Profile cutting path optimization method based on artificial intelligence

Abstract

The invention discloses an artificial intelligence-based profile cutting path optimization method, which comprises the following steps of: S1, acquiring a to-be-cut surface stress map and a part arrangement image, and generating a two-dimensional stress geometric model; s2, taking the two-dimensional stress geometric model as input, and constructing a non-Euclidean path diagram structure; s3, the non-Euclidean path diagram structure serves as target task input, and an initial path strategy suitable for current profile arrangement is output; s4, constructing a disturbance space by identifying a heat sensitive area in the path diagram, and outputting an optimized path strategy; s5, performing structure mapping on the optimized path strategy to obtain a permutation entropy analysis result; s6, performing structure rearrangement on the high-complexity path segment according to a permutation entropy analysis result, and outputting a path sequence with relatively high structure stability; s7, the control instruction is issued to the numerical control cutting equipment to be executed; s8, performing dynamic strategy updating based on the feedback data set; the application of the method improves the precision, stability and intelligent adaptability of path planning.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of profile processing intelligent optimization, and particularly relates to a profile cutting path optimization method based on artificial intelligence. BACKGROUND

[0002] In the existing profile manufacturing and processing flow, the cutting path planning of the profile mostly relies on a static method model or a heuristic method based on artificial experience to perform path sorting and tool path generation. These methods have certain feasibility when facing regular arrangement, single type profile and simple part geometry, but it is difficult to guarantee the overall optimization of the path and the cutting precision under complex part arrangement, variable profile stress distribution and dynamic physical constraint environment. At the same time, the current mainstream path optimization method generally takes the shortest path as the objective function, ignoring the actual processing problems such as material thermal deformation, stress concentration and thermal zone overlap in the cutting process, resulting in problems such as part warping, thermal stress cracking or path reconstruction failure in path execution.

[0003] The existing cutting path planning method generally lacks the perception ability of the physical processing environment, and shows the phenomenon of local optimization but global instability when dealing with heat-sensitive areas in the path, boundary safety and stress field distribution. Especially in the material type switching, arrangement style change or cross-task scene, the traditional path model is difficult to have effective migration ability, and often needs to rebuild the complete planning process, causing low efficiency. In addition, although the existing path optimization method based on neural network or genetic method has certain intelligence, it has poor generalization ability and insufficient adaptability to new tasks in actual deployment, and it is difficult to meet the growing demand for flexible manufacturing and intelligent processing.

[0004] Therefore, how to provide a profile cutting path optimization method based on artificial intelligence is a problem to be solved by those skilled in the art. SUMMARY

[0005] The purpose of the present application is to provide a profile cutting path optimization method based on artificial intelligence, which optimizes and rearranges the cutting path by combining stress modeling, non-Euclidean path graph construction, transfer learning, disturbance search and permutation entropy analysis, and realizes dynamic updating of the path strategy based on device feedback, thereby improving the precision, stability and intelligent adaptability of the path planning.

[0006] The specific scheme of the present application is: a profile cutting path optimization method based on artificial intelligence, comprising the following steps:

[0007] S1. Obtain the surface stress map and part arrangement image to be cut, generate a two-dimensional stress geometry model by combining laser interference imaging and elastic modeling technology;

[0008] S2. Construct a non-Euclidean path graph structure as input of the two-dimensional stress geometry model;

[0009] S3. Input the non-Euclidean path graph structure as a target task, load the path strategy data in the historical cutting task by using the collaborative fuzzy inductive transfer learning method, and output the initial path strategy suitable for the current profile arrangement;

[0010] S4. Construct a perturbation space by identifying the heat-sensitive area in the path graph, call the saddle point perturbation evolution search method to perform multiple rounds of perturbation and selection on the initial path strategy, and output the optimized path strategy;

[0011] S5. Perform structure mapping on the optimized path strategy, construct a lattice matrix representation of the path sequence, and perform analysis on the mapping result to obtain the permutation entropy analysis result;

[0012] S6. According to the permutation entropy analysis result, rearrange the structure of the high complexity path segment, output the path sequence with higher structural stability, and convert it into control instructions of the numerical control equipment;

[0013] S7. The control instructions are sent to the numerical control cutting equipment for execution, and the device data is collected in real time through the sensing system to generate a feedback data set;

[0014] S8. Dynamic strategy update based on the feedback data set, online correction and parameter update of the path strategy.

[0015] Further, the generation of the two-dimensional stress geometry model in step S1 in the application specifically includes:

[0016] The laser interference imaging device is used to scan the profile to be cut at multiple angles, collect interference image sequences of the profile surface under different excitation conditions, and obtain initial optical interference data containing surface micro-displacement characteristics;

[0017] The micro-displacement change of the initial optical interference data is extracted by image frame difference, and two-dimensional displacement distribution data representing the micro-deformation state of the profile is formed;

[0018] Based on the two-dimensional displacement distribution data, the stress response of each region of the profile is mapped, and a stress distribution map is generated;

[0019] The boundary of the part arrangement in the profile image is extracted to obtain the spatial contour of the part to be cut, and the spatial contour is spatially aligned and fused with the stress distribution map to form a two-dimensional stress geometry model containing stress characteristics and arrangement structure.

[0020] Further, the construction of the non-Euclidean path graph structure in step S2 in the application specifically includes:

[0021] The two-dimensional stress geometry model is regionally segmented, and significant geometry structures containing part boundaries and corresponding stress distribution maps are divided into multiple local regions, and key boundary points representing path transformation are selected as cutting nodes in each region to establish a cutting node set;

[0022] In the cutting node set, connection paths between nodes are constructed according to spatial proximity and profile structure morphology to form a candidate path segment set;

[0023] The physical constraint weight value of each path segment is calculated for each candidate path segment;

[0024] The cutting node set, candidate path segment set and corresponding physical constraint weight value are integrated to construct a non-Euclidean path graph structure.

