Linear cutting path intelligent optimization method based on deep learning image recognition

Abstract

The invention discloses a linear cutting path intelligent optimization method based on deep learning image recognition. The method comprises the following steps that a workpiece multi-view image is obtained and imported into an electronic drawing, the image is recognized, workpiece structure information is extracted, a unified coordinate model is established in combination with a datum point, and an element list is output; based on the element list and machine tool and material parameters, a machining graph is constructed with contour segments as nodes and feasible transition between the segments as edges; candidate cutting paths are generated on the machining drawing according to the machining constraint, and an initial cutting scheme is formed; inputting the processing graph into a quantum heuristic neural network, performing optimization selection on the initial cutting scheme, and outputting an optimal cutting scheme; and geometric and energy simulation is conducted on the optimal cutting scheme, the cutting seam crossing, residual area and material falling risks are detected, and local correction is executed when the constraint is not met. According to the invention, through combination of deep learning identification and quantum heuristic optimization, automatic generation and correction of the linear cutting path are realized.

Description

TECHNICAL FIELD

[0001] The present application relates to the field of numerical control machining and intelligent optimization, and particularly relates to a wire cutting path intelligent optimization method based on deep learning image recognition. BACKGROUND

[0002] As a common precision machining process, the machining quality and efficiency of wire cut electrical discharge machining (WEDM) largely depends on the generation and optimization of cutting paths. Existing path planning methods mostly rely on artificial experience or simple geometric rules to generate paths, often failing to fully consider the complexity of workpiece structures, which can easily lead to problems such as kerf intersection, path redundancy, and machining dead angles, thereby reducing machining precision and stability.

[0003] To improve automation, existing technologies have attempted to introduce intelligent optimization methods such as genetic algorithms and ant colony algorithms to improve cutting paths, but in large-scale path combination scenarios, there are problems such as slow convergence speed and easy falling into local optimum. At the same time, traditional path optimization generally lacks pre-prediction of machining constraints and potential risks, making it difficult to avoid residual uncut area risks or dropping risks during path planning, resulting in a large room for improvement in wire cutting path optimization.

[0004] Therefore, how to provide a wire cutting path intelligent optimization method based on deep learning image recognition is a problem that those skilled in the art need to solve. SUMMARY

[0005] One object of the present application is to provide a wire cutting path intelligent optimization method based on deep learning image recognition. The present application realizes joint recognition and modeling of workpiece images and electronic drawings, constructs a machining graph and introduces a quantum heuristic neural network for optimization, combines simulation and correction mechanisms, and outputs an optimal cutting scheme that meets process constraints and is safe and efficient.

[0006] According to the wire cutting path intelligent optimization method based on deep learning image recognition of the present application, the following steps are included: Obtain workpiece multi-view images and import electronic drawings, recognize the workpiece multi-view images, extract workpiece structure information, establish a unified coordinate model in combination with electronic drawing reference points, and output an element list; Based on the element list and machine tool and material parameters, a machining graph is constructed with contour segments as nodes and feasible transitions between segments as edges; Generate candidate cutting paths on the machining graph according to machining constraints, determine the lead wire position and bridge distribution, and form an initial cutting scheme; Input the machining graph into a quantum heuristic neural network to optimize and select the initial cutting scheme, and output an optimal cutting scheme; Geometric simulation and energy simulation are performed on the optimal cutting scheme to detect the cross of kerf, the residual uncut area and the risk of material falling, and local correction is performed when there are constraints that do not meet the requirements.

[0007] Further, the workpiece structure information extraction process comprises: The collected images are preprocessed by distortion correction and illumination normalization; An instance segmentation model combining convolutional neural network and attention mechanism is used to segment the workpiece contour boundary, inner hole, slot and step surface; The segmentation results are refined by edge detection and curve fitting; The workpiece structure information includes the workpiece contour boundary, inner hole position and aperture, slot center line and width, step surface geometric boundary and workpiece reference point coordinates.

[0008] Further, the step of establishing a unified coordinate model combined with the electronic drawing reference points comprises: The reference point coordinates marked in the electronic drawing are extracted; The reference points obtained by multi-view image recognition of the workpiece are matched correspondingly; The image recognition results and the electronic drawing are aligned in coordinates based on the reference point coordinate relationship to form a unified coordinate model.

[0009] Further, the element list includes a geometric element table, a process constraint table and a reference mapping table, wherein: The geometric element table records the geometric dimensions and position parameters of the outer contour section, inner hole, slot and step surface; The process constraint table records the size tolerance and surface roughness grade associated with each element; The reference mapping table records the absolute coordinates of the elements in the machine tool coordinate system and the corresponding relationship with the electronic drawing reference.

[0010] Further, the step of constructing a machining graph based on the element list and machine tool and material parameters comprises: The outer contour section, inner hole section, slot section and step boundary determined by image recognition and unified coordinate model are taken as nodes of the graph structure; The transition distance between sections calculated based on the node endpoint coordinates and the feasible movement relationship determined by combining the machine tool stroke constraint are taken as edges of the graph structure; The geometric dimensions, curvature and thickness parameters obtained from the recognition results and the corresponding tolerance grade are recorded in the node attributes; The transition distance calculated based on the endpoint coordinates, whether it needs to rise and fall according to the machine tool process rule, the estimated discharge energy consumption based on the material thickness and machine tool energy parameters, and the risk score calculated based on the geometric and material characteristics are recorded in the edge attributes; forming a directed weighted graph with machining constraint information, i.e., a machining graph.

[0011] Further, the machining constraint includes: dimensional tolerance and surface roughness requirement; machine tool stroke range and discharge energy threshold; bridge location minimum width, maximum spacing and forbidden area condition; wire lead-in angle and wire withdrawal safety condition.

