduced during feature extraction.

[0045] Furthermore, during the data preprocessing stage, the system skips the normalization process for nodes marked as invalid by masking. This is because the preset physical default values ​​(such as 25.0℃) typically differ significantly from the mean and variance when the sensor is operating normally. Forcibly performing Z-score normalization on these nodes along with normal data would cause the normalized values ​​to deviate from the normal distribution, introducing additional noise. By skipping normalization, the system preserves the physical meaning of the default values, while the masking mechanism teaches the model to ignore their numerical details.

[0046] Finally, during message passing in the graph neural network, for nodes marked as invalid due to masking, inference compensation is performed using the aggregated features of neighboring nodes. This is the core of the fault tolerance mechanism in this embodiment. Since the physical topology graph explicitly encodes the physical coupling relationships between device components, when sensor data of a node is missing, the states of its neighboring nodes often contain key information about that node's state. For example, if the temperature sensor in temperature zone T2 is disconnected, but the sensors in temperature zones T1 and T3 are working normally, since T1, T2, and T3 are physically adjacent and have a heat conduction relationship, the graph neural network can infer the hidden state of T2 by aggregating the features of T1 and T3. This neighbor compensation mechanism based on physical topology enables the model to make reasonable inferences using the spatial coupling relationships of devices even when some information is missing, thereby maintaining the stability of control decisions.

[0047] It should be understood that the specific values ​​of the aforementioned preset physical default values ​​are merely illustrative. In practical applications, these default values ​​can be adjusted according to the specific operating conditions of the equipment, environmental conditions, or historical statistical data. For example, for equipment in a high-temperature workshop, the default value of the temperature sensor can be set to the average workshop temperature of 40.0°C. This embodiment does not impose any limitations on this.

[0048] Furthermore, to enable the model to handle sensor failures during the training phase, this embodiment introduces a random masking strategy during cloud training. Specifically, in each training iteration, sensor data from some nodes is randomly replaced with preset physical default values ​​with a preset probability (e.g., 10%), and the corresponding masking markers are set to invalid, simulating a sensor disconnection scenario. This random masking strategy allows the model to fully learn how to use the aggregated features of neighboring nodes for inference compensation during the training phase, thus seamlessly switching to fault-tolerant inference mode when encountering sensor failures in actual operation, without the need for additional online adaptation. It should be understood that the probability of random masking can be adjusted according to the frequency of sensor failures in the actual scenario, and this embodiment does not impose any restrictions on this.

[0049] Step S4: Based on the time-series evolution characteristics, the multi-motor frequency fine-tuning vector and temperature zone convergence prediction time are synchronously output through the joint decision network. The start-up stage jump is dynamically triggered according to the convergence prediction time, and coordinated control is achieved according to the frequency fine-tuning vector.

[0050] After extracting the temporal evolution features containing spatiotemporal information, this step performs multi-task decision-making through a joint decision network (e.g., a multilayer perceptron, MLP). This network synchronously outputs two key control commands: first, a multi-motor frequency fine-tuning vector, used to fine-tune the operating frequencies of multiple motors such as the spindle, air pump, and mold head, to achieve coordinated control of multiple motors and avoid air pressure spikes or mechanical shocks caused by independent acceleration; second, a temperature zone convergence prediction time, used to predict the remaining time required for the current temperature zone to reach a steady state.

[0051] Preferably, the step of dynamically triggering the startup stage jump based on the convergence prediction time includes: when the convergence prediction time is less than a preset trigger threshold, skipping the fixed delay of the current stage and entering the next process stage, while setting a minimum protection time for the stage to prevent excessively rapid jump caused by false triggering.