[0025] Further, the output process of the initial path strategy suitable for the current profile arrangement in step S3 in the application specifically includes:

[0026] The nodes, path segments and physical constraint features in the non-Euclidean path graph structure are encoded into multi-dimensional vector data to construct a structure embedding matrix of the target task;

[0027] Select a completed cutting task from the historical task database, embed the structure embedding matrix into a source task structure representation set, extract node-level, edge-level and graph-level multi-granularity structure vectors, use K-medoids-based path style prototype extraction to form a prototype library, the selection of K-medoids is determined according to the dual conditions of maximum contour coefficient and minimum style coverage rate, and the prototype library is subjected to redundancy compression and execution stability check to generate a fuzzy weighted structure set;

[0028] In the collaborative fuzzy induction transfer learning module, a structure-semantic collaborative alignment loss function is established and minimized, and according to the consistency constraint of the superimposed path strategy label and historical execution feedback on the fuzzy weighted structure set, local mismatch correction and fine-grained mapping are performed to generate a structure-consistent mapping result;

[0029] The structure-consistent mapping result is applied to the historical path strategy vector space to project and reconstruct high-quality path strategies in the source task, generating an initial strategy candidate set aligned with the target task structure, and the selected initial strategy candidate set is subjected to a cycle consistency constraint from source to target to source to output an initial path strategy suitable for the current profile arrangement.

[0030] Further, the construction of the optimized path strategy in step S4 in the application specifically includes:

[0031] Based on the initial path strategy, segment annotation is performed according to path segments and key nodes, each path segment is scored according to temperature rising trend, stress change amplitude and historical error record, a heat-sensitive area map is generated, and high-risk area, transition area and safe area are marked, and a candidate disturbance area list is output;

[0032] Based on the candidate disturbance area list, a path disturbance operation set is constructed, and disturbance operations are listed for each area, applicable conditions and prohibited conditions are set for each type of operation, constraints such as not breaking the path connectivity, not crossing the device travel boundary and minimum safety distance are clearly defined, and operations meeting the constraints are combined into disturbance schemes, the influence range and expected change are recorded, and a disturbance scheme list is output;

[0033] Taking the disturbance scheme list and the initial path strategy as input, each disturbance scheme is applied to the corresponding area one by one to generate a candidate path set, and each candidate path is uniformly evaluated, the evaluation results are summarized into a candidate path index table, and the candidate path with obvious comprehensive improvement without triggering residual forbidden area is selected, and a candidate path survival set is output;

[0034] In each round of disturbance process, a change trend model is established in the physical loss function space for the candidate path survival set, the change sequence of the path strategy in heat accumulation, path energy consumption and structure stability index before and after disturbance is recorded, the cost surface of the path strategy is constructed, and the saddle point disturbance evolution search method is adopted to calculate the second-order change relationship between the local change directions of the cost surface. When it is detected that the change directions of the cost surface in multiple directions are contradictory to each other and there is a change trend outside the local minimum value, it is determined that the current path strategy is in the local saddle point area. The saddle point direction is used as the disturbance guide for multiple rounds of disturbance and selection, and an optimized path strategy is output.

[0035] Further, the process of obtaining the permutation entropy analysis result in step S5 in the application specifically includes:

[0036] The optimized path strategy is disassembled into a path segment sequence according to the device execution order, the device attributes are extracted segment by segment, and the device attributes are written into a two-dimensional grid-shaped data carrier in a fixed field order according to the order of the path segments, forming an initial grid matrix representation;

[0037] Uniformization processing and data neatness are performed on the initial grid matrix, the direction and corner are discretized into a finite state and stored in a fixed code, and consistency check is performed on the rows and columns, to obtain a path grid matrix;

[0038] Submatrices are sequentially intercepted in the path segment index direction of the path grid matrix, each column attribute in each submatrix is given a small-to-large order label, and the permutation pattern identifier is generated by concatenating in the predetermined column order to form a permutation pattern sequence;

[0039] The frequency of each arrangement mode in the whole path range is summarized, a global arrangement entropy metric based on the current frequency distribution is calculated, and a local arrangement entropy is calculated at each window position. The global and local arrangement entropies are backfilled to the path segment index to mark the continuous fragments with high complexity and the complexity mutation positions, and an arrangement entropy analysis result is generated.

[0040] Further, the conversion of the control instruction format of the numerical control equipment in step S6 includes:

[0041] The arrangement entropy analysis result is received, and the continuous fragments with high complexity and the mutation points are marked to generate a high complexity path segment list;

[0042] The high complexity path segment list is taken as input, and structure rearrangement is performed within the same segment. Local exchange and insertion are performed on adjacent segments with conflicts, and a unique rearrangement result is determined and backfilled to form a new path sequence;

[0043] The new path sequence is subjected to performability checking, and local rollback or enabling of a backup rearrangement scheme is performed on a segment that fails the checking until the checking is passed, and the path is converted into an absolute position point sequence in the machine tool coordinates;

[0044] Based on the absolute position point sequence, a control instruction file, a rearrangement log, and a key process parameter list are output, semantic and format checking is performed, and the control instruction of the numerical control equipment is converted.

[0045] Further, the generation of the feedback data set in step S7 includes:

[0046] The control instruction of the numerical control equipment is loaded into the to-be-executed queue of the numerical control equipment, and one-time proofreading is completed according to the plate thickness and tool parameters to generate an execution parameter package and a start token;

[0047] The start token and the execution parameter package are taken as input, and a timestamp and a value are recorded at a fixed sampling period to form an original time series data stream and an actual trajectory record;

[0048] The original time series data stream and the actual trajectory record are taken as input, and the planned path point sequence is compared to mark abnormal segments and generate a segmentation index table, and a segmentation evaluation result is output;

[0049] The segmentation evaluation result is taken as input, and the time series data and the trajectory record are summarized according to the path segment number to generate a feedback data set.