[0012] Further, the step of generating candidate cutting paths on the machining graph according to the machining constraint includes: limiting the machining level of each contour segment according to the dimensional tolerance and surface roughness requirement; According to the machine tool stroke range and discharge energy threshold, the transition edges between segments in the machining graph are screened, and the edges that do not meet the machine tool stroke range and discharge energy threshold are removed, and only the executable transition relationship is retained; According to the bridge location constraint condition, candidate bridge location points are automatically inserted on the outer contour and inner hole contour, and the number and spacing of the bridge locations are determined according to the minimum bridge location width and maximum bridge location spacing requirement, and no bridge location is set in the forbidden area; Under the premise of meeting the constraints, heuristic path generation rules are used to form candidate cutting paths, and the heuristic rules include: Generate the path according to the clockwise traversal of the contour closure order; Generate the path according to the nearest neighbor splicing of the minimum transition distance between nodes; In the complex geometric structure, the inner hole is preferentially processed, and the level traversal strategy of the outer contour is processed; Determine the wire lead position in the candidate path, which is set in the wire leadable area of the outer contour boundary and meets the constraint conditions of the wire lead-in angle and the wire withdrawal safety; Combine the path with the bridge location and the wire lead scheme to form an initial cutting scheme.

[0013] Further, the step of optimizing the initial cutting scheme to generate an optimal cutting scheme using a quantum heuristic neural network includes: Convert the initial cutting scheme into an input vector and perform normalization processing; Map the input vector to an amplitude representation in the form of a superposition of quantum states, with each candidate path corresponding to an amplitude component; During the evolution of the quantum state, a comprehensive evaluation function is constructed according to four types of objectives: machining time, energy consumption, surface roughness and material loss risk, and a penalty operation is applied to path components that do not meet the constraint conditions; Perform an enhancement operation on path components that meet the constraint conditions and have a comprehensive evaluation function value higher than a threshold value to increase the amplitude amplitude; After multiple rounds of evolution, the output stable converges to a quantum state distribution, and the path component with the largest probability amplitude is selected from the measurement operation to obtain the corresponding optimal cutting scheme.

[0014] Further, the quantum heuristic neural network comprises: The constraint-driven amplitude encoding layer maps the workpiece profile segment, bridge position and lead position to the amplitude components of independent registers respectively, and embeds the size tolerance, bridge position width and machine stroke constraint in the encoding process; The process constraint interference unit applies phase inversion to the path component that violates the stroke, energy or bridge position constraint during quantum state evolution, and keeps the phase of the path component that meets the condition unchanged; The risk adaptive amplitude regulation layer enhances or attenuates the amplitude of the corresponding path component according to the residual uncut area risk and the dropping risk; The measurement-reconstruction coupling mechanism performs local reconstruction on the candidate path with similar phase but insufficient amplitude before quantum state measurement, and then performs collapse output.

[0015] Further, the step of simulating and correcting the optimal cutting scheme comprises: In the geometric simulation stage, the cutting path geometric topology based on the unified coordinate model is established, and whether there is a slit intersection and an isolated area between different cutting paths is detected; In the energy simulation stage, the local energy distribution and temperature change in the cutting process are evaluated based on the cutting parameters and material thermal properties; The geometric simulation and energy simulation results are risk judged, and if the residual uncut area or the dropping risk is detected, the local correction operation is triggered; The local correction operation includes reassigning the bridge position, adjusting the cutting order or inserting an auxiliary transition segment.

[0016] The beneficial effects of the present application are: Through multi-view image recognition combined with electronic drawings, the complete structure information of the workpiece can be automatically extracted, manual intervention is reduced, the workpiece profile, transition segment and process constraint are explicitly expressed in the form of processing drawings, and the path generation is more flexible and expandable.

[0017] The quantum heuristic neural network is introduced, the quantum state evolution mechanism is used to process large-scale combination problems in the path optimization process, the convergence is faster than that of the traditional optimization algorithm, the ability to jump out of local optimum is stronger, combined with geometric simulation and energy simulation, potential risks can be predicted before processing, and scheme closed-loop optimization is realized through local correction, and the processing stability and safety are improved.

[0018] The overall scheme of the present application can be widely applied to mold manufacturing and complex part processing scenes, and has high industrial application value. BRIEF DESCRIPTION OF DRAWINGS

[0019] The accompanying drawings are included to provide a further understanding of the application and are incorporated in and constitute a part of this specification, illustrate embodiments of the application and are meant to explain the application without limiting the application to the exact embodiments shown. In the drawings:

[0020] Fig. 1 A flow chart of a linear cutting path intelligent optimization method based on deep learning image recognition is proposed for the present application; Fig. 2 A quantum heuristic neural network optimization structure diagram of a linear cutting path intelligent optimization method based on deep learning image recognition is proposed for the present application; Fig. 3 A candidate cutting path generation and constraint diagram of a linear cutting path intelligent optimization method based on deep learning image recognition is proposed for the present application. DETAILED DESCRIPTION

[0021] The application will now be described in further detail with reference to the drawings. These drawings are simplified schematic diagrams which show the basic structure of the application in a schematic manner only, and therefore only show the components relevant to the application.

