Aluminum veneer punching path optimization method based on artificial intelligence

Abstract

The invention discloses an aluminum veneer punching path optimization method based on artificial intelligence, and the method comprises the following steps: collecting hole site distribution information in an aluminum veneer design drawing, and generating standardized hole group data; constructing a hole group topological relation graph, and generating a hole group topological entropy feature vector; constructing a path state vector, and generating a plurality of groups of candidate punching path sequences through an OfficientZero reinforcement learning model; performing residual stress field prediction on the candidate punching path sequence to generate a corresponding stress stability index set; a comprehensive cost function is constructed, and iterative training is carried out on the OfficientZero model; and selecting the path sequence with the minimum comprehensive cost function value as an aluminum veneer punching path optimization scheme. According to the method, EfficitZero reinforcement learning and stress field prediction are adopted, intelligent optimization of the punching path of the aluminum veneer is achieved, and the method has the advantages of being high in efficiency, low in energy consumption and good in stability.

Description

TECHNICAL FIELD

[0001] The application relates to the field of numerical control processing of metal plates, and in particular to an aluminum veneer punching path optimization method based on artificial intelligence. BACKGROUND

[0002] Existing aluminum veneer punching path optimization methods mostly take the shortest geometric distance or the least processing time as the main target, and generate the punching sequence through simple heuristic algorithms or traversal methods. Such methods can improve the numerical control processing efficiency to a certain extent, but they often fail to effectively control the accumulation of residual stress due to the neglect of the stress changes of the plate during the punching process, resulting in quality problems such as warping and cracking of the aluminum veneer after processing, which affects the precision and stability of the finished product.

[0003] In recent years, although some scholars have attempted to introduce optimization algorithms into punching path design, most of them are limited to single-objective optimization and fail to comprehensively consider the path length, processing time, energy consumption, and residual stress distribution. In addition, existing methods lack the measurement and modeling of the geometric characteristics of complex hole group distribution, making it difficult to cope with high complexity and multiple hole positions in actual working conditions. Therefore, how to use artificial intelligence methods to build a multi-objective optimization framework to effectively control the residual stress distribution while ensuring the punching efficiency has become a technical problem to be solved. SUMMARY

[0004] One object of the application is to provide an aluminum veneer punching path optimization method based on artificial intelligence. The application uses EfficientZero reinforcement learning and stress field prediction to realize intelligent optimization of aluminum veneer punching path, and has the advantages of high efficiency, low energy consumption and good stability.

[0005] According to an aluminum veneer punching path optimization method based on artificial intelligence, the method comprises the following steps: Collect the hole position distribution information in the aluminum veneer design drawing, establish a hole position coordinate set, standardize the hole position coordinate set, obtain an effective hole position coordinate set, and generate standardized hole group data; Based on the standardized hole group data, a hole group topological relationship graph is constructed, the topological complexity index is calculated, the geometric complexity of the hole group distribution is measured, and a hole group topological entropy feature vector is generated; Define the path state and the candidate action set, construct the path state vector, input the path state vector and the topological entropy feature vector into the EfficientZero reinforcement learning model, and generate multiple groups of candidate punching path sequences in combination with the effective hole position coordinate set; Perform residual stress field prediction on each group of candidate punching path sequences, obtain the stress distribution in the path execution process, and generate a corresponding stress stability index set; The path length, processing time, energy consumption and stress stability index are taken as multi-objective constraint conditions, a comprehensive cost function is constructed, an EfficientZero reinforcement learning model is iteratively trained, a candidate punching path sequence is optimized, and a multi-objective punching path optimization result is generated. In the multi-objective punching path optimization result, the path sequence with the minimum comprehensive cost function value is selected as the aluminum plate punching path optimization scheme, and is output for calling and executing by a numerical control machining device.

[0006] Optionally, the generation of the standardized hole group data specifically includes: Data reading is performed on the aluminum plate design drawing, hole position distribution information in the design drawing is extracted, the transverse coordinates and longitudinal coordinates of each hole position are arranged into a hole position coordinate point set, and an initial hole position coordinate set is formed; The initial hole position coordinate set is normalized to generate a normalized hole position coordinate set; The normalized hole position coordinate set is denoised, and a spatial filtering method is used to remove duplicate hole points and abnormal hole points to obtain an effective hole position coordinate set; The hole group geometric center coordinates are calculated based on the effective hole position coordinate set, the geometric center transverse coordinate is calculated by dividing the sum of the transverse coordinates of all effective hole points by the number of effective hole points, the geometric center longitudinal coordinate is calculated by dividing the sum of the longitudinal coordinates of all effective hole points by the number of effective hole points, and thus the hole group geometric center coordinates are obtained; The relative position vector of each effective hole point with respect to the hole group geometric center coordinates is calculated based on the hole group geometric center coordinates to form a relative position vector set; The relative position vector set is taken as the standardized hole group data.

[0007] Optionally, the generation of the hole group topology entropy feature vector includes: The obtained standardized hole group data is called, each hole point is defined as a node in a graph structure, a connection rule of edges is established according to the spatial proximity relationship between nodes, the Euclidean distance is calculated based on the relative position vector, and when the Euclidean distance between any two hole points is less than a preset threshold, a non-directed edge is established between the corresponding nodes, thereby forming a hole group topology relationship graph including a node set and an edge set; In the hole group topology relationship graph, the node degree of each node is calculated, and all node degrees are sequentially arranged to form a node degree distribution set; The degree probability of each node is calculated based on the node degree distribution set; After obtaining the degree probability of all nodes, the topology entropy value of the hole group is calculated; The topological entropy value is taken as a quantitative index of the overall geometric complexity of the hole group, and the topological entropy value is combined with all node degree probabilities to form a topological entropy feature vector.