[0050] Further, the online correction and parameter update of the path strategy in step S8 include:

[0051] Based on the feedback data set, the path segment number corresponding to each path segment is extracted, and is aligned with the original optimized path strategy to form a path segment-execution feedback mapping table;

[0052] Based on the path segment-execution feedback mapping table, the penalty factor and the thermal constraint level of the path segment are adjusted according to the local residual change to generate a revised path segment attribute revision table;

[0053] The original path strategy in the path segment attribute revision table is disturbed online, and new physical constraint boundaries are set in combination with the overheat and overstroke path segments marked in the feedback;

[0054] Based on the new physical constraint boundaries, the feedback data after a new round of execution is recorded to form an iterative optimization closed loop. When the improvement amplitude of the residual error of two consecutive rounds of feedback is equal, a strategy freezing mechanism is triggered to output a stable path strategy and an update log.

[0055] The present application has the following beneficial effects:

[0056] 1. By introducing the collaborative fuzzy induction transfer learning method, the saddle point disturbance evolutionary search mechanism and the path structure rearrangement method based on permutation entropy analysis, the problems of poor transfer ability, insufficient response to thermal physical constraints and low path stability in the existing technology in profile cutting path optimization are overcome.

[0057] 2. The collaborative fuzzy induction transfer learning method introduces historical task strategy information, and through the structure inversion and attribute alignment mechanism, the effective transfer of path strategy between different tasks is realized, the generalization ability of the model in complex profile structure and multi-task scene is improved, thereby avoiding the inefficient problem of repeated modeling and training.

[0058] 3. The saddle point disturbance evolutionary search method can construct a disturbance space in the heat-sensitive area, search and select the initial path strategy for multiple rounds, avoid local optimal trap, and realize multi-objective dynamic weighting of thermal stability and path cost. Combined with the lattice matrix representation and permutation entropy analysis means, the high complexity area in the path is effectively identified, and the overall continuity and processing stability of the path are optimized through structure rearrangement.

[0059] 4. In combination with real-time sensing feedback during cutting, the path strategy is dynamically revised and iteratively updated, an intelligent system from stress modeling, path planning to execution feedback closed loop optimization is constructed, the adaptability, execution safety and processing precision of the path strategy are greatly improved, and the present application has significant engineering application value and popularization prospect. BRIEF DESCRIPTION OF DRAWINGS

[0060] Figure 1 A flowchart of a profile cutting path optimization method based on artificial intelligence is provided for the present application;

[0061] Figure 2 A schematic diagram of a profile cutting path optimization method based on artificial intelligence is provided for the present application. DETAILED DESCRIPTION

[0062] The technical solutions of the present application will be described clearly and completely below in combination with the drawings of the present application. Obviously, the described embodiments are only some of the embodiments of the present application, but not all the embodiments. Based on the embodiments in the present application, all other embodiments obtained by those skilled in the art without creative work fall within the protection scope of the present application.

[0063] In the description of the present application, it should be noted that, unless otherwise explicitly specified and limited, the terms "arrange", "mount", and "connect" should be understood broadly, for example, can be fixedly connected, or can be detachably connected, or integrally connected; can be mechanically connected, or can be electrically connected; can be directly connected, or can be indirectly connected through an intermediate medium, or can be the internal communication of two elements. For those skilled in the art, the specific meanings of the above terms in the present application can be understood according to the specific circumstances.

[0064] Referring to Figure 1 , Figure 2 The present application is a profile cutting path optimization method based on artificial intelligence, comprising the following steps:

[0065] S1. Obtain the surface stress map of the profile to be cut and the part arrangement image, generate a two-dimensional stress geometry model containing micro-deformation information and arrangement characteristics by combining laser interference imaging and elastic modeling technology;

[0066] S2. Take the two-dimensional stress geometry model as input, construct a non-Euclidean path graph structure containing cutting nodes, candidate path segments and physical constraint weights, which is used to describe the path connection relationship and stress distribution characteristics in the cutting area to be cut;

[0067] S3. Take the non-Euclidean path graph structure as the target task input, load the path strategy data in the historical cutting task by using the collaborative fuzzy induction transfer learning method, adjust the structural difference between the source task and the target task through the fuzzy entropy adjustment mechanism, and output the initial path strategy suitable for the current profile arrangement;

[0068] S4. Construct a perturbation space by identifying the heat-sensitive area in the path graph, call the saddle point perturbation evolution search method to perform multiple rounds of perturbation and selection on the initial path strategy, and output the optimized path strategy;

[0069] S5. Perform structure mapping on the optimized path strategy, construct a lattice matrix representation of the path sequence, and perform analysis on the mapping result to obtain the permutation entropy analysis result, and identify the complexity abnormal area in the path structure;

[0070] S6. According to the arrangement entropy analysis result, the high complexity path segment is rearranged in structure, a path sequence with high structural stability is output, and control instructions recognizable by a numerical control device are converted;

[0071] S7. The control instruction format is issued to the numerical control cutting device for execution, and device data including thermal field distribution, mechanical response and energy consumption data are collected in real time through a sensing system to generate a feedback data set of the cutting process;

[0072] S8. Based on the feedback data set, dynamic strategy updating is performed to correct and update the path strategy online, so as to realize adaptive closed-loop optimization of path planning and cutting execution.

