A method for dynamic adjustment and control of risk parameters in the production process of protective materials

Abstract

This invention relates to the field of production control technology, specifically to a method for dynamic adjustment and control of risk parameters in the production process of protective materials. The method involves acquiring operating state parameters and process constraint parameters, and assigning them as initial attributes to an undirected graph data structure. By combining the weaving density baseline value and the operating linear velocity to identify physical disturbance characteristics, equivalent coupling damping characteristics are derived and assigned to the undirected graph data structure to obtain a graph mapping structure. Multiple cycles of the graph mapping structure are extracted to construct a time-series state sequence. This time-series state sequence is input into a time-series prediction network model to deduce a dynamic runaway index. When the dynamic runaway index exceeds a safety threshold, the lateral width scale is extracted, and the control response hysteresis characteristics of the extrusion interface under multi-axis coupled servo force are identified. Based on the control response hysteresis characteristics, control commands, including end-effector pose compensation commands and feed servo adjustment commands, are identified for dynamic adjustment control.

Description

Technical Field

[0001] This invention relates to the field of production control technology, specifically to a method for dynamic adjustment and control of risk parameters in the production process of protective materials. Background Technology

[0002] In continuous production lines for protective materials, control systems need to perform high-frequency closed-loop regulation of multiple dynamic physical quantities, such as traction speed, mechanical actuator posture, and fluid medium feed. Traditional automation control architectures often employ serial adjustment logic based on error feedback, relying on single-dimensional sensor thresholds to trigger mechanical compensation actions.

[0003] However, when the controlled object becomes a wide-width flexible entity with a densely interwoven mesh, the entire control system faces complex challenges of nonlinear multi-field coupling. On the one hand, mesh fluctuations at high speeds introduce high-frequency microscopic physical disturbances into the control system, causing disproportionate abrupt changes in the fluid medium state. These high-frequency disturbances are often misjudged as environmental noise and filtered out by traditional controllers, resulting in the loss of underlying real physical interaction characteristics. On the other hand, the ultra-wide-width controlled object, under the interwoven tension and compression of multiple axes, produces lateral asymmetric elastic deformation. This causes the time for different regions of the execution interface to reach the target force state to vary significantly after the control center issues the execution command, resulting in extremely significant spatial response lag.

[0004] Faced with the aforementioned strongly coupled controlled objects with spatiotemporal lag characteristics, the existing single-variable control paradigm, lacking physical forward-looking prediction and multi-variable collaborative decoupling mechanisms, not only struggles to extract the dynamic trend of flow field distribution evolution over time, but also, when faced with sudden anomalies, often blindly increases the intensity of single-item compensation, causing serious mutual interference of control commands in the confrontation between gap reshaping and flow regulation, leading to the entire automated regulation system falling into long-period oscillations.

[0005] Therefore, a dynamic adjustment and control method for risk parameters in the production process of protective materials is proposed. Summary of the Invention

[0006] The purpose of this invention is to provide a method for dynamic adjustment and control of risk parameters in the production process of protective materials, which identifies control commands including end-effector pose compensation commands and feed servo adjustment commands based on the control response lag characteristics.

[0007] To achieve the above objectives, the present invention provides the following technical solution: A method for dynamically adjusting and controlling risk parameters in the production process of protective materials includes: Obtain operating status parameters and process constraint parameters. The operating status parameters include operating linear velocity, initial mechanical pressure, and medium state characteristics. The process constraint parameters include weaving density reference value and transverse width dimension. The operating state parameters and process constraint parameters are assigned as initial attributes to the undirected graph data structure; the physical disturbance characteristics are identified by combining the weaving density benchmark value and the operating linear velocity, and the equivalent coupling damping characteristics are derived as edge association characteristics; the edge association characteristics are assigned to the undirected graph data structure to obtain the graph mapping structure. Extract graph mapping structures from multiple cycles to construct a time-series state sequence; input the time-series state sequence into a time-series prediction network model to extract the dynamic nonlinear features of the state distribution of the controlled object evolving with the time array, and deduce the dynamic runaway index. When the dynamic runaway index exceeds the safety threshold, the lateral width scale is extracted, and the control response hysteresis characteristics of the extrusion interface under the action of multi-axis coupled servo force are identified. Based on the control response hysteresis characteristics, control commands including end pose compensation commands and feed servo adjustment commands are identified and output for dynamic adjustment control.

[0008] Preferably, the process of obtaining the operating status parameters and process constraint parameters includes: The system acquires operating status parameters through a sensor network deployed on the underlying physical equipment, reads the rotation pulse signal of the traction drive shaft and converts it into a digital linear velocity, collects the feedback level signal of the extrusion pneumatic component to extract the initial mechanical pressure, and uses the damping sensing unit built into the pipeline to acquire the viscous resistance signal of the fluid as a medium state characteristic. Retrieve the pre-set material attribute list for the current production batch and extract the base value of the weaving density and the transverse width of the substrate as process constraint parameters. The collected operating status parameters and process constraint parameters are aligned with timestamps based on a globally unified clock, and a multi-source synchronized data stream is output.

[0009] Preferably, the step of assigning initial attributes to an undirected graph data structure includes: Construct a topological network representing the spatial relationships of the physical production line; establish physical execution entities as spatial node entities in an undirected graph data structure; the physical execution entities include an extrusion assembly, a feeding assembly, and a traction assembly; Based on the operating status parameters and process constraint parameters, the initial mechanical pressure, operating linear velocity, and medium state characteristics are extracted and written into the corresponding physical execution entity nodes as dynamic operating attribute features; at the same time, the weaving density benchmark value and lateral width scale are extracted and written into the global environment node as static constraint attribute features. By defining the physical distance for material transfer and the response boundary for signal transmission between spatial node entities, an undirected connection path between node entities is established, and an undirected graph data structure containing multi-dimensional attribute features and spatial transmission topology is output.

[0010] Preferably, the step of deriving edge association features to obtain the graph mapping structure includes: Calculate the product of the running linear speed and the braiding density baseline value, identify the periodic physical disturbance frequency of the fluid generated by the bottom substrate when it passes through the extrusion working surface, and use the periodic physical disturbance frequency as the physical disturbance characteristic. The physical disturbance characteristics are compared with the inherent rheological relaxation time in the medium state characteristics to identify the sudden change in viscous resistance generated by the fluid medium when subjected to physical disturbance. Based on the abrupt change in viscous resistance, an equivalent coupled damping characteristic characterizing the fluid medium's ability to dissipate physical disturbances is derived. The equivalent coupling damping feature is used as the weight value of the physical coupling strength between nodes, i.e. the edge association feature, and assigned to the undirected connection path in the undirected graph data structure to output a graph mapping structure containing state nodes and edge weights.

[0011] Preferably, the process of deriving the dynamic runaway index includes: According to the time sequence, the graph mapping structure within the set number control period is continuously extracted, and the spatial dimension graph model is stacked and expanded in the time dimension to construct a temporal state sequence containing spatiotemporal coupling information. The time-series state sequence is input into a time-series prediction network model containing a memory-forgetting gating unit, and the acceleration and slope of the change of physical state and edge weights on the time axis are identified through the neuron nodes inside the model. Based on the changing acceleration and slope, the dynamic nonlinear characteristics of the non-proportional changes in the state distribution of the controlled object that will occur in future periods are extracted. By comparing the distance between the dynamic nonlinear characteristics and the built-in historical evolution samples, the absolute deviation trend of the expected deviation from the target set value is deduced while maintaining the current control action unchanged, and the output is the dynamic runaway index.

[0012] Preferably, the step of identifying the control response hysteresis characteristics includes: When the value of the dynamic runaway index is detected to be greater than the preset safety control threshold, the lateral width scale is retrieved, and combined with the current running linear speed, initial mechanical pressure, and the material's inherent stiffness mapped by the weaving density benchmark value, the elastic deformation gradient generated at different physical positions in the lateral direction of the wide flexible material under the interlacing tension and compression of multiaxial mechanical forces is identified. Based on the elastic deformation gradient, the absolute time difference between the central region and the two edge regions of the extrusion interface when the controller issues a synchronous execution command is evaluated. The distribution set of time difference values ​​at each horizontal node is used to construct a control response hysteresis characteristic curve that characterizes the physical space transmission delay, thus obtaining the control response hysteresis characteristic.