[0008] Optionally, the generation of the plurality of candidate punch path sequences comprises: The topological entropy feature vector and the effective hole coordinate set are input into an EfficientZero reinforcement learning model, and the EfficientZero reinforcement learning model comprises a representation network, a dynamic network, a prediction network and a Monte Carlo tree search module. A path state is defined, and a path state vector is generated; The input data and the path state vector are input into the representation network to generate a latent representation vector through encoding by the representation network, and the representation network comprises an embedding layer, a multi-layer fully connected network and a normalization layer. A candidate action set is constructed; The current punch position is taken as a starting point, a horizontal coordinate difference and a vertical coordinate difference between the candidate target hole and the effective hole coordinate set are calculated, a two-dimensional vector formed by the horizontal coordinate difference and the vertical coordinate difference is taken as basic displacement information, the Euclidean distance between the current punch position and the candidate target hole is calculated, the node degree probability corresponding to the candidate target hole is extracted from the topological entropy feature vector, and the basic displacement information, the Euclidean distance and the node degree probability are spliced to form a candidate action representation vector. The latent representation vector and the candidate action representation vector are input into the dynamic network, and the latent representation vector is updated through the dynamic network to obtain a new latent representation vector and immediate feedback information. The dynamic network is implemented by using a multi-layer perception network. The new latent representation vector is input into the prediction network, and a path strategy distribution and a path state evaluation are output by the prediction network, the path strategy distribution represents the probability of selection of each candidate action in the candidate action set, and the path state evaluation represents the overall advantages and disadvantages of the current path state in the overall punch task. The prediction network comprises a multi-layer perception network, and after receiving the new latent representation vector, the prediction network is divided into a policy branch and a value branch, the policy branch is used to output the selection probability of each candidate action in the candidate action set through a fully connected layer and a Softmax function, and the value branch is used to output the overall advantages and disadvantages of the current path state in the overall punch task through a fully connected layer and a nonlinear activation function. During the generation of the candidate punch path, the candidate action set is expanded and simulated by using the Monte Carlo tree search module, the path strategy distribution and the path state evaluation are taken as expansion bases of the Monte Carlo tree search module, and the output results of the representation network, the dynamic network and the prediction network are repeatedly expanded to search to generate a plurality of candidate punch path sequences. During the generation of candidate punching path sequences, the path length information of each candidate punching path sequence is calculated, and the path length information is stored in correspondence with the candidate punching path sequence to construct a path sequence set.

[0009] Optionally, the generation of the corresponding set of stress stability indices specifically includes: Use the candidate punching path sequence as input data; The aluminum single-panel material is discretized and modeled on the finite element mesh. The hole points in the effective hole position coordinate set are mapped to the node positions of the finite element model. Loads and boundary conditions are applied to the corresponding nodes in sequence according to the order of the candidate punching path sequence to obtain the mesh stress state under the execution of the path. The residual stress value is calculated in the finite element mesh. The residual stress value consists of three parts: the transverse residual stress component, the longitudinal residual stress component, and the thickness direction residual stress component. The three parts are added together to obtain the residual stress value of the node. During the execution of the candidate punching path sequence, the stress increment is calculated for each hole location; As the candidate punching path sequence gradually extends, the stress increments at all hole locations are accumulated sequentially to obtain the stress accumulation effect. The stress stability index is calculated based on the stress accumulation effect. The stress stability index is defined as the sum of the absolute values ​​of stress increments in all punching steps divided by one and added to it. The stress stability indices corresponding to each candidate punching path sequence are organized to generate a set of stress stability indices.

[0010] Optionally, the generation of the multi-objective punching path optimization result specifically includes: The path sequence set is fused with the stress stability index set; Calculate processing time; Calculate energy consumption; Construct a comprehensive cost function, which is path length information multiplied by path length weight, plus processing time multiplied by time weight, plus energy consumption multiplied by energy consumption weight, plus stress instability multiplied by stress weight, where stress instability is equal to one minus stress stability index. During training, the EfficientZero reinforcement learning model takes the candidate punching path sequence as input, uses the representation network to encode the path state and topological entropy feature vector to obtain the latent representation vector, uses the dynamic network to predict the new latent representation vector and real-time feedback information under the action of the latent representation vector and candidate actions, uses the prediction network to output the path policy distribution and path state evaluation, and uses the Monte Carlo tree search module to expand and simulate the candidate action set to generate new candidate punching path sequences. In each round of training, the calculation result of the comprehensive cost function is used as the evaluation signal, and the candidate punching path sequence is stored in the experience replay pool along with the corresponding path state vector, path length information, processing time, energy consumption, stress stability index and real-time feedback information of the dynamic network output, forming a set of data samples for training. Data samples are repeatedly sampled from the experience replay pool, and the parameters of the representation network, dynamic network and prediction network are continuously updated by the stochastic gradient descent method, so that the EfficientZero reinforcement learning model gradually converges. After multiple rounds of iterative training, the candidate punching path sequence output by the EfficientZero reinforcement learning model gradually tends to the direction where the comprehensive cost function is minimized, thus obtaining the multi-objective punching path optimization result.

[0011] Optionally, the generation of the aluminum panel punching path optimization scheme specifically includes: In the multi-objective punching path optimization results, the comprehensive cost function value of each candidate punching path sequence is extracted; A comparison is performed among the comprehensive cost function values ​​of all candidate punching path sequences to obtain the path sequence with the smallest comprehensive cost function value; The optimal path sequence is used as the optimization scheme for the punching path of aluminum single-panel.

[0012] The beneficial effects of this invention are: This invention optimizes the punching path of complex hole clusters in aluminum panels by introducing an artificial intelligence reinforcement learning model, effectively overcoming the limitations of traditional methods that only focus on the geometric shortest path or minimum processing time. By utilizing a hole cluster topology graph constructed based on standardized hole cluster data and calculating the topological entropy feature vector, a quantitative measure of the geometric complexity of the hole cluster is achieved. This allows punching path generation to no longer be limited to local geometric information but to comprehensively reflect the overall distribution characteristics of the hole cluster. Furthermore, the path state vector and the topological entropy feature vector are input into the EfficientZero reinforcement learning model, which, combined with the effective hole position coordinate set, generates multiple sets of candidate punching path sequences. This endows the path optimization process with learning and adaptive capabilities, enabling the automatic generation of reasonable candidate paths under different working conditions.