[0073] The present application proposes a profile cutting path optimization method based on artificial intelligence, which generates a two-dimensional stress geometry model by combining laser interference imaging and elastic modeling, and constructs a non-Euclidean path graph considering stress and thermal sensitivity. Through collaborative fuzzy induction transfer learning, an initial path strategy is obtained, and path optimization is realized by combining saddle point disturbance evolution search. The lattice matrix and permutation entropy analysis are introduced to identify complex regions and rearrange the path, and the executable control instructions are converted. Combined with device feedback, the closed-loop adaptive optimization of the path strategy is realized, and the cutting stability and precision are improved.

[0074] In the embodiment, the generation of the two-dimensional stress geometry model specifically includes:

[0075] The profile to be cut is scanned by a laser interference imaging device at multiple angles, and interference image sequences of the profile surface under different excitation conditions are collected to obtain initial optical interference data containing surface micro-displacement characteristics;

[0076] The micro-displacement change of the initial optical interference data is extracted by image frame difference to form two-dimensional displacement distribution data representing the micro-deformation state of the profile;

[0077] Based on the two-dimensional displacement distribution data, the stress response of each region of the profile is mapped in structure to generate a stress distribution atlas describing the micro-deformation trend of the profile;

[0078] The boundary of the part arrangement in the profile image is extracted to obtain the spatial contour of the part to be cut, and the spatial contour is spatially aligned and fused with the stress distribution atlas to form a two-dimensional stress geometry model containing stress characteristics and arrangement structure.

[0079] The application collects micro-displacement data of the profile surface by using laser interference imaging technology, extracts micro-deformation features by combining inter-frame difference, constructs a two-dimensional distribution map reflecting the micro stress state of the profile, and extracts and spatially fuses the part arrangement contour to generate a two-dimensional stress geometry model containing stress information and arrangement structure at the same time. The model provides a high-resolution, physically accurate basic expression for subsequent path graph construction and strategy optimization, improving the perception ability of the cutting path to deformation risk and structural features.

[0080] In the embodiment, the construction of the non-Euclidean path graph structure specifically includes:

[0081] The two-dimensional stress geometry model is regionally segmented, and the significant geometry structure containing the part boundary and the corresponding stress distribution map are divided into multiple local regions. Key boundary points representing path transformation are selected as cutting nodes in each region to establish a cutting node set;

[0082] In the cutting node set, the connection paths between nodes are constructed according to the spatial proximity relationship and the profile structure form to form a candidate path segment set. Each candidate path segment connects two cutting nodes, and whether the path crosses a high stress area, a heat-sensitive area or a narrow gap area is considered;

[0083] For each candidate path segment, the path segment is quantitatively analyzed in combination with the path length, the average stress value of the crossed region, the heat sensitivity index and the material structure density in the region. The physical constraint weight value of each path segment is calculated. The physical constraint weight value is the influence of stress disturbance and thermal deformation in the actual cutting process;

[0084] The cutting node set, the candidate path segment set and the corresponding physical constraint weight value are integrated to construct a non-Euclidean path graph structure. The path graph structure does not rely on the shortest distance in traditional Euclidean geometry, but takes physical stress, thermal influence and structural complexity as the path connection principle.

[0085] The application constructs a graph structure containing cutting nodes and candidate path segments by regionally dividing the two-dimensional stress geometry model and extracting key nodes, and quantifies the constraint weight of each path segment by comprehensively considering physical factors such as stress distribution, heat sensitivity and material density, forming a non-Euclidean path graph model reflecting the real cutting difficulty. This method breaks through the traditional graph construction method based on geometric shortest distance, improves the expression ability of the path graph to complex material structure and thermal effect, and provides a more physically realistic modeling basis for subsequent path strategy optimization.

[0086] In the embodiment, the output process of the initial path strategy suitable for the current profile arrangement specifically includes:

[0087] The nodes, path segments, and physical constraint features in the non-Euclidean path graph structure are encoded into multi-dimensional vector data to construct the structural embedding matrix of the target task. The structural embedding matrix represents the structural expression of the current profile cutting task.

[0088] Completed cutting tasks are selected from the historical task database, and the structure embedding matrix is ​​embedded into the source task structure representation set. Node-level, edge-level, and graph-level multi-granularity structure vectors are extracted. A prototype library is formed by extracting path style prototypes based on K-medoids. The central representative path of each type of path style is determined as the path style prototype. The selection of K-medoids is determined by the dual conditions of maximizing the contour coefficient and minimizing the lower limit of the style coverage. The prototype library is subjected to redundancy compression and execution stability checks. Path style prototypes that meet the requirements of structural connectivity, reasonableness of high-risk area crossing, and historical execution success rate are retained to generate a fuzzy weighted structure set.

[0089] In the collaborative fuzzy-induced transfer learning module, a structure-semantic collaborative alignment loss function is established and minimized.

[0090]

[0091] in, The total loss function for collaborative fuzzy-induced transfer learning, The structure mapping weight matrix, The number of source tasks selected to participate in the migration. For the first Embedment matrix of each source task For the first Samples in the source task Fuzzy membership degree, Embed a matrix for the target task. For the transpose of the mapping matrix, To minimize the solution, The regularization coefficient is . The Laplace matrix of the target task path graph. The trace of the matrix, The Frobenius norm is used to measure the difference between two matrices; consistency constraints between path strategy labels and historical execution feedback are superimposed on the fuzzy weighted structure set to perform local mismatch correction and fine-grained mapping, generating a structurally consistent mapping result;

[0092] The mapping result with consistent structure is applied to a historical path strategy vector space, high-quality path strategies in a source task are projected and reconstructed, an initial strategy candidate set aligned with a target task structure is generated, a source-to-target-to-source cyclic consistency constraint is implemented on the screened initial strategy candidate set, and an initial path strategy suitable for the current profile arrangement is output, and the current profile arrangement structure and stress state are adapted.