[0052] In traditional extrusion equipment startup procedures, a fixed delay strategy hard-coded by the PLC is typically used. For example, after the spindle starts, a fixed 60-second wait is required, regardless of whether the equipment is in a cold or hot state. While this strategy is simple, it has significant efficiency drawbacks: in a hot state, the equipment may only need 20 seconds to reach a stable state, making the remaining 40-second wait time a complete waste. Figure 4 This diagram illustrates a timing comparison between the dynamic jump start-up phase of this invention and traditional PLC fixed-delay control. The red bar at the top represents traditional hard-coded control: employing a fixed 60s delay strategy, regardless of whether the equipment temperature zone converges early, it must wait for the entire preset time to T=60s before determining that the operating condition meets the standard and proceeding to the next process. The green curve at the bottom represents the SSM dynamic prediction scheme of this invention: by using a model to predict the temperature zone convergence status in real time at high frequency, it can predict that the temperature convergence meets the standard at T≈22s (star position), thus triggering the phase jump in advance. The time difference between the two timing segments represents the ineffective standby time saved by the system, intuitively demonstrating the technical advantage of dynamic scheduling in shortening start-up time.

[0053] This embodiment fundamentally changes this situation by introducing a dynamic triggering mechanism based on convergence prediction time. The system utilizes the temporal evolution characteristics output by the selective state-space model to predict in real time the remaining time (i.e., convergence prediction time) required for each temperature zone to reach a stable state. Here, "stable state" refers to the temperature zone entering a safe convergence window, meaning temperature fluctuations are within ±3℃ and last for at least 10 seconds. When this prediction time is less than a preset trigger threshold, the system determines that the equipment is ready to enter the next process stage, thus skipping the fixed delay of the current stage and directly triggering the start-up stage transition. This shift in control logic from "fixed delay" to "intelligent prediction" significantly shortens start-up time and improves production efficiency.

[0054] Specifically, the preset trigger threshold is preferably 5 seconds. This value is set based on a balance between safety and response speed in industrial control. If the threshold is set too high (e.g., 30 seconds), the chance of dynamic jumps will decrease, and the efficiency improvement will not be significant; if the threshold is set too low (e.g., 1 second), false triggers may occur due to sensor noise or instantaneous fluctuations in model predictions. A threshold of 5 seconds ensures that the equipment has enough time to complete the final stabilization process while minimizing ineffective waiting time. It should be understood that 5 seconds is only an preferred value. In practical applications, it can be adjusted according to the inertia of the equipment, sensor accuracy, or process requirements. For example, for large extruders with high thermal inertia, this threshold can be set to 8 seconds or 10 seconds.

[0055] Furthermore, to prevent false triggering caused by model prediction errors or sensor malfunctions, this embodiment also introduces a minimum protection time mechanism for each stage. The minimum protection time for each stage is preferably 10 seconds. This means that even if the convergence prediction time is less than 5 seconds at the initial startup, the system will not immediately jump, but will force the current stage to run for at least 10 seconds. This mechanism is equivalent to adding a "safety lock" to dynamic jumps, preventing premature jumps due to model misjudgment of the cold start state, and ensuring that the equipment has the basic physical conditions before entering the next stage. For example, during the spindle startup stage, even if the model predicts that the air pressure has stabilized, at least 10 seconds are needed for the mechanical components to complete initial lubrication and engagement to avoid mechanical shock.

[0056] Furthermore, this embodiment also provides a predictive error-proofing intervention mechanism. Preferably, the method further includes: predicting the air pressure or temperature trend after a preset time period based on the temporal evolution characteristics; and outputting a smooth deceleration command in advance when the predicted value approaches the alarm threshold, thus achieving predictive error-proofing intervention. In the extrusion production process, abnormal fluctuations in air pressure or temperature often have a lag. When the sensor detects a value exceeding the alarm threshold, the equipment may already be on the verge of danger. Taking emergency stop or deceleration measures at this time can easily lead to product scrapping or even equipment damage. This embodiment utilizes the long-range predictive capability of the selective state-space model to predict the state trend over a future period. Specifically, the preset time period is preferably 15 seconds. The system predicts the air pressure or temperature value 15 seconds later. If the predicted value approaches the alarm threshold (e.g., reaches 90% of the alarm threshold), it outputs a smooth deceleration command in advance, gradually reducing the speed of the spindle or air pump, thereby eliminating the abnormality before it occurs. This "prevention before the event" intervention strategy greatly improves the safety and continuity of the production process compared to the traditional "post-event alarm" strategy.