[0013] This invention further introduces residual stress field prediction based on the candidate punching path sequence. Through finite element discretization modeling and loading solution, the stress distribution under different path execution conditions is obtained, and the stress stability index is calculated, thus fully considering the cumulative effect of residual stress during path selection. In this way, this invention can establish multi-objective optimization constraints among path length, processing time, energy consumption, and stress stability index, construct a comprehensive cost function, and input it as an evaluation signal into the iterative training process of the EfficientZero reinforcement learning model, enabling the model to gradually converge to the optimal path selection result under multi-objective conditions. Ultimately, the method not only outputs the punching path sequence with the minimum comprehensive cost function as the execution scheme for CNC machining equipment, but also reduces warping and deformation caused by residual stress during processing, improving the structural stability and processing accuracy of the finished aluminum panel. Attached Figure Description

[0014] The accompanying drawings are provided to further illustrate the invention and form part of the specification. They are used in conjunction with embodiments of the invention to explain the invention and do not constitute a limitation thereof. In the drawings: Fig. 1 This is a flowchart of an artificial intelligence-based method for optimizing the punching path of aluminum single-panel proposed in this invention. Fig. 2 This is a schematic diagram of the structure of the EfficientZero reinforcement learning model input and candidate punching path sequence generation for an artificial intelligence-based aluminum single-panel punching path optimization method proposed in this invention. Detailed Implementation

[0015] The present invention will now be described in further detail with reference to the accompanying drawings. These drawings are simplified schematic diagrams, illustrating only the basic structure of the invention, and therefore only show the components relevant to the invention.

[0016] refer to Figs. 1-2 An artificial intelligence-based method for optimizing the punching path of aluminum single panels includes the following steps: Collect hole distribution information from aluminum single-panel design drawings, establish a set of hole coordinates, perform standardization preprocessing on the set of hole coordinates to obtain an effective set of hole coordinates, and generate standardized hole group data; A topological relationship graph of a pore group is constructed based on standardized pore group data, a topological complexity index is calculated, the geometric complexity of the pore group distribution is measured, and a topological entropy feature vector of the pore group is generated. Define the path state and candidate action set, construct the path state vector, input the path state vector and the topological entropy feature vector into the EfficientZero reinforcement learning model, and generate multiple sets of candidate punching path sequences by combining the effective hole position coordinate set. For each group of candidate punching path sequences, residual stress field prediction is performed to obtain the stress distribution during the path execution process and generate the corresponding set of stress stability indices. The path length, processing time, energy consumption, and stress stability index are used as multi-objective constraints to construct a comprehensive cost function. The EfficientZero reinforcement learning model is used for iterative training to optimize the candidate punching path sequence and generate multi-objective punching path optimization results. The path sequence with the smallest comprehensive cost function value is selected from the multi-objective punching path optimization results and used as the aluminum single-panel punching path optimization scheme, which is then output for CNC machining equipment to call and execute.

[0017] In this embodiment, the generation of the standardized pore group data specifically includes: Data is read from the aluminum single-panel design drawings, the hole distribution information in the design drawings is extracted, and the horizontal and vertical coordinates of each hole are organized into a set of hole coordinate points to form an initial set of hole coordinates. The initial set of borehole coordinates is normalized by subtracting the minimum value of the horizontal coordinates of all boreholes from the horizontal coordinate of each borehole point, and then dividing by the difference between the maximum and minimum values ​​of the horizontal coordinates of all borehole points to obtain the normalized horizontal coordinate. Similarly, the minimum value of the vertical coordinates of all boreholes is subtracted from the vertical coordinates of each borehole point, and then divided by the difference between the maximum and minimum values ​​of the vertical coordinates of all borehole points to obtain the normalized vertical coordinate. This generates a set of normalized borehole coordinates. The normalized set of borehole coordinates is denoised, and a spatial filtering method is used to remove duplicate and abnormal borehole locations to obtain an effective set of borehole coordinates. The number of boreholes in the effective set of borehole coordinates is less than or equal to the total number of boreholes. Based on the set of effective borehole coordinates, the geometric center coordinates of the borehole group are calculated. The horizontal coordinate of the geometric center is calculated by summing the horizontal coordinates of all effective boreholes and dividing by the number of effective boreholes. The vertical coordinate of the geometric center is calculated by summing the vertical coordinates of all effective boreholes and dividing by the number of effective boreholes. Thus, the geometric center coordinates of the borehole group are obtained. Using the coordinates of the geometric center of the hole group as a reference, calculate the relative position vector of each effective hole point relative to the coordinates of the geometric center of the hole group. The horizontal component of the relative position vector is the horizontal coordinate of the hole point minus the horizontal coordinate of the geometric center of the hole group, and the vertical component of the relative position vector is the vertical coordinate of the hole point minus the vertical coordinate of the geometric center of the hole group, thus forming a set of relative position vectors. The set of relative position vectors is used as the standardized pore group data.