[0093] The application encodes a cutting task structure into a multi-dimensional embedding matrix, extracts and fuzzy weights a path style prototype of a historical task to construct a migration learning input basis, introduces a collaborative fuzzy induction migration mechanism, and realizes projection and reconstruction of high-quality path strategies through structure-semantic alignment and consistency constraint.

[0094] In the embodiment, the construction of the optimized path strategy specifically includes:

[0095] Based on the initial path strategy, the path segments and key nodes are segmented and labeled, each path segment is scored according to the temperature rising trend, stress change amplitude and historical error record, a heat-sensitive area map is generated, and high-risk areas, transition areas and safe areas are labeled, and a candidate disturbance area list is output;

[0096] Based on the candidate disturbance area list, a path disturbance operation set is constructed, and executable disturbance operations are listed for each area, including detour replacement, adjacent path segment order adjustment, key node microshift and corner smoothing; for each type of operation, the applicable conditions and prohibited conditions are set, the constraints of not breaking the path connectivity, not crossing the equipment travel boundary and minimum safety distance are clarified, and the operations meeting the constraints are combined into disturbance schemes, the influence range and expected change are recorded, and a disturbance scheme list is output;

[0097] Taking the disturbance scheme list and the initial path strategy as input, each disturbance scheme is applied to the corresponding area one by one to generate a candidate path set, and each candidate path is uniformly evaluated, and the path total length change, heat distribution uniformity change, energy consumption change, corner and acceleration and deceleration smoothness change, alignment parameter consistency change and residual error triggering are recorded one by one, the evaluation results are summarized into a candidate path index table, and the candidate paths with no residual error forbidden area triggered and obvious comprehensive improvement are selected, and a candidate path survival set is output;

[0098] In each round of disturbance process, a change trend model is established for the candidate path survival set in the physical loss function space, a change sequence of the path strategy in heat accumulation, path energy consumption and structure stability indexes before and after the disturbance is recorded, a path strategy cost surface is constructed, a saddle point disturbance evolution search method is used to calculate the second order change relationship between the local change directions of the cost surface, so that the position of the current strategy in the optimization space is judged, when it is detected that the change directions of the multiple directions are contradictory to each other and there is a change trend outside the local minimum, it is determined that the current path strategy is in the local saddle point area, the saddle point direction is used as the disturbance guide to carry out multiple rounds of disturbance and selection, and the optimized path strategy is output.

[0099] The application constructs a disturbance area list through heat sensitive area recognition and path segment scoring, generates multiple feasible disturbance schemes under physical constraints, selects high-quality candidate paths after evaluating path quality indexes. The saddle point disturbance evolution search method is introduced, the change trend in multiple directions is analyzed based on the cost surface, the local saddle point area is identified and directional disturbance is carried out to jump out of the local optimum. This method effectively improves the heat balance, energy efficiency and structure stability of the path strategy, and realizes the deep adaptive evolution of path optimization.

[0100] In the embodiment, the obtaining process of the permutation entropy analysis result specifically includes:

[0101] The optimized path strategy is disassembled into a path segment sequence according to the device execution order, and the device attributes including the start and end positions, segment length, direction change, corner size, heat risk level and stress gradient level of the region are extracted segment by segment. The device attributes are written into a two-dimensional grid data carrier in a fixed field order according to the order of the path segments, forming an initial grid matrix represented by path segments as rows and attributes as columns;

[0102] Uniformization processing and data neatness are performed on the initial grid matrix, all position quantities are represented by the same reference origin and the same dimension, the direction and corner are discretized into a finite state and stored by fixed coding, the heat risk and stress gradient are linearly stretched to a uniform interval according to the global statistical range of the task, the missing items are filled with the average of adjacent valid items, the abnormal values are replaced by the neighborhood median, and the rows and columns are consistency checked to obtain a path grid matrix that can be used for sequence analysis;

[0103] A fixed-width sliding window is set in the path segment index direction of the path grid matrix and the sliding step is one step, a sub-matrix composed of adjacent path segments is intercepted, and each column attribute in each sub-matrix is given an order label from small to large according to the numerical value, and the order labels of each column are concatenated to generate the permutation pattern identifier corresponding to the current window according to the predetermined column order, and the whole window is intercepted and the pattern identifier is generated along the whole path to form a permutation pattern sequence covering the whole path, while the mapping relationship between the window and the original path segment index is retained.

[0104] The global and local two-level statistics of the permutation mode sequence are performed, the frequency of each permutation mode in the whole path range is summarized, the global permutation entropy measure based on the current frequency distribution is calculated, and the local permutation entropy is calculated at each window position, the global and local permutation entropy is backfilled to the path segment index, the continuous segment with high complexity and the complexity mutation position are labeled, and the permutation entropy analysis result including the global measure, the local measure and the mode frequency record is generated.

[0105] The application encodes the optimized path strategy into a path segment attribute matrix, unifies the dimension and processes missing and abnormal data, and forms a consistent lattice matrix representation. The sliding window method is used to extract and encode the path segment sequence, generate a permutation mode sequence covering the whole path, and calculate the permutation entropy based on global and local frequency statistics to quantify the complexity change of the path structure. This method realizes accurate description and visual recognition of path complexity, and provides data basis for subsequent structure rearrangement and path stability improvement.

