An improved method based on bus duct production process parameters

Abstract

The application relates to the field of deep neural networks, in particular to an improved method based on bus duct production process parameters. The method defines each process of a bus duct production line as a graph network topology node, defines a physical medium flow path between processes as an edge, and defines a real-time multi-dimensional process parameter sequence of each process as a feature vector; through a graph convolution layer in a deep neural network, spatial coupling features between processes are extracted according to an adjacency matrix, and a dynamic drift feature of the feature vector with time evolution is extracted through a time convolution network layer; the spatial coupling features and the dynamic drift features are fused and mapped to generate a dynamic compensation amount of a key parameter of a downstream process, and the dynamic compensation amount is issued to a controller to correct an actuator parameter. The application breaks the isolation state between each process of the bus duct, realizes cross-process dynamic feedforward compensation, intercepts the conduction and multi-stage amplification of upstream deviation to the downstream, and inhibits product deformation and insulation failure.

Description

Technical Field

[0001] This invention relates to the field of deep neural networks, and more specifically to an improved method based on busbar production process parameters. Background Technology

[0002] Busbar trunking production involves multiple continuous processes, including stamping, bending, welding, and plating. In traditional process parameter control systems, each process employs an independent closed-loop feedback adjustment mechanism. The stamping process adjusts the stamping speed based on feedback from a stamping depth sensor; the bending process corrects the bending pressure based on angle measurements; the welding process adjusts the welding current using infrared temperature data; and the plating process controls the plating solution concentration using a thickness monitor. Each process's controller only receives real-time detection data for its own process and makes local adjustments according to preset static parameter thresholds or single-variable proportional-integral-derivative control laws, forming isolated single-point feedback loops that lack consideration for the transfer of physical quantities between upstream and downstream processes.

[0003] With the development of data processing technology, some existing solutions have begun to incorporate conventional deep neural networks for process parameter optimization. These solutions typically construct fully connected network models for a single, specific process, using the real-time parameters of the current process as network input and outputting control adjustments for that process. This purely data-driven, single-point prediction model still treats each process as a separate, independent object; its network architecture essentially only captures the mapping relationship between input and output parameters within a single process. During model training and inference, no information transmission channels are established along the physical media flow paths between stamping and bending, bending and welding, etc., making it impossible to depict the dynamic evolution of upstream process parameter fluctuations spreading downstream along the physical transport trajectory.

[0004] Because the existing technology breaks down the continuous busbar production line into isolated adjustment nodes, when slight drifts in process parameters occur in the upstream stamping or bending processes due to equipment vibration or batch differences in materials, these deviations are transmitted to the downstream welding and plating processes through the physical flow of the busbar conductor material. The downstream independent closed-loop control loop lacks the ability to perceive and predict the transmission path of upstream deviations, and cannot perform dynamic feedforward intervention across processes before the deviations accumulate and amplify. This results in the nonlinear superposition and continuous amplification of small upstream disturbances through multiple processes, ultimately leading to problems such as out-of-tolerance conductor dimensional deformation and insulation layer adhesion failure in the finished busbar. Summary of the Invention

[0005] The purpose of this invention is to provide an improved method based on busbar trunking production process parameters, which can solve the problems mentioned in the background art.

[0006] To achieve the above objectives, the technical solution adopted by the present invention is as follows:

[0007] An improved method based on busbar trunking production process parameters includes: defining each process of the busbar trunking production line as a topological node of a graph network, defining the physical medium flow path between the processes as an edge of the graph network, and defining the real-time multidimensional process parameter sequence corresponding to each process as a feature vector of the topological node; extracting spatial coupling features between the processes using a graph convolutional layer in a deep neural network based on the adjacency matrix of the graph network, and extracting dynamic drift features of the feature vectors over time using a temporal convolutional network layer in the deep neural network; fusing and mapping the spatial coupling features and the dynamic drift features to generate dynamic compensation amounts for key parameters of downstream processes corresponding to each process, and sending the dynamic compensation amounts to the controller to correct the parameters of the actuator.

[0008] Preferably, each process of the busbar trunking production line is defined as a topological node of a graph network, the physical medium flow path between the processes is defined as an edge of the graph network, and the real-time multidimensional process parameter sequence corresponding to each process is defined as a feature vector of the topological node. This includes: mapping the stamping, bending, welding, and plating processes in the busbar trunking production line to a first, second, third, and fourth topological node of the graph network, respectively; constructing a first edge between the first and second topological nodes based on the physical transport trajectory of the busbar trunking conductor material between the stamping and bending processes, and similarly constructing a second edge between the second and third topological nodes and a third edge between the third and fourth topological nodes; collecting the stamping speed and stamping depth parameters of the stamping process, the bending angle and bending speed parameters of the bending process, the welding current and welding temperature parameters of the welding process, and the plating thickness and plating solution concentration parameters of the plating process, respectively, as the feature vectors of the first, second, third, and fourth topological nodes.

[0009] Preferably, the spatial coupling features between the various processes are extracted using the graph convolutional layer in the deep neural network based on the adjacency matrix of the graph network. This includes: constructing the adjacency matrix of the graph network, where the element values ​​represent the spatial influence weights of process parameters between adjacent topological nodes; performing graph convolution operations on the feature vector and the adjacency matrix to obtain cross-process spatial interaction information of process parameters between adjacent topological nodes; performing nonlinear activation mapping on the cross-process spatial interaction information to capture the spatial transmission path and attenuation coefficient of parameter deviations of upstream topological nodes to downstream topological nodes, thereby generating the spatial coupling features.

[0010] Preferably, the dynamic drift features of the feature vector evolving over time are extracted through the temporal convolutional network layer in the deep neural network, including: performing a sliding convolution operation on the feature vector of each topological node along the time dimension to extract the local temporal dependencies of the feature vector within a preset time window; expanding the receptive field of the local temporal dependencies through the dilated convolutional kernel in the temporal convolutional network layer to capture the dynamic drift trend of the feature vector over a long period of time; fusing the local temporal dependencies and the dynamic drift trend through residual connections to filter out high-frequency process noise interference in the feature vector and output the dynamic drift features.

[0011] Preferably, the spatial coupling feature and the dynamic drift feature are fused and mapped to generate dynamic compensation amounts for key parameters of downstream processes corresponding to each process, and the dynamic compensation amounts are sent to the controller to correct the parameters of the actuator. This includes: performing feature concatenation operations on the spatial coupling feature and the dynamic drift feature in the fully connected layer of the deep neural network to obtain a joint feature representation; inputting the joint feature representation into the feedforward output layer of the deep neural network, using the elimination of potential cumulative deviations in the downstream process as the objective function, to calculate the dynamic compensation amounts for the key parameters of the downstream processes; converting the dynamic compensation amounts into parameter correction instructions that the controller can recognize, and sending the parameter correction instructions to the controller, which then adjusts the operating parameters of the actuator according to the parameter correction instructions.

[0012] Preferably, constructing the adjacency matrix of the graph network, wherein the element values ​​in the adjacency matrix represent the spatial influence weights of process parameters between adjacent topological nodes, includes: initializing the element values ​​in the adjacency matrix as static basic weights determined based on physical mechanisms; introducing an attention mechanism to dynamically modulate the static basic weights during the training and inference process of the deep neural network, calculating the real-time attention score of the feature vector of the upstream topological node to the feature vector of the downstream topological node; weighting and fusing the static basic weights and the real-time attention score to generate dynamic spatial influence weights, and updating the dynamic spatial influence weights in the adjacency matrix.