[0018] In this embodiment, the generation of the pore group topological entropy feature vector includes: The obtained standardized hole group data is called, and each hole point is defined as a node in the graph structure. The connection rules of the edges are established according to the spatial proximity relationship between the nodes. The Euclidean distance is calculated based on the relative position vector. When the Euclidean distance between any two hole points is less than a preset threshold, an undirected edge is established between the corresponding nodes, thus forming a hole group topology graph containing a set of nodes and a set of edges. The Euclidean distance between the two aperture points is equal to the square of the horizontal component of the relative position vector of the first aperture point minus the square of the horizontal component of the relative position vector of the second aperture point, plus the square of the vertical component of the relative position vector of the first aperture point minus the square of the vertical component of the relative position vector of the second aperture point, and then the square root is taken. The preset threshold is an important parameter for determining whether two hole points are connected. It is obtained by: statistically analyzing the Euclidean distances between all hole points in the aluminum panel design drawings, and then calculating the average Euclidean distance between each pair of hole points. The average value is then used as a basic reference, and an adjustment coefficient is set in combination with the thickness of the aluminum panel, the hole size, and the precision requirements of the processing equipment. The preset threshold is determined by multiplying the average Euclidean distance by the adjustment coefficient. The adjustment coefficient is usually between 0.5 and 1, which is used to achieve a balance between overly dense and overly sparse connections. The final preset threshold can ensure that the hole group topology diagram can accurately reflect the spatial proximity of the hole points, and avoid the situation where the graph structure is too sparse due to the threshold being too small or too dense due to the threshold being too large, thus providing a reasonable data basis for subsequent topological entropy calculation. In the topological graph of the hole group, the node degree is calculated for each node. The node degree is defined as the number of edges connecting the node to the remaining nodes. All node degrees are arranged in order to form a node degree distribution set. The degree probability of each node is calculated based on the set of node degree distributions. Specifically, the node degree value of a certain node is taken as the numerator, the sum of the degree values ​​of all nodes is taken as the denominator, and the node degree value of that node is divided by the sum of the degree values ​​of all nodes to obtain the degree probability of that node. After obtaining the degree probabilities of all nodes, the topological entropy of the hole group is calculated. Specifically, the degree probability of each node is multiplied by the logarithm of the degree probability, the negative value of the multiplication result is taken, and then all nodes are added together. The sum is the topological entropy of the hole group. The topological entropy value is used as a quantitative indicator of the overall geometric complexity of the hole group. The topological entropy value is combined with the degree probabilities of all nodes to form a topological entropy feature vector, which consists of a topological entropy value and the degree probabilities of each node arranged in sequence.

[0019] In this embodiment, the generation of the multiple sets of candidate punching path sequences includes: The topological entropy feature vector and the set of effective aperture coordinates are used as input data and input into the EfficientZero reinforcement learning model, which includes a representation network, a dynamic network, a prediction network and a Monte Carlo tree search module. This invention improves upon the EfficientZero reinforcement learning model, moving beyond the traditional model which only uses environmental state and immediate reward as input. Instead, it introduces a set of effective hole position coordinates, a hole group topological entropy feature vector, and a path state vector as joint input, enabling the model to comprehensively represent the geometric distribution and complexity of the hole group. Simultaneously, it unifies path length, processing time, energy consumption, and stress stability indices into a comprehensive cost function, storing it as a supervisory signal in the experience replay pool to guide the model in balancing efficiency and stress control during training. During the Monte Carlo tree search expansion process, this invention maintains the mechanism of immediate feedback from the dynamic network output, while introducing residual stress prediction results as a supervisory signal, comparing and correcting them with the immediate feedback prediction values ​​of the dynamic network. This allows the model to balance path efficiency and stress stability during learning, achieving multi-objective optimization of candidate punching paths. Through these improvements, the model output of this invention not only ensures path optimization in both geometric and temporal dimensions but also effectively reduces residual stress and warping risk in the sheet metal, significantly improving the intelligence level and processing quality of aluminum single-panel punching path planning. Define the path state and generate a path state vector. The path state is a triplet consisting of three parts: the sequence of completed punching positions, the set of incomplete punching positions, and the current punching position. It is dynamically updated as the path is generated and is used to characterize the dynamic progress of the punching task during the path generation process. It can reflect the overall information of the executed part and the part to be executed. The path state vector consists of three parts: the sequence of completed punching positions, the set of incomplete punching positions, and the current punching position. The sequence of completed punching positions is represented by a sequence of hole point numbers, which is normalized and arranged according to the actual execution order. The set of incomplete punching positions is represented by the set of hole point numbers that have not yet been completed, and is also normalized. The current punching position is represented by the hole point number of the most recently completed punching operation, and is also normalized. The three parts are concatenated to form the path state vector, which is dynamically updated as the punching sequence progresses during the path generation process. The input data and path state vector are input into the representation network, where they are encoded to generate a latent representation vector. Specifically, this includes: performing an embedding transformation on the path state vector input into the embedding layer to encode it into a path state feature representation; performing a fully connected mapping on the set of effective aperture coordinates to extract the geometric distribution feature representation of the apertures on the two-dimensional plane; simultaneously performing normalization and fully connected mapping on the topological entropy feature vector to generate a topological entropy feature representation; concatenating and fusing the path state feature representation, geometric distribution feature representation, and topological entropy feature representation, and inputting them into a multi-layer nonlinear mapping network to output a latent representation vector. The latent representation vector is used to express the geometric distribution information and aperture group complexity information of the path state. The representation network includes an embedding layer, a multi-layer fully connected network, and a normalization layer. Constructing a candidate action set specifically includes: determining all uncompleted punching points based on the effective punching point coordinate set, recording the current punching position by the path state, taking the current punching position as the starting point, and treating the remaining punching points one by one as potential targets. Each potential target corresponds to a candidate action, which is defined as moving from the current punching position to the target punching position and performing the punching operation. All action combinations formed by the current punching position and the remaining punching points are organized into a candidate action set. The candidate action set completely covers all possible punching path extension directions under the current path state and is used as part of the dynamic network input to simulate the update and extension of the path state when generating the candidate punching path sequence. Starting from the current punching position, calculate the difference between the horizontal and vertical coordinates of the current punching position and the candidate target hole position in the effective hole position coordinate set. Use the two-dimensional vector formed by the difference between the horizontal and vertical coordinates as the basic displacement information. At the same time, calculate the Euclidean distance between the current punching position and the candidate target hole position. Extract the node degree probability of the corresponding candidate target hole position from the topological entropy feature vector. Concatenate the basic displacement information, Euclidean distance and node degree probability to form the candidate action representation vector. The latent representation vector and candidate action representation vector are input into a dynamic network. The dynamic network updates the latent representation vector to obtain a new latent representation vector and real-time feedback information. The dynamic network is implemented using a multilayer perceptron network. The latent representation vector and the candidate action representation vector are concatenated and then input into a multilayer fully connected layer. Through nonlinear activation functions, the network performs layer-by-layer mapping and feature extraction, and outputs a new latent representation vector and real-time feedback information. The new latent representation vector is used to describe the state changes after the path is extended, and the real-time feedback information is used to reflect the real-time performance after the path is extended. The new latent representation vector is input into the prediction network, and the prediction network outputs the path policy distribution and the path state evaluation. The path policy distribution represents the probability of each candidate action being selected in the candidate action set, and the path state evaluation represents the overall quality of the current path state in the overall punching task. The prediction network includes a multilayer perceptron, which is divided into a policy branch and a value branch after receiving a new latent representation vector. The policy branch is used to output the selection probability of each candidate action in the candidate action set through a fully connected layer and a Softmax function. The value branch is used to output the comprehensive advantages and disadvantages of the current path state in the overall punching task through a fully connected layer and a nonlinear activation function. During the generation of candidate punching paths, the Monte Carlo tree search module is used to expand and simulate the set of candidate actions. The distribution of path strategies and the evaluation of path states are used as the basis for the expansion of the Monte Carlo tree search module. The output results of the representation network, dynamic network and prediction network are used to repeatedly expand the search and generate multiple candidate punching path sequences. Each candidate punching path sequence consists of the starting hole position, the path extension sequence and the ending hole position, and the path generation process is completely recorded. The generation of multiple candidate punching path sequences specifically includes: inputting the latent representation vector generated by the representation network, the latent representation vector updated by the dynamic network, and the path strategy distribution and path state evaluation output by the prediction network into the Monte Carlo tree search module. The Monte Carlo tree search module takes the path state as the root node at the initial moment, performs probability-weighted selection on each candidate action in the candidate action set according to the path strategy distribution, and combines the path state evaluation as the evaluation basis for node expansion to determine the expansion priority. During the expansion process, the Monte Carlo tree search module uses the representation network and the dynamic network to continuously simulate path extension, generate new latent representation vectors, and calls the prediction network to output new path strategy distribution and path state evaluation after each expansion to update the node weights in the search tree. Through this continuous expansion, simulation, and backtracking process, the Monte Carlo tree search module can repeatedly search on multiple possible paths in the candidate action set, gradually splicing together to form a complete candidate punching path sequence. Each candidate punching path sequence consists of the starting hole position, the path extension order, and the ending hole position, and completely records the execution process of path generation. During the candidate punching path generation process, the Monte Carlo tree search module receives the path strategy distribution and path state evaluation output by the prediction network. It then calculates the initial selection weight for each candidate action in the candidate action set based on its probability value in the path strategy distribution. Subsequently, it weights and combines the path state evaluation with the initial selection weight corresponding to the candidate action to form a comprehensive evaluation value. This comprehensive evaluation value reflects both the probability of the candidate action in the path strategy distribution and the overall quality of the current path state. The Monte Carlo tree search module sorts the candidate actions in the candidate action set according to the comprehensive evaluation value. Candidate actions with higher comprehensive evaluation values ​​have a higher expansion priority. In the actual expansion process, the Monte Carlo tree search module prioritizes candidate actions with higher comprehensive evaluation values ​​for node expansion, while retaining a small number of candidate actions with lower comprehensive evaluation values ​​for exploration. This ensures the quality of the candidate punching path sequence while maintaining the diversity of path generation.