[0106] In the embodiment, the conversion of the control instruction format recognizable by the numerical control equipment specifically includes:

[0107] The permutation entropy analysis result is received, the continuous segment with high complexity and the mutation point are labeled, and the high complexity path segment list with direction change, length difference, heat risk superposition and stress jump mark is generated;

[0108] Taking the high complexity path segment list as input, structure rearrangement is performed within the same segment, path segments with similar direction, compatible length, staggered heat risk and smooth stress change are preferentially combined, local exchange and insertion are performed on adjacent segments with conflicts, the unique rearrangement result is determined and backfilled to form a new path sequence;

[0109] Executable verification is performed on the new path sequence, the head and tail connection, the minimum safety distance, the device travel and tool lifting accessibility, the corner acceleration and deceleration smoothness are checked, local rollback or enabling of the standby rearrangement scheme is performed on the segments that do not pass the verification until the verification is passed, and the path is converted into an absolute position point sequence in the machine tool coordinates and added with tool compensation, introduction of tool withdrawal segment and feed height and speed parameters;

[0110] Based on the absolute position point sequence, the control instruction file, the rearrangement log and the key process parameter list are output, the program header, the coordinate system and the process parameter setting, the main cutting segment, the abnormal processing and the program tail are arranged, the semantic and format verification is performed, and the control instruction recognizable by the numerical control equipment is converted.

[0111] The application identifies high complexity path segments based on permutation entropy analysis results, rearranges path structures in combination with direction, length, thermal risk and stress characteristics, improves path continuity and processing stability, and after rearrangement, the path sequence is checked for executability and parameters are corrected to ensure that the execution conditions of the numerical control equipment are met, and is converted into absolute coordinate control instructions. The final output is a complete control program and a list of key parameters, realizing closed-loop optimization of path reconstruction and instruction generation, and ensuring processing precision and cutting safety.

[0112] In the embodiment, the generation of the feedback data set specifically includes:

[0113] The control instructions recognizable by the numerical control equipment are loaded into the to-be-executed queue of the numerical control equipment, and one-time proofreading is completed according to the plate thickness and tool parameters, an execution parameter package containing the program version, equipment parameters and safety inspection results and a start token are generated;

[0114] The start token and the execution parameter package are taken as inputs, and a timestamp and a value are recorded at a fixed sampling period to form an original time series data stream and an actual trajectory record;

[0115] The original time series data stream and the actual trajectory record are taken as inputs, and compared with the planned path point sequence, position deviation, speed deviation and temperature peak position offset are calculated segment by segment, abnormal segments are marked and a segmented index table is generated, and a segmented evaluation result containing segment number, deviation value and abnormal mark is output;

[0116] The segmented evaluation result is used to collect time series data and trajectory records according to the path segment number, and a feedback data set containing segment number, start and end time, position residual sequence, temperature change sequence, energy consumption change sequence and abnormal mark is generated.

[0117] The application realizes processing task initialization by loading control instructions into the numerical control equipment and completing parameter verification, collects equipment operation data during execution and constructs a time series data stream and an actual trajectory record, calculates key performance deviations and identifies abnormal areas in combination with the planned path and the execution result, generates detailed segmented evaluation indexes, and finally forms a complete feedback data set containing path segment residual, thermal response and energy consumption change, etc., providing high-precision data support for subsequent dynamic adjustment of path strategy.

[0118] In the embodiment, the online correction and parameter update of the path strategy specifically include:

[0119] Based on the feedback data set, the path segment number corresponding to each path segment is extracted and aligned with the original optimized path strategy to form a path segment-execution feedback mapping table;

[0120] Based on the path segment-execution feedback mapping table, a feature comparison sequence containing historical prediction values and actual feedback values is constructed and used as input training data to update the bias estimation sub-model for path prediction, a sliding window mechanism is used to fit the error accumulation trend in each window, and the penalty factor and thermal constraint level of the path segment are adjusted according to the local residual change to generate a revised path segment attribute correction table;

[0121] The original path strategy in the path segment attribute correction table is subjected to online local disturbance, and the maximum gradient point of the current residual curve is introduced as the disturbance priority area in the disturbance generation, the continuity of the saddle point disturbance evolution mechanism is maintained, and new physical constraint boundaries are set in combination with the overheat and overstroke path segments marked in the feedback;

[0122] Based on the new physical constraint boundary, the feedback data after a new round of execution is recorded to form an iterative optimization closed loop, and when the improvement amplitude of the feedback residual is equal for two consecutive rounds, the strategy freezing mechanism is triggered to output the stable path strategy and update log.

[0123] The application forms a feature comparison sequence by constructing the mapping relationship between the path segment and the feedback data, updates the path prediction bias model, dynamically adjusts the path segment attribute, and realizes the local disturbance optimization of the path strategy. The residual curve gradient is introduced as the disturbance priority area, the physical constraint is updated in combination with the overheat and overstroke information, and a closed loop optimization mechanism based on feedback data is constructed. When the strategy is stable, it is automatically frozen, and the final path and update record are output, improving the adaptability and stability of the path strategy.

[0124] Example 1:

[0125] In order to verify the feasibility of the application in implementation, the application is applied to the intelligent cutting production line of a heavy section processing enterprise. The production line is mainly responsible for the processing tasks of H-shaped steel, box girder and special-shaped steel components for high-rise buildings. The thickness distribution of the section raw materials is wide, the part arrangement is dense and complex, and it is limited by structural thermal deformation and arrangement interference. The traditional path optimization strategy has problems such as local path crossing high stress area, high path execution error rate, large cutting energy consumption, long processing time, etc.

[0126] The enterprise selected 6 typical steel component processing tasks for measurement, each section size was 12m x 2.5m, the original part number was 93 to 157, the section arrangement density was high, and the traditional optimization strategy was difficult to balance the path efficiency and cutting accuracy. The project team embedded the path optimization system of the application into its control platform, collected the micro-deformation interference images of the section before the task started through the laser interference imaging system, and generated a two-dimensional stress geometry model combined with elastic modeling to accurately identify the micro-deformation area.

