Adaptive collaborative control method and system for multi-medium weaving equipment

Abstract

The application relates to the technical field of collaborative regulation, and discloses a multi-medium weaving equipment adaptive collaborative regulation method and system. The application obtains running state data of a weaving equipment and intrinsic attribute vectors of weaving media, and outputs a control parameter combination adaptive to a current medium combination through forward reasoning of a graph neural network. When it is identified that the weaving media are changed, a compensation amount is generated according to the difference between the elastic modulus and the thermal shrinkage rate of the media before and after the change, combined with the sensitivity relationship of the output layer of the graph neural network relative to the input layer, and added to the original control parameter combination before being issued and executed. When no change occurs, the control parameter combination is directly reused and issued and executed. The actual weaving defect rate after the issuance and execution is monitored, and when the actual weaving defect rate exceeds the standard, the control parameter combination is iteratively optimized with the minimum weaving defect rate as the target, the graph neural network parameters are incrementally trained and updated, and the robustness and long-term stability of process regulation under the multi-medium collaborative weaving condition are greatly improved.

Description

Technical Field

[0001] This application relates to the field of collaborative control technology, specifically to an adaptive collaborative control method and system for multi-media weaving equipment. Background Technology

[0002] Modern weaving equipment is characterized by diverse media types, significant differences in physical properties, and high coupling correlation between equipment process units. Currently, the common approach is to use a fixed process parameter table matched with manual experience for control. Some systems introduce simple neural networks to fit process parameters, relying on historical operating data to perform parameter interpolation and linear mapping output. This enables basic weaving process control under single-media, steady-state conditions, reducing the workload of manual parameter tuning to some extent. This has become the mainstream technical approach for multi-media weaving process control at present.

[0003] However, in practical applications, several inherent technical defects have gradually been exposed. First, traditional process parameter control methods rely on manual experience to solidify parameter ranges. When faced with frequent switching of multiple media combinations and small fluctuations in media properties, the process parameters have poor adaptability, which can easily lead to an increase in defects in woven products and a decline in quality consistency. Second, ordinary neural networks mostly adopt a flat data input structure, which only performs simple fitting mapping on equipment operation data and media property data. The model lacks prior support for process mechanisms, has weak generalization ability, and will exhibit parameter inference bias once it deviates from historical training conditions.

[0004] Meanwhile, if the material or specifications of the medium are changed, the original fixed process parameters are still used or only rough manual parameter fine-tuning is made, which cannot offset the process coupling shift caused by the disturbance of the medium's physical properties. Based on the above-mentioned shortcomings of the existing technology, the technical problem to be solved by this application is how to achieve precise coordinated control and adaptive evolution of the corresponding process of weaving equipment under multi-medium and variable working conditions. Summary of the Invention

[0005] To address the shortcomings of existing technologies, the purpose of this application is to provide an adaptive and coordinated control method and system for multi-media weaving equipment, which can adapt to multi-category, variable working conditions and long-cycle multi-media weaving scenarios.

[0006] To achieve the above objectives, this application adopts the following technical solution:

[0007] In the first aspect, this application provides an adaptive collaborative control method for multi-media weaving equipment, which obtains the operating status data of the weaving equipment and the intrinsic attribute vector of the weaving medium, inputs them into a graph neural network, and outputs the control parameter combination under the current medium combination through forward inference;

[0008] When any change in the weaving medium is detected, the corresponding compensation amount is generated based on the difference in elastic modulus and thermal shrinkage rate of the weaving medium before and after the change, according to the sensitivity relationship between the output layer and the input layer of the graph neural network. This compensation amount is then superimposed on the control parameter combination and sent to the actuator.

[0009] When no change in any weaving medium is detected, the control parameter combination is directly sent to the actuator;

[0010] Monitor the actual defect rate after execution. When the actual defect rate is greater than the preset quality threshold, iteratively optimize the combination of control parameters with the goal of minimizing the defect rate, and incrementally train and update the parameters of the graph neural network.

[0011] Furthermore, the control parameter combination under the current media combination includes:

[0012] Based on the physical property correlation constraints of the weaving medium and the process matching constraints of the weaving equipment, a topological correlation diagram adapted to multi-media collaborative weaving is constructed.

[0013] The operating status data of the weaving equipment and the intrinsic attribute vector of the weaving medium are respectively processed by feature normalization and assigned as node features of the corresponding graph nodes.

[0014] The graph neural network is used to aggregate neighborhood features and optimize adaptive weights of node features in the topological association graph, and then fuses them to generate global features.

[0015] The global features are input into the inference layer of the graph neural network for forward inference, and the output is mapped to a combination of control parameters that are adapted to the current medium combination and operating conditions.

[0016] Furthermore, the topological association graph includes the medium feature nodes of the weaving medium and the equipment topological nodes of the corresponding process execution units in the weaving equipment; the neighborhood feature aggregation is obtained by traversing the neighborhood nodes of the topological association graph through multi-layer graph convolutional layers to aggregate the physical property association features between the weaving medium and the process coupling features between the weaving medium and the process execution units in the weaving equipment; the adaptive weight optimization is to adaptively allocate the feature contribution weight of the graph nodes according to the degree of influence of different graph nodes on the quality of the woven product through a graph attention mechanism.

[0017] Further, generating the compensation amount includes:

[0018] When any change in the weaving medium is detected, the difference in elastic modulus and thermal shrinkage rate of the weaving medium before and after the change is converted into physical property disturbance features and embedded into the corresponding medium feature node in the topology graph.

[0019] Based on the interlayer gradient relationship of graph neural networks, the feature sensitivity of the graph neural network output parameters to the feature nodes of the medium under the current working condition is solved.

[0020] By combining the feature contribution weights of graph nodes, the physical property disturbance characteristics and feature sensitivity are coupled and calculated to obtain the single-node compensation component corresponding to the changed weaving medium.

[0021] Aggregate all single-node compensation components to obtain the compensation amount adapted to the current working condition.

[0022] Furthermore, the physical property disturbance feature is a dimension-aligned feature obtained by normalizing the difference in elastic modulus and the difference in thermal shrinkage rate and then concatenating them; the feature sensitivity is a sensitivity coefficient obtained by solving the gradient value of the output parameter of the graph neural network relative to the feature node of the medium through backpropagation; the coupling calculation is to perform an inner product operation with the physical property disturbance feature after weighted fusion of the feature sensitivity and the corresponding feature contribution weight, to obtain a single-node compensation component.

[0023] Furthermore, when no change in any weaving medium is detected, the node structure and feature contribution weights of the current topological association graph are kept fixed, thus solidifying the global features corresponding to the current medium combination.

[0024] The control parameter combination of the current inference output of the graph neural network is reused and directly sent to the actuator.

[0025] Furthermore, the parameters of the incrementally trained and updated graph neural network include:

[0026] After the monitoring and control parameter combination is issued and executed, the actual defect rate of the woven product is triggered. When the actual defect rate is greater than the preset quality threshold, the parameter update of the graph neural network is triggered.

[0027] Based on the node characteristics of the topological association graph, the current combination of control parameters is iteratively optimized with the goal of minimizing the defect rate.

[0028] The training samples are formed by combining the current operating status data, intrinsic attribute vectors and optimized control parameters.

[0029] The graph neural network is incrementally trained based on the training samples. The weights of the graph neural network are iteratively corrected through gradient backpropagation, and the parameters of the graph neural network are updated through incremental training.