[0057] In terms of multi-motor cooperative control, this embodiment implements strict safety limits on the output frequency fine-tuning vector. Preferably, the multi-motor frequency fine-tuning vector is compressed to [-1, 1] using a Tanh activation function and then multiplied by a physical limiting coefficient, thus limiting the single-cycle change to ±0.5 Hz / s. Specifically, the frequency fine-tuning vector output by the joint decision network is a continuous real value. If directly applied to the inverter, it may lead to excessive frequency adjustment, causing mechanical shock or sudden changes in air pressure. This embodiment first compresses the fine-tuning vector to the range [-1, 1] using a Tanh activation function to achieve normalization; then it multiplies it by a physical limiting coefficient (e.g., 0.5 Hz / s) to limit the frequency change within a single cycle to a safe range. For example, if the model output fine-tuning vector is 1.2, after Tanh compression it becomes approximately 0.99, and then multiplied by 0.5 Hz / s, the final output frequency adjustment is 0.495 Hz / s. This limiting logic ensures that the actuator responds smoothly and safely regardless of the magnitude of the adjustment command output by the model, avoiding equipment damage caused by excessive model output. Furthermore, the superimposed target frequency must be within the physical limits of the inverter, i.e., the maximum absolute frequency range is 0.0–60.0 Hz. Simultaneously, considering the physical coupling relationship between multiple motors, this embodiment also introduces a coupling compensation mechanism: the air pump-spindle coupling compensation coefficient is 0.3, meaning that when the spindle accelerates, causing an increase in air pressure, the air pump performs deceleration compensation at a compensation ratio of 0.3; the mold head lag compensation time is 2.0 seconds, used to compensate for the mold head's response delay to air pressure changes, achieving timing alignment.

[0058] By combining the aforementioned dynamic triggering mechanism, protection time setting, predictive error prevention intervention, and frequency fine-tuning limiting logic, this embodiment achieves refined and intelligent control of the extrusion equipment's start-up and operation processes, improving production efficiency while ensuring the safety and stability of equipment operation.

[0059] Furthermore, this embodiment employs a multi-task joint loss function to perform end-to-end optimization of the model during the cloud training phase. The multi-task joint loss function includes four components: convergence time prediction loss. Mean squared error (MSE) is used to measure the deviation between the predicted convergence time and the actual convergence time; motor cooperative loss. Mean square error is used to measure the deviation between the frequency fine-tuning vector and the target frequency; smoothness regularization. ,use The method measures the magnitude of frequency fine-tuning between adjacent time steps to prevent drastic motor frequency fluctuations; it also masks consistency losses. Mean squared error is used to measure the deviation between the inferred features and the true features of the occluded nodes, ensuring stable output in scenarios with missing sensors. The four loss components are weighted and summed using configurable weighting coefficients: ; As a preferred option, the weighting coefficients are respectively =1.0、 =2.0、 =0.5、 =0.3. Wherein The setting of =2.0 prioritizes motor coordination loss the highest because the smooth operation of the motor is directly related to the physical safety of the equipment and the product yield, and its importance is higher than the optimization of start-up time. =0.5 is used as a key regularization term to prevent the model from outputting high-frequency jittering commands in pursuit of convergence speed. It should be understood that the above weight coefficients are only illustrative and can be adjusted according to the actual training results.

[0060] Regarding edge inference performance, the model in this embodiment exhibits a single-step inference latency of less than 50 milliseconds on an edge industrial control computer (e.g., an industrial PC equipped with an ARM architecture processor and 4GB of memory), a total end-to-end inference time of less than 2 seconds for a 30-minute sequence, and a memory usage of less than 200MB. This performance is attributed to the linear computational complexity of the selective state-space model and the sparse message passing mechanism of the graph neural network. This enables the technical solution of this embodiment to achieve real-time and efficient intelligent control at resource-constrained industrial edge environments, without relying on cloud-based inference, thus meeting the stringent requirements for real-time performance and reliability in industrial settings.

[0061] Through the above steps, this embodiment realizes intelligent control of the entire process from data acquisition, spatial feature aggregation, temporal feature extraction to control decision output. It makes full use of the physical topology information and long-term historical information of the equipment, and achieves high-precision spatiotemporal joint reasoning and control under low computing power resources at the edge.