[0127] The system automatically constructs a non-Euclidean path graph, calls the structural embedding data of a total of 143 historical tasks in the past by using a collaborative fuzzy inductive migration learning model, performs cross-task migration on the path strategy, and outputs an initial path in combination with the current stress distribution. In addition, the system also initially converts the path into a grid matrix and performs permutation entropy analysis, rearranges the structure of high complexity paragraphs, and corrects disturbances to output control instructions for processing equipment to execute. During the execution process, the system synchronously collects data such as thermal field, position, and energy consumption, and dynamically corrects the strategy in combination with the feedback.

[0128] In the profile cutting scene, the CAM strategy value is usually the data performance output by the mainstream industrial CAM system during path planning and instruction generation to guide the actual cutting task. We compare the effects of different schemes according to the CAM strategy value.

[0129]

[0130] As can be seen from Table 1, the total path length is shortened by about 4.6%, which means that the tool motion is more efficient, reducing the idle travel and redundant path. The maximum thermal peak position offset is reduced by 48.5%, indicating that the optimized path can better avoid the thermal deformation concentration area, thereby improving the machining precision. The average thermal risk weight of the path segment is reduced by nearly 30%, and the thermal stability during cutting is significantly improved. And the tool lift frequency is reduced by 12.3%, and the total cutting energy consumption is reduced by 12.4%, showing that the present application not only improves the path continuity, but also is more economical in energy consumption. The average cutting residual error is reduced by 16.7%, and the permutation entropy mean is reduced by nearly 20%, indicating that the path structure is more stable and the complexity is reduced. The path stability score is improved by more than 20%, and the proportion of abnormal path segments is greatly reduced by 79.4%, showing that the system can still maintain high stability under multi-source disturbance.

[0131] The above describes only the preferred specific embodiments of the present application, but the protection scope of the present application is not limited thereto, and any skilled person in the art can make equivalent substitutions or changes to the technical solutions and inventive concepts of the present application within the technical scope disclosed by the present application, which should be covered within the protection scope of the present application.

Claims

1. A profile cutting path optimization method based on artificial intelligence, characterized by, The method comprises the following steps: S1. Obtain the surface stress map to be cut and the part arrangement image, combine laser interference imaging and elastic modeling technology to generate a two-dimensional stress geometry model; S2. The two-dimensional stress geometry model is used as input to construct a non-Euclidean path graph structure; S3. The non-Euclidean path graph structure is used as the target task input, and the collaborative fuzzy induction transfer learning method is used to load the path strategy data in the historical cutting task, and the initial path strategy suitable for the current profile arrangement is output; S4. Construct a perturbation space by identifying the heat-sensitive area in the path graph, call the saddle point perturbation evolution search method to perform multiple rounds of perturbation and selection on the initial path strategy, and output the optimized path strategy; S5. Map the optimized path strategy to construct a lattice matrix representation of the path sequence, and analyze the mapping results to obtain the permutation entropy analysis results; S6. According to the permutation entropy analysis result, the high complexity path segment is rearranged, the path sequence with high structural stability is output, and the control instruction of the numerical control equipment is converted; S7. The control instruction is sent to the numerical control cutting equipment for execution, and the device data is collected in real time through the sensing system to generate a feedback data set; S8. Based on the feedback data set, dynamically update the strategy, and correct and update the parameters of the path strategy online.

2. The profile cutting path optimization method based on artificial intelligence according to claim 1, characterized in that, The generation of the two-dimensional stress geometry model in step S1 specifically includes: Scan the profile to be cut by a laser interference imaging device at multiple angles, collect interference image sequences of the profile surface under different excitation conditions, and obtain initial optical interference data containing surface micro-displacement characteristics; Extract the micro-displacement change of the initial optical interference data through image frame difference, and form two-dimensional displacement distribution data representing the micro-deformation state of the profile; Map the stress response of each region of the profile based on the two-dimensional displacement distribution data, and generate a stress distribution map; Extract the boundary of the part arrangement in the profile image to obtain the spatial contour of the part to be cut, and align and fuse the spatial contour with the stress distribution map to form a two-dimensional stress geometry model containing stress characteristics and arrangement structure.

3. The profile cutting path optimization method based on artificial intelligence according to claim 1, characterized in that, The construction of the non-Euclidean path graph structure in step S2 specifically includes: Divide the two-dimensional stress geometry model into multiple local regions by dividing the significant geometric structure containing the part boundary and the corresponding stress distribution map, and select key boundary points representing path transformation in each region as cutting nodes to establish a cutting node set; In the cutting node set, construct the connection path between the nodes according to the spatial proximity relationship and the profile structure form to form a candidate path segment set; Calculate the physical constraint weight value of each path segment for each candidate path segment; Integrate the cutting node set, the candidate path segment set and the corresponding physical constraint weight value to construct a non-Euclidean path graph structure.

4. The profile cutting path optimization method based on artificial intelligence according to claim 1, characterized in that, The output process of the initial path strategy suitable for the current profile arrangement in step S3 specifically includes: Encode the nodes, path segments and physical constraint features in the non-Euclidean path graph structure into multi-dimensional vector data to construct a structure embedding matrix of the target task; Select completed cutting tasks from the historical task database, embed the structure into a matrix to form a source task structure representation set, extract node-level, edge-level, and graph-level multi-granularity structure vectors, and use K-medoids-based path pattern prototype extraction to form a prototype library, wherein the selection of K-medoids is determined based on the maximization of the contour coefficient and the minimum lower limit of the pattern coverage rate, the prototype library is subjected to redundancy compression and execution stability checking, and a fuzzy weighted structure set is generated; In the collaborative fuzzy induction transfer learning module, a structure-semantic collaborative alignment loss function is established and minimized, local mismatch correction and fine-grained mapping are performed based on the consistency constraints of the fuzzy weighted structure set superimposed with path strategy labels and historical execution feedback, and a structure-consistent mapping result is generated; The structure-consistent mapping result is applied to the historical path strategy vector space to project and reconstruct high-quality path strategies in the source task, generating an initial strategy candidate set aligned with the target task structure, and implementing a source-to-target-to-source cyclic consistency constraint on the screened initial strategy candidate set, and outputting an initial path strategy suitable for the current profile arrangement.