[0013] Preferably, the steps for dynamic adjustment and control include: Based on the control response hysteresis characteristics and the control objective of reducing the dynamic runaway index, a multi-dimensional state matrix including mechanical space geometric parameters and medium supply kinetic energy parameters is established. Through a multi-objective constraint solving algorithm, multiple sets of comprehensive candidate control schemes with different numerical ratios are calculated and output simultaneously. Each set of candidate control schemes includes an end pose compensation command for reshaping the lateral spatial geometric gap of the extrusion channel, and a feed servo adjustment command for changing the total amount of fluid medium injected. The comprehensive candidate control scheme set is input into the control scheme evaluation module, and the comprehensive fitness score of each group of candidate control schemes is calculated using the minimization of fluid lateral distribution variance and servo drive energy peak as dual evaluation indicators. The candidate control scheme with the highest score is selected, converted into electrical-level drive commands, and sent to the corresponding physical execution components to complete the dynamic closed-loop regulation under the strong coupling state of multiple physics fields.

[0014] Compared with the prior art, the beneficial effects of the present invention are as follows: 1. This invention integrates the linear velocity and weaving density benchmark values ​​to identify the frequency of periodic physical disturbances generated by the underlying substrate on the fluid medium. By comparing the inherent rheological relaxation time, it derives the equivalent coupling damping characteristic characterizing the energy dissipation capability of the fluid medium and assigns it as an edge-related feature to the undirected graph data structure. This design successfully quantifies the complex nonlinear physical coupling relationship between mechanical excitation and non-Newtonian fluid rheological abrupt changes at the macroscopic topological dimension, eliminating blind spots caused by the loss of underlying real physical interaction features, and constructing a high-fidelity data foundation for subsequent risk simulation.

[0015] 2. This invention extracts graph mapping structures sequentially over time to construct temporal state sequences, which are then imported into a temporal prediction network model containing memory-forgetting gating units. By deeply exploring the acceleration and slope of change in physical states and edge weights on the time axis, it can capture the non-proportional dynamic nonlinear characteristics of the controlled object's state distribution that will occur in future periods. This allows the control center to deduce the dynamic runaway index in advance before substantial defects form, reserving ample smooth intervention windows for physical actuators, and upgrading the decision-making logic of automated manufacturing from passive post-event response to proactive defense.

[0016] 3. This invention constructs a multi-dimensional state matrix based on control response hysteresis characteristics and risk reduction objectives. It utilizes a multi-objective constraint-solving algorithm to synchronously and parallelly output a comprehensive set of candidate control schemes covering end-effector pose compensation and feed servo adjustment. Combining dual evaluation metrics aimed at eliminating fluid lateral distribution variance and minimizing servo drive energy peaks, the electrical drive command with the best overall fitness is selected for execution. This collaborative decoupling strategy resolves the command interference contradiction between spatial gap reshaping and fluid total volume adjustment, ensuring the extreme uniformity of the coating surface's lateral quality and the long-term stable operation of the entire system. Attached Figure Description

[0017] Figure 1 This is a flowchart illustrating a method for dynamically adjusting and controlling risk parameters in the production process of protective materials according to the present invention. Figure 2 This is a schematic diagram of the logic of a method for dynamically adjusting and controlling risk parameters in the production process of protective materials according to the present invention. Figure 3 This is a schematic diagram illustrating the construction process of the undirected graph data structure of the present invention. Detailed Implementation

[0018] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0019] Example 1:

[0020] A method for dynamically adjusting and controlling risk parameters in the production process of protective materials, the method flow is as follows: Figure 1 As shown, the logic of the method is as follows: Figure 2 As shown, it specifically includes: Obtain operating status parameters and process constraint parameters. The operating status parameters include operating linear velocity, initial mechanical pressure, and medium state characteristics. The process constraint parameters include weaving density reference value and transverse width dimension. The operating state parameters and process constraint parameters are assigned as initial attributes to the undirected graph data structure; the physical disturbance characteristics are identified by combining the weaving density benchmark value and the operating linear velocity, and the equivalent coupling damping characteristics are derived as edge association characteristics; the edge association characteristics are assigned to the undirected graph data structure to obtain the graph mapping structure. Extract graph mapping structures from multiple cycles to construct a time-series state sequence; input the time-series state sequence into a time-series prediction network model to extract the dynamic nonlinear features of the state distribution of the controlled object evolving with the time array, and deduce the dynamic runaway index. When the dynamic runaway index exceeds the safety threshold, the lateral width scale is extracted, and the control response hysteresis characteristics of the extrusion interface under the action of multi-axis coupled servo force are identified. Based on the control response hysteresis characteristics, control commands including end pose compensation commands and feed servo adjustment commands are identified and output for dynamic adjustment control.

[0021] Preferably, the process of obtaining the operating status parameters and process constraint parameters includes: The system acquires operating status parameters through a sensor network deployed on the underlying physical equipment, reads the rotation pulse signal of the traction drive shaft and converts it into a digital linear velocity, collects the feedback level signal of the extrusion pneumatic component to extract the initial mechanical pressure, and uses the damping sensing unit built into the pipeline to acquire the viscous resistance signal of the fluid as a medium state characteristic. Retrieve the pre-set material attribute list for the current production batch and extract the base value of the weaving density and the transverse width of the substrate as process constraint parameters. The collected operating status parameters and process constraint parameters are aligned with timestamps based on a globally unified clock, and a multi-source synchronized data stream is output.

[0022] The acquisition of operating status parameters and process constraint parameters is achieved through a sensor network deployed on the underlying physical equipment.

[0023] For acquiring the linear velocity, an incremental rotary encoder is installed at the shaft position of the traction drive shaft. The encoder continuously outputs equally spaced pulse signals as the shaft rotates. The number of pulses is counted according to a fixed sampling period. Combining the encoder's pulse count per revolution parameter with the effective transmission diameter of the drive shaft, the count value is converted into a digital linear velocity in meters per second through a speed conversion relationship.

[0024] To collect the initial mechanical pressure, a pressure sensor is introduced into the cylinder control circuit of the extrusion pneumatic component. This sensor converts the gas pressure in the cylinder chamber into a level signal output within the corresponding range. The feedback level signal is then processed by analog-to-digital conversion, and combined with the effective cross-sectional area of ​​the cylinder and the pre-calibrated force value mapping relationship, the level value is converted into an initial mechanical pressure value in Newtons.

[0025] To collect media state characteristics, a damping sensing unit is built into a key node of the fluid delivery pipeline. The damping sensing unit senses the viscous resistance signal of the fluid based on the pressure difference change generated when the fluid flows in a fixed cross-section channel. The original viscous resistance signal is low-pass filtered to remove high-frequency interference noise and then the statistical characteristics such as the mean, variance and spectral peak value of the signal are extracted to form the media state characteristics, which include the estimated value of the fluid's inherent rheological relaxation time required for subsequent derivation.

[0026] To obtain process constraint parameters, based on the batch number of the current production task, the material attribute list of the corresponding batch is retrieved from the preset material attribute database, and the base value of the weaving density of the substrate (unit: number of warp and weft threads per centimeter) and the transverse width dimension are extracted as process constraint parameters.

[0027] After completing the data acquisition, a timestamp alignment method based on a globally unified clock is used to synchronize multiple data streams from different physical devices. The globally unified clock provides unified time synchronization through a network time protocol, and each acquisition node synchronously records a timestamp when writing data. During the alignment process, the greatest common divisor sampling period is used as the reference time step, and linear interpolation compensation is performed on data with misaligned timestamps, outputting a multi-source synchronized data stream aligned under a unified time reference.