[0062] Example 2: This example provides an intelligent control system for an extrusion equipment. This system corresponds to the intelligent control method described in the method example. Through a modular architecture design, it achieves fully automated processing from data acquisition and feature extraction to control decision-making. See the appendix for the overall system architecture. Figure 1 .like Figure 1 As shown, the system is divided into three layers from top to bottom: cloud brain, edge control terminal, and underlying hardware devices, and relies on the data flywheel to achieve closed-loop optimization of the entire process.

[0063] The top layer is the cloud brain layer, which sequentially stores data in a time-series database and cleans the data to complete the training of the GNN+SSM spatiotemporal graph network model. After training, a lightweight ONNX model and Z-Score statistical parameter file are exported. The cloud sends the model and parameters to the edge control terminal via the OTA protocol. The operating condition data collected by the edge terminal is asynchronously transmitted back to the cloud via MQTT, forming an iterative closed loop.

[0064] The middle layer is the edge control terminal. It first collects bus data through a timing sliding window, and after parameter loading and Masking fault-tolerant preprocessing, it is sent to the SSM+GNN inference model. The model output is used for dynamic start-up scheduling and multi-motor collaborative control, and is connected to the HMI human-machine interface.

[0065] The bottom layer is the hardware execution layer. All devices are interconnected with the edge control terminal via Modbus / OPCUA industrial bus, including plastic extruders, T1-T11 temperature zones, spindle motors, and air pump motors, to realize on-site data acquisition and control command implementation.

[0066] Specifically, the system's modules include a topology construction module, a spatial feature aggregation module, a temporal feature extraction module, and a control decision module.

[0067] The topology construction module is used to obtain the adjacency matrix of the physical topology graph. The adjacency matrix is ​​constructed based on the mechanical assembly distance and thermodynamic conduction coupling strength of each component of the extrusion equipment, and maps multimodal sensor data into graph node features.

[0068] This module forms the "physical prior" foundation of the entire intelligent control system. Specifically, the topology building module is configured to construct the edge weights of the adjacency matrix based on the following relationships: ; in, For nodes With nodes The mechanical assembly distance between them For nodes With nodes Thermodynamic conduction coupling strength between them and Configurable scaling factor; The edge weights of the adjacency matrix are symmetrically normalized to obtain a normalized adjacency matrix, which is used for spatial dimension aggregation in the spatial feature aggregation module. Before system operation, the topology construction module first reads the 3D design drawings or configuration files of the extrusion equipment, extracts the spatial positional relationships between various components (such as temperature zones T1-T11, main shaft, air pump, and die head), and calculates the mechanical assembly distance. Simultaneously, it determines the thermodynamic conduction coupling strength between components by combining thermodynamic simulation data or expert experience. Through weighted summation and symmetric normalization, the topology construction module generates a normalized adjacency matrix that accurately reflects the physical structure of the equipment and stores it in system memory for subsequent modules to access. It should be understood that the topology construction module can be deployed on a cloud server to calculate and update the adjacency matrix offline, or it can be deployed on an edge industrial control computer to dynamically adjust the topology based on actual equipment wear or modification. This physical topology-based construction method enables the system to have physical interpretability from the data input stage, avoiding the blind reliance on data distribution in traditional black-box models.

[0069] The spatial feature aggregation module is used to aggregate the graph node features in the spatial dimension using a graph neural network to obtain a spatially coupled feature sequence.

[0070] This module receives a normalized adjacency matrix from the topology construction module and real-time acquired multimodal sensor data (i.e., graph node features). Specifically, the spatial feature aggregation module encapsulates a graph neural network model (such as GCN, GAT, etc.). Through a message passing mechanism, each node aggregates the feature information of its neighboring nodes. For example, the node features of temperature zone T1 not only include its own temperature reading but also incorporate the features of adjacent temperature zones T2, T3, and the main axis node, thereby capturing the coupling relationship between heat conduction and mechanical vibration within the local space of the device. The output of this module is a spatial coupling feature sequence, where each feature vector contains the spatial context information of the device at the current moment. The spatial feature aggregation module can be executed by an edge GPU or CPU. Its computational complexity is related to the sparsity of the graph. With a reasonable graph structure design, millisecond-level spatial feature extraction can be achieved.