[0013] Preferably, the receptive field of the local temporal dependency is expanded by the dilated convolutional kernel in the temporal convolutional network layer to capture the dynamic drift trend of the feature vector over a long period of time. This includes: setting multiple dilated convolutional kernels with different dilation rates, performing multi-scale temporal convolution processing on the feature vector to obtain initial temporal features at multiple different time scales; concatenating the initial temporal features at multiple different time scales to construct a multi-scale temporal feature matrix; performing feature selection on the multi-scale temporal feature matrix to select the target scale temporal feature with the highest matching degree to the current process drift mode of the busbar, and confirming the target scale temporal feature as the dynamic drift trend.

[0014] Preferably, the joint feature representation is input into the feedforward output layer of the deep neural network to calculate the dynamic compensation amount of the key parameters of the downstream process, with the objective function of eliminating potential cumulative deviations in the downstream process. This includes: predicting the prediction deviation vector of the downstream process under the action of the current upstream process parameters in the feedforward output layer based on the joint feature representation; inputting the prediction deviation vector into a preset quadratic programming solver, with the minimization of the L2 norm of the dynamic compensation amount as a constraint, and with the objective of offsetting the prediction deviation vector, to solve for the dynamic compensation amount of the key parameters of the downstream process.

[0015] Preferably, after sending the dynamic compensation amount to the controller to correct the parameters of the actuator, the method further includes: collecting physical performance test data of the final product of the busbar trunking, the physical performance test data including conductor size deformation tolerance data and insulation resistance test data; comparing the physical performance test data with preset standard performance indicators to calculate the performance deviation value of the finished product; constructing a global loss gradient based on the performance deviation value of the finished product, and adaptively updating the adjacency matrix weights of the graph convolutional layer of the deep neural network and the convolution kernel parameters of the temporal convolutional network layer through a backpropagation algorithm.

[0016] Preferably, before defining the real-time multidimensional process parameter sequence corresponding to each process as the feature vector of the topology node, the method further includes: performing outlier removal and smoothing filtering on the collected real-time multidimensional process parameter sequence to eliminate sensor pulse noise in the real-time multidimensional process parameter sequence; performing a normalization mapping operation on the real-time multidimensional process parameter sequence after smoothing filtering to uniformly map heterogeneous process parameters to the same numerical range, eliminating numerical differences between different dimensions, and confirming the normalized and mapped real-time multidimensional process parameter sequence as the feature vector of the topology node.

[0017] Compared with the prior art, the beneficial effects of the present invention are as follows: 1. This invention constructs a complete cross-process graph network structure by defining each process of the busbar production line as a topological node of a graph network and defining the physical medium flow path between processes as edges. Based on this, the graph convolutional layer in a deep neural network extracts the spatial coupling features between processes according to the adjacency matrix, establishing a spatial transmission path mapping from upstream to downstream processes along the physical transport trajectory, enabling downstream processes to perceive parameter deviations transmitted from upstream. Simultaneously, the temporal convolutional network layer in the deep neural network extracts the dynamic drift features of feature vectors evolving over time, capturing the continuous changes of process parameters in the time dimension. The spatial coupling features and dynamic drift features are fused and mapped to generate dynamic compensation quantities for key parameters of downstream processes and sent to the controller, realizing cross-process dynamic feedforward compensation based on the real-time state of upstream processes. This control method breaks the isolation between processes, blocks the transmission and multi-level amplification of upstream deviations to downstream processes, overcomes the shortcomings of single-process local closed-loop control in dealing with the accumulation of cross-process disturbances, and suppresses the deformation of the conductor size and insulation layer adhesion failure of the finished busbar due to parameter mismatch.

[0018] 2. In the execution of dynamic compensation, this invention introduces a mechanism of weighted fusion of static basic weights and real-time attention scores to generate dynamic spatial influence weights that are updated in the adjacency matrix. This allows the graph convolutional layer to adaptively adjust the degree of spatial influence between processes according to the current working conditions, improving the accuracy of cross-process spatial coupling feature extraction. In the time dimension, multiple convolutional kernels with different dilation rates are set to obtain multi-scale temporal features and perform feature selection operations to filter out target scale features that match the current drift pattern, enhancing the adaptability to process fluctuations in different cycles. When generating the dynamic compensation amount, the prediction deviation vector is used to solve the problem under the constraint of minimizing the L2 norm of the compensation amount, ensuring the smoothness of the compensation action and avoiding mechanical impact on the actuator due to excessive compensation amount. A global loss gradient is constructed based on the conductor size deformation tolerance data and insulation resistance detection data of the final product of the busbar trunking. The parameters of the graph convolutional layer and the temporal convolutional layer are adaptively updated through the backpropagation algorithm, enabling the network model to adapt to the slow time-varying characteristics of the production environment. By performing outlier removal, smoothing filtering, and normalization mapping operations on the real-time multidimensional process parameter sequence, the interference of sensor impulse noise and the dimensional differences of heterogeneous parameters on feature extraction is eliminated, ensuring the stability of the graph network feature vector. Attached Figure Description

[0019] Figure 1 This is a flowchart of the preprocessing of original process parameters for the bus trunking and the construction of the production line process topology network for the present invention. Figure 2 This is a flowchart of the graph convolutional adjacency matrix static weight initialization + attention dynamic weight correction and update process of the present invention; Figure 3 This is a flowchart of the multi-scale dilated convolution and temporal feature filtering and dynamic drift feature extraction of the present invention; Figure 4 This is a flowchart of the spatial temporal feature fusion + quadratic programming solution for downstream process dynamic compensation in this invention; Figure 5 This is a flowchart illustrating the dynamic compensation command conversion, controller issuance, and field actuator parameter correction processes of the present invention. Figure 6 This is a flowchart of the finished product performance testing, performance deviation calculation, and network parameter reverse iterative closed-loop optimization process of the present invention. Detailed Implementation

[0020] In one embodiment, each process of the busbar trunking production line is defined as a topological node of a graph network, the physical medium flow paths between processes are defined as edges of the graph network, and the real-time multidimensional process parameter sequences corresponding to each process are defined as feature vectors of the topological nodes. The busbar trunking production line comprises multiple continuous production stages, each corresponding to an independent production process. Conductor materials flow between processes via physical conveyor devices such as conveyor belts and robotic arms. The topological structure of the graph network completely corresponds to the physical layout of the busbar trunking production line. Each topological node uniquely corresponds to a production process, and edges between nodes exist only between adjacent processes with a direct physical medium flow relationship. The feature vector of a topological node consists of multiple real-time process parameters corresponding to that process. Each process parameter is collected at fixed time intervals, forming continuous time-series data.

[0021] Spatial coupling features between different processes are extracted using graph convolutional layers in a deep neural network based on the adjacency matrix of the graph network. The adjacency matrix is ​​a two-dimensional matrix whose number of rows and columns equals the number of topological nodes in the graph network. Each element in the matrix corresponds to the connection relationship and influence degree between two topological nodes. The graph convolutional layer performs convolution operations on the feature vectors and the adjacency matrix, fusing the feature information of each topological node with the feature information of its neighboring nodes to capture the mutual influence relationships of process parameters between different processes. Spatial coupling features include information on the potential impact of upstream process parameter changes on downstream processes, as well as the transmission path and intensity of parameter deviations between different processes.