[0028] This invention collects three types of physical quantities—mechanical motion state, pneumatic actuation state, and fluid medium state—using categorized sensors. Through calibration mapping and feature extraction, the raw signals are converted into parameter values ​​with clear physical meaning. The introduction of a globally unified clock synchronization mechanism and linear interpolation effectively eliminates time deviations caused by differences in hardware response at different acquisition nodes, ensuring strict consistency of multi-channel data in the time dimension. This provides a high-quality, low-noise data foundation for subsequent undirected graph-based state modeling, guaranteeing the data reliability of the overall control method's input.

[0029] Preferably, the construction process for assigning initial attributes to an undirected graph data structure is as follows: Figure 3 As shown, it specifically includes: Construct a topological network representing the spatial relationships of the physical production line; establish physical execution entities as spatial node entities in an undirected graph data structure; the physical execution entities include an extrusion assembly, a feeding assembly, and a traction assembly; Based on the operating status parameters and process constraint parameters, the initial mechanical pressure, operating linear velocity, and medium state characteristics are extracted and written into the corresponding physical execution entity nodes as dynamic operating attribute features; at the same time, the weaving density benchmark value and lateral width scale are extracted and written into the global environment node as static constraint attribute features. By defining the physical distance for material transfer and the response boundary for signal transmission between spatial node entities, an undirected connection path between node entities is established, and an undirected graph data structure containing multi-dimensional attribute features and spatial transmission topology is output.

[0030] The construction of the undirected graph data structure is based on the actual physical spatial relationship of the protective material coating production line.

[0031] An undirected graph is a graph theory data structure consisting of a set of nodes and a set of edges. The connection between any two nodes in the graph is directional; that is, if there is a path between node A and node B, the path is valid in both directions: from A to B and from B to A. The reason for choosing an undirected graph over a directed graph is that the physical influence between the various physical components in the protective material coating process is bidirectional. Changes in traction speed simultaneously affect the fluid spreading state at the extrusion interface, while changes in fluid pressure at the extrusion interface, in turn, create a resistance load on the traction components. Therefore, using undirected connection paths can more accurately represent this bidirectional physical coupling relationship.

[0032] In the node construction phase, the physical execution components on the production line are abstracted as spatial node entities in an undirected graph. Specifically, this includes three types of physical execution entity nodes: the extrusion component node corresponds to the blade coating head assembly, responsible for controlling the geometric gap and pressure state of the extrusion channel; the feeding component node corresponds to the fluid metering pump and injection system, responsible for controlling the injection volume and rate of the polymer fluid; and the traction component node corresponds to the traction roller drive system, responsible for controlling the linear velocity of the substrate. In addition, a global environment node is established to store static constraint parameters related to the overall physical space state of the production line.

[0033] During the dynamic operation attribute writing phase, the initial mechanical pressure is written as a numerical attribute to the dynamic attribute domain of the extrusion component node; the medium state feature vector is written as a multi-dimensional attribute to the dynamic attribute domain of the feeding component node; and the running linear velocity is written as a numerical attribute to the dynamic attribute domain of the traction component node. These writing operations are synchronously refreshed with sensor data updates in each control cycle, ensuring that the node attributes remain consistent with the real-time physical state.

[0034] During the static constraint attribute writing phase, the weaving density baseline value and the lateral width dimension are written to the global environment node as fixed attributes. These two parameters do not change with the production line's operating status within the same production batch, but are only updated when switching batches. Therefore, they are characterized as static constraint attributes, distinguishing them from the operating status parameters that are dynamically refreshed with the control cycle, reflecting the physical nature of the time-varying nature of parameters.

[0035] In the undirected connection path construction phase, undirected connection paths are established between corresponding nodes based on the actual physical distance of material transfer and the response time boundary of control signal transmission between components in the physical production line. The physical distance of material transfer refers to the path length of the substrate between two adjacent functional components; the signal transmission response boundary refers to the maximum time delay required from when a component issues a control action to when an adjacent component perceives the effect of that action. These two types of quantified values ​​are recorded as initial structural attributes of the connection path in the attribute domain of the edges, providing a path carrier for subsequent edge weight assignment.

[0036] The final output undirected graph data structure contains multi-dimensional node attribute features and a set of connection paths with spatial transmission topological relationships, providing a structured data carrier for subsequent edge association feature calculation and graph mapping structure generation.

[0037] The undirected graph modeling method proposed in this invention allocates the operating state parameters and process constraint parameters of each physical execution component in the protective material coating production line to corresponding nodes according to their physical affiliation, and explicitly expresses the spatial transmission relationship between components through undirected connection paths. The classification and inclusion of dynamic and static attributes in the design ensures both accurate perception of the real-time physical state by the control system and preserves the global availability of process constraints for production batches. Transforming unstructured multi-source physical quantities into structured graph data with topological semantics provides a physically self-consistent data foundation for subsequent edge feature derivation and time-series prediction based on graph data, preserving a complete expression of the physical coupling relationship between components at a macroscopic scale.

[0038] Preferably, the step of deriving edge association features to obtain the graph mapping structure includes: Calculate the product of the running linear speed and the braiding density baseline value, identify the periodic physical disturbance frequency of the fluid generated by the bottom substrate when it passes through the extrusion working surface, and use the periodic physical disturbance frequency as the physical disturbance characteristic. The physical disturbance characteristics are compared with the inherent rheological relaxation time in the medium state characteristics to identify the sudden change in viscous resistance generated by the fluid medium when subjected to physical disturbance. Based on the abrupt change in viscous resistance, an equivalent coupled damping characteristic characterizing the fluid medium's ability to dissipate physical disturbances is derived. The equivalent coupling damping feature is used as the weight value of the physical coupling strength between nodes, i.e. the edge association feature, and assigned to the undirected connection path in the undirected graph data structure to output a graph mapping structure containing state nodes and edge weights.

[0039] The derivation of edge association features is based on the fluid-structure interaction physical mechanism in the production process of protective materials. In the coating process, polymer fluid is coated on the surface of the underlying interwoven grid substrate. The substrate moves continuously with the traction system, and the intersecting nodes of the grid periodically extrude through the working surface, forming periodic mechanical disturbances on the fluid layer at the working surface.

[0040] The principle for calculating the physical disturbance frequency is as follows: For every warp and weft spacing advanced by the substrate, a complete grid node passage event occurs at the extrusion working surface. This event generates a physical disturbance pulse of microscopic compression and release in the fluid. Therefore, the number of disturbance pulses occurring per unit time, i.e., the physical disturbance frequency, is equal to the product of the running linear velocity and the weaving density reference value. The running linear velocity is measured in centimeters per second, and the weaving density reference value is measured in warp and weft threads per centimeter. The product of these two values ​​is measured in times per second, meaning the physical disturbance frequency is measured in Hertz (Hz).

[0041] The identification of abrupt changes in viscous drag is based on the linear viscoelasticity theory of non-Newtonian fluids. Polymer coatings are non-Newtonian fluids, and their rheological properties are determined by both elastic and viscous moduli. The inherent rheological relaxation time of a fluid is the characteristic time required for it to recover to equilibrium after deformation, reflected by the viscoelastic parameters in the medium's state characteristic vector. When the disturbance period corresponding to the physical disturbance frequency is much greater than the inherent rheological relaxation time, the fluid has sufficient time to relax within the disturbance gap, and the change in viscous drag is relatively gradual. When the disturbance period is close to or less than the inherent rheological relaxation time, the fluid does not have enough time to relax before being subjected to the next disturbance, the elastic effect is significantly enhanced, and the viscous drag increases abruptly. In this embodiment, the ratio of the reciprocal of the physical disturbance frequency (disturbance period) to the inherent rheological relaxation time is defined as the dimensionless rheological response ratio. When the ratio is less than 1, the fluid is identified as being in the elastic-dominant response range, and the viscous drag abrupt change is quantified according to the inverse relationship with the ratio. When the ratio is greater than or equal to 1, the fluid is identified as being in the viscous-dominant response range, and the viscous drag abrupt change is small, and is calculated according to an approximately linear relationship.