[0022] Reference Figs. 1-3 A linear cutting path intelligent optimization method based on deep learning image recognition, comprising the following steps: Obtain multi-view images of a workpiece and import an electronic drawing, identify the multi-view images of the workpiece, extract workpiece structure information, establish a unified coordinate model in combination with the electronic drawing reference points, output an element list, realize multi-source information fusion of complex workpieces, make the identified geometric structure consistent with the design reference, and improve the accuracy and repeatability of subsequent path generation; Based on the element list and machine tool and material parameters, a machining graph is constructed with contour segments as nodes and feasible transitions between segments as edges, the geometric features of the workpiece and the process conditions are converted into a calculable graph structure, the cutting problem can be expressed in a unified data model, and subsequent constraint screening and optimization processing are facilitated; According to the machining constraints, candidate cutting paths are generated on the machining graph, the lead wire position and bridge site distribution are determined, an initial cutting scheme is formed, it is ensured that the generated candidate paths meet the size tolerance, energy threshold and bridge site safety constraints, etc., to avoid generating paths that cannot be executed or have safety hazards, and the feasibility of the scheme is improved; The machining graph is input into a quantum heuristic neural network, the initial cutting scheme is optimized and selected, and the optimal cutting scheme is output, the large-scale candidate path space is searched efficiently through the quantum heuristic mechanism, the traditional algorithm is prevented from falling into local optimization, and the global optimality and convergence speed of the cutting path planning are improved; Geometric and energy simulations are performed on the optimal cutting scheme to detect risks such as kerf intersections, residual uncut areas, and material loss. Local corrections are performed when there are situations that do not meet the constraints. Risk prediction and correction of the cutting scheme are carried out before actual processing to eliminate potential processing defects in advance and ensure the stability and processing safety of the final scheme.

[0023] In this embodiment, the workpiece structure information extraction process includes: The acquired images were preprocessed with distortion correction and illumination normalization: A pinhole imaging model based on camera calibration parameters is adopted. The intrinsic and extrinsic parameter matrices are obtained by Zhang's orthogonal calibration method, and barrel distortion and pincushion distortion are corrected by the inverse mapping method. In terms of illumination normalization processing, an adaptive histogram equalization-based brightness equalization algorithm is used to enhance the image, while Gamma correction is combined to perform nonlinear compression of brightness to eliminate interference caused by shadows and local overexposure.

[0024] An instance segmentation model combining convolutional neural networks and attention mechanisms is used to segment the workpiece's external boundary, internal holes, slots, and stepped surfaces. In this embodiment, the instance segmentation model adopts a deep segmentation architecture combining convolutional neural networks and attention mechanisms, and is composed of a feature extraction layer, a region proposal layer, an attention enhancement module, and a mask generation layer. Feature extraction layer: The backbone network uses an improved ResNet-50 (which can be replaced by ResNet-34 / 101), removing the final fully connected layer; the convolutional kernel size is mainly 3×3, with a stride of 1 or 2, and batch normalization and ReLU / SiLU activation. To obtain multi-scale representations, a feature pyramid (FPN) is introduced, generating feature layers P2 to P5 from C2 / C3 / C4 / C5, with the number of channels set to 128 / 256 / 256 / 256 (adjustable within ±64).

[0025] Shallow enhancement: Add an edge enhancement branch (3×3 Sobel / Scharr learnable filter + 1×1 fusion convolution) to the C2 / P2 bypass to enhance the response of the shape boundary and hole corner.

[0026] Deep semantics: Use dilated convolution (dilation rate 2~4) at C4 / C5 to expand the receptive field, which is beneficial for the overall perception of steps and slots.

[0027] Regional proposal layer: Anchor boxes are generated on each feature pyramid, with sizes preferably {16, 32, 64, 128} pixels, and aspect ratios {1:1, 1:2, 2:1}; each RPN classification and regression head contains 3x3 convolution + 1x1 output. The candidate box screening threshold is: positive sample IoU >= 0.7, negative sample IoU <= 0.3, and the intermediate interval is ignored; the top 1000 candidates (when training) and the top 300 candidates (when reasoning) are taken for each image, and NMS (IoU threshold 0.5~0.7) is performed on the overlapping boxes. The candidate region is normalized to a 14x14 (or 28x28) feature tensor by RoIAlign.

[0028] Attention enhancement module: Channel attention (SE / EMA, compression ratio 16~32) and spatial attention (CBAM or 7x7 convolution attention) are concatenated on the RoI features; channel attention gives higher weight to edge / corner-related channels, and spatial attention suppresses background / reflection areas. The attention module is placed before the mask head, and the calculation overhead is controllable.

[0029] Mask generation layer: The mask head adopts a full convolution structure: 4 layers of 3x3 convolution (each with 256 channels) + up-sampling (x2) + 1x1 classification convolution output K-class mask (K e {outline, inner hole, slot, step}). The mask resolution is 28x28 or 56x56 during training, and is bilinearly interpolated to the original image size in the inference stage. The loss function is composed of: class cross-entropy / focus loss + L1 / smooth L1 bounding box regression loss + mask binary cross-entropy or Dice loss (each loss weight is recommended to be 1:1:2, allowing a ±50% fine-tuning). To improve the boundary quality, a boundary alignment loss (BoundaryIoU or Level-set approximation) can be introduced at the edge of the mask.

[0030] In the embodiment, the processing flow of the instance segmentation model is: The multi-view images collected are first corrected for distortion to restore the actual scale relationship. For example, barrel distortion can cause straight lines on the edge to bend, and after correction, the geometric straight line feature remains unchanged. The essence of illumination normalization is to map the gray scale distribution of different pixels to a unified statistical space, so that the brightness of the same material area is consistent under different shooting conditions, avoiding deviations in subsequent feature extraction. The normalized image is input into the deep learning model in the form of a tensor, and the value of each pixel point corresponds to the reflectivity of the workpiece surface.

[0031] The improved residual convolutional network performs multi-layer convolution on the input tensor, and each layer of convolution kernel extracts geometric textures of different scales. The shallow layer feature retains the high gradient change of the workpiece edge, which is used for detecting the edge of straight line or circular hole; the deep layer feature aggregates global information, which is used for identifying large-scale step area and complex profile. The feature pyramid network ensures that the same position has both local detail features and overall semantic features in this process, which is convenient for subsequent identification of narrow slot and hole diameter and other small structures.