[0022] The dynamic drift features of feature vectors evolving over time are extracted using temporal convolutional network layers in deep neural networks. These layers perform convolution operations on the feature vectors of each topological node along the time dimension, extracting the variation patterns of process parameters at different time scales. The dynamic drift features reflect the slow changing trends of process parameters over time, typically caused by slow time-varying factors such as equipment wear, batch-to-batch material variations, and environmental temperature fluctuations. The temporal convolutional network layer effectively separates normal fluctuations and abnormal drifts in process parameters, providing accurate temporal data for subsequent parameter compensation.

[0023] Spatial coupling features and dynamic drift features are fused and mapped to generate dynamic compensation quantities for key parameters of downstream processes corresponding to each process. These dynamic compensation quantities are then sent to the controller to correct the parameters of the actuators. Spatial coupling features and dynamic drift features describe the state information of the busbar production process from spatial and temporal dimensions, respectively. Fusing them yields a more comprehensive and accurate representation of the production state. The fused features are then processed by a multi-layer neural network to convert them into adjustment quantities for key parameters of downstream processes. Upon receiving the dynamic compensation quantities, the controller converts them into control signals for the actuators, adjusting their operating parameters to offset the impact of upstream process parameter deviations on downstream processes.

[0024] In this embodiment, the construction process of the graph network maintains a strict correspondence with the physical structure of the busbar production line. Each topology node contains a unique identifier, which is associated with the controller address of the corresponding process. The edges between topology nodes are directional, with the direction of the edges consistent with the flow direction of the physical medium, i.e., from the upstream process to the downstream process. This directed graph structure can accurately reflect the transmission direction of process parameter deviations and avoid interference from irrelevant information. The dimension of the feature vector is the same as the number of process parameters collected for the corresponding process, with each dimension corresponding to a specific process parameter, and each element of the feature vector being the value of that process parameter at the current sampling time.

[0025] The deep neural network adopts an end-to-end architecture, with the input being the feature vectors of each topological node and the output being the dynamic compensation amount for key parameters in the downstream process. The input layer receives the feature vectors of each topological node and converts them into a unified tensor format. The graph convolutional layer, located after the input layer, is responsible for extracting spatial coupling features. The temporal convolutional network layer is connected in parallel with the graph convolutional layer, simultaneously processing the input feature vectors to extract dynamic drift features. The fusion layer concatenates the spatial coupling features output from the graph convolutional layer and the dynamic drift features output from the temporal convolutional network layer to form a joint feature representation. The fully connected layer performs further nonlinear transformations on the joint feature representation, and finally, the output layer generates the dynamic compensation amount.

[0026] In this embodiment, real-time process parameters for each step are first collected at a fixed sampling frequency to construct a feature vector for each topological node. The feature vectors of all topological nodes are then input into a deep neural network. A graph convolutional layer performs spatial convolution on the feature vectors based on the adjacency matrix to extract spatial coupling features between different steps. Simultaneously, a temporal convolutional network layer performs temporal convolution on the feature vectors of each topological node to extract the dynamic drift features of process parameters over time. A fusion layer concatenates the spatial coupling features and dynamic drift features along the channel dimension to obtain joint features containing both spatial and temporal information. A fully connected layer performs multi-layer nonlinear mapping on the joint features, converting them into dynamic compensation values ​​for key parameters in downstream processes. Finally, the dynamic compensation values ​​are sent to the controller of the corresponding process. The controller adjusts the operating parameters of the actuators based on the dynamic compensation values ​​to achieve real-time control of the production process.

[0027] In one embodiment, the stamping, bending, welding, and plating processes in the busbar trunking production line are mapped to the first, second, third, and fourth topological nodes of a graph network, respectively. The stamping process is the first step in busbar trunking production, primarily responsible for stamping copper or aluminum busbar raw materials into specified shapes and sizes. The bending process bends the stamped conductor material to form the specific structure required for the busbar trunking. The welding process welds multiple bent conductor components together to form a complete conductor unit. The plating process applies a surface coating to the welded conductor unit to improve its conductivity and corrosion resistance. These four processes constitute the core flow of busbar trunking production, and changes in the process parameters of each process directly affect the quality of the final product.

[0028] Based on the physical transport trajectory of the conductor material in the busbar trunking between the stamping and bending processes, a first edge is constructed between the first and second topological nodes. Similarly, a second edge is constructed between the second and third topological nodes, and a third edge between the third and fourth topological nodes. The first edge corresponds to the conveyor belt between the stamping and bending processes; after stamping, the conductor material is transported to the bending process for further processing. The second edge corresponds to the robotic arm transport device between the bending and welding processes; the bent conductor component is picked up by the robotic arm and placed at the welding station. The third edge corresponds to the overhead conveyor chain between the welding and plating processes; the welded conductor unit is transported to the plating workshop for surface treatment via the overhead conveyor chain. Each edge has a unique identifier used to locate the corresponding element in the adjacency matrix.

[0029] The stamping speed and depth parameters of the stamping process, the bending angle and bending speed parameters of the bending process, the welding current and welding temperature parameters of the welding process, and the plating thickness and plating solution concentration parameters of the plating process are collected and used as feature vectors for the first, second, third, and fourth topological nodes, respectively. Stamping speed refers to the movement speed of the stamping press slide, measured in millimeters per second. Stamping depth refers to the maximum distance the stamping press slide travels downwards, measured in millimeters. Bending angle refers to the angle formed after bending the conductor material, measured in degrees. Bending speed refers to the movement speed of the bending machine slide, measured in millimeters per second. Welding current refers to the current passing through the welding electrode during welding, measured in amperes. Welding temperature refers to the highest temperature in the welding area, measured in degrees Celsius. Plating thickness refers to the average thickness of the plating layer on the conductor surface, measured in micrometers. Plating solution concentration refers to the mass concentration of metal ions in the plating solution, measured in grams per liter.

[0030] In this embodiment, the feature vector of each topology node is a two-dimensional tensor, with the first dimension being the time dimension and the second dimension being the parameter dimension. The length of the time dimension of the feature vector is determined by the size of a preset time window, and the parameter dimension has a length of 2, corresponding to the two process parameters collected in each process step. The feature vector of the first topology node is represented as follows: ,in The time window size is represented by the first column, which corresponds to the stamping speed parameter, and the second column corresponds to the stamping depth parameter. The feature vector of the second topological node is represented as follows: The first column corresponds to the bending angle parameter, and the second column corresponds to the bending speed parameter. The feature vector of the third topological node is represented as follows: The first column corresponds to the welding current parameters, and the second column corresponds to the welding temperature parameters. The eigenvector of the fourth topological node is represented as follows: The first column corresponds to the coating thickness parameter, and the second column corresponds to the plating solution concentration parameter.

[0031] Table 1. Correspondence between topology nodes and process parameters

[0032] Table 1 shows the mapping relationship between each topology node and its corresponding process and parameters. As can be seen from Table 1, each topology node corresponds to a specific production process, and each process collects two key process parameters as components of its feature vector. This mapping method can comprehensively reflect the operating status of each process while avoiding increased computational complexity caused by introducing too many irrelevant parameters. (Time window size) The settings are configured based on the changing characteristics of production rhythm and process parameters to ensure that the feature vector contains sufficient historical information to accurately capture the dynamic changing trends of process parameters.