[0071] The temporal feature extraction module is used to input the spatially coupled feature sequence into a structured temporal state evolution model with linear computational complexity to extract temporal evolution features.

[0072] This module is the core of the system's long-term memory and efficient reasoning. Specifically, the temporal feature extraction module encapsulates a Selective State-Space Model (SSM), which uses a hardware-aware parallel scanning algorithm to process ultra-long temporal data with a linear computational complexity of O(N). The module receives the spatially coupled feature sequence output by the spatial feature aggregation module, evolves and compresses the features in the time dimension, and extracts the dynamic patterns of equipment state changes over time, such as temperature rise curves and air pressure fluctuation trends. Due to the use of a linear complexity model, this module can process historical data of more than 30 minutes in real time on resource-constrained edge industrial control computers without the long-range forgetting problem of traditional LSTM models. The output of the temporal feature extraction module is a temporal evolution feature, which integrates the spatial coupling information and historical state information of the equipment, providing comprehensive input for subsequent control decisions.

[0073] The control decision module is used to synchronously output the multi-motor frequency fine-tuning vector and temperature zone convergence prediction time through a joint decision network based on the time-series evolution characteristics, dynamically trigger the start-up stage jump according to the convergence prediction time, and realize coordinated control according to the frequency fine-tuning vector.

[0074] This module is the "brain" of the system, responsible for converting the extracted spatiotemporal features into specific control commands. Specifically, the control decision module encapsulates a joint decision network (such as a multilayer perceptron, MLP), which synchronously outputs two key control quantities: first, a multi-motor frequency fine-tuning vector, used to fine-tune the operating frequencies of multiple motors such as the spindle, air pump, and mold head, achieving coordinated control of multiple motors and eliminating pressure spikes or mechanical shocks caused by independent acceleration; second, a temperature zone convergence prediction time, used to predict the remaining time required for the current temperature zone to reach a stable state. The control decision module dynamically triggers start-up phase transitions based on this prediction time. For example, if the predicted convergence time is less than 5 seconds, it directly skips the fixed delay of the current phase and proceeds to the next process phase. Furthermore, the control decision module integrates safety limiting logic to ensure that the output frequency fine-tuning vector remains within physically permissible limits, guaranteeing equipment operational safety. The control decision module communicates with underlying frequency converters, PLCs, and other actuators via fieldbus (such as Modbus RTU or OPC UA) to achieve real-time issuance of control commands.

[0075] Through the collaborative work of the four modules described above, the system in this embodiment achieves full-process automation from physical topology construction, spatial feature aggregation, temporal feature extraction to control decision output. Each module of the system corresponds one-to-one with each step in the method embodiment: the topology construction module corresponds to step S1, the spatial feature aggregation module corresponds to step S2, the temporal feature extraction module corresponds to step S3, and the control decision module corresponds to step S4.

[0076] Preferably, the system in this embodiment adopts a cloud-edge collaborative deployment architecture, deploying model training and topology construction in the cloud and feature extraction and control decision-making at the edge, thereby meeting the dual requirements of real-time performance and reliability in industrial settings. Specifically, the cloud-edge collaborative architecture includes the following key mechanisms: Cloud-edge data flywheel: The edge collects high-frequency data on the operating status of devices, executes forward inference of the lightweight ONNX computation graph and issues control commands, and simultaneously reports downsampled data asynchronously to the cloud time-series database via the MQTT protocol; the cloud platform periodically (e.g., monthly) uses newly added production data to trigger incremental training, generating a new generation of GNN+SSM models, forming a closed-loop data flywheel of "data collection, reporting, training, issuing, inference, and re-collection", so that the model can be continuously optimized as production data accumulates.

[0077] Shadow Mode Validation Mechanism: After the cloud platform generates a new generation model, it initiates a shadow mode in the cloud for security validation. The specific process of shadow mode is as follows: the new model and the previous generation model perform concurrent inference using the latest production data in the cloud; the mean absolute error (MAE) of the two models is calculated; if the MAE of the new model is better than that of the old model for a preset number of days (e.g., 7 days), the new model is considered a valid iteration; if the validation fails, monitoring continues and deployment is not executed. This mechanism ensures that only fully validated models can be deployed to the production site, eliminating the risk to equipment operation safety posed by unvalidated models.