[0020] During the generation of candidate punching path sequences, the path length information of each candidate punching path sequence is calculated, and the path length information is stored in correspondence with the candidate punching path sequence to construct a path sequence set.

[0021] In this embodiment, the generation of the corresponding set of stress stability indices specifically includes: Use the candidate punching path sequence as input data; The aluminum single-panel material is discretized and modeled on the finite element mesh. The hole points in the effective hole position coordinate set are mapped to the node positions of the finite element model. Loads and boundary conditions are applied to the corresponding nodes in sequence according to the order of the candidate punching path sequence to obtain the mesh stress state under the execution of the path. When discretizing the aluminum single-panel sheet on the finite element mesh, the geometric region of the aluminum single-panel sheet is divided into a finite element mesh structure composed of multiple elements and nodes. The spatial coordinates of all nodes are marked in the mesh. Each hole point in the effective hole point coordinate set is mapped to the nearest node position in the finite element mesh according to the positional relationship between its horizontal and vertical coordinates, ensuring that the geometric correspondence of the hole point in the finite element model is accurate. Next, the nodes corresponding to the hole points are selected sequentially according to the order of the candidate punching path sequence. Punching load is applied at the node position, and boundary conditions are set at the edge of the finite element model to constrain the overall equilibrium. After each load is applied, the finite element solution is performed to obtain the residual stress value of the current node and the mesh stress state. The mesh stress state is then used as the initial condition for loading the next hole point until the entire loading process of the candidate punching path sequence is completed, thus obtaining the mesh stress state evolution result under the execution of the candidate punching path sequence. The load is the punching force applied to the candidate hole node to simulate the external force generated during the punching process. The boundary condition is the geometric constraint applied to the edge node of the aluminum single panel to maintain the mechanical balance of the overall structure. The residual stress value is calculated in the finite element mesh. The residual stress value consists of three parts: the transverse residual stress component, the longitudinal residual stress component, and the thickness direction residual stress component. The three parts are added together to obtain the residual stress value of the node. During the execution of the candidate punching path sequence, the stress increment is calculated for each hole position. The stress increment is defined as the residual stress value after the punching operation of the current hole position minus the residual stress value after the punching operation of the previous hole position. As the candidate punching path sequence gradually extends, the stress increments at all hole locations are accumulated sequentially to obtain the stress accumulation effect. The stress accumulation effect is used to reflect the overall accumulation of residual stress during the execution of the candidate punching path sequence. The stress stability index is calculated based on the stress accumulation effect. The stress stability index is defined as the sum of the absolute values ​​of the stress increments of all punching steps divided by one. The value of the stress stability index is between zero and one. The closer the value is to one, the more stable the residual stress distribution is. The stress stability indices corresponding to each candidate punching path sequence are organized to generate a set of stress stability indices.