[0032] The region proposal network generates candidate regions on the feature map and standardizes the candidate regions through multi-scale resampling and RoIAlign operation. The classification branch outputs the region category, and the positioning branch outputs the bounding box coordinates. The role of the region proposal network is to automatically generate candidate regions that may contain geometric elements of the workpiece according to the feature map. For example, for the area with a circular hole, RPN will generate a corresponding candidate box according to the local gradient distribution with high response. The generation of the candidate box not only depends on the pre-defined anchor box, but also combines the feature response strength to ensure that small holes are not covered by large boxes. After the candidate regions are aligned by RoIAlign, they are converted into tensors of uniform size and then input into the subsequent classification and segmentation branches.

[0033] The channel attention module calculates the global average value of each channel feature to obtain the importance of the channel to the overall recognition. For example, the channel related to edge detection will obtain a higher weight; the spatial attention module determines the importance of different spatial positions according to the gradient distribution of the pixels, thereby automatically highlighting the workpiece edge lines and reference point regions and suppressing background or light noise.

[0034] In the mask branch, the network outputs a binary mask for each candidate region, which represents the pixel-level "belongs to / does not belong to the geometric element". For example, for the inner hole area, the mask will form a nearly circular distribution. Then, the Canny operator is used to detect the mask boundary to obtain the edge pixel set; and based on the geometric fitting method, the edge points are converted into parameterized curves or circle equations, thereby outputting the center coordinates and diameter of the hole. The slot region is fitted by two parallel boundary lines to obtain the slot center line and width parameters.

[0035] The model outputs the workpiece contour boundary, inner hole position and diameter, slot center line and width, step face geometric boundary, and workpiece reference point coordinates, and stores the above information as an element list for subsequent unified coordinate model construction and machining map generation.

[0036] The segmentation result is refined through edge detection and curve fitting; The workpiece structure information includes the workpiece contour boundary, inner hole position and diameter, slot center line and width, step face geometric boundary, and workpiece reference point coordinates.

[0037] In the embodiment, the step of establishing a unified coordinate model based on the reference points of the electronic drawing includes: The pre-labeled reference point coordinates are parsed and extracted from the electronic drawing. The reference points are the geometric centers, clamping positioning holes, or process reference edges in the workpiece design drawing. The extracted reference points are one-to-one matched with the actual reference points identified from the multi-view image of the workpiece. The correspondence between them is determined by the nearest neighbor matching algorithm. On this basis, according to the spatial mapping relationship of the reference point coordinates, the image recognition result is geometrically aligned by using the coordinate transformation method, so that it is consistent with the electronic drawing. Finally, a unified coordinate model is established, so that the subsequent machining graph construction and cutting path generation are based on the same geometric reference system.

[0038] In the embodiment, the element list includes a geometric element table, a process constraint table, and a reference mapping table, wherein: The geometric element table is used to record the basic geometric information of the workpiece, including the line segment or curve parameters of the outer contour segment, the center coordinates and hole diameter of the inner hole, the center line equation and width of the slot, the boundary curve of the step surface and its spatial position, etc. Each type of geometric element is described in a numerical form.

[0039] The process constraint table supplements the process requirements related to machining accuracy based on the geometric elements, including the size tolerance range and corresponding surface roughness grade of each element, so as to be called during path planning and cutting parameter setting.

[0040] The reference mapping table is used to unify the design coordinates and machine tool coordinates. It records the absolute position of each geometric element in the machine tool coordinate system and establishes the correspondence with the reference points of the electronic drawing, so as to ensure that the machining path generated subsequently can be accurately positioned in the actual machining equipment.

[0041] In the embodiment, the step of constructing a machining graph based on the element list and machine tool and material parameters includes: The outer contour segment, inner hole segment, slot segment, and step boundary determined by image recognition and unified coordinate model are taken as nodes in the graph structure. Each node corresponds to a processable geometric unit. Based on the transition distance between segments calculated from the spatial coordinates of the endpoints of each node, combined with the machine tool stroke constraint condition, it is determined whether there is a feasible movement relationship between different nodes, and the relationship that meets the condition is defined as the edge of the graph structure. In the node attribute, the geometric size, curvature, and thickness parameters obtained from the recognition result are recorded, and the corresponding size tolerance grade is associated, so as to constrain the machining accuracy. In the edge attribute, the transition distance calculated according to the endpoint coordinates, whether the landing operation is required determined by the machine tool process rule, the discharge energy consumption estimated based on the material thickness and the machine tool energy parameter, and the risk score calculated according to the geometric characteristics and material properties are recorded: Geometric complexity risk factor: when the radius of curvature of the area where the transition edge is located is less than a preset threshold (for example, 0.5 mm), there is a sharp angle less than 60° or a slot with a depth-width ratio greater than 5, it is defined as a high-risk area, and the risk score is valued at 0.7-1.0; if the geometric feature size is within a safe range, the score is valued at 0.3 or less.

[0042] Material load risk factor: when the material thickness exceeds 80% of the upper limit of the rated cutting thickness of the machine tool, or the Brinell hardness of the material is greater than a certain threshold (such as HB200), or the electrical conductivity is lower than the standard value, the risk score is increased to 0.6 or more; within the normal processing range, the score is maintained at 0.2-0.4.

[0043] Energy coupling risk factor: when the estimated discharge energy requirement approaches 90% or more of the upper limit of the machine tool energy, or the transition distance exceeds the recommended value (such as 20 mm), the risk score is increased to 0.8-1.0; if the energy requirement and the transition distance are within a safe range, the score is controlled at 0.3 or less.

[0044] In actual application, the three types of risk factors are normalized and linearly combined by preset weights to obtain the comprehensive risk score of the transition edge. The comprehensive risk score is stored in the edge attribute of the processing graph and used as the basis for edge screening and priority determination in the candidate cutting path generation and optimization process.