5. The profile cutting path optimization method based on artificial intelligence according to claim 1, characterized in that, The construction of the optimized path strategy in step S4 specifically includes: Based on the initial path strategy, segment annotation is performed according to path segments and key nodes, each path segment is scored based on temperature rise trend, stress change amplitude, and historical error record, a heat-sensitive area map is generated, and high-risk areas, transition areas, and safe areas are labeled, and a candidate disturbance area list is output; Based on the candidate disturbance area list, a path disturbance operation set is constructed, and for each area, a disturbance operation is listed, and for each type of operation, applicable conditions and prohibited conditions are set, and constraints such as not breaking the path connectivity, not crossing the device travel boundary, and minimum safety distance are specified, and operations that meet the constraints are combined into disturbance schemes, and the impact range and expected change are recorded, and a disturbance scheme list is output; Taking the disturbance scheme list and the initial path strategy as input, each disturbance scheme is applied to the corresponding area one by one to generate a candidate path set, and each candidate path is uniformly evaluated, the evaluation results are summarized into a candidate path index table, and the candidate paths that do not trigger residual forbidden zones and have obvious comprehensive improvement are selected, and a candidate path survival set is output; In each round of disturbance process, a change trend model is established for the candidate path survival set in the physical loss function space, the change sequence of the path strategy in the heat accumulation, path energy consumption, and structure stability index before and after disturbance is recorded, the cost surface of the path strategy is constructed, and the saddle point disturbance evolutionary search method is adopted to calculate the second-order change relationship between the local change directions of the cost surface. When it is detected that the change directions of the cost surface in multiple directions are contradictory and there is a change trend outside the local minimum value, it is determined that the current path strategy is in the local saddle point area, and the saddle point direction is used as the disturbance guide for multiple rounds of disturbance and selection, and an optimized path strategy is output.

6. The profile cutting path optimization method based on artificial intelligence according to claim 1, wherein, The process of obtaining the permutation entropy analysis result in step S5 specifically includes: The optimized path strategy is disassembled into a path segment sequence according to the execution order of the equipment, the equipment attributes are extracted segment by segment, the equipment attributes are written into a two-dimensional grid-shaped data carrier in a fixed field order according to the order of the path segments, and an initial grid matrix representation is formed; Uniformization processing and data neatness are performed on the initial grid matrix, the direction and corner are discretized into a finite state and stored in a fixed code, and consistency checking is performed on the rows and columns to obtain a path grid matrix; Submatrices are sequentially intercepted in the path segment index direction of the path grid matrix, an order label from small to large is given to each column attribute in each submatrix, and a permutation pattern identification is generated by concatenating in a predetermined column order to form a permutation pattern sequence; The frequency of each permutation pattern in the permutation pattern sequence within the entire path range is summarized, a global permutation entropy metric based on the current frequency distribution is calculated, and a local permutation entropy is calculated at each window position. The global and local permutation entropies are backfilled to the path segment index to mark the continuous segments with high complexity and the complexity mutation positions, and a permutation entropy analysis result is generated.

7. The profile cutting path optimization method based on artificial intelligence according to claim 1, characterized in that, The conversion of the control instruction format of the numerical control equipment in step S6 specifically includes: Receiving the permutation entropy analysis result, marking the continuous segments with high complexity and the mutation points, and generating a high complexity path segment list; Taking the high complexity path segment list as input, performing structure rearrangement within the same segment, performing local exchange and insertion on adjacent segments with conflicts, determining a unique rearrangement result and backfilling to form a new path sequence; Performing execution verification on the new path sequence, performing local rollback or enabling a backup rearrangement scheme for the segments that do not pass the verification until they pass the verification, and converting the path into an absolute position point sequence in the machine tool coordinates; Based on the absolute position point sequence, outputting a control instruction file, a rearrangement log and a key process parameter list, performing semantic and format checking, and converting into control instructions of the numerical control equipment. 8.The profile cutting path optimization method based on artificial intelligence according to claim 1, wherein, The generation of the feedback data set in step S7 specifically includes: Loading the control instructions of the numerical control equipment into the execution queue of the numerical control equipment, performing one-time proofreading according to the plate thickness and tool parameters, generating an execution parameter package and a start token; Taking the start token and the execution parameter package as input, recording the timestamp and the value at a fixed sampling period to form an original time series data stream and an actual trajectory record; Taking the original time series data stream and the actual trajectory record as input, marking abnormal segments and generating a segmented index table by comparing with the planned path point sequence, and outputting a segmented evaluation result; Summarizing the time series data and the trajectory record according to the path segment number in the segmented evaluation result to generate a feedback data set. 9.The profile cutting path optimization method based on artificial intelligence of claim 1, wherein, The online correction and parameter update of the path strategy in step S8 specifically include: Based on the feedback data set, extracting the path segment number corresponding to each path segment, and aligning with the original optimized path strategy to form a path segment-execution feedback mapping table; Based on the path segment-execution feedback mapping table, adjusting the penalty factor and the thermal constraint level of the path segment according to the local residual change to generate a corrected path segment attribute correction table; Performing online local disturbance on the original path strategy in the path segment attribute correction table, and setting new physical constraint boundaries in combination with the overheating and overstroke path segments marked in the feedback; On the basis of new physical constraint boundary, record the feedback data after a new round of execution, form the iterative optimization closed loop, when the improvement amplitude of residual error of two consecutive rounds of feedback is equal, trigger the strategy freezing mechanism, output the stable path strategy and update log.

Images