[0030] Furthermore, the iterative optimization of the current combination of control parameters includes:

[0031] Construct a loss function adapted to multi-media collaborative weaving based on the feature contribution weights of graph nodes;

[0032] The control parameter combination is corrected round by round using the gradient descent method. After each round of iteration, the loss value of the actual fabric defect rate is calculated. When the loss value of the actual fabric defect rate is less than the loss threshold in multiple consecutive rounds of iteration, the iteration is considered to have converged, and the optimized control parameter combination is output.

[0033] Furthermore, the training samples are input into the current graph neural network to complete forward inference, and the predicted combination of control parameters is output. The inference loss of the training samples is calculated by combining the optimized combination of control parameters.

[0034] The parameters of the multi-layer graph convolutional layer, attention weight layer, and inference layer are updated via backpropagation.

[0035] After each round of incremental iteration, the validity of the parameters is verified, and the parameters that make the inference loss less than the inference threshold are retained. Once the inference loss converges iteratively, the parameters of the current graph neural network are fixed.

[0036] Secondly, this application provides an adaptive collaborative control system for multi-media weaving equipment, comprising a sensing module, an execution module, and a feedback module;

[0037] The perception module is used to acquire the operating status data of the weaving equipment and the intrinsic attribute vector of the weaving medium. After inputting the data into the graph neural network, it outputs the combination of control parameters under the current medium combination through forward inference.

[0038] The execution module is used to generate a corresponding compensation amount based on the difference in elastic modulus and thermal shrinkage rate of the fabric before and after the change, according to the sensitivity relationship between the output layer and the input layer of the graph neural network, when any change in the fabric is detected. This compensation amount is then superimposed on the control parameter combination and sent to the execution mechanism.

[0039] When no change in any weaving medium is detected, the control parameter combination is directly sent to the actuator;

[0040] The feedback module is used to monitor the actual defect rate after the execution is issued. When the actual defect rate is greater than the preset quality threshold, the combination of control parameters is iteratively optimized with the goal of minimizing the defect rate, and the parameters of the graph neural network are incrementally trained and updated.

[0041] Compared with the prior art, the beneficial effects achieved by this application are as follows:

[0042] This application obtains the operating status data of the weaving equipment and the intrinsic attribute vector of the weaving medium, and outputs a combination of control parameters adapted to the current medium combination through graph neural network forward inference. It structurally describes the physical property correlation constraints between weaving media and the process matching constraints between the weaving media and the equipment process unit, thereby improving the accuracy and adaptability of process parameter inference under complex multi-media conditions.

[0043] When a change in the weaving medium is detected, this application generates a compensation amount based on the difference in elastic modulus and thermal shrinkage rate of the medium before and after the change, combined with the sensitivity relationship between the output layer and the input layer of the graph neural network. This compensation amount is then superimposed on the original control parameter combination and executed. This accurately quantifies the process disturbance caused by the change in medium properties, achieving directional and quantitative adaptive compensation for the disturbance of medium properties. It effectively suppresses the generation of weaving defects caused by medium replacement and ensures a smooth transition of the weaving process and stable finished product quality during the switching of multiple media.

[0044] This application directly reuses the control parameter combination output by graph neural network inference and issues it for execution under steady-state conditions where no change in weaving medium is detected, thereby reducing the computing power consumption of the weaving equipment and maintaining the consistency of process parameters and finished product quality during continuous weaving.

[0045] This application monitors the actual defect rate after the combination of monitoring and control parameters is issued and executed. Using a preset quality threshold as the judgment benchmark, the combination of control parameters is iteratively optimized with the minimum defect rate as the target after the actual defect rate exceeds the standard. The graph neural network parameters are updated incrementally to form a closed-loop control mechanism of defect quality feedback, process parameter optimization and network weight self-updating. This adaptively offsets the adverse effects of equipment operating condition drift, natural decay of media properties and fluctuations in the production environment, and realizes the full life cycle adaptive evolution of graph neural network and weaving process, which greatly improves the robustness and long-term stability of process control under multi-media collaborative weaving conditions.

[0046] Overall, this application takes into account the multiple technical effects of multi-media topology coupling characteristic representation, dynamic compensation for media changes, steady-state efficient reuse, closed-loop optimization of weaving defects and incremental network updates, and is suitable for multi-media weaving production scenarios with multiple product categories, changing working conditions and long cycles. Attached Figure Description

[0047] Figure 1 A flowchart illustrating the adaptive and coordinated control method for multi-media weaving equipment;

[0048] Figure 2 The logic flowchart for outputting the control parameter combination under the current media combination;

[0049] Figure 3 This is a schematic diagram of an adaptive and coordinated control system for multi-media weaving equipment. Detailed Implementation

[0050] The technical solution of this application will be described in detail below with reference to the accompanying drawings and specific embodiments. It should be understood that the embodiments and specific features in the embodiments are detailed descriptions of the technical solution of this application, rather than limitations thereof.

[0051] Example 1:

[0052] like Figure 1 As shown, this embodiment provides an adaptive collaborative control method for multi-media weaving equipment, including:

[0053] S1. Obtain the operating status data of the weaving equipment and the intrinsic attribute vector of the weaving medium, input them into the graph neural network, and output the control parameter combination under the current medium combination through forward inference.

[0054] Specifically, such as Figure 2 As shown, the control parameter combinations for the current media combination include:

[0055] Based on the physical property correlation constraints of the weaving medium and the process matching constraints of the weaving equipment, a topological correlation diagram adapted to multi-media collaborative weaving is constructed.

[0056] The operating status data of the weaving equipment and the intrinsic attribute vector of the weaving medium are respectively processed by feature normalization and assigned as node features of the corresponding graph nodes.

[0057] The graph neural network is used to aggregate neighborhood features and optimize adaptive weights of node features in the topological association graph, and then fuses them to generate global features.

[0058] The global features are input into the inference layer of the graph neural network for forward inference, and the output is mapped to a combination of control parameters that are adapted to the current medium combination and operating conditions.

[0059] The topology graph includes the medium feature nodes of the weaving medium and the equipment topology nodes of the corresponding process execution units in the weaving equipment; the neighborhood feature aggregation is obtained by traversing the neighborhood nodes of the topology graph through multi-layer graph convolutional layers to obtain the physical property association features between the weaving medium and the process coupling features between the weaving medium and the process execution units in the weaving equipment; the adaptive weight optimization is to adaptively allocate the feature contribution weight of the graph nodes according to the degree of influence of different graph nodes on the quality of the woven product through the graph attention mechanism.

[0060] In multi-media weaving scenarios, the elastic deformation and thermal shrinkage characteristics of different weaving media are mutually restrictive. The adjustment range and working parameters of each process execution unit of the equipment must match the fluctuation range of the media's physical properties to form rigid process constraints. Traditional data fitting methods cannot characterize this kind of implicit coupling relationship. The physical property relationship constraint is the inherent physical rule that different weaving media affect each other during the process of deformation under force and thermal shrinkage. The process matching constraint is the weaving process limitation rule that the adjustment parameters of the equipment's process execution unit must be adapted to the range of media physical properties.

[0061] Specifically, three types of commonly used industrial weaving media are selected to form a multi-media combination: pure cotton fiber media, polyester filament media, and viscose blended media. Simultaneously, four types of process execution units of weaving equipment are selected: warp feed adjustment unit, warp tension control unit, drying temperature setting unit, and weft insertion speed adjustment unit. Each of the three types of weaving media is defined as a media feature node, for a total of three media feature nodes. Each of the four types of process execution units is defined as a device topology node, for a total of four device topology nodes, forming a total of seven basic topology nodes.