[0033] In one embodiment, an adjacency matrix is ​​constructed for the graph network, where the element values ​​represent the spatial influence weights of process parameters between adjacent topological nodes. (Adjacency matrix) ,in The number of topology nodes, in this embodiment Elements in the adjacency matrix Indicates the first The topology node pair of the th The process parameters of each topological node have an impact weight. When there is no direct physical medium flow relationship between two topological nodes, the corresponding adjacency matrix element value is 0. When there is a direct physical medium flow relationship between two topological nodes, the corresponding adjacency matrix element value is a real number greater than 0, and the magnitude of this real number reflects the degree of influence of changes in upstream process parameters on downstream processes.

[0034] By performing graph convolution on the feature vectors and the adjacency matrix, cross-process spatial interaction information of process parameters between adjacent topological nodes is obtained. The graph convolution operation can be represented as:

[0035] in, For the first The input feature matrix of the layered graph convolutional layer, For the first The output feature matrix of the layer graph convolutional layer, To add self-loops to the adjacency matrix, It is the identity matrix. for The degree matrix, , For the first The trainable weight matrix of a layered graph convolutional layer The ReLU activation function is used. In this embodiment, the input feature matrix... It is obtained by averaging the feature vectors of each topological node over the time dimension, i.e. , of which The row corresponds to the first The average eigenvector of the topological nodes.

[0036] Nonlinear activation mapping is applied to the spatial interaction information across processes to capture the spatial propagation path and attenuation coefficient of parameter deviations from upstream topological nodes to downstream topological nodes, generating spatial coupling features. The ReLU activation function expression is as follows: The ReLU function introduces a nonlinear transformation, enhancing the ability of graph convolutional layers to fit complex nonlinear relationships. After graph convolution and nonlinear activation mapping, the output feature matrix... ,in Let be the number of output channels of the convolutional layer. Spatial coupling characteristics are... , of which The row contains all upstream topology node pairs. The impact information of each topological node, and the attenuation of this impact information during the propagation process.

[0037] In this embodiment, the initial value of the adjacency matrix is ​​set according to the physical mechanism of busbar production. Since changes in the parameters of the stamping process affect the processing accuracy of the bending process, changes in the parameters of the bending process affect the welding quality of the welding process, and changes in the parameters of the welding process affect the coating adhesion of the plating process, the initial value of the adjacency matrix is ​​set as follows:

[0038] in, , , These are static basic weights determined based on physical mechanisms. This indicates the degree of influence of changes in stamping process parameters on the bending process. This indicates the degree of influence of changes in bending process parameters on the welding process. This indicates the degree of impact of changes in welding process parameters on the coating process. These static basic weights were obtained through statistical analysis of a large amount of historical production data, reflecting the average level of parameter influence between different processes.

[0039] Figure: Number of output channels of the convolutional layer The settings are configured according to the requirements of feature extraction, as described in this embodiment. The trainable weight matrix of the convolutional layer. The algorithm is optimized using backpropagation. During graph convolution, the feature vector of each topological node is weighted and summed with the feature vectors of its neighboring nodes; the weights are determined by the corresponding elements in the adjacency matrix. This method effectively integrates process parameter information across processes, capturing the transmission pattern of upstream process parameter deviations to downstream processes.

[0040] In one embodiment, a sliding convolution operation is performed along the time dimension on the feature vectors of each topological node to extract the local temporal dependencies of the feature vectors within a preset time window. The temporal convolutional network layer consists of multiple stacked one-dimensional convolutional layers, each of which performs a convolution operation on the input feature vector along the time dimension. The size of the convolution kernel determines the range of the local temporal dependencies. In this embodiment, the kernel size of the first temporal convolutional layer is 3, which can capture the parameter change relationship between three adjacent sampling times. The stride of the sliding convolution operation is 1, ensuring that the output temporal features and the input feature vectors maintain the same length in the time dimension.

[0041] Dilated convolutional kernels in temporal convolutional networks expand the receptive field of local temporal dependencies, capturing the dynamic drift trend of feature vectors over long-term time spans. Dilated convolution is a convolutional operation that inserts holes into the convolutional kernel, expanding the receptive field without increasing the number of kernel parameters. The formula for calculating the receptive field size of dilated convolution is:

[0042] in, To feel the size of the wild, This is the original size of the convolution kernel. The dilation rate is 2. In this embodiment, the second temporal convolutional layer uses dilated convolution with a kernel size of 3 and a dilation rate of 2, and its receptive field size is 2. The third temporal convolutional layer also uses dilated convolution, with a kernel size of 3 and a dilation rate of 4, and its receptive field size is [missing value]. By stacking multiple dilated convolutional layers with different dilation rates, temporal convolutional networks can capture the variation patterns of process parameters at different time scales.

[0043] Residual connections are used to fuse local temporal dependencies and dynamic drift trends, filtering out high-frequency process noise interference in the feature vectors and outputting dynamic drift features. Residual connections are a connection method that directly adds input features to output features, effectively alleviating the gradient vanishing problem during deep neural network training. In this embodiment, a residual connection is added after each temporal convolutional layer, adding the input and output features of that layer element-wise. Residual connections preserve the original information in the input features while introducing higher-order features extracted by the convolutional layers, improving the accuracy of feature extraction. After fusing multiple temporal convolutions and residual connections, the output dynamic drift features are: ,in This represents the number of output channels for the temporal convolutional network layer.

[0044] In this embodiment, the temporal convolutional network layer employs a causal convolutional structure, ensuring that during convolution operations, the output at the current time step depends only on the inputs at the current and previous time steps, and not on the inputs at future time steps. Causal convolution is implemented by zero-padding the input features before the convolution operation, with the padding length being the kernel size minus 1. This structure aligns with the causal relationships in the production process, because changes in process parameters at the current time step can only affect future production states, not past production states.

[0045] Number of output channels of temporal convolutional network layers The settings are configured according to the requirements of temporal feature extraction, as described in this embodiment. The trainable weight matrix of each temporal convolutional layer is optimized using the backpropagation algorithm. During the convolution operation, each convolutional kernel slides along the temporal dimension, performing a dot product operation with the corresponding region of the input feature vector to generate an output feature map. Different convolutional kernels can capture different types of temporal variation patterns, such as periodic fluctuations, linear drift, and step changes.

[0046] In one embodiment, spatial coupling features and dynamic drift features are concatenated in a fully connected layer of a deep neural network to obtain a joint feature representation. Spatial coupling features It includes spatial interaction information between various processes and dynamic drift characteristics. It contains dynamic information about the evolution of process parameters over time. To fuse these two different dimensions of features, a global average pooling operation is first performed on the dynamic drift feature along the time dimension to obtain... The formula for calculating the global average pooling operation is:

[0047] in, For the pooling process The topological node's first The characteristic values ​​of each channel, The first of the original dynamic drift features The topology node at the ... The moment of the first The characteristic values ​​of each channel.

[0048] The dynamic drift characteristics after pooling Spatial coupling characteristics The features are concatenated along the channel dimension to obtain a joint feature representation. The calculation formula for the splicing operation is:

[0049] in, For the joint feature representation of the th The topological node's first The feature values ​​of each channel. In this embodiment, , Therefore, the number of channels represented by the joint feature is 32.