[0022] In this embodiment, the generation of the multi-objective punching path optimization result specifically includes: The path sequence set is fused with the stress stability index set, wherein the path sequence set includes multiple candidate punching path sequences and corresponding path length information; The processing time is calculated as the path length information divided by the processing speed parameter. The path length information is accumulated sequentially from the effective hole position coordinates in the candidate punching path sequence. The processing speed parameter is determined by the feed speed of the CNC machining equipment. Calculate energy consumption, which is defined as path length information multiplied by energy consumption parameter per unit length. The energy consumption parameter per unit length is determined by the rated power and energy efficiency ratio of the CNC machining equipment. A comprehensive cost function is constructed, which is path length information multiplied by path length weight, plus processing time multiplied by time weight, plus energy consumption multiplied by energy consumption weight, plus stress instability multiplied by stress weight. The stress instability is equal to one minus the stress stability index. The path length weight, time weight, energy consumption weight, and stress weight are used to balance the relative importance of path length, processing time, energy consumption, and stress stability index in multi-objective optimization. During training, the EfficientZero reinforcement learning model takes the candidate punching path sequence as input, uses the representation network to encode the path state and topological entropy feature vector to obtain the latent representation vector, uses the dynamic network to predict the new latent representation vector and real-time feedback information under the action of the latent representation vector and candidate actions, uses the prediction network to output the path policy distribution and path state evaluation, and uses the Monte Carlo tree search module to expand and simulate the candidate action set to generate new candidate punching path sequences. In each round of training, the calculation result of the comprehensive cost function is used as the evaluation signal, and the candidate punching path sequence is stored in the experience replay pool along with the corresponding path state vector, path length information, processing time, energy consumption, stress stability index and real-time feedback information of the dynamic network output, forming a data sample set for training. Data samples are repeatedly sampled from the experience replay pool, and the parameters of the representation network, dynamic network and prediction network are continuously updated by the stochastic gradient descent method, so that the EfficientZero reinforcement learning model gradually converges. The data samples are used as input states and supervision signals during iterative training to drive the parameter updates of the representation network, dynamic network and prediction network, where the path state vector is used as the input state and the rest are used as supervision signals. The supervision signal in the data sample generates mean square error by comparing the predicted value with the true value in the experience replay pool. The mean square error is backpropagated to the representation network, dynamic network and prediction network, driving the parameters to be continuously updated, so that the EfficientZero reinforcement learning model gradually learns to generate the punching path with the minimum comprehensive cost function. After multiple rounds of iterative training, the candidate punching path sequence output by the EfficientZero reinforcement learning model gradually tends to the direction where the comprehensive cost function is minimized, thus obtaining the multi-objective punching path optimization result.

[0023] In this embodiment, the generation of the aluminum single-panel punching path optimization scheme specifically includes: In the multi-objective punching path optimization results, the comprehensive cost function value of each candidate punching path sequence is extracted; A comparison is performed among the comprehensive cost function values ​​of all candidate punching path sequences to obtain the path sequence with the smallest comprehensive cost function value. The path sequence is determined as the optimal path sequence, which simultaneously achieves a trade-off between path length, processing time, energy consumption, and stress stability index. The optimal path sequence is used as the optimization scheme for the punching path of aluminum single-panel. Example

[0024] To verify the feasibility of this invention in practical applications, it was applied to the processing of aluminum panels for the exterior facade of a public building. In this scenario, the design of the decorative aluminum panels for the building facade involves a complex distribution of holes. These holes are irregular in shape, numerous in number, and exhibit significant density variations. Traditional punching path planning methods, when faced with such highly complex hole groups, often only consider the shortest processing distance or time, lacking a comprehensive understanding of the overall complexity of the hole distribution and the evolution of residual stress in the sheet material. This results in the finished aluminum panels frequently exhibiting varying degrees of warping or edge cracking, even after the CNC machine tool completes the punching task, severely impacting subsequent installation quality and structural stability. Therefore, a new technical approach is urgently needed that can balance processing efficiency with reducing residual stress accumulation, thereby improving the quality of the finished product from the source.

[0025] In this scenario, the design drawings of the aluminum single-panel are first read digitally, and the coordinate information of all holes in the drawings is extracted to form a set of hole coordinates. To ensure the accuracy of subsequent calculations, these hole coordinates are standardized and preprocessed to remove duplicate and abnormal hole positions, resulting in a set of valid hole coordinates. Based on this set, the geometric center coordinates of the hole group and the relative position vector of each hole position are calculated to form standardized hole group data. Subsequently, a topological relationship graph of the hole group is constructed based on the standardized hole group data, and the node degree and degree probability are calculated according to the proximity relationship between hole positions. Finally, the topological entropy feature vector of the hole group is obtained, which is used to characterize the overall complexity of the hole group.

[0026] After processing the basic data, the effective hole position coordinate set, topological entropy feature vector, and path state vector constructed from the path state are input into the EfficientZero reinforcement learning model. This model consists of a representation network, a dynamic network, a prediction network, and a Monte Carlo tree search module. The representation network is responsible for encoding the input data, the dynamic network is responsible for simulating the potential state after the path is extended, the prediction network outputs the path policy distribution and state evaluation, and the Monte Carlo tree search module combines these outputs to expand and backtrack, generating multiple candidate punching path sequences. In this process, the model considers the geometric distance between holes, and also integrates the hole group complexity and the dynamic state of the current path, making the candidate paths more consistent with the overall optimal expectation.

[0027] To evaluate the actual performance of different candidate punching paths, each candidate path was input into the residual stress field prediction module. The aluminum panel was discretized and modeled on a finite element mesh, mapping each punching point to a mesh node. Loads and boundary conditions were then applied sequentially to the nodes to simulate the stress changes throughout the punching process. Based on the residual stress distribution calculated in this process, a stress stability index was further formed to quantitatively reflect the balance of stress accumulation during path execution. Thus, each candidate punching path not only has geometric and temporal indicators but also constraints on energy consumption and mechanical stability.

[0028] In the actual optimization process, path length, processing time, energy consumption, and stress stability indicators are uniformly incorporated into the comprehensive cost function. This comprehensive cost function balances the importance of different factors through weighting, ensuring that the optimization result is not limited to a single objective but achieves a balance among multiple objectives. The EfficientZero reinforcement learning model uses the comprehensive cost function as an evaluation signal for iterative training and continuously updates the parameters of the representation network, dynamic network, and prediction network through an experience replay mechanism. After multiple rounds of training, the model gradually learns to select the path sequence that minimizes the comprehensive cost function and outputs the multi-objective punching path optimization results.