[0042] The derivation of the equivalent coupling damping characteristic employs a physical modeling method based on energy dissipation. Damping is a physical quantity describing the ability of a physical system to dissipate vibrational energy. In this embodiment, the equivalent coupling damping characteristic is defined as an index of the equivalent ability of a fluid medium to dissipate mechanical vibration energy within a single disturbance pulse cycle. Specifically, the calculation method is as follows: based on the abrupt change in viscous drag, combined with the fluid density parameter and fluid velocity estimate in the medium's state characteristic vector, and following the calculation principle of viscous dissipation power, the dissipation power of the fluid medium on mechanical excitation energy per unit time is derived. This dissipation power is then normalized to the standard disturbance amplitude to obtain the dimensionless value of the equivalent coupling damping characteristic.

[0043] When assigning equivalent coupling damping features to the corresponding undirected connection paths in the undirected graph data structure, for the connection path between the extrusion component node and the feeding component node, the equivalent coupling damping feature calculated at the fluid medium conduction interface between them is used as the weight value; for the connection path between the traction component node and the extrusion component node, the equivalent coupling damping feature calculated at the material conduction interface between them is used as the weight value. After the assignment is completed, each connection path in the undirected graph data structure carries edge weight information that quantifies the physical coupling strength, forming the final graph mapping structure, providing a complete state snapshot of each control cycle for the subsequent construction of the time-series state sequence.

[0044] This invention interprets the physical product of the running linear velocity and the weaving density benchmark as the mechanical excitation frequency, and combines this with dimensionless ratio analysis using the fluid's inherent rheological relaxation time. This achieves a quantitative characterization of the complex coupling relationship between mechanical excitation characteristics and the non-Newtonian rheological properties of the fluid. The introduction of equivalent coupling damping features as edge weights in the undirected graph transforms the graph data structure from a purely topological description into a weighted graph structure carrying the true physical coupling strength. This provides a physically meaningful high-level feature representation for subsequent time-series prediction networks, solving the regulation effect problem caused by traditional control systems neglecting the coupling relationship between underlying disturbance signals and rheological responses.

[0045] Preferably, the process of deriving the dynamic runaway index includes: According to the time sequence, the graph mapping structure within the set number control period is continuously extracted, and the spatial dimension graph model is stacked and expanded in the time dimension to construct a temporal state sequence containing spatiotemporal coupling information. The time-series state sequence is input into a time-series prediction network model containing a memory-forgetting gating unit, and the acceleration and slope of the change of physical state and edge weights on the time axis are identified through the neuron nodes inside the model. Based on the changing acceleration and slope, the dynamic nonlinear characteristics of the non-proportional changes in the state distribution of the controlled object that will occur in future periods are extracted. By comparing the distance between the dynamic nonlinear characteristics and the built-in historical evolution samples, the absolute deviation trend of the expected deviation from the target set value is deduced while maintaining the current control action unchanged, and the output is the dynamic runaway index.

[0046] The temporal state sequence is constructed as follows: Using the control period as the time step, graph mapping structures generated within the most recent set number of control periods are extracted sequentially. Each graph mapping structure consists of two parts: a node attribute matrix and an edge weight matrix. The number of rows in the node attribute matrix equals the number of nodes, and the number of columns equals the attribute dimension of each node. The number of rows and columns in the edge weight matrix equals the number of nodes, and the matrix elements are the equivalent coupling damping feature values ​​of the connection paths between corresponding node pairs. The node attribute matrices and edge weight matrices from the set number of control periods are stacked sequentially along the third dimension to form a three-dimensional tensor containing the time step, number of nodes, and attribute dimension, which is the temporal state sequence. This sequence simultaneously encodes spatial topological features and temporal evolution information, providing a spatiotemporally coupled input representation for the model.

[0047] The architecture of the time-series prediction network model is as follows: This model is a multi-layered recurrent neural network, with the core computational unit being the Long Short-Term Memory (LSTM) gating unit. The LSM gating unit is a type of recurrent neural network computational unit that solves the gradient vanishing and gradient exploding problems inherent in traditional recurrent neural networks when processing long sequence data by introducing three gating mechanisms: an input gate, a forget gate, and an output gate. The forget gate determines which information from the previous time-series memory state needs to be forgotten; the input gate determines which input information from the current time-series needs to be written into the memory state; and the output gate extracts information useful for the current time-series output from the current memory state. These three gating mechanisms work together to enable the model to selectively retain historical state information, effectively capturing the evolution of the controlled object's state over a longer time span.

[0048] The model input layer receives a 3D temporal state sequence, unfolds the sequence step-by-step, and feeds it into a Long Short-Term Memory (LSTM) gating unit for sequence processing. At each time step, the model processes the flattened vectors of the node attribute matrix and edge weight matrix at the current time step, and updates the memory state at the current time step by combining them with the hidden state vector from the previous time step. After iterative processing through all time steps, the model extracts the hidden state vector at the final time step and feeds it into the fully connected output layer. The output layer has two parallel output branches: the first branch outputs the slope vector of the physical state and edge weights on the time axis, representing the current rate of change of each state variable; the second branch outputs the acceleration vector, representing the trend of the rate of change of each state variable itself (i.e., the derivative of the slope with respect to time). The acceleration can reflect the turning point signal of the state variable's evolution trend in advance, which is of great significance for predicting impending nonlinear deterioration.

[0049] Specifically, the temporal prediction network model comprises two stacked LSTM layers. The first LSTM layer has a hidden unit dimension of 128, and the second layer has a hidden unit dimension of 64. Residual connections are used between layers to mitigate gradient decay in deep networks. The input processing method is as follows: the node attribute matrix of each time step in the temporal state sequence is flattened row-wise into a node attribute vector, and the upper triangular elements of the edge weight matrix are flattened into an edge weight vector. The node attribute vector and the edge weight vector are concatenated and used as the input feature vector for that time step, which is then fed into the first LSTM layer. The second LSTM layer outputs the hidden state vector at the final time step, which is fed into two parallel fully connected output branches: the first branch outputs a slope vector (with the same dimension as the input feature vector), and the second branch outputs an acceleration vector (with the same dimension as the input feature vector). The slope vector and acceleration vector are concatenated and fed into a dual-hidden-layer fully connected feature mapping sub-network: the first hidden layer contains 32 neurons, and the second hidden layer contains 16 neurons, both using the leaky linear rectified activation function, with a fixed slope of 0.01 in the negative region, and outputting a dynamic nonlinear feature vector with a dimension of 8. The model's training data is constructed as follows: data from multiple production batches are extracted from a historical production database. Each batch contains continuous control cycle records and corresponding offline quality inspection results, divided proportionally into training, validation, and test sets. The quality inspection result label is the measured standard deviation of the coating's lateral thickness, taking consecutive real numbers greater than 0. The model employs an end-to-end joint training approach, using mean squared error as the loss function and an adaptive moment estimation optimizer for parameter updates. The initial learning rate is set to 0.001, and the batch size is set to 32. Training terminates when the validation set loss no longer decreases after 10 consecutive iterations. After training, the model parameters are compressed using 8-bit integer quantization to ensure that the inference time within a single control cycle does not exceed 50% of the control cycle's time.

[0050] The process of extracting dynamic nonlinear features from changing acceleration and slope includes: concatenating the slope vector representing the current rate of change with the acceleration vector representing the evolution trend to construct a feature concatenation vector; inputting the feature concatenation vector into a dual-hidden-layer fully connected feature mapping subnetwork, and performing nonlinear mapping processing through a leakage linear rectified activation function with a small gradient preservation mechanism to output a high-dimensional dynamic nonlinear feature vector; each data dimension of the dynamic nonlinear feature vector is constrained to correspond to specific physical evolution trend components such as fluid shear thickening trend and substrate edge tension attenuation.