[0062] Based on the process mechanism, directional correlation edges are arranged. The characteristic nodes of pure cotton fiber medium are simultaneously associated with the drying temperature unit node and the tension control unit node. The characteristic nodes of polyester filament medium are associated with the weft insertion speed unit node and the warp feed adjustment unit node. The characteristic nodes of viscose blended medium are fully associated with all four types of equipment topology nodes, thus forming a topology correlation diagram for multi-medium collaborative weaving.

[0063] The specific modeling rules for the physical property correlation constraints are as follows: when the difference in elastic modulus between two weaving media is less than 15 MPa or the difference in thermal shrinkage rate is less than 1%, it is determined that there is a strong physical property correlation between the two, and a bidirectional undirected correlation edge is established in the topology correlation graph; the modeling rules for the process matching constraints are as follows: when the control range of the equipment process execution unit covers the physical property parameter requirement range of the corresponding weaving media, a directed correlation edge is established from the media feature node to the equipment topology node; the above modeling rules ensure that each correlation edge in the topology correlation graph has a clear physical process basis, rather than being an arbitrary connection.

[0064] Specifically, the physical property correlation determination was performed on each pair of the three types of weaving media. The difference in elastic modulus between pure cotton fiber and polyester filament is about 22MPa (>15MPa), and the difference in heat shrinkage rate is about 1.2% (>1%). Neither of these two conditions meets the strong correlation condition, so no physical property correlation edge is established. The difference in elastic modulus between pure cotton fiber and viscose blend is about 8MPa (<15MPa), which meets the strong correlation condition, so a bidirectional undirected correlation edge is established. The difference in elastic modulus between polyester filament and viscose blend is about 14MPa (<15MPa), which meets the strong correlation condition, so a bidirectional undirected correlation edge is established. The final topology correlation graph includes 7 nodes, 7 equipment process correlation edges, and 2 media physical property correlation edges.

[0065] This provides a clear physical scope and topological logic for subsequent feature embedding and graph convolution traversal, and also provides a unique carrier for the corresponding node of subsequent feature data division, supporting the accurate implementation of subsequent feature normalization and node assignment steps.

[0066] The dimensions of equipment operating parameters and the physical properties of the medium are different, and the numerical ranges vary greatly. Directly inputting the raw data into the network will result in the submergence of effective features with low amplitude and an imbalance in feature weight allocation. Normalization is used to eliminate the difference in dimensions and numerical magnitudes. Node feature assignment is to match and bind the standardized feature vectors to the corresponding graph nodes. Specifically, the operating status data of the weaving equipment is fixed as a four-dimensional parameter dimension, namely weaving speed, warp basic tension, drying set temperature, and warp feed rate. The intrinsic property vector of the weaving medium is set as a three-dimensional parameter dimension, namely elastic modulus, room temperature heat shrinkage rate, and breaking elongation.

[0067] An extreme value normalization algorithm is used to uniformly map all original features to the interval between 0 and 1. For example, the original working range of the equipment drying temperature is 90℃ to 150℃, and the actual measured value of 126℃ is taken. After extreme value normalization, the standardized feature is 0.6. The original range of the weaving speed is 10m / min to 30m / min, and the measured value of 22m / min is taken. After normalization, the standardized feature is 0.6. The three-dimensional intrinsic attribute vectors of the three types of weaving media are simultaneously aligned and normalized. All equipment working condition features and media physical property features are uniformly mapped to 16-dimensional fixed-length node feature vectors to ensure that the feature dimensions of all topological nodes are completely consistent.

[0068] The dimension mapping adopts a fully connected linear transformation method. The normalized features of the four-dimensional equipment operating conditions are input into a 4×16-dimensional fully connected linear layer, and after transformation by the weight matrix, 16-dimensional features are output. The normalized features of the three-dimensional medium properties are input into a 3×16-dimensional fully connected linear layer, and after transformation by the weight matrix, 16-dimensional features are output. The weight matrices of the two sets of linear layers are learned and determined synchronously during the initial training stage of the graph neural network to ensure the effective preservation of feature information and semantic consistency during the dimension expansion process.

[0069] The semantic structure of the 16-dimensional node feature vector is as follows: the first 4 dimensions retain the principal component features of the original physical parameters, corresponding to the main influencing factors such as elastic modulus / vehicle speed; the 5th to 8th dimensions encode the cross-correlation features between parameters, such as the interaction effect between elastic modulus and thermal shrinkage rate; the 9th to 12th dimensions encode the time-series operating condition features, such as the current operating stage and parameter change trend; and the 13th to 16th dimensions encode the topological location features, such as the correlation degree and centrality index of nodes in the topological graph. The semantic division is determined by the self-learning of the weight matrix of the fully connected layer during the initial training stage.

[0070] The 16-dimensional feature vector generated by mapping the normalized four-dimensional equipment operating condition features is assigned to the four equipment topology nodes one by one. The 16-dimensional feature vector generated by mapping the normalized three-dimensional physical property vectors of the three types of media is assigned to the three media feature nodes one by one, thus completing the filling and initialization of all initial features of the seven topology nodes. This provides a fixed input dimension for the subsequent layer-by-layer feature aggregation of the graph convolutional layer.

[0071] A single node only carries its own characteristics and cannot explore the physical property interaction and process coupling relationship between neighboring nodes. Different nodes have different degrees of influence on weaving quality. Neighboring feature aggregation is achieved by traversing neighboring nodes through multiple graph convolutional layers, aggregating the physical property correlation features between media and the process coupling features between media and equipment execution units. Adaptive weight optimization allocates feature contribution weights according to the degree of influence of graph nodes on the quality of woven products.

[0072] Specifically, a fixed configuration of three cascaded graph convolutional layers is used. The first graph convolutional layer has an input dimension of 16 and an output dimension of 32, which completes the aggregation of first-order neighborhood node features. The second graph convolutional layer maintains 32-dimensional input and output, and deeply mines the physical property correlation features between feature nodes of different media. The third graph convolutional layer has an input dimension of 32 and an output dimension of 64, which extracts the process coupling features between media feature nodes and equipment topology nodes.

[0073] The graph attention mechanism employs a 4-head attention structure, calculating feature contribution weights for each graph node. The weights for core medium feature nodes and core equipment topology nodes, which significantly impact weaving defects, are set between 0.8 and 0.95, while the weights for secondary related nodes range from 0.2 to 0.4. Higher feature contribution weights are assigned to medium feature nodes and core equipment topology nodes that significantly impact weaving quality, while the weight ratio of redundant nodes with low correlation is weakened. This completes the differential weighting correction for all node features. The 64-dimensional node features output from the three-layer graph convolution are then weighted by attention, and all node features are spliced ​​and fused to generate a 128-dimensional fixed-length global feature vector, fully encompassing the implicit information of multi-medium property relationships, equipment operating conditions, and process coupling.

[0074] The forward propagation of the convolutional layer is H. (l+1) =σ(D -1 / 2 ×Ã×D -1 / 2 ×H (l) ×W (l) ), H (l) Let W be the feature matrix of the nodes in the l-th layer. (l) Let A be the trainable weight matrix of the l-th layer, ∠A = A + I be the adjacency matrix with added self-connections, I be the identity matrix with the same dimension as the adjacency matrix A, used to add self-connection edges to each node, D be the degree matrix of ∠A, and σ(·) be the ReLU activation function; α is the attention coefficient α of node i to node j in the graph attention mechanism. ij =softmax(LeakyReLU(a T [W·h i ‖W·h j The feature of node j represents the degree of contribution of the feature aggregation of node i. The sum of the attention coefficients of all neighboring nodes after softmax normalization is 1. iand h j Let represent the features of nodes i and j, W be the trainable weight matrix, a be the attention parameter vector, and || be the vector concatenation operation. The negative slope of LeakyReLU is set to 0.2. The outputs of the 4-head attention are fused by concatenation (first 3 heads) and averaging (4th head).