[0029] The path sequence with the smallest comprehensive cost function value in the optimization results is taken as the final punching path optimization scheme and converted into a path instruction file that can be recognized by CNC machining equipment. The path instruction file not only contains the machining sequence of hole position coordinates, but also includes relevant information such as machining time, energy consumption and stress stability index, so that the equipment can call parameters and compare in real time during the execution process.

[0030] To verify the performance of the present invention in practice, it was compared with traditional methods, and the results are shown in Table 1.

[0031] Table 1. Comparison of the punching path optimization effects between the method of the present invention and the traditional method. In terms of average path length, the optimal punching path generated by the method of this invention is shortened by nearly 10% compared with the traditional method. This shows that by introducing the combination of the topological entropy feature vector of the hole group and the EfficientZero reinforcement learning model, redundant path extension can be avoided and more compact and reasonable path planning can be achieved.

[0032] In terms of processing time, the method of this invention shortens the processing time by more than ten minutes compared with the traditional method, with an improvement of more than 10%. This is directly related to the shortening of the path and the optimization of the action sequence. At the same time, due to the improvement of path efficiency, the idle time of the machine tool movement during processing is reduced, thereby further shortening the total processing time.

[0033] In terms of energy consumption, the method of this invention reduces energy consumption by about 15% compared with the traditional method. This is mainly due to the introduction of energy consumption factors in the comprehensive cost function, which enables the optimization process to achieve energy-saving goals while ensuring processing accuracy.

[0034] The comparison results of residual stress index are particularly outstanding. The residual stress stability of the method of the present invention is improved by nearly 30% compared with the traditional method. This means that the internal stress distribution of the aluminum single panel after processing is more balanced and less prone to sudden structural deformation. Combined with the residual stress prediction module, the present invention can avoid the risk of high stress accumulation in advance when selecting the path, which is difficult to achieve with the traditional method.

[0035] Regarding the warpage of the aluminum panels, the method of this invention reduces the warpage of the finished aluminum panels by 40%, which fundamentally improves the stability of the finished product during installation and long-term use. The reduction in warpage is due to the fact that this invention comprehensively considers the dynamic accumulation effect of residual stress and maintains a balanced mechanical distribution during path execution through multi-objective optimization constraints, thereby significantly reducing deformation after processing.

[0036] The above description is only a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any equivalent substitutions or modifications made by those skilled in the art within the scope of the technology disclosed in the present invention, based on the technical solution and inventive concept of the present invention, should be covered within the scope of protection of the present invention.

Claims

1. A method for optimizing the punching path of aluminum single-panel based on artificial intelligence, characterized in that, Includes the following steps: Collect hole distribution information from aluminum single-panel design drawings, establish a set of hole coordinates, perform standardization preprocessing on the set of hole coordinates to obtain an effective set of hole coordinates, and generate standardized hole group data; A topological relationship graph of a pore group is constructed based on standardized pore group data, a topological complexity index is calculated, the geometric complexity of the pore group distribution is measured, and a topological entropy feature vector of the pore group is generated. Define the path state and candidate action set, construct the path state vector, input the path state vector and the topological entropy feature vector into the EfficientZero reinforcement learning model, and generate multiple sets of candidate punching path sequences by combining the effective hole position coordinate set. For each group of candidate punching path sequences, residual stress field prediction is performed to obtain the stress distribution during the path execution process and generate the corresponding set of stress stability indices. The path length, processing time, energy consumption, and stress stability index are used as multi-objective constraints to construct a comprehensive cost function. The EfficientZero reinforcement learning model is used for iterative training to optimize the candidate punching path sequence and generate multi-objective punching path optimization results. The path sequence with the smallest comprehensive cost function value is selected from the multi-objective punching path optimization results and used as the aluminum single-panel punching path optimization scheme, which is then output for CNC machining equipment to call and execute.

2. The method for optimizing the punching path of aluminum single-panel based on artificial intelligence according to claim 1, characterized in that, The generation of the standardized pore group data specifically includes: Data is read from the aluminum single-panel design drawings, the hole distribution information in the design drawings is extracted, and the horizontal and vertical coordinates of each hole are organized into a set of hole coordinate points to form an initial set of hole coordinates. The initial set of borehole coordinates is normalized to generate a normalized set of borehole coordinates. The normalized set of pore location coordinates is denoised, and a spatial filtering method is used to remove duplicate and abnormal pore locations to obtain an effective set of pore location coordinates. Based on the set of effective borehole coordinates, the geometric center coordinates of the borehole group are calculated. The horizontal coordinate of the geometric center is calculated by summing the horizontal coordinates of all effective boreholes and dividing by the number of effective boreholes. The vertical coordinate of the geometric center is calculated by summing the vertical coordinates of all effective boreholes and dividing by the number of effective boreholes. Thus, the geometric center coordinates of the borehole group are obtained. Using the coordinates of the geometric center of the hole group as a reference, calculate the relative position vector of each effective hole point relative to the coordinates of the geometric center of the hole group, and form a set of relative position vectors; The set of relative position vectors is used as the standardized pore group data.

3. The method for optimizing the punching path of aluminum single-panel based on artificial intelligence according to claim 1, characterized in that, The generation of the topological entropy feature vector of the pore group includes: The obtained standardized hole group data is called, and each hole point is defined as a node in the graph structure. The connection rules of the edges are established according to the spatial proximity relationship between the nodes. The Euclidean distance is calculated based on the relative position vector. When the Euclidean distance between any two hole points is less than a preset threshold, an undirected edge is established between the corresponding nodes, thus forming a hole group topology graph containing a set of nodes and a set of edges. In the topological relationship graph of the hole group, the node degree is calculated for each node, and all node degrees are arranged in order to form a set of node degree distributions. Calculate the degree probability of each node based on the set of node degree distributions; After obtaining the degree probabilities of all nodes, the topological entropy of the hole group is calculated. The topological entropy value is used as a quantitative indicator of the overall geometric complexity of the hole group, and the topological entropy value is combined with the degree probability of all nodes to form the topological entropy feature vector.