[0050] The joint feature representation is input into the feedforward output layer of a deep neural network. With the objective function of eliminating potential cumulative biases in downstream processes, the dynamic compensation amounts for key parameters in each downstream process are calculated. The feedforward output layer consists of two fully connected layers. The first fully connected layer maps the number of channels in the joint feature representation from 32 to 16, and the second fully connected layer maps the number of channels from 16 to 2, corresponding to the dynamic compensation amounts for the two key parameters in each downstream process. The activation function of the feedforward output layer is a linear activation function, ensuring that the output dynamic compensation amount can be any real number.

[0051] refer to Figure 5 The dynamic compensation value is converted into a parameter correction instruction that the controller can recognize, and the parameter correction instruction is sent to the controller. The controller adjusts the operating parameters of the actuator according to the parameter correction instruction. The dynamic compensation value is a continuous value and needs to be converted into a digital signal that the controller can recognize. In this embodiment, the Modbus communication protocol is used to send the dynamic compensation value to the programmable logic controller (PLC) of each process. After receiving the parameter correction instruction, the PLC adds it to the current set value to obtain the new control parameters, and sends the new control parameters to the actuator to adjust the operating state of the actuator.

[0052] In this embodiment, each topology node corresponds to a dynamic compensation amount for a downstream process, except for the last topology node (plating process), which has no downstream processes. The dynamic compensation amount corresponding to the first topology node (stamping process) is used to adjust the bending angle and bending speed parameters of the second topology node (bending process). The dynamic compensation amount corresponding to the second topology node (bending process) is used to adjust the welding current and welding temperature parameters of the third topology node (welding process). The dynamic compensation amount corresponding to the third topology node (welding process) is used to adjust the plating thickness and plating solution concentration parameters of the fourth topology node (plating process). This correspondence ensures that parameter deviations in each upstream process can be compensated for by adjusting parameters in the downstream process.

[0053] refer to Figure 2 In one embodiment, the elements in the initial adjacency matrix are static base weights determined based on physical mechanisms. These static base weights are obtained by analyzing the physical mechanisms of the busbar production process and combining this with statistical results from a large amount of historical production data. In this embodiment, the stamping depth parameter of the stamping process has a significant impact on the bending angle parameter of the bending process, and the stamping speed parameter has a certain impact on the bending speed parameter. Therefore… The initial value is set to 0.7. The bending angle parameter in the bending process has a significant impact on the welding current parameter in the welding process, and the bending speed parameter has a certain impact on the welding temperature parameter. The initial value is set to 0.6. The welding temperature parameter in the welding process has a significant impact on the coating thickness parameter in the plating process, and the welding current parameter has a certain impact on the plating solution concentration parameter. The initial value is set to 0.5. These static basic weights reflect the average level of parameter influence between different processes, providing a reasonable initial value for the adjacency matrix.

[0054] In the training and inference process of deep neural networks, an attention mechanism is introduced to dynamically modulate the static base weights, calculating the real-time attention score between the feature vectors of upstream topological nodes and the feature vectors of downstream topological nodes. This attention mechanism can adaptively adjust the influence of different upstream processes on downstream processes based on the current production status. In this embodiment, an additive attention mechanism is used to calculate the real-time attention score, and its calculation formula is as follows:

[0055] in, For the first The topology node pair of the th Attention score of each topological node For the first Feature vectors of topological nodes For the first Feature vectors of topological nodes , For a trainable weight matrix, These are trainable weight vectors. Attention score. This reflects the current production status, the first The feature vector of the i-th topological node is related to the i-th The importance of the feature vectors of each topological node.

[0056] The static base weights and real-time attention scores are weighted and fused to generate dynamic spatial influence weights, which are then updated in the adjacency matrix. The formula for weighted fusion is as follows:

[0057] in, The updated dynamic space affects the weights. For static base weights, This is a balancing factor used to control the relative importance of static base weights and real-time attention scores. For the first The set of all upstream neighbor nodes of a topological node. In this embodiment, the balance coefficient... Setting it to 0.5 ensures that the static base weights and real-time attention scores contribute equally to the dynamic space influence weights. The attention score is normalized using the softmax function to ensure that the sum of the dynamic space influence weights of all upstream neighbor nodes for each downstream topology node is 1.

[0058] In this embodiment, the introduction of an attention mechanism enables the adjacency matrix to be dynamically adjusted according to the current production status. When the parameters of an upstream process deviate significantly, the attention score of that upstream process for the downstream process will increase accordingly, increasing its weight in the adjacency matrix. This dynamic adjustment mechanism can improve the accuracy of graph convolutional layers in extracting cross-process spatial coupling features, allowing the model to better adapt to different production conditions.

[0059] Table 2. Weighting of Dynamic Spatial Influence under Different Operating Conditions

[0060] Table 2 shows the changes in the dynamic spatial influence weights in the adjacency matrix under different operating conditions. As can be seen from Table 2, under normal operating conditions, the dynamic spatial influence weights are quite close to the static basic weights. When the stamping depth deviation is large, The value increased from 0.68 to 0.82, indicating a significant increase in the influence of the stamping process on the bending process. When the bending angle deviation is large, The value increased from 0.59 to 0.78, indicating a significant increase in the impact of the bending process on the welding process. When the welding temperature deviation is large, The value increased from 0.48 to 0.76, indicating a significant increase in the impact of the welding process on the coating process. This dynamic adjustment mechanism allows the model to pay closer attention to the processes currently experiencing deviations and their impact on downstream processes, improving the targeting and effectiveness of parameter compensation.

[0061] refer to Figure 3 In one embodiment, multiple dilated convolution kernels with different dilation rates are set to perform multi-scale temporal convolution processing on the feature vectors to obtain initial temporal features at multiple different time scales. In this embodiment, three dilated convolution kernels with different dilation rates of 1, 2, and 4 are set. The dilated convolution kernel with a dilation rate of 1 is a normal convolution kernel with a receptive field size of 3, which can capture process parameter fluctuations at short time scales. The dilated convolution kernel with a dilation rate of 2 has a receptive field size of 5, which can capture process parameter changes at medium time scales. The dilated convolution kernel with a dilation rate of 4 has a receptive field size of 9, which can capture process parameter drift trends at long time scales. Each dilated convolution kernel has 8 output channels, so each topology node obtains three initial temporal features after multi-scale temporal convolution processing, and the dimension of each initial temporal feature is... .

[0062] A multi-scale temporal feature matrix is ​​constructed by concatenating the initial temporal features from multiple time scales. This is achieved by concatenating the initial temporal features output from three dilated convolutional kernels with different dilation rates along the channel dimension. The calculation formula for the splicing operation is:

[0063] in, The first element in the multi-scale time series feature matrix The moment of the first The characteristic values ​​of each channel, , , These are the initial temporal features of the outputs of dilated convolutional kernels with dilation rates of 1, 2, and 4, respectively.

[0064] Feature selection is performed on the multi-scale time-series feature matrix to filter out the target-scale time-series features that have the highest matching degree with the current process drift pattern of the busbar, and these target-scale time-series features are confirmed as dynamic drift trends. The feature selection operation employs a channel attention module based on an attention mechanism, which can automatically learn the importance weights of different channel features. The calculation formula for the channel attention module is:

[0065] in, This refers to the feature vector obtained by performing global average pooling on the multi-scale temporal feature matrix along the time dimension. , This is the weight matrix of the first fully connected layer. It is the ReLU activation function. This is the weight matrix for the second fully connected layer. It is the sigmoid activation function. This is the channel attention weight vector.