[0075] This forms a solidifiable and reproducible feature aggregation process, which filters and strengthens the contribution of core node features, suppresses interference from redundant nodes, and provides global features with dimensionally regular and information-complete standard inputs for forward inference, determining the matching accuracy of subsequent control parameter outputs.

[0076] The global features obtained through fusion are high-dimensional abstract topological features, which cannot directly correspond to the process control parameters that can be executed by the weaving equipment. Forward inference is defined as locking all trained parameters of the graph neural network, not performing backpropagation parameter updates, and only completing the forward layer-by-layer propagation of features, weighted transformation and parameter fitting operations. The inference layer is a multi-layer fully connected regression layer at the end, used to realize the nonlinear mapping of high-dimensional global features to weaving process control parameters. Specifically, the inference layer is fixedly configured with a three-layer fully connected network structure. The first fully connected layer has an input dimension of 128 dimensions and 128 neurons, the second fully connected layer has 64 neurons, the third fully connected layer has 4 neurons, and finally outputs a four-dimensional combination of control parameters.

[0077] The forward inference process is as follows: the 128-dimensional global feature vector generated in the previous step is input into the first layer of the inference layer. After nonlinear feature transformation and dimension preservation by 128 neurons, it is passed to the second layer of 64 neurons for feature dimensionality reduction and key feature recombination. Finally, the fourth neuron in the third layer directly outputs four types of core control parameters: weaving speed, warp tension setting, drying reference temperature, and warp feed rate. The parameter output is limited to the industrial standard adjustable range. The output range of weaving speed is fixed at 15m / min to 25m / min, the output range of warp tension setting is fixed at 800N to 1200N, the output range of drying reference temperature is fixed at 110℃ to 140℃, and the output range of warp feed rate is adjusted to match the fixed range of the medium weight setting.

[0078] This allows it to adapt to the coupling constraints of current multi-media collaborative weaving, enabling adaptive and precise output of process parameters under complex working conditions. The combination of control parameters serves as the benchmark process parameters for steady-state weaving, media change disturbance compensation, weaving defect quality monitoring feedback, and incremental iterative training of the network, providing core parameter support for the closed-loop operation of the entire control method.

[0079] S2. When any change in the weaving medium is detected, the corresponding compensation amount is generated based on the difference in elastic modulus and thermal shrinkage rate of the weaving medium before and after the change, according to the sensitivity relationship between the output layer and the input layer of the graph neural network. This compensation amount is then superimposed on the control parameter combination and sent to the actuator.

[0080] If no change in any weaving medium is detected, the control parameter combination will be directly sent to the actuator;

[0081] Specifically, the generated compensation amount includes:

[0082] When any change in the weaving medium is detected, the difference in elastic modulus and thermal shrinkage rate of the weaving medium before and after the change is converted into physical property disturbance features and embedded into the corresponding medium feature node in the topology graph.

[0083] Based on the interlayer gradient relationship of graph neural networks, the feature sensitivity of the graph neural network output parameters to the feature nodes of the medium under the current working condition is solved.

[0084] By combining the feature contribution weights of graph nodes, the physical property disturbance characteristics and feature sensitivity are coupled and calculated to obtain the single-node compensation component corresponding to the changed weaving medium.

[0085] All single-node compensation components are aggregated to obtain a compensation amount adapted to the current working condition, which is then superimposed on the control parameter combination and sent to the actuator.

[0086] Among them, the physical property disturbance feature is the dimension-aligned feature obtained by normalizing the difference in elastic modulus and the difference in thermal shrinkage rate; the feature sensitivity is the sensitivity coefficient obtained by solving the gradient value of the output parameter of the graph neural network relative to the feature node of the medium through backpropagation; the coupling calculation is to perform an inner product operation with the physical property disturbance feature after weighted fusion of the feature sensitivity and the corresponding feature contribution weight, to obtain the single-node compensation component.

[0087] Meanwhile, when no change in any weaving medium is detected, the node structure and feature contribution weight of the current topology graph remain unchanged, thus solidifying the global features corresponding to the current medium combination.

[0088] The control parameter combination of the current inference output of the graph neural network is reused and directly sent to the actuator.

[0089] In the continuous operation of multi-media collaborative weaving, various weaving media rely on their own elastic modulus, thermal shrinkage rate and other physical properties to participate in the feature coupling calculation within the topological association graph. When any type of weaving media changes its material specifications or physical properties, only the two core physical properties of the medium itself are deflected, and it cannot be directly connected to the node feature calculation space of the graph neural network. If the original difference is directly used in the feature iteration calculation, the node feature will be disturbed and distorted due to the dimensional barrier and the difference in numerical amplitude, and it will be impossible to accurately map the process coupling disturbance state caused by the change of medium.

[0090] The essence of physical property perturbation features is to normalize the difference in elastic modulus and the difference in thermal shrinkage rate, and then form a dimension-aligned standardized feature by concatenating feature vectors. The normalization process maps the original difference data with physical dimensions to a unified dimensionless standard range of 0~1, eliminating the computational interference caused by the difference in dimensions and numerical span of different physical property parameters. The feature vector concatenation is used to integrate the two-dimensional physical property change information and construct feature vectors with regular dimensions and adapted to the embedding rules of graph nodes.

[0091] Specifically, an actual industrial weaving scenario was selected as an application example. The original steady-state weaving used conventional polyester fiber as the base medium. After the working conditions were changed, it was replaced with high-elasticity modified polyester fiber. According to the physical property parameters specified by the factory of the two types of media, the original difference in elastic modulus was determined to be 35MPa, and the original difference in heat shrinkage rate was 1.8%. The normalized effective range of elastic modulus difference was defined as 0 to 50MPa, and the normalized effective range of heat shrinkage rate difference was defined as 0 to 3%. The normalized value was calculated as (actual difference - minimum value of the range) ÷ (maximum value of the range - minimum value of the range). The normalized value of elastic modulus was calculated as (35-0) ÷ (50-0) = 0.7, and the normalized value of heat shrinkage rate was calculated as (1.8-0) ÷ (3-0) = 0.6. The two sets of one-dimensional standardized scalar features were spliced ​​together to generate the physical property disturbance features of the fixed structure [0.7, 0.6].

[0092] The effective range is determined based on the range of physical property parameters of selectable media in the workshop weaving media category database. The effective range of elastic modulus is the difference between the minimum and maximum elastic modulus of all media in the database. In this embodiment, the elastic modulus range of all selectable media is 80MPa to 130MPa, with a range of 50MPa, so the effective range is set to [0, 50MPa]. The effective range of heat shrinkage rate is determined similarly. In this embodiment, the heat shrinkage rate range of all selectable media is 0.5% to 3.5%, with a range of 3%, so the effective range is set to [0, 3%.

[0093] When embedding the topology association graph, the overall topology architecture that has been built before, including medium feature nodes and device topology nodes, is used to accurately locate the medium feature node corresponding to the current change in medium physical properties. Only the constructed physical property disturbance feature is embedded into the additional feature dimension position of the node. The original initial feature parameters and the topology connection relationship between the nodes of the other medium feature nodes that have not changed physical property parameters are kept completely constant. Only local fixed-point feature disturbance injection operation is performed, without destroying the inherent coupling logic and original feature distribution state of the overall topology association structure.