4. The method for optimizing the punching path of aluminum single-panel based on artificial intelligence according to claim 1, characterized in that, The generation of the multiple sets of candidate punching path sequences includes: The topological entropy feature vector and the set of effective aperture coordinates are used as input data and input into the EfficientZero reinforcement learning model, which includes a representation network, a dynamic network, a prediction network and a Monte Carlo tree search module. Define the path state and generate the path state vector; Input data and path state vectors are input into a representation network, which encodes them to generate latent representation vectors. The representation network includes an embedding layer, a multi-layer fully connected network, and a normalization layer. Construct a set of candidate actions; Starting from the current punching position, calculate the difference between the horizontal and vertical coordinates of the current punching position and the candidate target hole position in the effective hole position coordinate set. Use the two-dimensional vector formed by the difference between the horizontal and vertical coordinates as the basic displacement information. At the same time, calculate the Euclidean distance between the current punching position and the candidate target hole position. Extract the node degree probability of the corresponding candidate target hole position from the topological entropy feature vector. Concatenate the basic displacement information, Euclidean distance and node degree probability to form the candidate action representation vector. The latent representation vector and candidate action representation vector are input into a dynamic network. The dynamic network updates the latent representation vector to obtain a new latent representation vector and real-time feedback information. The dynamic network is implemented using a multilayer perceptron network. The new latent representation vector is input into the prediction network, and the prediction network outputs the path policy distribution and the path state evaluation. The path policy distribution represents the probability of each candidate action being selected in the candidate action set, and the path state evaluation represents the overall quality of the current path state in the overall punching task. The prediction network includes a multilayer perceptron, which is divided into a policy branch and a value branch after receiving a new latent representation vector. The policy branch is used to output the selection probability of each candidate action in the candidate action set through a fully connected layer and a Softmax function. The value branch is used to output the comprehensive advantages and disadvantages of the current path state in the overall punching task through a fully connected layer and a nonlinear activation function. In the process of generating candidate punching paths, the Monte Carlo tree search module is used to expand and simulate the candidate action set. The path strategy distribution and path state evaluation are used as the basis for expanding the Monte Carlo tree search module. The output results of the representation network, dynamic network and prediction network are used to repeatedly expand the search and generate multiple candidate punching path sequences. During the generation of candidate punching path sequences, the path length information of each candidate punching path sequence is calculated, and the path length information is stored in correspondence with the candidate punching path sequence to construct a path sequence set.

5. The method for optimizing the punching path of aluminum single-panel based on artificial intelligence according to claim 1, characterized in that, The generation of the corresponding set of stress stability indices specifically includes: Use the candidate punching path sequence as input data; The aluminum single-panel material is discretized and modeled on the finite element mesh. The hole points in the effective hole position coordinate set are mapped to the node positions of the finite element model. Loads and boundary conditions are applied to the corresponding nodes in sequence according to the order of the candidate punching path sequence to obtain the mesh stress state under the execution of the path. The residual stress value is calculated in the finite element mesh. The residual stress value consists of three parts: the transverse residual stress component, the longitudinal residual stress component, and the thickness direction residual stress component. The three parts are added together to obtain the residual stress value of the node. During the execution of the candidate punching path sequence, the stress increment is calculated for each hole location; As the candidate punching path sequence gradually extends, the stress increments at all hole locations are accumulated sequentially to obtain the stress accumulation effect. The stress stability index is calculated based on the stress accumulation effect. The stress stability index is defined as the sum of the absolute values ​​of stress increments in all punching steps divided by one and added to it. The stress stability indices corresponding to each candidate punching path sequence are organized to generate a set of stress stability indices.

6. The method for optimizing the punching path of aluminum single-panel based on artificial intelligence according to claim 1, characterized in that, The generation of the multi-objective punching path optimization results specifically includes: The path sequence set is fused with the stress stability index set; Calculate processing time; Calculate energy consumption; Construct a comprehensive cost function, which is path length information multiplied by path length weight, plus processing time multiplied by time weight, plus energy consumption multiplied by energy consumption weight, plus stress instability multiplied by stress weight, where stress instability is equal to one minus stress stability index. During training, the EfficientZero reinforcement learning model takes the candidate punching path sequence as input, uses the representation network to encode the path state and topological entropy feature vector to obtain the latent representation vector, uses the dynamic network to predict the new latent representation vector and real-time feedback information under the action of the latent representation vector and candidate actions, uses the prediction network to output the path policy distribution and path state evaluation, and uses the Monte Carlo tree search module to expand and simulate the candidate action set to generate new candidate punching path sequences. In each round of training, the calculation result of the comprehensive cost function is used as the evaluation signal, and the candidate punching path sequence is stored in the experience replay pool along with the corresponding path state vector, path length information, processing time, energy consumption, stress stability index and real-time feedback information of the dynamic network output, forming a set of data samples for training. Data samples are repeatedly sampled from the experience replay pool, and the parameters of the representation network, dynamic network and prediction network are continuously updated by the stochastic gradient descent method, so that the EfficientZero reinforcement learning model gradually converges. After multiple rounds of iterative training, the candidate punching path sequence output by the EfficientZero reinforcement learning model gradually tends to the direction where the comprehensive cost function is minimized, thus obtaining the multi-objective punching path optimization result.

7. The method for optimizing the punching path of aluminum single-panel based on artificial intelligence according to claim 1, characterized in that, The generation of the aluminum panel punching path optimization scheme specifically includes: In the multi-objective punching path optimization results, the comprehensive cost function value of each candidate punching path sequence is extracted; A comparison is performed among the comprehensive cost function values ​​of all candidate punching path sequences to obtain the path sequence with the smallest comprehensive cost function value; The optimal path sequence is used as the optimization scheme for the punching path of aluminum single-panel.

Images