[0078] Hot update mechanism: After the shadow mode verification passes, the system automatically and silently pushes the new generation ONNX model file and the accompanying Z-Score statistical parameter file (JSON format) to the edge terminal via an encrypted channel; the edge terminal automatically reloads the model handle during device standby or shift handover, completing a seamless hot replacement. The entire process requires no downtime or manual intervention, ensuring production continuity. The complete process of seamless upgrade from cloud verification to edge terminal for the new model of this invention is as follows: Figure 5 As shown, the steps are as follows: STEP 1: The next-generation GNN+SSM model is trained in the cloud using data transmitted back from the edge. STEP 2: Start the Shadow Mode verification mechanism in the cloud; STEP3: The new and old generation models perform concurrent inference on the same batch of production data; STEP4: Compare the inference accuracy of the two models using the Mean Absolute Error (MAE) index; STEP5 Judgment: If the new model's error index is better than the old model for 7 consecutive days, the iteration is deemed successful; if it does not meet the standard, continue monitoring and retain the original old model for further iteration and testing. STEP5 - Final: The qualified model is encrypted and packaged into an ONNX model file + Z-Score parameter JSON file, and silently distributed to the edge industrial control computer via the OTA encrypted channel. The device reloads the model handle during production standby intervals, and completes the seamless model upgrade without stopping the machine.

[0079] Deployment Specifications: As a preferred option, the input and output tensor specifications of the ONNX computation graph model deployed at the edge are as follows: The input tensor sensor_seq has a shape of (B, T, F), where B=1 is the single-device inference batch size, T=30 is the 30-second historical time series length, F=20 is the number of sensor features, and the data type is float32; the input tensor mask_seq has a shape of (B, T, F), a data type of float32, and a value of 0 or 1, where 0 indicates a missing sensor (using the default value), and 1 indicates normal operation; the output tensor delta_freq has a shape of (B, 3), a data type of float32, and represents the frequency fine-tuning of the three motors [Δspindle, Δpump, Δmold], with a range of [-0.5, +0.5] Hz / s; the input tensor adjacency has a shape of (N, N), where N=14 is the number of nodes, the data type is float32, and it is a pre-computed normalized adjacency matrix that does not participate in gradient calculation during inference; the input tensors zscore_mean and zscore_std are both of shape (N, 1), the data type is float32, and they are the Z-Score mean and standard deviation of each node's features, respectively, used for edge standardization.

[0080] As a preferred option, the ONNX computation graph model deployed at the edge meets the following engineering specifications: model file size not exceeding 5MB, containing the full graph of GNN+SSM+joint decision network; end-to-end inference latency not exceeding 100 milliseconds (test environment: Intel i5-10400 processor, 16GB memory, ONNXRuntime CPU mode, 30-second timing input); runtime memory usage not exceeding 200MB, including sliding window queue, model weights, and adjacency matrix caching; OTA silent update time not exceeding 3 seconds. These specifications ensure that the model can still meet millisecond-level real-time control requirements on resource-constrained edge industrial control computers.

[0081] Example 3: To more clearly illustrate the present invention, the following example uses the startup preheating and multi-motor coordinated control scenario of a plastic extruder to describe the application of the technical solution of the present invention in a real industrial scenario. It should be understood that this scenario is merely exemplary and not restrictive, and the technical solution of the present invention is also applicable to other types of industrial equipment control scenarios.

[0082] In this scenario, the plastic extruder comprises 11 temperature zones (T1-T11), a main shaft motor, an air pump motor, and a die head motor, totaling 14 physical nodes. The system executes the following processing flow upon cold start: First, based on the 3D design drawings of the extrusion equipment, the system obtains the mechanical assembly distances and thermodynamic conduction coupling strengths between various components, constructs the adjacency matrix of the physical topology graph (see step S100 in Example 1), and collects real-time data from each sensor via a fieldbus at high frequency, mapping it to graph node features. At this time, the equipment is in a cold state, with the temperature of each temperature zone close to the ambient temperature of 25.0℃ and the rotation speed at 0.0 Hz.