[0045] Through the construction of the above nodes and edges, a directed weighted graph with processing constraint information, i.e. a processing graph, is finally formed, providing complete data support for the generation and optimization of subsequent cutting paths.

[0046] In this embodiment, the processing constraints include: In terms of precision constraints, the size tolerance range and surface roughness grade of different geometric elements are specified to limit the processing precision that the cutting path must meet; In terms of equipment capability constraints, the maximum travel range of the machine tool and the energy threshold of single discharge are limited to ensure that the cutting path does not exceed the executable capability of the machine tool; In terms of bridge site constraints, the minimum width of the bridge site is required to be greater than a preset lower limit, the spacing between adjacent bridge sites is required to be less than a preset upper limit, and the bridge site is prohibited in the prohibited area with process restrictions; In terms of lead-in and wire withdrawal constraints, the angle of lead-in is required to meet the machine tool wire in angle range, and the safety is ensured in the wire withdrawal process to avoid damage to the workpiece or wire due to unreasonable wire withdrawal path.

[0047] By setting the above constraints, it is ensured that the candidate cutting path is generated under the premise of meeting the geometric, process and equipment requirements.

[0048] In this embodiment, the step of generating a candidate cutting path on the machining graph according to the machining constraints comprises: Classify each contour segment according to dimensional tolerance and surface roughness requirements, and sort them according to machining precision requirements to ensure reasonable cutting order; According to the stroke range of the machine tool and the discharge energy threshold, the transition edges between the segments in the machining graph are filtered one by one, and the edges exceeding the maximum stroke or energy limit of the machine tool are removed, and only the transition relationships that can be executed under the equipment conditions are retained; According to the bridge position constraint condition, automatically insert bridge position candidate points on the outer contour and inner hole contour, and the number and spacing of the bridge positions are determined according to the minimum bridge width and maximum bridge spacing requirements, and no bridge position is set in the prohibited area; Under the premise of meeting the above constraints, a heuristic path generation rule is used to form a candidate cutting path, wherein the heuristic rule comprises: Generate the path according to the clockwise traversal method of the contour closure sequence; Generate the path according to the nearest neighbor splicing method of the minimum transition distance between nodes; And the hierarchical traversal strategy of preferentially machining the inner hole and then machining the outer contour in complex geometric structures.

[0049] Determine the lead position in the candidate path, and the lead position is arranged in the tool accessible area of the outer contour boundary and meets the constraint conditions of the lead-in angle and the lead-out safety; Combine the above path scheme with the configuration of the bridge and the lead to form an initial cutting scheme.

[0050] In this embodiment, the step of optimizing the initial cutting scheme to generate an optimal cutting scheme using a quantum heuristic neural network comprises: The initial cutting scheme composed of candidate paths, bridge locations and lead positions is encoded as an input vector and normalized to ensure that different dimensional features are in the same numerical range. For processing time, energy consumption, surface roughness and material loss risk, the mean-variance standardization method is used, i.e. by calculating the mean and standard deviation of the index in the sample set, each index value is subtracted from the mean and then divided by the standard deviation, so that the index is converted into a dimensionless variable with a mean of zero and a variance of one, thereby avoiding weight deviation caused by value magnitude differences. For geometric size, transition distance and other edge attribute parameters, the interval normalization method is used, i.e. by calculating the maximum and minimum values of the parameters in the sample set, each parameter value is subtracted from the minimum value and then divided by the difference between the maximum and minimum values, thereby scaling the parameters to the interval of zero to one to ensure that all input features are in the same dimension range.

[0051] On this basis, for discrete elements such as bridge layout scheme and lead position selection, a one-hot encoding method is used to convert them into vector representation, which is then concatenated with continuous features after normalization to form a complete input vector for subsequent mapping and evolution in the quantum-inspired neural network.

[0052] The input vector is mapped to an amplitude representation in the form of a superposition of quantum states, with each candidate path corresponding to an amplitude component of a quantum state, thereby enabling parallel expression in the quantum-inspired neural network. In this embodiment, the step of mapping the input vector to an amplitude representation in the form of a superposition of quantum states includes dividing the normalized input vector into sub-vectors corresponding one-to-one to the candidate paths, each sub-vector containing feature parameters such as processing time, energy consumption, surface roughness and material loss risk. The comprehensive evaluation value of each sub-vector is converted into a probability amplitude, specifically: The proportion of each sub-vector in the total sum of the comprehensive evaluation values of all sub-vectors is calculated, which is used as the initial amplitude size of the sub-vector in the quantum state. The amplitudes of all candidate paths are normalized according to the square additivity rule, so that the sum of the squares of the amplitudes of all candidate paths is equal to one, thereby forming a legal superposition representation of quantum states. Each candidate path corresponds to an amplitude component of a quantum state, and all candidate paths are expressed in parallel in the form of a superposition in the quantum-inspired neural network.

[0053] In the evolution process of the quantum state, a comprehensive evaluation function is constructed, which considers four types of optimization objectives: processing time, energy consumption, surface roughness and material loss risk. Comprehensive evaluation function (candidate path ) ; wherein, ... is a weighting coefficient, set by the administrator.

[0054] 1) Processing time score ; wherein, represents the total length of the path ( ), represents the nominal cutting speed (determined from the process table according to material / thickness); represents the upper limit of the reference time (high quantile of historical qualified samples or given by the process specification).

[0055] 2) Energy consumption score ; wherein, represents the energy consumption per unit length (obtained from the material and thickness in the energy consumption table, missing entries can be interpolated from the adjacent ones); : upper limit of the reference energy consumption (process upper limit).