[0094] This enables the precise regularization and transformation of engineering physical parameters into computationally standardized feature vectors for graph networks. It provides dimensionally regular, positionally unique, and numerically standardized node perturbation input conditions for solving the feature sensitivity of the inter-layer gradient relationship in the graph neural network, limits the target node range and feature input format for gradient solving, and ensures that the subsequent sensitivity calculation has a computational benchmark.

[0095] Multi-media weaving processes exhibit typical nonlinear coupling characteristics. The forward inference mapping logic of graph neural networks relies on the construction of inter-layer weight associations in multi-layer networks. Under different weaving conditions and different node characteristic states, the response amplitude of the process control parameters output by the network end to changes in the characteristic nodes of each medium is not a fixed constant. If the traditional control mode of manually preset fixed compensation coefficients is adopted, it cannot adapt to the dynamic nonlinear change law of multi-media coupling systems. The feature sensitivity is obtained by solving the gradient value of the graph neural network output parameters relative to the characteristic nodes of the medium through the backpropagation algorithm, and then normalizing it to obtain the sensitivity coefficient.

[0096] During the operation, all network weight parameters of the graph neural network that have been trained and converged need to be locked throughout the process. Only the local backpropagation gradient calculation logic is enabled. The medium feature node with embedded physical perturbation features is used as the independent variable for gradient calculation, and the combination of control parameters output by the graph neural network inference layer is used as the dependent variable. The partial derivative gradient values ​​of the output parameter vector with respect to each dimension component of the medium feature node are calculated one by one. Then, the gradient values ​​are normalized and uniformly mapped to the interval between 0 and 1 to obtain a single sensitivity coefficient that represents the strength of the response.

[0097] Let the input feature of the medium feature node be x∈R 16 The combination of control parameters output by the inference layer of the graph neural network is y∈R. 4 Then the Jacobian matrix J∈R of the output parameters with respect to the input features (4×16) for: m∈{1,2,3,4} corresponds to the four control parameter dimensions, and n∈{1,2,…,16} corresponds to the node feature dimensions. Partial derivatives are calculated dimension-by-dimensionally through backpropagation. The L2 norm of each row of the Jacobian matrix is ​​then calculated, and the average value is taken to obtain the feature sensitivity S = (1 / 4)×Σ. m ||J[m,:]||2, ultimately normalizing S to the interval [0,1], i.e., S normalized =(SS min ) / (S max -S min ), S min and S max These are the historical sensitivity extremes statistically analyzed during the network training phase.

[0098] For example, after gradient solving and normalization, the sensitivity coefficient of the current medium feature node is fixed at 0.76. This coefficient can objectively characterize the adaptive response amplitude of the process control parameters when the medium feature node generates a unit disturbance change, and is generated autonomously by the inherent correlation between network layers.

[0099] This adaptively matches the parameter response differences under different media combinations and weaving conditions, making the compensation control logic conform to the real mapping law of the network; the sensitivity coefficient serves as the dynamic correction weight for subsequent coupled calculations, establishing a quantitative correlation between the physical disturbance of node characteristics and the response of process parameters, providing a core correction basis for the accurate quantitative calculation of single-node compensation components.

[0100] Relying solely on the feature sensitivity obtained through network calculation can only characterize the dynamic response amplitude of network parameters and cannot take into account the inherent influence weight of the medium feature nodes on the quality of the woven product. The use of a single factor in the calculation is prone to process deviations such as over-compensation or under-compensation. Coupled calculation involves weighted fusion of feature sensitivity and the feature contribution weight of the corresponding medium feature nodes to generate a composite correction coefficient. Then, the composite correction coefficient is used to perform a vector inner product operation with the constructed physical property disturbance feature to obtain the single-node compensation component.

[0101] Specifically, the feature contribution weight of the current changed medium feature node in the topological association graph, which has been trained and solidified, is taken as 0.82. Combined with the already solved sensitivity coefficient of 0.76, the two types of weight factors are fused using an arithmetic average weighting method, that is, the composite correction coefficient = (0.76 + 0.82) ÷ 2 = 0.79. Then, according to the vector inner product operation rules, the composite correction coefficient is multiplied by the values ​​of each dimension of the physical property disturbance feature in turn and summed. Substituting this sum into the feature vector [0.7, 0.6] completes the full calculation. The single-node compensation component corresponding to this medium change is calculated as 0.79 × 0.7 + 0.79 × 0.6 = 0.553 + 0.474 = 1.027.

[0102] This enables a deep integration of the weaving process quality mechanism and the characteristics of the graph neural network model, allowing the single-node compensation component to accurately match the triple constraints of the medium property disturbance amplitude, the node quality influence level, and the network dynamic response characteristics. The single-node compensation component provides a quantized basic unit with unified dimensions and consistent units for the global aggregation operation of multiple components under the condition of synchronous change of multiple media, ensuring the operational regularity and logical consistency of the superposition and aggregation process of multiple node compensation components.

[0103] In actual multi-media weaving production scenarios, there are simple working conditions where a single weaving medium changes independently, as well as complex working conditions where two or more types of weaving media undergo simultaneous changes in their physical property parameters. Each media feature node that has changed generates a single-node compensation component. There is a natural process coupling relationship between these components, making it impossible to correct the process parameters individually. Specifically, we follow the aggregation operation principle of linear aggregation as the main method and weight correction as a supplement. We traverse all media feature nodes with changed physical property parameters in the topology graph and obtain the single-node compensation component generated by each node one by one. Under normal working conditions, we use a linear accumulation method to complete global aggregation. If there are complex working conditions where multiple nodes have opposite disturbance trends, we perform weighted aggregation based on the feature contribution weight of each node to weaken the component proportion of nodes with low quality impact.

[0104] For example, in addition to the single-node compensation component of 1.027 already calculated, another single-node compensation component generated by the synchronously changing medium feature node is 0.583. After linear accumulation and aggregation, the compensation amount is 1.027 + 0.583 = 1.61. The aggregated compensation amount is then superimposed on the baseline control parameter combination output by the graph neural network forward inference according to the process parameter dimension matching rules, thereby completing the adaptive correction of process parameters under multi-media disturbance conditions. The corrected control parameter combination is then sent to the process execution mechanism of the weaving equipment for implementation.

[0105] The process parameter dimension matching rule is to proportionally allocate the aggregated scalar compensation amount according to the feature contribution weights of the nodes corresponding to the control parameters, generating a compensation amount consistent with the dimension of the control parameter combination, i.e., Δp. b =C total ×(w b / Σw k ), C total w represents the total compensation amount after aggregation. b The feature contribution weight Σw is assigned to the device topology node corresponding to the b-th control parameter. k The sum of the weights contributing to the topology node features of all devices, Δp b This is the compensation component allocated to the b-th control parameter; during compensation superposition, it is scaled and adapted according to the dimensions of the control parameter and the adjustment range, i.e., Δp. b-scaled =Δp b ×R b , Δp b-scaled The final compensation component of the b-th control parameter after range scaling is directly superimposed on the reference value of that control parameter, i.e., the corrected control parameter value = reference control parameter value + Δp. b-scaled ;R bThe adjustable range width of the b-th control parameter is as follows: weaving speed R1 = 25-15 = 10 m / min, warp tension R2 = 1200-800 = 400 N, drying temperature R3 = 140-110 = 30 ℃, and warp feed amount R4 is determined according to the weight of the medium, to ensure that the dimensions of the compensation component are consistent with the corresponding control parameters.