[0051] The specific implementation process for extracting dynamic nonlinear features from changing acceleration and slope involves firstly aggregating information features within the time-series prediction network model, concatenating the output vectors of the two branches to form a concatenated feature vector. This concatenated feature vector is then fed into a two-layer fully connected feature mapping sub-network. The first fully connected layer breaks the independence of individual physical quantities, performing deep cross-combination and physical correlation analysis of mechanical slope and rheological acceleration; the second fully connected layer is responsible for refining and dimensionality-reducing the correlated features. During the mapping process in the dual-hidden-layer network, a leaky linear rectified activation function with a small gradient preservation mechanism is used. After processing by the nonlinear activation function, a dynamic nonlinear feature vector is output. Compared to conventional nonlinear activation mechanisms, this mechanism does not produce gradient vanishing when receiving weak negative physical disturbance signals, ensuring that even extremely small early fluctuations in non-Newtonian fluid rheology can be accurately mapped by the network. This feature vector, in a high-dimensional feature space, represents the non-proportional change pattern of the controlled object's state distribution that will occur over several future control cycles. Ultimately, each dimension corresponds to a component of the evolution trend of a type of physical state. For example, the first dimension of the network output precisely corresponds to the microscopic shear thickening surge trend in the central region of the coating, while the second dimension corresponds to the sudden relaxation trend of tension at the edge of the wide substrate. This achieves a precise translation from high-dimensional mathematical tensors to specific physical fault phenomena in the workshop.

[0052] This invention eliminates the potential for neuronal feature loss in the early stages of handling nonlinear variations in microfluidics by introducing an activation function with a small gradient preservation mechanism. This allows for the sensitive and complete capture of very early precursors of minute coating defects. Simultaneously, by forcibly binding the abstract high-dimensional output vector dimension to specific physical evolution components such as shear thickening and tension decay, the interpretability and targeted prevention capabilities of the industrial automation control system are enhanced, providing a pre-decision basis with absolutely clear physical meaning for multi-variable collaborative decoupled control.

[0053] The dynamic runaway index is derived by comparing its distance with a historical evolution sample database. This database is pre-constructed from historical production data after model training. Each sample record in the database corresponds to a dynamic nonlinear feature vector in the historical production process and its corresponding subsequent measured quality deviation value. During distance comparison, the Euclidean distance between the current dynamic nonlinear feature vector and the feature vector of each sample in the database is calculated. The closest samples are used as the search results. A weighted average of the subsequent quality deviation values ​​corresponding to these samples is then performed, with the weights calculated as the reciprocal of the distance. This yields an estimate of the absolute deviation trend from the target value. This estimate is then normalized to the range of 0 to 100, and the output is the dynamic runaway index.

[0054] The historical evolution sample library was initially constructed to cover three operating conditions: normal operation, minor anomalies, and severe anomalies. The sample library employs a rolling update mechanism: after each production batch, the actual dynamic nonlinear feature vectors and corresponding quality deviation values ​​from that batch are added to the library as new samples; simultaneously, old samples older than 12 months are discarded according to a first-in, first-out (FIFO) principle. For distance comparison, an approximate nearest neighbor retrieval method based on a KD-tree is used to retrieve the five closest samples from the sample library, and their subsequent quality deviation values ​​are weighted using the inverse of the distance as the weight.

[0055] The model training process is as follows: Time-series graph mapping structure sequences labeled with quality inspection results are extracted from historical production data as the training sample set. During training, the time-series state sequences are used as input, and the measured quality deviation values ​​of the corresponding production batches are used as supervision labels. Mean squared error is used as the loss function, and an adaptive gradient optimization algorithm is used to iteratively update all model parameters. The convergence of the loss function is monitored on the validation set, and training terminates when the loss no longer decreases. After training, the model is quantized and compressed to meet the latency requirements of real-time inference within the control cycle.

[0056] The proposed temporal prediction network model based on a memory-forgetting gating unit integrates spatial topology and time-series information into a joint model of temporal state sequences, overcoming the limitation of traditional control methods that rely solely on static parameters at the current moment for judgment. By simultaneously outputting two types of features—the slope and acceleration of change—it can identify the nonlinear deterioration trend of the controlled object's state distribution before quality deviations actually occur. Furthermore, by comparing historical sample distances, the trend is transformed into a quantifiable runaway risk index, providing a sufficient smooth intervention window for preventative action by the actuator, thus upgrading control decisions from passive response to proactive defense.

[0057] Preferably, the step of identifying the control response hysteresis characteristics includes: When the value of the dynamic runaway index is detected to be greater than the preset safety control threshold, the lateral width scale is retrieved, and combined with the current running linear speed, initial mechanical pressure, and the material's inherent stiffness mapped by the weaving density benchmark value, the elastic deformation gradient generated at different physical positions in the lateral direction of the wide flexible material under the interlacing tension and compression of multiaxial mechanical forces is identified. Based on the elastic deformation gradient, the absolute time difference between the central region and the two edge regions of the extrusion interface when the controller issues a synchronous execution command is evaluated. The time difference distribution set at each horizontal node is used to construct a control response hysteresis characteristic curve characterizing the physical space transmission delay, thus obtaining the control response hysteresis characteristic. The preset safety control threshold is determined based on the 85th percentile of the historical runaway index statistical distribution of this production batch.

[0058] The principle of elastic deformation gradient identification is based on the elastic mechanical analysis of flexible wide-width materials under multiaxial mechanical loading conditions. The protective material substrate is a flexible mesh structure with interwoven warp and weft, exhibiting anisotropic elastic characteristics. During the coating process, the substrate simultaneously bears two types of mechanical loads: one is the constant traction tension applied longitudinally by the traction system, and the other is the mechanical compressive force applied normally by the extrusion assembly. Under the interplay of these two types of loads, the substrate does not exhibit a uniform stress state in the transverse direction. Instead, due to the distribution characteristics of the elastic constraints within the material, the normal displacement response at different transverse positions differs, i.e., an elastic deformation gradient exists. The elastic deformation gradient is defined as the ratio of the difference in elastic deformation at adjacent transverse positions to the distance between those positions, reflecting the rate of change of elastic deformation along the transverse direction.

[0059] The quantitative calculation method for the elastic deformation gradient is as follows: The transverse width is discretized into several equally spaced transverse sampling nodes along the width direction. The number of sampling nodes is determined according to the transverse width scale and the required calculation accuracy, typically one sampling node is set every 100 mm. For each transverse sampling node, based on the weaving density benchmark value, the weaving structure is mapped to anisotropic elastic modulus parameters according to the principle of equivalent warp and weft structural stiffness, thus obtaining a transverse distribution characterization of the material's inherent stiffness. Combining the distance of the node from the transverse center position and the estimated value of the transverse component of the traction tension, the elastic deformation of the substrate at that node under the action of compressive force, i.e., the normal displacement value, is calculated according to the elastic catenary model.

[0060] The elastic catenary model is a classic mechanical model describing the deformation distribution of a flexible strip under the combined action of a uniformly distributed transverse load and concentrated end tension. Its model parameters include the equivalent bending stiffness of the substrate, the traction tension per unit width, and the normal pressure per unit area. These parameters can all be calculated from known operating parameters and process constraint parameters. Dividing the difference in elastic deformation between adjacent transverse nodes by the node spacing yields the elastic deformation gradient value for that node pair. The elastic deformation gradient values ​​of all transverse nodes constitute the elastic deformation gradient distribution, which typically exhibits a non-uniform distribution characteristic, being smaller in the middle and larger at both ends.

[0061] The evaluation method for control response time difference is as follows: When the control system issues a synchronous execution command, the command objective is to adjust the force state of each lateral position at the extrusion interface to the same target set value. Due to the different elastic deformation of each lateral position, the elastic restoring force required to overcome from the current force state to the target force state varies at each position. Edge regions with larger elastic deformation require overcoming greater elastic restoring forces, and the time required for the servo drive mechanism to reach the target force state under limited driving force conditions is correspondingly longer. Based on the elastic deformation gradient values ​​of each lateral node, combined with the rated output force parameters of the servo drive system and the rotational inertia parameters of the transmission mechanism, the response time required for each lateral node to reach the target set value from the issuance of the command is calculated according to the dynamic response model of the second-order mass spring damping system. The core parameters of the dynamic response model of the second-order mass spring damping system include equivalent mass (derived from the inertia of the transmission mechanism), equivalent spring stiffness (derived from the elastic deformation gradient of the material), and equivalent damping coefficient (determined by a combination of friction and material viscoelasticity). All of the above parameters have clear physical correspondences and can be obtained from known equipment parameters. Using the response time of the horizontal center node of the extrusion interface as a benchmark, the difference between the response time of each horizontal node and the benchmark response time is calculated to obtain the response time difference of each node.