[0056] 3) Surface roughness score ; wherein, represents the main parameter entry used along the whole path (the discharge parameter combination with the highest length coverage); represents the target roughness corresponding to this entry (given by the process table); : upper limit of the allowed roughness.

[0057] 4) Material loss risk score ; ; wherein, represents the maximum island area obtained from the simulation according to the path order; represents the total area of the workpiece; represents the longest unsupported span along the outer / inner contour; represents the recommended upper limit of the span; represents the overall rigidity of the bridge (summed up as a simple beam; material elastic modulus, bridge width, thickness, span); represents whether there is a high-risk order of "inner hole first and outer contour later" (1 for yes, 0 for no); represents the amount of deficiency in the relative target number / distribution of bridges; : Regression coefficients (fitted from historical / simulation data); This indicates the maximum allowable probability threshold for material loss.

[0058] For path components that do not meet the processing constraints, a penalty term is introduced into the evolution operator to weaken them, so that their amplitude gradually decreases. For path components that meet the constraints and whose comprehensive evaluation function value is higher than the preset threshold, their probability amplitude is increased through an amplitude enhancement mechanism.

[0059] In this embodiment, a penalty mechanism and an enhancement mechanism are introduced into the evolution process of candidate cutting path components to weaken unqualified paths and highlight excellent paths.

[0060] The penalty is implemented by introducing a scaling factor less than 1 into the evolution operator. This applies to path components that do not meet the processing constraints. Its amplitude is updated during the iteration process according to the following relationship:

[0061] ; in, Representing a path In the Amplitude during wheel iteration The penalty factor has a value range of 1. Through continuous iteration, the amplitude of the path component gradually decreases and eventually approaches zero, thus being naturally eliminated in the final measurement stage.

[0062] The enhancement mechanism is implemented by introducing an amplification factor greater than 1 in conjunction with a comprehensive evaluation function. This applies to cases where the constraints are met and the comprehensive evaluation function value is higher than a preset threshold. Path components Its amplitude is updated during the iteration process according to the following relationship:

[0063] ; in, Representing a path The comprehensive evaluation function value, This is the enhancement factor, used to control the enhancement magnitude. When Significantly greater than When, the amplitude is significantly enhanced; when When the threshold is just met, the enhancement effect is limited, thus ensuring that the enhancement operation has discriminative power.

[0064] By introducing the above-mentioned punishment and enhancement mechanism in the evolution operator, the amplitude of the unqualified path is gradually weakened, and the amplitude of the excellent path is gradually enhanced. After multiple rounds of iterative evolution, the overall quantum state distribution gradually converges and stabilizes, and finally the path component with the largest probability amplitude is obtained in the measurement operation, and the cutting path corresponding to the component is output as the optimal cutting scheme.

[0065] In the embodiment, the quantum heuristic neural network comprises: The constraint-driven amplitude encoding layer is used to map the workpiece outer contour segment, inner hole contour segment, bridge position and lead position to the amplitude components of independent registers, respectively. In the mapping process, the constraint conditions such as dimensional tolerance, minimum width of bridge position and machine tool stroke range are embedded in advance, so that the encoding result itself satisfies the basic geometric and process boundary conditions, thereby reducing invalid solutions in subsequent evolution.

[0066] The process constraint interference unit is used to apply a π phase inversion operation to the path component that violates the stroke constraint, discharge energy threshold or bridge position prohibited area condition, so that it is destructively interfered in the subsequent interference superposition; and the phase of the path component satisfying the constraint condition is kept unchanged, so as to be strengthened in the quantum interference process.

[0067] The risk adaptive amplitude regulation layer is used to apply a decay factor to the path component with residual uncut area risk or dropping risk based on the results of geometric simulation and energy simulation, so that the amplitude gradually decreases; for the path component with low residual risk and high processing stability, an enhancement factor is applied to increase the amplitude. Through the dynamic regulation of this layer, the network can adaptively highlight the low-risk path and suppress the high-risk path, thereby improving the reliability of the final output solution.

[0068] The measurement-reconstruction coupling mechanism is used to perform a local reconstruction operation on the candidate path with similar phase distribution but insufficient amplitude, that is, to adjust and recombine the path segments in the local register to form a new candidate path with more balanced amplitude. After local reconstruction, quantum state collapse measurement is performed, thereby improving the probability of obtaining an approximately optimal path and ensuring that the final output cutting scheme balances between global optimality and local feasibility.

[0069] In the embodiment, the step of simulating and correcting the optimal cutting scheme comprises: In the geometric simulation stage, the geometric topology of the cutting path is established based on the unified coordinate model, whether there is a kerf intersection is judged by calculating the distance relationship between path segments, and whether there is a residual isolated area is judged by counting the uncoated area.

[0070] In the energy simulation stage, the energy input per unit time is estimated according to the cutting parameters and the thermal properties of the material, and is compared with the heat dissipation capacity of the material. If the energy input is significantly greater than the heat dissipation capacity, it is determined that there is a risk of local overheating, and if the energy input is insufficient, it is determined that there is a risk of incomplete cutting.

[0071] When the simulation results show that there are residual uncut areas or material dropping risks, local correction operations are performed, including: reassigning bridge position locations to maintain workpiece stability, adjusting the cutting sequence to reduce path conflicts, or adding transition sections as necessary to ensure cutting continuity.

[0072] Embodiment 1 In order to verify the feasibility of the application in implementation, the application is applied to the wire cutting machining test of a certain mold part. The workpiece material is Cr12MoV mold steel, the thickness is 40mm, and the design drawing contains one outer contour, two round holes and one slot. The machine tool model is DK7750 numerical control wire cutting machine, the wire diameter is 0.18mm, the machine tool stroke range is 500mm×400mm, and the upper limit of single pulse discharge energy is 0.3mJ.