[0106] For example, compensation amount C total =1.61, the feature contribution weights of the four equipment topology nodes are 0.85 (warp feed adjustment unit), 0.9 (tension control unit), 0.88 (drying temperature unit), and 0.82 (weft insertion speed unit), respectively, with a total weight of 3.45. Therefore, the parameter compensation components are Δp1=1.61×(0.85 / 3.45)=0.396, Δp2=1.61×(0.9 / 3.45)=0.418, Δp3=1.61×(0.88 / 3.45)=0.411, Δp4=1.61×(0.82 / 3.45)=0.383; after range scaling, the weaving speed compensation component Δp 1-scaled =0.396×10=3.96m / min, warp tension compensation component Δp 2-scaled =0.418×400=167.2N, drying temperature compensation component Δp 3-scaled =0.411×30=12.3℃, and the feed rate compensation component is determined according to the range ratio.

[0107] This allows for comprehensive adaptation to all types of weaving conditions involving single-medium and multi-medium synchronous changes. By standardizing component aggregation, it achieves integrated and balanced correction of multi-source physical property disturbances, maintaining the stability of the entire process of multi-medium coupled weaving. It also provides real and effective working condition sample data support for subsequent real-time monitoring of weaving defect rates and incremental iterative training of graph neural networks.

[0108] In the steady-state weaving stage, where the weaving medium combination remains fixed and the equipment operating conditions do not fluctuate drastically, the node partitioning form of the topology graph, the topological connection relationship between nodes, and the feature contribution weights assigned to each graph node after training and convergence through the graph attention mechanism are all fully adapted to the inherent law of process coupling of the current fixed medium combination. There is no need to repeatedly reconstruct the topology structure, re-perform neighborhood feature aggregation and adaptive weight optimization calculations. At the same time, repeatedly executing the complete forward inference process of the graph neural network will generate a lot of redundant computing power consumption, and frequent refresh of process parameters can easily cause continuous fluctuations in the quality of the woven products. This is the core reason for adopting the structure locking, feature solidification, and parameter reuse control logic in steady-state operation.

[0109] The system continuously monitors the physical properties and status of various weaving media in real time. Once it is determined that no material, specification, or physical property parameter changes have occurred in any weaving media, the overall architecture of the completed topology graph is immediately locked. The number, position, and topological relationship of the media feature nodes and equipment topology nodes are fixed. At the same time, the feature contribution weights of each graph node after training and convergence are frozen, and no weight iteration or parameter update operations are performed. Based on the constant topology structure and node weights, the global features generated by matching the current media combination with the equipment operating conditions are solidified. The operation process of feature normalization, neighborhood feature aggregation, and adaptive weight optimization is no longer repeated. The system directly reads the control parameter combination output by the real-time forward inference of the stored graph neural network. No physical property disturbance features are introduced throughout the process, and no compensation calculation or parameter superposition correction operations are performed. The cached control parameter combination is sent to the weaving equipment actuator as is.

[0110] Taking continuous weaving with a fixed cotton-polyester blend as a practical example, the system determines that the blend remains unchanged for a long time, maintains a constant topology and node weights throughout the process, solidifies steady-state global features, and continuously reuses inference output parameters to ensure the continuous and stable operation of the weaving process.

[0111] This significantly reduces the frequency of real-time iterative computations in the graph neural network, decreases the computational load on weaving equipment, and avoids fluctuations in finished product quality caused by frequent parameter fine-tuning under steady-state conditions, maintaining the continuity and consistency of weaving process parameters. Together with the aforementioned dynamic compensation conditions for changes in media, it forms a complete coverage of the full-scenario control system, providing a stable and reliable foundation of process operation data for the routine monitoring of weaving defect rates throughout the day, while ensuring the integrity of the operating conditions and the validity of the data in the subsequent incremental training sample set of the graph neural network.

[0112] S3. Monitor the actual defect rate after execution. When the actual defect rate is greater than the preset quality threshold, iteratively optimize the combination of control parameters with the goal of minimizing the defect rate, and incrementally train and update the parameters of the graph neural network.

[0113] Furthermore, the parameters of the graph neural network are updated incrementally through training, including:

[0114] After the monitoring and control parameter combination is issued and executed, the actual defect rate of the woven product is triggered. When the actual defect rate is greater than the preset quality threshold, the parameter update of the graph neural network is triggered.

[0115] Based on the node characteristics of the topological association graph, the current combination of control parameters is iteratively optimized with the goal of minimizing the defect rate.

[0116] The training samples are formed by combining the current operating status data, intrinsic attribute vectors and optimized control parameters.

[0117] The graph neural network is incrementally trained based on the training samples. The weights of the graph neural network are iteratively corrected through gradient backpropagation, and the parameters of the graph neural network are updated through incremental training.

[0118] Specifically, iterative optimization of the current combination of control parameters includes:

[0119] Construct a loss function adapted to multi-media collaborative weaving based on the feature contribution weights of graph nodes;

[0120] The control parameter combination is corrected round by round using the gradient descent method. After each round of iteration, the loss value of the actual fabric defect rate is calculated. When the loss value of the actual fabric defect rate is less than the loss threshold in multiple consecutive rounds of iteration, the iteration is considered to have converged, and the optimized control parameter combination is output.

[0121] The training samples are input into the current graph neural network to complete forward inference, and the predicted combination of control parameters is output. The inference loss of the training samples is calculated by combining the optimized combination of control parameters.

[0122] The parameters of the multi-layer graph convolutional layer, attention weight layer, and inference layer are updated via backpropagation.

[0123] After each round of incremental iteration, the validity of the parameters is verified, and the parameters that make the inference loss less than the inference threshold are retained. Once the inference loss converges iteratively, the parameters of the current graph neural network are fixed.

[0124] During the long-term continuous operation of multi-media weaving, fluctuations in equipment operating conditions, slight drifts in media properties, and shifts in environmental temperature and humidity can cause the combination of control parameters output by the fixed graph neural network to gradually deviate from the optimal process state, resulting in persistent weaving defects. The control method based solely on compensation for media changes and reuse of steady-state parameters cannot achieve self-evolutionary adaptation between the process and the network. The actual weaving defect rate is the proportion of defective samples in the current batch of woven products to the total samples. The preset quality threshold is the qualified weaving defect rate critical value pre-calibrated by the weaving process, used to determine whether the current network inference parameters are adapted to the real-time operating conditions.

[0125] Specifically, the online visual inspection unit of the equipment acquires images of the woven product surface frame by frame, and counts the number of defect samples in the current continuous woven batch and the total number of woven samples in real time to calculate the real-time actual woven defect rate. For example, if the preset quality threshold is set to 2‰, and the detected current actual woven defect rate is 2.9‰, which is greater than the preset quality threshold, it is determined that there is a working condition adaptation deviation in the combination of control parameters output by the current graph neural network inference, and the subsequent parameter iteration optimization and model incremental update process is formally triggered.

[0126] This allows for the accurate identification of failure points in the adaptation of operating conditions, ensuring the targeted and effective control and optimization, and providing precise triggering opportunities and operating condition data for subsequent iterative optimization of process parameters, construction of training samples, and updates of network weights.

[0127] After triggering the update process, the conventional fixed parameter iteration method does not distinguish the differences in the impact of different medium feature nodes and equipment topology nodes on weaving defects. Iterating all process parameters equally is prone to problems such as slow convergence speed, local optima, and poor optimization targeting. The loss function is a node-weighted weaving defect loss function adapted to multi-medium collaborative weaving. The loss constraint strength is configured differently according to the feature contribution weight of each graph node to the quality of the woven product. The gradient descent method is an iterative optimization algorithm that corrects the process control parameters round by round. The loss threshold is the critical standard for determining the convergence of process parameter iteration and the weaving defect loss.