[0066] The channel attention weight vector is multiplied channel-by-channel with the multi-scale temporal feature matrix to obtain the weighted multi-scale temporal feature matrix. This weighted multi-scale temporal feature matrix represents the dynamic drift trend, where channel features with a high degree of matching to the current process drift pattern are enhanced, while those with a low degree of matching are suppressed. In this embodiment, when process parameters experience short-term fluctuations, the channel attention module increases the weight of the output features of convolutional kernels with an expansion rate of 1; when process parameters experience long-term drifts, the channel attention module increases the weight of the output features of convolutional kernels with an expansion rate of 4.

[0067] In this embodiment, the combination of multi-scale temporal convolution and feature selection operations enables the temporal convolutional network layer to adaptively capture process fluctuations at different periods. Different process parameter drift patterns have different temporal scale characteristics. For example, parameter fluctuations caused by equipment vibration usually have a short period, while parameter drifts caused by equipment wear usually have a long period. By obtaining features at different time scales through multi-scale temporal convolution and filtering out features that match the current drift pattern through feature selection operations, the accuracy of dynamic drift feature extraction can be improved.

[0068] Table 3 Channel attention weights under different drift modes

[0069] Table 3 shows the average attention weights of channels at various scales under different drift modes. As can be seen from Table 3, when the process parameters exhibit short-period fluctuations, the channel with an expansion rate of 1 has the highest average weight, reaching 0.72, indicating that the model mainly focuses on features at short timescales. When the process parameters exhibit medium-period changes, the channel with an expansion rate of 2 has the highest average weight, reaching 0.68, indicating that the model mainly focuses on features at medium timescales. When the process parameters exhibit long-period drift, the channel with an expansion rate of 4 has the highest average weight, reaching 0.70, indicating that the model mainly focuses on features at long timescales. This adaptive feature selection mechanism enables the model to better adapt to different types of process parameter changes, improving the effectiveness of dynamic drift feature extraction.

[0070] refer to Figure 4 In one embodiment, based on joint feature representation, the prediction deviation vector of the downstream process under the current upstream process parameters is predicted in the feedforward output layer. The first element corresponds to the prediction deviation of the first key parameter in the downstream process, and the second element corresponds to the prediction deviation of the second key parameter in the downstream process. The calculation process of the prediction deviation vector is as follows:

[0071] in, For the first The joint feature representation of the topological nodes, , This is the weight matrix of the first fully connected layer. This is the bias vector for the first fully connected layer. It is the ReLU activation function. This is the weight matrix for the second fully connected layer. This is the bias vector for the second fully connected layer.

[0072] The prediction deviation vector is input into a pre-defined quadratic programming solver. Minimizing the L2 norm of the dynamic compensation amount is used as a constraint, and the objective is to offset the prediction deviation vector, to solve for the dynamic compensation amount of key parameters in the downstream process. The mathematical description of the quadratic programming problem is:

[0073] in, For dynamic compensation vectors, This is the compensation coefficient matrix, used to convert dynamic compensation amounts into actual parameter adjustment effects. Compensation coefficient matrix This was obtained through systematic identification of historical production data, reflecting the degree of impact of actuator parameter adjustments on process parameters. In this embodiment, the compensation coefficient matrix... It is a diagonal matrix, and its diagonal elements are the compensation coefficients for the corresponding parameters.

[0074] The analytical solution to the above quadratic programming problem is:

[0075] When the compensation coefficient matrix When the matrix is ​​invertible, the analytical solution exists. In this embodiment, since the compensation coefficient matrix is ​​a diagonal matrix and all diagonal elements are non-zero, its inverse matrix exists. By solving the above quadratic programming problem, the obtained dynamic compensation amount can completely offset the prediction deviation vector under the condition of minimizing the L2 norm of the compensation amount, thus achieving accurate compensation for downstream process parameters.

[0076] In this embodiment, a quadratic programming solver is used to calculate the dynamic compensation amount, ensuring the smoothness of the compensation action. The L2 norm minimization constraint of the compensation amount avoids excessive compensation, preventing mechanical shock to the actuator due to excessive compensation and extending the service life of the equipment. Simultaneously, by aiming to offset the predicted deviation vector, it ensures that the downstream process parameters after compensation are as close as possible to the standard values, improving the stability of product quality.

[0077] refer to Figure 6 In one embodiment, physical performance test data of the final busbar trunking product is collected. This data includes conductor dimensional deformation tolerance data and insulation resistance test data. Conductor dimensional deformation tolerance data is obtained by measuring the conductor portion of the finished busbar trunking product using a coordinate measuring machine, including the deviation between the actual and standard values ​​of the conductor's length, width, height, and various bending angles. Insulation resistance test data is obtained by measuring the insulation layer of the finished busbar trunking product using an insulation resistance tester, reflecting the insulation performance of the insulation layer. The physical performance test data is collected in fixed batches, with a certain number of finished product samples collected for each batch. The average test data of the samples is calculated as the final test result for that batch.

[0078] The physical performance test data is compared with the preset standard performance indicators to calculate the performance deviation value of the finished product. The preset standard performance indicators are determined according to the national and enterprise standards for busbar trunking products, including the allowable range of conductor dimensional deformation tolerance and the minimum value of insulation resistance. The formula for calculating the performance deviation value of the finished product is:

[0079] in, This represents the performance deviation value of the finished product. In this embodiment, the number of physical performance test indicators is [number]. These correspond to the conductor dimensional deformation tolerance and insulation resistance, respectively. For the first The weight of each detection indicator, For the first The actual measured value of each detection indicator For the first Standard values ​​for each detection indicator. Weighting. The importance of each testing indicator to the product quality is set. In this embodiment, the weight of conductor size deformation tolerance is set to 0.6, and the weight of insulation resistance is set to 0.4.

[0080] A global loss gradient is constructed based on the performance deviation values ​​of the finished product. The adjacency matrix weights of the graph convolutional layers and the convolutional kernel parameters of the temporal convolutional layers are adaptively updated using the backpropagation algorithm. The global loss function is the mean squared error loss function, whose expression is:

[0081] During the training of a deep neural network, the gradient of the loss function with respect to each network parameter is calculated using the backpropagation algorithm, and then the network parameters are updated using the gradient descent algorithm. The formula for parameter update is:

[0082] in, These are network parameters, including the adjacency matrix weights of graph convolutional layers and the convolution kernel parameters of temporal convolutional network layers. The learning rate controls the step size for parameter updates. In this embodiment, the learning rate is set to 0.001, and the Adam optimizer is used for parameter updates. The Adam optimizer can adaptively adjust the learning rate, improving training stability and convergence speed.

[0083] In this embodiment, the closed-loop update mechanism based on finished product performance testing data enables the deep neural network to adapt to the slow time-varying characteristics of the production environment. As production time progresses, equipment wears out, raw material batches change, and the temperature and humidity of the production environment fluctuate; all these factors cause the characteristics of the production process to change slowly. By periodically collecting finished product performance testing data and updating the network parameters based on this data, the model can always maintain an accurate model of the production process, ensuring the effectiveness of parameter compensation.