[0073] Firstly, the workpiece shape image is collected by using a multi-view industrial camera, and the CAD electronic drawing is imported. Distortion correction (based on pinhole model calibration parameter correction) and illumination normalization (through Retinex algorithm to balance illumination) are performed on the collected image to obtain a standardized image suitable for segmentation.

[0074] The preprocessed image is input into the instance segmentation model of the improved residual convolutional network combined with the region proposal mechanism. The model outputs the pixel-level mask of the workpiece shape boundary, the inner hole position, the slot contour and the step surface area. Then, edge detection and B-spline fitting are used to obtain the vectorized expression of the geometric elements. Finally, the workpiece structure information obtained includes: the outer contour curve, the center coordinates and hole diameter of each hole, the center line and width of the slot, and the geometric boundary of the step.

[0075] The reference point coordinates marked in the electronic drawing are extracted and matched with the reference points obtained by image recognition, and a unified coordinate model of the drawing and the real object image is established through rigid coordinate transformation. From this, the element list is generated, including:

[0076] Geometric element table: records the size and position parameters of the outer contour, inner hole, slot and step; Process constraint table: records the size tolerance and surface roughness requirements related to each element; Reference mapping table: records the absolute coordinates of the elements in the machine tool coordinate system and the corresponding relationship with the drawing reference.

[0077] The geometric elements are taken as nodes, the distance between segments is calculated according to the coordinates of end points, and the edge connection relationship is established by combining with the machine tool stroke constraint to form a directed weighted graph with machining constraint information. The curvature, thickness and tolerance level are written in the node attribute; the transition distance, whether the wire needs to be lifted, the discharge energy consumption estimation and the risk score calculated based on the geometric complexity and material parameters are written in the edge attribute.

[0078] According to the size tolerance, surface roughness, bridge width and spacing requirements in the process constraint table, the machining graph is screened. The bridge candidate points are automatically inserted in the outer contour and the inner hole, and the lead position is determined in the tool accessible area of the outer contour boundary. Under the premise of meeting the constraints, the heuristic rules of clockwise traversal, nearest neighbor splicing and hierarchical priority are adopted to form multiple candidate cutting paths.

[0079] The candidate paths are combined with the bridge and lead configuration to obtain an initial cutting scheme. The scheme covers all workpiece contours, the path is complete and meets the basic machine tool constraints, but has not been globally optimized.

[0080] The initial scheme is encoded into an input vector, which is mapped to the amplitude of a quantum state after normalization. During the evolution process, the comprehensive evaluation function introduces four types of indicators: machining time, energy consumption, surface roughness and material loss risk; phase penalty is applied to path components that do not meet the constraints to gradually weaken their amplitudes; enhancement operations are performed on path components with a comprehensive evaluation better than the threshold. After multiple iterations, the quantum state distribution converges, and the measurement result selects the path with the maximum probability as the optimal cutting scheme.

[0081] Geometric simulation is performed on the optimal scheme to check the intersection of kerf and isolated areas; then energy simulation is performed to estimate the local temperature distribution using cutting parameters and material thermal properties. If residual uncut areas or material loss risks are found, local modifications are made by redistributing the bridge, adjusting the cutting order or inserting auxiliary transition segments.

[0082] Finally, the optimized scheme is compared with the traditional manual experience path. The results are shown in Table 1.

[0083] Table 1 Comparison of workpiece wire cutting path optimization experimental results

[0084] As shown in Table 1, compared with the manual path scheme, the method of the present application reduces the path length by about 9.5%, shortens the processing time by 13.4%; the unit energy consumption is reduced by 14.3%, and the surface roughness is improved from 2.1 μm to 1.5 μm. The geometric and energy simulation verifies that the residual area and material loss risk are significantly reduced, which proves the effectiveness and superiority of the method in actual processing.

[0085] The above merely describes preferred specific embodiments of the present application, but the protection scope of the present application is not limited thereto, and any person skilled in the art, according to the technical solution and inventive concept of the present application, makes equivalent replacement or change within the technical range disclosed by the present application, which should be covered within the protection scope of the present application.

Claims

1. A method for intelligent optimization of wire cutting paths based on deep learning image recognition, characterized in that, Includes the following steps: Acquire multi-view images of the workpiece and import them into electronic drawings. Recognize the multi-view images of the workpiece, extract the structural information of the workpiece, establish a unified coordinate model with the reference points of the electronic drawings, and output a list of elements. Based on the aforementioned list of elements and machine tool and material parameters, a machining diagram is constructed using contour segments as nodes and feasible transitions between segments as edges. Based on the processing constraints, candidate cutting paths are generated on the processing diagram to determine the lead wire positions and bridge location distribution, thus forming an initial cutting scheme. The processing diagram is input into a quantum heuristic neural network to optimize the initial cutting scheme and output the optimal cutting scheme. Geometric and energy simulations are performed on the optimal cutting scheme to detect the risks of kerf intersection, residual uncut areas, and material loss. Local corrections are performed when there are situations that do not meet the constraints.

2. The intelligent optimization method for wire cutting paths based on deep learning image recognition according to claim 1, characterized in that, The process of extracting workpiece structural information includes: The acquired images are preprocessed with distortion correction and illumination normalization. An instance segmentation model combining convolutional neural networks and attention mechanisms is used to segment the workpiece's external boundary, internal holes, slots, and stepped surfaces. The segmentation results are refined through edge detection and curve fitting; The workpiece structural information includes the workpiece's external boundary, the position and diameter of the inner hole, the center line and width of the slot, the geometric boundary of the step surface, and the coordinates of the workpiece's reference point.