[0128] Specifically, the fixed feature contribution weights of all media feature nodes and device topology nodes in the topology graph are retrieved. Higher loss penalty coefficients are assigned to core nodes with high defect sensitivity, while the penalty weight coefficients are weakened for low-impact redundant nodes, thus constructing the loss function for multi-media collaborative weaving. The specific mathematical expression of the loss function L is as follows: R defect p represents the actual fabric defect rate. b Let b be the current iterative value of the b-th control parameter. w is the previous iteration value of the b-th control parameter. b The feature contribution weight of the node corresponding to the b-th control parameter is λ, which is the regularization coefficient. In this embodiment, λ is set to 0.01. The first term of the loss function guides the minimization of the weaving defect rate, and the second term constrains the parameter change range through node weighted regularization to avoid iterative oscillation. The iteration step size is set as follows: weaving speed step size 0.5m / min, warp tension step size 20N, drying temperature step size 2℃, and warp feed step size is determined proportionally according to the medium weight.

[0129] During the iterative calculation, the minimum defect rate is used as the convergence target. The gradient descent algorithm is used to solve the gradient value of the defect loss with respect to the current combination of control parameters in reverse iterations. The four core process parameters of weaving speed, warp tension, drying temperature and warp feed are fine-tuned in each iteration according to the above iteration step size. After each round of parameter fine-tuning, weaving simulation calculation is performed to calculate the actual defect rate loss value corresponding to the current parameters. For example, if the iteration loss threshold is set to 0.3, when the system monitors the loss values ​​corresponding to five consecutive iterations, which are 0.28, 0.27, 0.26, 0.26 and 0.25 respectively, and the loss values ​​of five consecutive iterations are all less than the preset loss threshold of 0.3, it is determined that the current process parameter iteration has converged, the iteration process is terminated and the optimized combination of control parameters is output.

[0130] This allows for precise alignment with the quality impact patterns of multi-media coupled weaving, effectively avoiding optimization deviations and local optima issues inherent in traditional uniform iteration, and rapidly converging to the process parameter state corresponding to the global minimum weaving defect rate. The combination of control parameters output during iteration serves as the standard label data for subsequent incremental training samples, determining the optimization baseline accuracy of the graph neural network's incremental learning.

[0131] After the process parameters have converged iteratively, conventional model updates often use random samples or samples from poor working conditions for training, which can easily lead to a degradation in the model's generalization ability. Instead, we select the original equipment operation data and intrinsic material property data of the medium under the current real poor working conditions and pair them with the optimal parameters after iterative optimization to construct samples, ensuring that the sample data fully conforms to the deviation patterns of the real working conditions. Specifically, we synchronously acquire the real-time operating status data of the weaving equipment at the time of triggering the update and the normalized intrinsic attribute vector corresponding to the current multi-medium combination. We then pair and bind the two types of original topological feature data with the control parameter combination obtained by the aforementioned iterative convergence to form a single incremental training sample. The sample data retains the material property disturbances, equipment deviations, and optimal process adaptation relationships of the current working conditions.

[0132] This ensures the authenticity, working condition matching, and mechanism integrity of the incremental training samples, accurately represents the optimal parameter mapping relationship corresponding to the current working condition deviation, and provides standardized input data for subsequent graph neural network forward inference, inference loss calculation, and weight backward update. It is the core data carrier for incremental model learning.

[0133] After constructing incremental training samples, traditional global parameter update methods will overwrite the multi-media coupling process mechanism learned by the model, causing catastrophic forgetting problems. Forward inference is the process of fixing the current network weights and outputting the predicted control parameters by inputting training sample features. The inference loss is the deviation between the network predicted parameters and the optimized parameters. Backpropagation update is the gradient iteration process of targeted correction of the weights of the core computation layer of the network. Parameter validity verification is a screening mechanism that selects and retains high-quality network parameters.

[0134] First, the training samples are input into the currently converged graph neural network to complete topological feature aggregation, attention weighting, and forward operation of the inference layer. The network then outputs the control parameter combination predicted by the network. The deviation calculation is performed between the predicted control parameter combination and the optimized control parameter combination bound to the sample to obtain the inference loss value corresponding to the current sample. For example, the inference threshold is set to 0.25, which is used as the criterion for determining the convergence of network parameters.

[0135] Specific inference loss L infer =(1 / N)×Σ‖p pred(b) -p opt(b) || 2 / R b 2 p pred(b)Let p be the predicted value of the b-th control parameter output by the graph neural network forward inference. opt(b) For the b-th optimized control parameter in the training samples, R b The adjustable range width of the b-th control parameter is used to normalize and eliminate the dimensional differences between different parameters. N is the number of dimensions of the control parameter. In this embodiment, N=4.

[0136] For example, the network prediction output is [20m / min, 1050N, 128℃, corresponding warp feed rate], and the optimized parameters are [21m / min, 1080N, 130℃, corresponding warp feed rate]. Assuming the warp feed rate deviation is negligible, then L infer =[(20-21) 2 / 10 2 +(1050-1080) 2 / 400 2 +(128-130) 2 / 30 2 +0] / 4≈(0.01+0.0056+0.0044+0) / 4≈0.005.

[0137] In the parameter update phase, the topological feature embedding layer parameters of the fixed graph neural network remain unchanged, preserving the original learned multi-media property correlation and process coupling general mechanism. Only the trainable parameters of the multi-layer graph convolutional layer, attention weight layer, and inference layer are updated by gradient backpropagation to accurately correct the mapping deviation of the current network condition. After each round of gradient backpropagation iteration, the inference loss value corresponding to the current network parameters is recalculated, and the parameter validity is checked. Only network parameters with inference loss less than the preset inference threshold are retained, while invalid parameters that degrade the network accuracy are discarded. The iteration continues until the inference loss stabilizes and converges and remains below the inference threshold for a long period of time. The latest graph neural network weight parameters are then fixed, completing the entire incremental update process.

[0138] The specific hyperparameter configuration for incremental training is as follows: the learning rate is set to 0.001, using the Adam optimizer; the batch size is all incremental training samples in the current batch, usually 1 sample; the maximum number of incremental training rounds is 50 rounds; the convergence criterion for inference loss is that the change in inference loss is less than 0.001 for 5 consecutive rounds; the learning rate of the graph convolutional layer and the attention weight layer is multiplied by a decay factor of 0.5, while the inference layer maintains a base learning rate of 0.001, ensuring that the underlying general mechanism parameters are finely adjusted and the top-level mapping parameters are quickly adapted.

[0139] This ensures that each iteration positively optimizes the network accuracy, enabling the updated network to accurately adapt to the latest weaving conditions, completing the self-evolutionary iterative update of the graph neural network. The updated network will then participate in subsequent real-time weaving parameter inference, achieving full lifecycle adaptive evolution of multi-media weaving control.

[0140] Example 2:

[0141] like Figure 3 As shown, this embodiment provides an adaptive collaborative control system for multi-media weaving equipment, including a sensing module, an execution module, and a feedback module.

[0142] The perception module is used to acquire the operating status data of the weaving equipment and the intrinsic attribute vector of the weaving medium. After inputting the data into the graph neural network, it outputs the combination of control parameters under the current medium combination through forward inference.

[0143] The execution module is used to generate a corresponding compensation amount when any change in the weaving medium is detected, based on the difference in elastic modulus and thermal shrinkage rate of the weaving medium before and after the change, and according to the sensitivity relationship between the output layer and the input layer of the graph neural network, and then superimpose it on the control parameter combination and send it to the execution mechanism.