[0062] The control response hysteresis characteristic curve is constructed as follows: Using the lateral position coordinates (with the center point of the lateral width scale as the origin) as the independent variable and the response time difference of each lateral node as the dependent variable, cubic spline interpolation is performed on the discrete point set formed by the values ​​of the independent and dependent variables of each sampling node. Cubic spline interpolation is a piecewise polynomial interpolation method that ensures the interpolation curve has continuous second derivatives at each data point. It can generate smooth and continuous curves and is insensitive to sample noise, making it suitable for describing continuous distribution characteristics in physical space. After interpolation, a continuous control response hysteresis characteristic curve is obtained. This curve fully describes the spatial physical transmission delay distribution law along the lateral direction, providing quantitative spatial constraint boundary conditions for the subsequent generation of multi-objective cooperative control schemes.

[0063] This invention, through quantitative analysis of the lateral elastic deformation gradient of wide-width flexible materials, explicitly characterizes the spatial response non-uniformity of the control system in the form of a response time difference curve. It breaks through the simplistic assumption of treating the extrusion interface as a uniform rigid body in traditional methods, truly reflecting the non-uniform lateral deformation characteristics of ultra-wide-width flexible materials under multi-axis mechanical loading. It reveals the cause of control response lag in the edge region from a physical mechanism perspective, providing a quantitative basis for subsequently generating precise control commands that can compensate for spatial response lag, and solving the quality defect problem of coating material shortage in the edge region caused by ignoring lateral deformation differences.

[0064] Preferably, the steps for dynamic adjustment and control include: Based on the control response hysteresis characteristics and the control objective of reducing the dynamic runaway index, a multi-dimensional state matrix including mechanical space geometric parameters and medium supply kinetic energy parameters is established. Through a multi-objective constraint solving algorithm, multiple sets of comprehensive candidate control schemes with different numerical ratios are calculated and output simultaneously. Each set of candidate control schemes includes an end pose compensation command for reshaping the lateral spatial geometric gap of the extrusion channel, and a feed servo adjustment command for changing the total amount of fluid medium injected. The comprehensive candidate control scheme set is input into the control scheme evaluation module, and the comprehensive fitness score of each group of candidate control schemes is calculated using the minimization of fluid lateral distribution variance and servo drive energy peak as dual evaluation indicators. The candidate control scheme with the highest score is selected, converted into electrical-level drive commands, and sent to the corresponding physical execution components to complete the dynamic closed-loop regulation under the strong coupling state of multiple physics fields.

[0065] The multidimensional state matrix is ​​constructed as follows: the rows of the matrix correspond to the sampling nodes in the lateral width direction, and the number of rows is equal to the number of sampling nodes in the control response hysteresis characteristic curve; the columns of the matrix correspond to two types of control parameter dimensions. The first type is mechanical spatial geometric parameters, specifically the target geometric gap adjustment amount of the extrusion channel at each lateral sampling node; the second type is medium supply kinetic energy parameters, specifically the injection flow rate adjustment amount of the fluid medium. The initial values ​​of each element in the mechanical spatial geometric parameter column of the matrix are obtained by looking up the response time difference of the corresponding node in the control response hysteresis characteristic curve according to a pre-calibrated table of correspondence between response time difference and pose compensation amount. The mapping principle is that nodes with larger response time differences correspond to larger end pose pre-compensation amounts to offset the error caused by the response hysteresis at that node, which prevents the force state from reaching the target value.

[0066] The correspondence table is based on the closed-loop response speed of the extrusion component and the stiffness coefficient of the servo drive motor. The mapping relationship satisfies that the pose pre-compensation amount equals the product of the node response time difference and the average action rate of the extrusion component. The initial values ​​of each element in the medium supply kinetic energy parameter column of the matrix are estimated based on the difference between the current dynamic runaway index and the target runaway index. These initial values ​​constitute the initial population of the multi-objective constraint solving algorithm, serving as the starting basis for algorithm iteration.

[0067] The specific implementation of the multi-objective constraint solution algorithm is as follows: The algorithm uses each element of the multi-dimensional state matrix as decision variables, with the hard constraint being to reduce the dynamic runaway index below the safety threshold, and the boundary constraints being the physical feasible range of the adjustment amount of the lateral geometric gap of the extrusion channel and the upper and lower limits of the fluid injection adjustment amount. Within the feasible solution space that satisfies the above constraints, the algorithm simultaneously searches for multiple sets of candidate solutions with different numerical ratios using the Pareto multi-objective optimization framework. Pareto multi-objective optimization is a mathematical optimization method that can find a set of trade-off solutions among multiple competing optimization objectives. Its output is a set of Pareto optimal solutions, meaning that no single solution is superior to another solution in all objectives. Each Pareto optimal solution represents a decision scheme with a specific trade-off relationship among different objectives.

[0068] In this embodiment, an initial candidate solution set is generated according to a preset population size. An iterative evolutionary method with an elite retention strategy is used to simulate crossover and random mutation updates on the candidate solutions. After each iteration, all candidate solutions are screened for feasibility based on constraint satisfaction. Feasible solutions that satisfy the constraints are retained for the next iteration. After a set number of iterations, a comprehensive candidate control scheme set is output. Each scheme in the scheme set includes the numerical distribution of end-position pose compensation commands for each lateral node (i.e., the target position deviation of the extrusion head end at each node) and the numerical values ​​of feed servo adjustment commands (i.e., the target speed adjustment of the fluid metering pump and the opening adjustment of the lateral sub-control valve group).

[0069] Specifically, the multi-objective constraint solving algorithm sets the population size to 50 and the maximum number of iterations to 100. The crossover operation uses a simulated binary crossover operator with a crossover probability of 0.9 and a distribution exponent of 20; the mutation operation uses a multinomial mutation operator with a mutation probability of 0.1 and a distribution exponent of 20. Constraint handling employs a penalty function method, adding solutions that exceed boundary constraints by multiplying the square of the excess by a penalty factor of 1000 to the objective function value. An elite retention strategy is adopted, retaining the top 10 solutions with the largest crowding distance from the Pareto front in each iteration, directly advancing to the next generation. After iteration terminates, the top 10 solutions from the Pareto front, ranked from largest to smallest crowding distance, form a comprehensive candidate control scheme set.

[0070] The dual evaluation indexes of the control scheme evaluation module are designed as follows: The first evaluation index is the fluid lateral distribution variance, defined as the statistical variance of the predicted fluid coating thickness at each node in the lateral direction of the extrusion interface after the candidate control scheme is issued and executed. The smaller the variance, the better the lateral uniformity of the coating layer; this index is optimized by minimizing it. The predicted fluid coating thickness is estimated forward based on the pose compensation and flow adjustment of the candidate scheme, combined with the fluid thin-layer flow dynamics. Specifically, the geometric gap adjustment and flow adjustment of each node are mapped proportionally to the coating thickness change, and then superimposed on the current measured thickness baseline value to obtain the predicted value. The second evaluation index is the servo drive energy peak, defined as the maximum instantaneous power of all servo drive motors when executing the candidate control scheme. The smaller the energy peak, the smaller the impact on the electrical drive system during execution, which is beneficial for extending equipment life and reducing grid disturbance; this index is also optimized by minimizing it. For each candidate control scheme, the values ​​of the two indicators mentioned above are calculated separately, and the values ​​are summed according to the preset weight coefficients to obtain the comprehensive fitness score of the scheme. The lower the score, the better the comprehensive performance of the scheme. The sum of the weight coefficients is 1, and the default is 0.5.