[0084] refer to Figure 1 In one embodiment, outlier removal and smoothing filtering are performed on the acquired real-time multidimensional process parameter sequence to eliminate sensor impulse noise in the real-time multidimensional process parameter sequence. Outlier removal employs... The criterion, based on the statistical properties of the normal distribution, considers a data point as an outlier when it deviates from the mean by more than three standard deviations. The specific steps for outlier removal are as follows: First, calculate the average value of the process parameter sequence. and standard deviation Then, all sequences that satisfy the following conditions are considered. The data points are marked as outliers, and finally the average of the two normal data points before and after the data point is used to replace the outlier.

[0085] The smoothing filtering process uses moving average filtering, which effectively removes high-frequency noise while preserving the low-frequency trend of the signal. The formula for calculating moving average filtering is:

[0086] in, For the filtered first The parameter values ​​at each moment. For the first time before filtering The parameter values ​​at each moment. The size of the moving window is 5. In this embodiment, the size of the moving window is set to 5, meaning that each filtered parameter value is obtained by averaging the five data points: the two data points before and after it, and itself.

[0087] A normalization mapping operation is performed on the real-time multidimensional process parameter sequence after smoothing filtering to unify heterogeneous process parameters into the same numerical range, eliminating numerical differences between different dimensions. The normalized real-time multidimensional process parameter sequence is then identified as the feature vector of the topological nodes. The normalization mapping adopts the min-max normalization method, which linearly maps the data to the [0,1] interval. The calculation formula for min-max normalization is as follows:

[0088] in, For the normalized first The parameter values ​​at each moment. For the filtered first The parameter values ​​at each moment. This is the minimum value of the process parameter sequence. This represents the maximum value of the process parameter sequence.

[0089] In this embodiment, outlier removal and smoothing filtering effectively eliminate the interference of sensor impulse noise on the process parameter sequence. During data acquisition, the sensor may be affected by factors such as electromagnetic interference and mechanical vibration, generating abnormal pulse signals. If these abnormal signals are not processed, they will severely affect the subsequent feature extraction and parameter compensation effects. By eliminating outliers and smoothing the data using moving average filtering, a more stable and reliable sequence of process parameters can be obtained.

[0090] Normalization mapping can eliminate dimensional differences between different process parameters. Different process parameters have different dimensions and numerical ranges. For example, the numerical range of stamping speed is usually tens to hundreds of millimeters per second, while the numerical range of welding temperature is usually hundreds to thousands of degrees Celsius. If these heterogeneous parameters are directly input into a deep neural network, parameters with larger numerical ranges will dominate the model training, while the influence of parameters with smaller numerical ranges will be ignored. By mapping all parameters uniformly to the [0,1] interval through min-max normalization, the contribution of each parameter to the model can be more balanced, improving the model's training effect and prediction accuracy.

[0091] Table 4 Comparison of Process Parameters Before and After Pretreatment

[0092] Table 4 shows the data range and number of outliers for each process parameter before and after preprocessing. The table shows that a certain number of outliers existed in the original data. After outlier removal and smoothing filtering, the data range was reduced, and the outliers were effectively eliminated. After normalization mapping, the data range of all process parameters was unified to the [0,1] interval, eliminating dimensional differences and providing a good data foundation for subsequent feature extraction and model training.

[0093] In one embodiment, the training process of the deep neural network is divided into two stages: offline pre-training and online fine-tuning. The offline pre-training stage uses a large amount of historical production data to train the model, enabling it to learn the basic rules of the busbar production process. The historical production data includes process parameter data for each step and corresponding finished product performance test data. In the offline pre-training stage, the historical process parameter data is used as the model input, and the corresponding finished product performance deviation values ​​are used as the model output. The model parameters are trained through supervised learning. The number of iterations in the offline pre-training is set to 1000, and the batch size used in each iteration is 32.

[0094] The online fine-tuning phase takes place during actual production, using real-time collected production data to continuously fine-tune the model. The frequency of online fine-tuning is set according to the production batch. After each production batch, the performance test data of the finished product in that batch is collected, and the model parameters are updated based on this data. The learning rate for online fine-tuning is set to 1 / 10 of the offline pre-training learning rate, i.e., 0.0001, to avoid drastic changes in model parameters and ensure the stability of the production process.

[0095] In this embodiment, the training method combining offline pre-training and online fine-tuning enables the model to possess good performance before deployment and continuously adapt to changes in the production environment during use. Offline pre-training utilizes a large amount of historical data, allowing the model to learn the general patterns of the production process. Online fine-tuning utilizes real-time production data, enabling the model to learn the specific characteristics of the current production environment, improving the model's adaptability and accuracy.

[0096] In one embodiment, the controller employs a distributed control architecture, with each process equipped with an independent programmable logic controller (PLC). The PLCs communicate with each other via industrial Ethernet. A deep neural network runs on a central server, which collects real-time process parameter data uploaded by each PLC, performs feature extraction and dynamic compensation calculation, and then sends the calculated dynamic compensation values ​​to the corresponding PLCs. Upon receiving the dynamic compensation value, the PLC adds it to the current setpoint to generate new control parameters, which are then sent to the actuators.

[0097] In this embodiment, the distributed control architecture improves system reliability and response speed. Each PLC controls the actuator of one process; even if the central server fails, each PLC can still independently control according to preset static parameters, ensuring the continuity of the production process. Industrial Ethernet communication latency is typically in the millisecond range, meeting the requirements of real-time control. The central server handles complex computational tasks, fully utilizing its computing resources to improve the model's running speed and processing power.

[0098] In one embodiment, the actuators include a servo motor for a stamping machine, a hydraulic system for a bending machine, a power controller for a welding machine, and a plating solution circulation pump for the plating process. The servo motor for the stamping machine adjusts the movement speed and downward distance of the slide block according to the stamping speed and stamping depth parameters issued by the controller. The hydraulic system for the bending machine adjusts the pressure and movement speed of the hydraulic cylinder according to the bending angle and bending speed parameters issued by the controller. The power controller for the welding machine adjusts the output current and voltage of the welding power supply according to the welding current and welding temperature parameters issued by the controller. The plating solution circulation pump for the plating process adjusts the circulation flow rate and replenishment amount of the plating solution according to the plating solution concentration parameters issued by the controller.

[0099] In this embodiment, each actuator possesses high-precision control capabilities, enabling it to accurately execute parameter adjustment commands issued by the controller. The control precision of the servo motor can reach 0.1 mm, the angle control precision of the hydraulic system can reach 0.1 degrees, the current control precision of the welding power supply can reach 1 ampere, and the flow control precision of the plating solution circulation pump can reach 1 liter per minute. This high-precision actuator ensures that dynamic compensation is accurately converted into actual process parameter adjustments, achieving precise control of the production process.

[0100] The foregoing has shown and described the basic principles, main features, and advantages of the present invention. Those skilled in the art should understand that the present invention is not limited to the above embodiments. The embodiments and descriptions in the specification are merely illustrative of the principles of the invention. Various changes and modifications can be made to the invention without departing from its spirit and scope, and all such changes and modifications fall within the scope of the present invention as claimed. The scope of protection of this invention is defined by the appended claims and their equivalents.