[0083] Subsequently, the system utilizes a graph neural network to aggregate graph node features in the spatial dimension to obtain a spatial coupling feature sequence (see step S200 in Example 1). During the heating stage, when the temperature of temperature zone T5 rises first, the GNN senses the influence of T5 on the heat conduction of adjacent temperature zones by aggregating the features of neighboring nodes, and pre-encodes this spatial coupling information in the feature vectors of T4 and T6.

[0084] Next, the system inputs the spatially coupled feature sequence into the selective state-space model (SSM) to extract temporal evolution features (see step S300 and feature masking and fault tolerance mechanism in Example 1). During the initial power-on phase when device states change drastically, the SSM dynamically adjusts the discretization step size and input matrix through an input adaptive mechanism, efficiently encoding key state changes into hidden states. When the device stabilizes, the state is updated smoothly, achieving lossless context memory for power-on cycles exceeding 30 minutes.

[0085] Then, the system synchronously outputs the multi-motor frequency fine-tuning vector and the temperature zone convergence prediction time through a joint decision-making network (see step S400 in Example 1). When the system predicts that temperature zones T1-T11 will all enter the stable range within 4 seconds, and the convergence prediction time is less than the preset trigger threshold of 5 seconds, the system directly skips the fixed delay trigger start-up phase jump, while forcing the current phase to run for at least 10 seconds as the minimum protection time. In terms of multi-motor cooperative control, after Tanh compression and physical limiting, the frequency fine-tuning vector ensures that the single-cycle frequency adjustment amount does not exceed ±0.5 Hz / s, achieving smooth cooperative acceleration. In addition, the system predicts the air pressure or temperature trend 15 seconds later, and outputs a smooth deceleration command in advance when the predicted value approaches the alarm threshold, achieving predictive error prevention intervention.

[0086] In the case of sensor failure, assuming that the temperature sensor in temperature zone T3 is disconnected, the system immediately fills in the preset physical default value of 25.0℃ and sets the masking mark to invalid, skipping the normalization process. At the same time, the GNN uses the aggregated features of neighboring nodes such as T2 and T4 to infer the hidden state of T3 (see the feature masking and fault tolerance mechanism in step S3 of Example 1) to maintain the inference ability of the node and avoid the whole machine downtime due to a single sensor failure.

[0087] At the system level, the above scenario is executed by the intelligent control system described in Example 2. The topology construction module completes the offline calculation and update of the adjacency matrix in the cloud, while the spatial feature aggregation module, temporal feature extraction module, and control decision module run in real time at the edge. The model performance is continuously optimized through the cloud-edge data flywheel mechanism, shadow mode verification ensures the safety of model iteration, and OTA hot updates enable seamless model replacement, ensuring production continuity.

[0088] Through the above scenarios, this invention fully demonstrates its application in the preheating and multi-motor collaborative control of plastic extruders: the system perceives the spatial coupling relationship between equipment components through a physical topology diagram, and memorizes long-term states through a selective state space model, thereby achieving dynamic prediction of convergence time to skip invalid waiting, smooth collaborative acceleration of multiple motors, and fault-tolerant operation of sensor faults. This fully demonstrates the technical effects of this invention in improving production efficiency, ensuring equipment safety, and enhancing system robustness.

[0089] The above description is merely a specific embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the technical scope disclosed in the present invention, as long as they do not depart from the spirit and scope of the technical solution of the present invention, should be covered within the scope of protection of the present invention.

Claims

1. An intelligent control method for an extrusion equipment, characterized in that, Includes the following steps: Step S1: Obtain the adjacency matrix of the physical topology graph. The adjacency matrix is ​​constructed based on the mechanical assembly distance and thermodynamic conduction coupling strength of each component of the extrusion equipment, and the multimodal sensor data is mapped into graph node features. Step S2: Use a graph neural network to aggregate the graph node features in the spatial dimension to obtain a spatially coupled feature sequence; Step S3: Input the spatial coupling feature sequence into a structured temporal state evolution model with linear computational complexity to extract temporal evolution features; Step S4: Based on the time-series evolution characteristics, the multi-motor frequency fine-tuning vector and temperature zone convergence prediction time are synchronously output through the joint decision network. The start-up stage jump is dynamically triggered according to the convergence prediction time, and coordinated control is achieved according to the frequency fine-tuning vector.