[0071] The process of selecting the optimal solution and issuing commands is as follows: The optimal integrated control solution is selected from the set of comprehensive candidate control solutions, based on the overall fitness score. The end-effector pose compensation command in this solution is converted into the target position deviation of each axis servo motor according to the lateral node positions. This deviation is then converted into the corresponding drive current command by the position loop regulator in the electrical control layer and issued to the servo drives of each axis of the extrusion assembly. The flow rate adjustment in the feed servo adjustment command is converted into the speed command signal for the metering pump drive motor and issued to the drive controller of the feeding assembly. Both types of commands are simultaneously issued to the corresponding physical execution components to complete the dynamic closed-loop adjustment under the strong coupling state of multiple physics fields. The adjustment effect is verified through sensor feedback in the next control cycle, and the cycle continues until the dynamic runaway index falls below the safety threshold.

[0072] The multi-objective collaborative control scheme generation and evaluation mechanism proposed in this invention uses the control response hysteresis characteristic curve as the spatial constraint boundary. Through a Pareto multi-objective optimization framework, it synchronously generates a comprehensive candidate scheme set that considers both end-effector pose compensation and feed servo adjustment. Scheme selection is then performed using both lateral distribution uniformity and driving energy peak values. This achieves precise coordination of mechanical geometric parameters and fluid supply parameters in the lateral space, effectively resolving the inherent physical contradiction between ultra-wide material total quantity adjustment and material shortage in edge regions. While ensuring the lateral uniformity of the coating layer, it reduces the peak energy consumption during dynamic adjustment, thereby improving the overall operational stability of the production line and the service life of the equipment.

[0073] Example 2:

[0074] This invention proposes a method for dynamically adjusting and controlling risk parameters in the production process of protective materials, comprising: Obtain operating status parameters and process constraint parameters. The operating status parameters include operating linear velocity, initial mechanical pressure, and medium state characteristics. The process constraint parameters include weaving density reference value and transverse width dimension. The operating state parameters and process constraint parameters are assigned as initial attributes to the undirected graph data structure; the physical disturbance characteristics are identified by combining the weaving density benchmark value and the operating linear velocity, and the equivalent coupling damping characteristics are derived as edge association characteristics; the edge association characteristics are assigned to the undirected graph data structure to obtain the graph mapping structure. Extract graph mapping structures from multiple cycles to construct a time-series state sequence; input the time-series state sequence into a time-series prediction network model to extract the dynamic nonlinear features of the state distribution of the controlled object evolving with the time array, and deduce the dynamic runaway index. When the dynamic runaway index exceeds the safety threshold, the lateral width scale is extracted, and the control response hysteresis characteristics of the extrusion interface under the action of multi-axis coupled servo force are identified. Based on the control response hysteresis characteristics, control commands including end pose compensation commands and feed servo adjustment commands are identified and output for dynamic adjustment control.

[0075] Furthermore, the step of inputting the time-series state sequence into the time-series prediction network model containing the memory-forgetting gating unit also includes a dynamic bias adjustment process: synchronously calculating the absolute difference of the dynamic operating attribute characteristics of the traction component nodes in the undirected graph data structure within adjacent control cycles, characterizing the mechanical transient impact intensity generated when the splicing seam of the underlying substrate passes through the extrusion component; converting the mechanical transient impact intensity into a negative bias adjustment amount, dynamically injecting it into the memory-forgetting gating unit inside the time-series prediction network model, increasing the forgetting weight allocation of the historical state sequence; in the first prediction cycle after the mechanical transient impact occurs, making the model rely on the distribution of newly formed states after the impact to perform dynamic nonlinear feature extraction, weakening the interference weight of the failure history evolution samples before the impact.

[0076] During the production of protective materials in rolls, there are physical seams between two rolls of fabric. When this seam passes through the traction roller, it generates huge instantaneous tension fluctuations (a dramatic increase in the absolute value of the difference). At this point, the stable operating data before the seam becomes meaningless for predicting the state after the seam. For the specific implementation of the dynamic bias adjustment process, the real-time traction resistance scalar of the traction component node in the undirected graph data structure is extracted as a dynamic operating attribute feature in each control cycle. The real-time traction resistance scalar of the current control cycle is subtracted from the same scalar of the previous adjacent control cycle, and the absolute value of the difference is obtained. Subsequently, this absolute value of the difference is divided by a preset normal fluctuation extreme value for normalization, and the output quotient is the quantified mechanical transient impact intensity. Next, this mechanical transient impact intensity is multiplied by a preset negative proportional gain constant, converting it into a negative bias adjustment amount, specifically used for internal intervention in the time-series prediction network model.

[0077] The preset extreme value of normal fluctuation is taken as the 95th percentile of the absolute value of the difference between adjacent values ​​of the traction resistance scalar within 100 consecutive control cycles under seamless and stable operation of the production line. The preset negative proportional gain constant is set to -2.0, so that when the normalized value of the mechanical transient impact intensity reaches 1.0, the output value of the forget gate is attenuated to approximately 0.12 times the original output value.

[0078] The temporal prediction network model employs a recurrent neural network architecture comprising cellular memory states, an input gate, a forget gate, and an output gate. For the internal data flow of the forget gate, the conventional operation rule is to linearly weight and sum the current input feature sequence with the previous hidden state sequence, then pass it through a nonlinear mapping function to output a retained weight between zero and one. After introducing adaptive intervention, the data processing module adds the aforementioned negative bias adjustment as an additional mandatory penalty term directly to the pre-linear summation polynomial of the nonlinear mapping function within the forget gate. When the physical seam of the protective material experiences a strong mechanical transient impact through the drive shaft, the absolute value of this negative bias adjustment increases sharply, forcing the output value of the nonlinear mapping function of the forget gate to approach zero infinitely. This operation allows the model to actively and completely sever the reading weights of the cellular memory state from the previous time step in the current control cycle, completing the physical clearing of historical contamination data. Subsequently, the model reconstructs the network memory solely based on the newly generated sensor input after the seam passes through, outputting pure dynamic nonlinear features that discard historical contamination fluctuations.

[0079] This invention provides an adaptive forgetting gating adjustment mechanism based on mechanical shock, which connects sudden mechanical anomalies in industrial settings with the neuronal gating mechanism at the bottom layer of deep learning networks. Traditional time-series prediction models exhibit a smooth decay characteristic in their historical memory weights. When production lines encounter strong, instantaneous physical impacts such as those at substrate splicing seams, they are prone to making sluggish or even contradictory loss-of-control warnings due to over-reliance on stable historical data prior to the impact. This invention transforms the absolute value of the difference in traction resistance into a trigger signal that forcibly opens the forgetting gate, endowing the control center with the ability to rapidly forget and reshape its state when encountering sudden changes in the physical environment. This eliminates historical data contamination caused by sudden disturbances and maintains a high dynamic prediction accuracy.

[0080] Furthermore, in the step of identifying the elastic deformation gradient generated at different physical locations in the lateral direction, an asymmetric fluid mass load correction mechanism can be introduced: extract the medium state characteristics in the global environment nodes and the historical cumulative amount of feed servo adjustment in the previous cycle, and combine them with the actual thickness profile of the extrusion interface to calculate the mass non-uniform distribution load matrix of the fluid medium in the lateral space; superimpose the mass non-uniform distribution load matrix as an additional vertical gravity component into the multi-axis mechanical intertwined tension model to perform asymmetric perturbation correction on the ideal symmetrical elastic deformation gradient distribution; based on the elastic deformation gradient after asymmetric perturbation correction, re-evaluate the absolute time difference of each node and output the true control response hysteresis characteristics.