Claims

1. An improved method based on busbar trunking production process parameters, characterized in that, include: Each process of the busbar trunking production line is defined as a topological node of a graph network, the physical medium flow path between each process is defined as an edge of the graph network, and the real-time multidimensional process parameter sequence corresponding to each process is defined as a feature vector of the topological node. The spatial coupling features between the various processes are extracted by the graph convolutional layer in the deep neural network based on the adjacency matrix of the graph network, and the dynamic drift features of the feature vector as it evolves over time are extracted by the temporal convolutional network layer in the deep neural network. The spatial coupling feature and the dynamic drift feature are fused and mapped to generate dynamic compensation amounts for key parameters of downstream processes corresponding to each process, and the dynamic compensation amounts are sent to the controller to correct the parameters of the actuator.

2. The improved method based on busbar trunking production process parameters according to claim 1, characterized in that, Each process of the busbar trunking production line is defined as a topology node of a graph network, the physical medium flow path between each process is defined as an edge of the graph network, and the real-time multidimensional process parameter sequence corresponding to each process is defined as a feature vector of the topology node. This includes mapping the stamping process, bending process, welding process and plating process in the busbar trunking production line to the first topology node, second topology node, third topology node and fourth topology node of the graph network, respectively. Based on the physical transfer trajectory of the busbar conductor material between the stamping process and the bending process, a first side between the first topology node and the second topology node is constructed, and so on, a second side between the second topology node and the third topology node and a third side between the third topology node and the fourth topology node are constructed; The stamping speed and stamping depth parameters of the stamping process, the bending angle and bending speed parameters of the bending process, the welding current and welding temperature parameters of the welding process, and the coating thickness and plating solution concentration parameters of the plating process are collected and used as the feature vectors of the first topological node, the second topological node, the third topological node, and the fourth topological node, respectively.

3. The improved method based on busbar trunking production process parameters according to claim 1, characterized in that, The spatial coupling features between the various processes are extracted by the graph convolutional layer in the deep neural network based on the adjacency matrix of the graph network, including: constructing the adjacency matrix of the graph network, wherein the element values ​​in the adjacency matrix represent the spatial influence weights of process parameters between adjacent topological nodes; The feature vector is convolved with the adjacency matrix to obtain cross-process spatial interaction information of process parameters between adjacent topological nodes; The cross-process spatial interaction information is nonlinearly activated and mapped to capture the spatial transmission path and attenuation coefficient of the parameter deviation of the upstream topological node corresponding to the process to the downstream topological node corresponding to the process, thereby generating the spatial coupling feature.

4. The improved method based on busbar trunking production process parameters according to claim 1, characterized in that, Extracting the dynamic drift features of the feature vector over time by using the temporal convolutional network layer in the deep neural network includes: performing a sliding convolution operation on the feature vector of each topological node along the time dimension to extract the local temporal dependencies of the feature vector within a preset time window; The receptive field of the local temporal dependency is expanded by the dilated convolution kernel in the temporal convolutional network layer, thereby capturing the dynamic drift trend of the feature vector over a long period of time. The local temporal dependencies and the dynamic drift trend are fused by residual connection to filter out high-frequency process noise interference in the feature vector and output the dynamic drift features.

5. The improved method based on busbar trunking production process parameters according to claim 1, characterized in that, The spatial coupling feature and the dynamic drift feature are fused and mapped to generate dynamic compensation amounts for key parameters of downstream processes corresponding to each process. The dynamic compensation amounts are then sent to the controller to correct the parameters of the actuator. This includes: performing feature concatenation operations on the spatial coupling feature and the dynamic drift feature in the fully connected layer of the deep neural network to obtain a joint feature representation. The joint feature representation is input into the feedforward output layer of the deep neural network, and the dynamic compensation amount of the key parameters of the downstream process is calculated with the objective function of eliminating potential cumulative deviations in the downstream process. The dynamic compensation amount is converted into a parameter correction instruction that the controller can recognize, and the parameter correction instruction is sent to the controller. The controller adjusts the operating parameters of the actuator according to the parameter correction instruction.

6. The improved method based on busbar trunking production process parameters according to claim 3, characterized in that, Constructing the adjacency matrix of the graph network, wherein the element values ​​in the adjacency matrix represent the spatial influence weights of process parameters between adjacent topological nodes, includes: initializing the element values ​​in the adjacency matrix as static basic weights determined based on physical mechanisms; During the training and inference process of the deep neural network, an attention mechanism is introduced to dynamically modulate the static basic weights and calculate the real-time attention score of the feature vector of the upstream topological node to the feature vector of the downstream topological node. The static basic weights are weighted and fused with the real-time attention score to generate dynamic spatial influence weights, and the dynamic spatial influence weights are updated in the adjacency matrix.

7. An improved method based on busbar trunking production process parameters according to claim 4, characterized in that, The receptive field of the local temporal dependency is expanded by the dilated convolution kernel in the temporal convolutional network layer to capture the dynamic drift trend of the feature vector over a long period of time. This includes setting multiple dilated convolution kernels with different dilation rates, performing multi-scale temporal convolution on the feature vector respectively, and obtaining initial temporal features at multiple different time scales. The initial time-series features at multiple different time scales are concatenated to construct a multi-scale time-series feature matrix. A feature selection operation is performed on the multi-scale time-series feature matrix to filter out the target-scale time-series feature that has the highest matching degree with the current process drift mode of the busbar trunking, and the target-scale time-series feature is confirmed as the dynamic drift trend.

8. An improved method based on busbar trunking production process parameters according to claim 5, characterized in that, The joint feature representation is input into the feedforward output layer of the deep neural network. The dynamic compensation amount of the key parameters of the downstream process is calculated with the objective function of eliminating potential cumulative deviation in the downstream process. This includes: predicting the prediction deviation vector of the downstream process under the action of the current upstream process parameters in the feedforward output layer based on the joint feature representation. The predicted deviation vector is input into a preset quadratic programming solver. With the minimization of the L2 norm of the dynamic compensation amount as a constraint and the goal of offsetting the predicted deviation vector, the dynamic compensation amount of the key parameters of the downstream process is solved.

9. An improved method based on busbar trunking production process parameters according to claim 1, characterized in that, After the dynamic compensation amount is sent to the controller to correct the parameters of the actuator, the method further includes: collecting physical performance test data of the final product of the busbar trunking, wherein the physical performance test data includes conductor size deformation tolerance data and insulation resistance test data; The physical performance test data is compared with the preset standard performance indicators to calculate the performance deviation value of the finished product. Based on the performance deviation value of the finished product, a global loss gradient is constructed, and the adjacency matrix weights of the graph convolutional layer of the deep neural network and the convolution kernel parameters of the temporal convolutional network layer are adaptively updated through the backpropagation algorithm.

10. An improved method based on busbar trunking production process parameters according to claim 1, characterized in that, Before defining the real-time multidimensional process parameter sequence corresponding to each process as the feature vector of the topology node, the method further includes: performing outlier removal and smoothing filtering on the collected real-time multidimensional process parameter sequence to eliminate sensor impulse noise in the real-time multidimensional process parameter sequence. The real-time multidimensional process parameter sequence after smoothing filtering is normalized and mapped to the same numerical range to eliminate numerical differences between different dimensions. The normalized and mapped real-time multidimensional process parameter sequence is then identified as the feature vector of the topological node.

Images