[0144] If no change in any weaving medium is detected, the control parameter combination will be directly sent to the actuator;

[0145] The feedback module is used to monitor the actual defect rate after the execution is issued. When the actual defect rate is greater than the preset quality threshold, the combination of control parameters is iteratively optimized with the goal of minimizing the defect rate, and the parameters of the graph neural network are incrementally trained and updated.

[0146] Based on the above embodiments, this application is implemented.

[0147] The above description is merely a preferred embodiment of this application. The scope of protection of this application is not limited to the above embodiments. All technical solutions falling within the scope of this application's concept are within the scope of protection of this application. It should be noted that for those skilled in the art, any improvements and modifications made without departing from the principles of this application should also be considered within the scope of protection of this application.

Claims

1. An adaptive and coordinated control method for multi-media weaving equipment, characterized in that, include: The system acquires the operating status data of the weaving equipment and the intrinsic attribute vector of the weaving medium, inputs them into a graph neural network, and outputs the combination of control parameters under the current medium combination through forward inference. When any change in the weaving medium is detected, a corresponding compensation amount is generated based on the difference in elastic modulus and thermal shrinkage rate of the weaving medium before and after the change, as well as the sensitivity relationship between the output layer and the input layer of the graph neural network. This compensation amount is then superimposed on the control parameter combination and sent to the actuator. Generating the compensation amount includes: The difference in elastic modulus and thermal shrinkage rate of the fabric before and after the change is converted into physical property perturbation features and embedded into the corresponding medium feature nodes in the topology graph. Based on the interlayer gradient relationship of graph neural networks, the feature sensitivity of the graph neural network output parameters to the feature nodes of the medium under the current working condition is solved. By combining the feature contribution weights of graph nodes, the physical property disturbance characteristics and feature sensitivity are coupled and calculated to obtain the single-node compensation component corresponding to the changed weaving medium. Aggregate all single-node compensation components to obtain a compensation amount adapted to the current working condition; the physical property disturbance feature is a dimension alignment feature obtained by normalizing the difference in elastic modulus and the difference in thermal shrinkage rate and then splicing them together; the feature sensitivity is a sensitivity coefficient obtained by solving the gradient value of the output parameter of the graph neural network relative to the feature node of the medium through backpropagation. When no change in any weaving medium is detected, the control parameter combination is directly sent to the actuator; Monitor the actual defect rate after execution. When the actual defect rate is greater than the preset quality threshold, iteratively optimize the combination of control parameters with the goal of minimizing the defect rate, and incrementally train and update the parameters of the graph neural network.

2. The adaptive collaborative control method for multi-media weaving equipment according to claim 1, characterized in that, The control parameter combination under the current media combination includes: Based on the physical property correlation constraints of the weaving medium and the process matching constraints of the weaving equipment, a topological correlation diagram adapted to multi-media collaborative weaving is constructed. The operating status data of the weaving equipment and the intrinsic attribute vector of the weaving medium are respectively processed by feature normalization and assigned as node features of the corresponding graph nodes. The graph neural network is used to aggregate neighborhood features and optimize adaptive weights of node features in the topological association graph, and then fuses them to generate global features. The global features are input into the inference layer of the graph neural network for forward inference, and the output is mapped to a combination of control parameters that are adapted to the current medium combination and operating conditions.

3. The adaptive collaborative control method for multi-media weaving equipment according to claim 2, characterized in that, The topological association graph includes the medium feature nodes of the weaving medium and the equipment topological nodes of the corresponding process execution units in the weaving equipment; the neighborhood feature aggregation is obtained by traversing the neighborhood nodes of the topological association graph through a multi-layer graph convolutional layer, and aggregating the physical property association features between the weaving medium and the process coupling features between the weaving medium and the process execution units in the weaving equipment. The adaptive weight optimization is achieved by using a graph attention mechanism to adaptively allocate the feature contribution weights of graph nodes based on the degree of influence of different graph nodes on the quality of the woven product.

4. The adaptive collaborative control method for multi-media weaving equipment according to claim 3, characterized in that, The coupling calculation involves weighting and fusing the feature sensitivity with the corresponding feature contribution weight, and then performing an inner product operation with the physical property disturbance feature to obtain a single-node compensation component.

5. The adaptive collaborative control method for multi-media weaving equipment according to claim 2, characterized in that, When no change in any weaving medium is detected, the node structure and feature contribution weights of the current topological graph remain unchanged, and the global features corresponding to the current medium combination are solidified. The control parameter combination of the current inference output of the graph neural network is reused and directly sent to the actuator.

6. The adaptive collaborative control method for multi-media weaving equipment according to claim 3, characterized in that, The parameters of the incrementally trained and updated graph neural network include: After the monitoring and control parameter combination is issued and executed, the actual defect rate of the woven product is triggered. When the actual defect rate is greater than the preset quality threshold, the parameter update of the graph neural network is triggered. Based on the node characteristics of the topological association graph, the current combination of control parameters is iteratively optimized with the goal of minimizing the defect rate. The training samples are formed by combining the current operating status data, intrinsic attribute vectors and optimized control parameters. The graph neural network is incrementally trained based on the training samples. The weights of the graph neural network are iteratively corrected through gradient backpropagation, and the parameters of the graph neural network are updated through incremental training.

7. The adaptive collaborative control method for multi-media weaving equipment according to claim 6, characterized in that, The iterative optimization of the current control parameter combination includes: Construct a loss function adapted to multi-media collaborative weaving based on the feature contribution weights of graph nodes; The control parameter combination is corrected round by round using the gradient descent method. After each round of iteration, the loss value of the actual fabric defect rate is calculated. When the loss value of the actual fabric defect rate is less than the loss threshold in multiple consecutive rounds of iteration, the iteration is considered to have converged, and the optimized control parameter combination is output.

8. The adaptive collaborative control method for multi-media weaving equipment according to claim 7, characterized in that, The training samples are input into the current graph neural network to complete forward inference, and the predicted combination of control parameters is output. The inference loss of the training samples is calculated by combining the optimized combination of control parameters. The parameters of the multi-layer graph convolutional layer, attention weight layer, and inference layer are updated via backpropagation. After each round of incremental iteration, the validity of the parameters is verified, and the parameters that make the inference loss less than the inference threshold are retained. Once the inference loss converges iteratively, the parameters of the current graph neural network are fixed.

9. An adaptive collaborative control system for multi-media weaving equipment, used to implement the adaptive collaborative control method for multi-media weaving equipment as described in any one of claims 1-8, characterized in that, include: The module consists of a perception module, an execution module, and a feedback module. The perception module is used to acquire the operating status data of the weaving equipment and the intrinsic attribute vector of the weaving medium. After inputting the data into the graph neural network, it outputs the combination of control parameters under the current medium combination through forward inference. The execution module is used to generate a corresponding compensation amount when any change in the weaving medium is detected, based on the difference in elastic modulus and thermal shrinkage rate of the weaving medium before and after the change, as well as the sensitivity relationship between the output layer of the graph neural network and the input layer, and then superimpose it on the control parameter combination and send it to the execution mechanism. When no change in any weaving medium is detected, the control parameter combination is directly sent to the actuator; The feedback module is used to monitor the actual defect rate after the execution is issued. When the actual defect rate is greater than the preset quality threshold, the combination of control parameters is iteratively optimized with the goal of minimizing the defect rate, and the parameters of the graph neural network are incrementally trained and updated.

Images