[0081] To implement the asymmetric fluid mass load correction mechanism, the inherent density constant of the fluid is first retrieved from the global environment node, and the cumulative volume injection at each lateral injection port within a set number of control cycles is extracted from the controller's historical log. Simultaneously, the actual coating thickness at each lateral sampling node is obtained using a laser thickness gauge positioned behind the extrusion assembly. During the data extrapolation phase, the theoretically consumed volume carried away by the substrate is subtracted from the local cumulative volume injection, and combined with the actual coating thickness profile, the transient fluid accumulation volume in each lateral sampling interval in front of the extrusion interface is calculated using a spatial integration algorithm. Subsequently, the transient fluid accumulation volume in each interval is multiplied by the fluid's inherent density constant to calculate the local transient gravity value at each lateral node, which is then arranged and combined according to the lateral coordinate order to form a one-dimensional non-uniform mass load matrix. This non-uniform mass load matrix is ​​used as an additional continuously distributed load in the vertical gravity direction, and is added node by node to the torque balance differential equation of the mechanical model for joint solution. After numerically solving the partial differential equation, a true profile curve of gravitational tilting and secondary physical sinking towards the fluid mass accumulation side is output. Finally, based on this reconstructed asymmetric profile curve, the spatial position is re-differentiated to generate an asymmetric elastic deformation gradient corrected for gravity, thereby calculating the absolute time difference of each node with mass hysteresis effect.

[0082] In the actual production of wide-width protective materials, polymer fluids possess a certain mass and are prone to lateral local aggregation during rheological fluctuations. This non-uniform distribution of fluid mass directly disrupts the original force symmetry of the flexible substrate, leading to localized secondary concave deformation. This technical solution transforms the historical accumulated mass of the fluid medium into a real lateral spatial mass load, reconstructing the theoretical mechanical model with asymmetric physical perturbation. This allows the accurately evaluated control response hysteresis characteristics to precisely address the microscopic execution blind spot caused by the fluid's own weight, providing a reliable and realistic geometric basis for subsequent end-effector pose compensation commands.

[0083] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. A method for dynamically adjusting and controlling risk parameters in the production process of protective materials, characterized in that, include: Obtain operating status parameters and process constraint parameters. The operating status parameters include operating linear velocity, initial mechanical pressure, and medium state characteristics. The process constraint parameters include weaving density reference value and transverse width dimension. The operating state parameters and process constraint parameters are assigned as initial attributes to the undirected graph data structure; the physical disturbance characteristics are identified by combining the weaving density benchmark value and the operating linear velocity, and the equivalent coupling damping characteristics are derived as edge association characteristics; Edge association features are assigned to an undirected graph data structure to obtain a graph mapping structure; Extract graph mapping structures from multiple cycles to construct a time-series state sequence; input the time-series state sequence into a time-series prediction network model to extract the dynamic nonlinear features of the state distribution of the controlled object evolving with the time array, and deduce the dynamic runaway index. When the dynamic runaway index exceeds the safety threshold, the lateral width scale is extracted to identify the control response hysteresis characteristics of the extrusion interface under the action of multi-axis coupled servo force. Based on the control response hysteresis characteristics, control commands including end-effector pose compensation commands and feed servo adjustment commands are identified and output for dynamic adjustment control.

2. The method for dynamic adjustment and control of risk parameters in the production process of protective materials according to claim 1, characterized in that, The process of obtaining the operating status parameters and process constraint parameters includes: The system acquires operating status parameters through a sensor network deployed on the underlying physical equipment, reads the rotation pulse signal of the traction drive shaft and converts it into a digital linear velocity, collects the feedback level signal of the extrusion pneumatic component to extract the initial mechanical pressure, and uses the damping sensing unit built into the pipeline to acquire the viscous resistance signal of the fluid as a medium state characteristic. Retrieve the pre-set material attribute list for the current production batch and extract the base value of the weaving density and the transverse width of the substrate as process constraint parameters. The collected operating status parameters and process constraint parameters are aligned with timestamps based on a globally unified clock, and a multi-source synchronized data stream is output.

3. The method for dynamic adjustment and control of risk parameters in the production process of protective materials according to claim 1, characterized in that, The steps for assigning initial attributes to an undirected graph data structure include: Construct a topological network representing the spatial relationships of the physical production line; establish physical execution entities as spatial node entities in an undirected graph data structure; the physical execution entities include an extrusion assembly, a feeding assembly, and a traction assembly; Based on the operating status parameters and process constraint parameters, the initial mechanical pressure, operating linear velocity, and medium state characteristics are extracted and written into the corresponding physical execution entity nodes as dynamic operating attribute features; at the same time, the weaving density benchmark value and lateral width scale are extracted and written into the global environment node as static constraint attribute features. By defining the physical distance for material transfer and the response boundary for signal transmission between spatial node entities, an undirected connection path between node entities is established, and an undirected graph data structure containing multi-dimensional attribute features and spatial transmission topology is output.

4. The method for dynamic adjustment and control of risk parameters in the production process of protective materials according to claim 1, characterized in that, The steps for deriving edge association features to obtain the graph mapping structure include: Calculate the product of the running linear speed and the braiding density baseline value, identify the periodic physical disturbance frequency of the fluid generated by the bottom substrate when it passes through the extrusion working surface, and use the periodic physical disturbance frequency as the physical disturbance characteristic. The physical disturbance characteristics are compared with the inherent rheological relaxation time in the medium state characteristics to identify the sudden change in viscous resistance generated by the fluid medium when subjected to physical disturbance. Based on the abrupt change in viscous resistance, an equivalent coupled damping characteristic characterizing the fluid medium's ability to dissipate physical disturbances is derived. The equivalent coupling damping feature is used as the weight value of the physical coupling strength between nodes, i.e. the edge association feature, and assigned to the undirected connection path in the undirected graph data structure to output a graph mapping structure containing state nodes and edge weights.

5. The method for dynamic adjustment and control of risk parameters in the production process of protective materials according to claim 1, characterized in that, The process of deriving the dynamic runaway index includes: According to the time sequence, the graph mapping structure within the set number control period is continuously extracted, and the spatial dimension graph model is stacked and expanded in the time dimension to construct a temporal state sequence containing spatiotemporal coupling information. The time-series state sequence is input into a time-series prediction network model containing a memory-forgetting gating unit, and the acceleration and slope of the change of physical state and edge weights on the time axis are identified through the neuron nodes inside the model. Based on the changing acceleration and slope, the dynamic nonlinear characteristics of the non-proportional changes in the state distribution of the controlled object that will occur in future periods are extracted. By comparing the distance between the dynamic nonlinear characteristics and the built-in historical evolution samples, the absolute deviation trend of the expected deviation from the target set value is deduced while maintaining the current control action unchanged, and the output is the dynamic runaway index.

6. The method for dynamic adjustment and control of risk parameters in the production process of protective materials according to claim 1, characterized in that, The steps for identifying the control response hysteresis characteristics include: When the value of the dynamic runaway index is detected to be greater than the preset safety control threshold, the lateral width scale is retrieved, and combined with the current running linear speed, initial mechanical pressure, and the material's inherent stiffness mapped by the weaving density benchmark value, the elastic deformation gradient generated at different physical positions in the lateral direction of the wide flexible material under the interlacing tension and compression of multiaxial mechanical forces is identified. Based on the elastic deformation gradient, the absolute time difference between the central region and the two edge regions of the extrusion interface when the controller issues a synchronous execution command is evaluated. The distribution set of time difference values ​​at each horizontal node is used to construct a control response hysteresis characteristic curve that characterizes the physical space transmission delay, thus obtaining the control response hysteresis characteristic.

7. The method for dynamic adjustment and control of risk parameters in the production process of protective materials according to claim 1, characterized in that, The steps for dynamic adjustment and control include: Based on the control response hysteresis characteristics and the control objective of reducing the dynamic runaway index, a multi-dimensional state matrix including mechanical space geometric parameters and medium supply kinetic energy parameters is established. Through a multi-objective constraint solving algorithm, multiple sets of comprehensive candidate control schemes with different numerical ratios are calculated and output simultaneously. Each set of candidate control schemes includes an end pose compensation command for reshaping the lateral spatial geometric gap of the extrusion channel, and a feed servo adjustment command for changing the total amount of fluid medium injected. The comprehensive candidate control scheme set is input into the control scheme evaluation module, and the comprehensive fitness score of each group of candidate control schemes is calculated using the minimization of fluid lateral distribution variance and servo drive energy peak as dual evaluation indicators. The candidate control scheme with the highest score is selected, converted into electrical-level drive commands, and sent to the corresponding physical execution components to complete the dynamic closed-loop regulation under the strong coupling state of multiple physics fields.

Images