2. The method according to claim 1, characterized in that, The edge weights of the adjacency matrix are constructed according to the following relationship: ； in, For nodes With nodes The mechanical assembly distance between them For nodes With nodes Thermodynamic conduction coupling strength between them and Configurable scaling factor; The edge weights of the adjacency matrix are symmetrically normalized to obtain a normalized adjacency matrix, which is used for spatial dimension aggregation in graph neural networks.

3. The method according to claim 1, characterized in that, The structured temporal state evolution model is a selective state space model. The discretization step size, discretization input matrix, and output projection matrix of the selective state space model are all adaptively generated based on the current time step input through linear projection and nonlinear activation, realizing content-dependent memory selection. The selective state-space model completes state recursion and output projection with linear computational complexity through a hardware-aware parallel scanning algorithm.

4. The method according to claim 1, characterized in that, The method further includes: Binary masking tags are injected into multimodal sensor data. When a sensor disconnection is detected, a preset physical default value is filled in and the corresponding masking tag is set to invalid. Nodes with invalid masking tags are skipped from the normalization process. During the message passing process of the graph neural network, for nodes whose masking is invalid, the aggregation features of neighboring nodes are used for inference compensation to maintain the hidden state reasoning of that node.

5. The method according to claim 4, characterized in that, The preset physical default values ​​include: For the temperature sensor, the ambient temperature is set to 25.0℃, corresponding to the initial state assumption of the chiller; For the speed sensor, set it to zero speed 0.0 Hz, which corresponds to the default value in the shutdown state; For the barometric pressure sensor, set it to standard atmospheric pressure of 0.101 MPa.

6. The method according to claim 1, characterized in that, Dynamically triggering the startup phase jump based on the convergence prediction time includes: When the convergence prediction time is less than the preset trigger threshold, the fixed delay of the current stage is skipped and the process enters the next stage. At the same time, a minimum protection time for the stage is set to prevent premature jumps caused by false triggers.

7. The method according to claim 6, characterized in that, The preset trigger threshold is 5 seconds, and the minimum protection time for the stage is 10 seconds; The multi-motor frequency fine-tuning vector is compressed to [-1, 1] by the Tanh activation function and then multiplied by the physical limiting coefficient, so that the single-cycle change is limited to ±0.5 Hz / s.

8. The method according to claim 1, characterized in that, The method further includes: based on the time-series evolution characteristics, predicting the air pressure or temperature trend after a preset time period; when the predicted value approaches the alarm threshold, outputting a smooth deceleration command in advance to achieve predictive error prevention intervention; the preset time period is 15 seconds.

9. The method according to claim 1, characterized in that, The adjacency matrix is ​​pre-generated by the cloud server based on the mechanical assembly distance and thermodynamic conduction coupling strength of each component of the extrusion equipment and then sent to the edge. The graph neural network and the structured temporal state evolution model are trained by the cloud using historical operating data and then deployed to the edge.

10. An intelligent control system for an extrusion equipment, characterized in that, include: The topology construction module is used to obtain the adjacency matrix of the physical topology graph. The adjacency matrix is ​​constructed based on the mechanical assembly distance and thermodynamic conduction coupling strength of each component of the extrusion equipment, and maps multimodal sensor data into graph node features. The spatial feature aggregation module is used to aggregate the graph node features in the spatial dimension using a graph neural network to obtain a spatially coupled feature sequence. The temporal feature extraction module is used to input the spatially coupled feature sequence into a structured temporal state evolution model with linear computational complexity to extract temporal evolution features; The control decision module is used to synchronously output the multi-motor frequency fine-tuning vector and temperature zone convergence prediction time through a joint decision network based on the time-series evolution characteristics, dynamically trigger the start-up stage jump according to the convergence prediction time, and realize coordinated control according to the frequency fine-tuning vector.

Images