3. The intelligent optimization method for wire cutting paths based on deep learning image recognition according to claim 1, characterized in that, The steps for establishing a unified coordinate model by combining electronic drawing reference points include: Extract the coordinates of the reference points marked in the electronic drawing; Match the corresponding reference points obtained from multi-view image recognition of the workpiece; Based on the coordinate relationship of the reference points, the image recognition results are aligned with the electronic drawings to form a unified coordinate model.

4. The intelligent optimization method for wire cutting paths based on deep learning image recognition according to claim 1, characterized in that, The element list includes a geometric element table, a process constraint table, and a datum mapping table, wherein: The geometric element table records the geometric dimensions and positional parameters of the outer contour segment, inner hole, slot and step surface; The process constraint table records the dimensional tolerances and surface roughness grades associated with each element; The reference mapping table records the absolute coordinates of elements in the machine tool coordinate system and their correspondence with the reference of the electronic drawing.

5. The intelligent optimization method for wire cutting paths based on deep learning image recognition according to claim 1, characterized in that, The steps for constructing the machining drawing based on the element list and machine tool and material parameters include: The outer contour segment, inner hole segment, slot segment, and step boundary determined by image recognition and unified coordinate model are used as nodes of the graph structure. The transition distance between segments, calculated using the coordinates of the node endpoints, and the feasible movement relationship determined by the machine tool travel constraints are used as the edges of the graph structure. Record the geometric dimensions, curvature and thickness parameters obtained from the recognition results, as well as the corresponding tolerance levels, in the node attributes; The edge attributes record the transition distance calculated based on the endpoint coordinates, whether wire lifting or lowering is required as determined by the machine tool process rules, the discharge energy consumption estimated based on material thickness and machine tool energy parameters, and the risk score calculated based on geometry and material properties. This forms a directed weighted graph with processing constraints, i.e., a processing graph.

6. The intelligent optimization method for wire cutting paths based on deep learning image recognition according to claim 1, characterized in that, The processing constraints include: Dimensional tolerances and surface roughness requirements; Machine tool travel range and discharge energy threshold; Minimum width, maximum spacing, and prohibited area conditions for bridge sites; Lead wire feed angle and wire unwinding safety conditions.

7. The intelligent optimization method for wire cutting paths based on deep learning image recognition according to claim 1, characterized in that, The step of generating candidate cutting paths on the machining drawing based on machining constraints includes: The machining levels for each contour segment are defined by dimensional tolerances and surface roughness requirements; Based on the machine tool travel range and discharge energy threshold, the transition edges between segments in the machining drawing are screened, and the edges that are within the machine tool travel range and discharge energy threshold are removed, leaving only the executable transition relationships. Based on the bridge location constraints, candidate bridge locations are automatically inserted on the outer contour and inner hole contour. The number and spacing of bridge locations are determined according to the minimum bridge location width and maximum bridge location spacing requirements. No bridge locations are set in the prohibited area. Under the premise of satisfying constraints, heuristic path generation rules are used to form candidate cutting paths, wherein the heuristic rules include: Generate a path by traversing clockwise according to the closing order of the contours; Paths are generated by splicing the nearest neighbors with the minimum transition distance between nodes. A hierarchical traversal strategy that prioritizes processing internal holes and then processes the outer contour in complex geometric structures. The lead wire position is determined in the candidate path. The lead wire position is set in the infeed area of ​​the outer contour boundary and meets the constraints of wire feed angle and wire retraction safety. The path is combined with the bridge location and lead wire scheme to form the initial cutting scheme.

8. The intelligent optimization method for wire cutting paths based on deep learning image recognition according to claim 1, characterized in that, The steps of using quantum heuristic neural networks to optimize the initial cutting scheme and generate the optimal cutting scheme include: The initial cutting scheme is transformed into an input vector and then normalized. The input vector is mapped to an amplitude representation in the form of a quantum state superposition, with each candidate path corresponding to an amplitude component; During the quantum state evolution process, a comprehensive evaluation function is constructed based on four objectives: processing time, energy consumption, surface roughness, and material loss risk. Penalty operations are applied to path components that do not meet the constraints. For path components that meet the constraints and whose comprehensive evaluation function value is higher than the threshold, an enhancement operation is performed to increase the amplitude. After multiple rounds of evolution, a stable and convergent quantum state distribution is output, and the path component with the largest probability amplitude is selected from it through measurement operations to obtain the corresponding optimal cutting scheme.

9. The intelligent optimization method for wire cutting paths based on deep learning image recognition according to claim 1, characterized in that, The quantum heuristic neural network includes: The constraint-driven amplitude encoding layer maps the workpiece contour segment, bridge position, and lead wire position to the amplitude components of independent registers, and embeds dimensional tolerances, bridge position width, and machine tool travel constraints during the encoding process. The process-constrained interference unit applies phase reversal to path components that violate path, energy, or bridge position constraints during quantum state evolution, while keeping the phase unchanged for path components that meet the conditions. The risk-adaptive amplitude control layer enhances or attenuates the amplitude of the corresponding path component based on the risk of residual uncut area and the risk of material loss. The measurement-reconstruction coupling mechanism performs local reconstruction on candidate paths with similar phases but insufficient amplitudes before quantum state measurement, and then performs collapse output.

10. The intelligent optimization method for wire cutting paths based on deep learning image recognition according to claim 1, characterized in that, The steps for simulating and correcting the optimal cutting scheme include: In the geometric simulation stage, a cutting path geometric topology based on a unified coordinate model is established to detect whether there are kerf intersections and isolated areas between different cutting paths; In the energy simulation phase, the local energy distribution and temperature changes during the cutting process are evaluated based on the cutting parameters and material thermal properties. The results of geometric simulation and energy simulation are used to assess risks. If residual uncut areas or material loss risks are detected, a local correction operation is triggered. The local correction operations include reallocating bridge positions, adjusting the cutting sequence, or inserting auxiliary transition segments.

Images