Ship steel plate digital cutting and typesetting optimization method and system

Abstract

The invention discloses a ship steel plate digital cutting and typesetting optimization method and system, and the method comprises the steps: collecting the morphology data of a to-be-cut steel plate, and generating a machining region model; based on the machining area model and the specification data of the target part, a deep reinforcement learning-genetic hybrid algorithm is adopted as a solving strategy, and an optimal typesetting scheme is output; generating a numerical control cutting path and a corresponding numerical control code according to the preferable typesetting scheme, and carrying out kerf gap constraint and thermal deformation compensation on the numerical control cutting path; based on the deviation between the ideal coordinates of the reference mark points in the machining area and the real coordinates of the reference mark points projected on the to-be-cut steel plate, the to-be-cut steel plate is positioned and calibrated; and executing a cutting instruction, evaluating a cutting result, and feeding back deviation data to adjust and optimize strategy parameters. On the premise of ensuring the process feasibility, limit utilization of the steel plate is achieved, the cost is reduced, the cutting quality is improved, and the method is suitable for shipbuilding scenes with various complex parts and high cutting precision requirements.

Description

TECHNICAL FIELD

[0001] The present application relates to the field of ship manufacturing, in particular to a ship steel plate digital cutting layout optimization method and system. BACKGROUND

[0002] In the process of ship production, most of the hull parts are produced by cutting the required size of the plate from a large steel plate and then welding / bolting, which involves the problem of how to arrange different shapes and sizes of target plates on a large steel plate. The existing ship steel plate cutting layout technology mainly has the following defects: 1. Low material utilization rate: Traditional layout mainly relies on manual experience or simple heuristic algorithms to arrange parts on the steel plate, which is difficult to fully utilize the steel plate material, and a lot of scattered residual materials are left after cutting operation, especially when dealing with irregular-shaped plates, leaving large irregular-shaped areas on the steel plate, resulting in more material waste and higher production cost.

[0003] 2. Insufficient algorithm optimization: Common automatic layout algorithms include heuristic rules, genetic algorithms, simulated annealing, etc., but these algorithms are prone to local optimization for complex layout problems (NP difficult problem), and cannot obtain the highest material utilization rate; in addition, these algorithms usually output the scheme at one time, lack self-learning ability, and cannot continuously improve the layout strategy according to the feedback of the cutting result.

[0004] 3. Non-optimized cutting path: The existing numerical control cutting path is usually generated directly according to the layout result, without fully optimizing the cutting sequence, resulting in long empty path and low cutting efficiency; in addition, there is also a lack of compensation mechanism for the thermal deformation of the steel plate during cutting, which will expand and deform after being cut at high temperature. If not compensated, it will affect the cutting accuracy and cause the actual size deviation of the parts.

[0005] 4. Positioning accuracy problem: The steel plate is usually placed on the cutting platform with position and angle deviation, if only manual positioning or single camera visual positioning is used, the cutting path may not match the actual plate position, affecting the cutting quality and production safety, and the existing technology lacks high-precision automatic calibration means, which leads to the increase of part gap or the reservation of excess edge, further hindering the realization of zero excess cutting.

[0006] 5. Lack of feedback mechanism: The traditional cutting system rarely analyzes and feeds back the result after each cutting, so it cannot adjust the subsequent layout strategy according to the actual excess material; for example, if it is found that the actual excess material is more than expected after cutting, the system will not "learn" this deviation, so that similar problems may still occur in subsequent layout.

[0007] 6、System integration is low: different links of steel plate cutting (such as scanning measurement, layout calculation, numerical control cutting, calibration, feedback) are often completed by different equipment and software respectively, the data interface is not unified, which affects the production efficiency; for example, the plate shape obtained by scanning needs to be manually imported into the layout software, and the cutting code is generated by another system, and there is a lack of unified communication protocol and data format to realize real-time cooperation of each module.

[0008] In view of the above problems, the industry urgently needs a new technical solution that can automatically obtain the shape of the steel plate, intelligently optimize the part layout to maximize material utilization, and generate an optimized cutting path, while also calibrating the actual cutting position, monitoring and feeding back the cutting result, forming a closed loop self-learning with the optimization algorithm, and continuously improving the efficiency and accuracy of layout and cutting. SUMMARY

[0009] The purpose of the present application is to provide a digital cutting and layout optimization technology for ship steel plates to overcome the shortcomings of existing ship steel plate nesting and numerical control cutting technology.

[0010] To achieve the above purpose, the present application discloses a digital cutting and layout optimization method for ship steel plates, comprising: Step 1, collecting the topographic data of the steel plate to be cut, and generating a processing area model; Step 2, based on the processing area model and the specification data of the target part, using a deep reinforcement learning-genetic hybrid algorithm as a solving strategy to output an optimized layout scheme; Step 3, generating a numerical control cutting path and corresponding numerical control code according to the optimized layout scheme, and performing kerf gap constraint and thermal deformation compensation on the numerical control cutting path; Step 4, based on the deviation between the ideal coordinates of the reference marker points in the processing area and the real coordinates of their projections on the steel plate to be cut, positioning and calibrating the steel plate to be cut; Step 5, executing the cutting instruction, evaluating the cutting result, and feeding back the deviation data to adjust the optimization strategy parameters.

[0011] Preferably, the step 1 comprises: The steel plate to be cut is a whole plate or a leftover plate after the last cutting, and the plate contour coordinate point cloud is obtained by a laser scanner or machine vision to generate a processing area model, and the expression is as follows: ; Wherein, is the boundary coordinate of the processing area.

[0012] Preferably, the step 2 comprises: Input the processing area model and the specification data of the target part, adopt a deep reinforcement learning agent and a genetic algorithm to explore the optimal layout mode of the target part on the steel plate, i.e., all the laid-out parts do not overlap with each other and the unused space in the processing area is the least, and the specific process is as follows: The reinforcement learning agent arranges the parts in the processing area one by one to form an action sequence , N is the total number of parts to be laid out, and a blank processing area without part arrangement is taken as an initial state , and the action is to arrange the first part in the processing area, and the new state is recorded after the action. In the early stage of the action, the parts are arranged randomly or based on heuristic rules, and the reward is calculated according to the material utilization rate and the part conflict situation after each arrangement, wherein a positive reward is given if the material utilization rate increases or there is no conflict, and a negative reward is given if the material utilization rate decreases or a conflict occurs, and the reinforcement learning agent adjusts the strategy accordingly to try different layout modes; The genetic algorithm is used to update the strategy network parameters of the feasible layout mode given by the reinforcement learning agent, and after a certain number of iterations, the better layout modes are crossed and mutated to generate new layout modes, and after a certain number of iteration training and evaluation, the optimized part layout parameters are output as the preferred layout scheme , wherein the part layout parameters include the rotation angle , the horizontal translation amount , and the vertical translation amount .

[0013] Preferably, the preferred layout scheme satisfies the zero tolerance condition, and the expression is as follows: ; wherein, is the area where the tth part is located after rotation and translation, and Area is a function of the area of the region.

[0014] It should be noted that the definition of zero tolerance in the present application does not mean that there is really no tolerance, but that all parts are placed in the steel plate processing area , do not overlap with each other, and the unused area is only a very small grid skeleton formed by laser cutting, and the excess material area is almost zero, and the specific judgment threshold is set according to the actual production situation.

[0015] Preferably, the step 3 comprises: According to the preferred layout scheme, the position of each part in the machining area is determined, the contour line of each part is divided into multiple line segments according to the nodes, and all line segments are arranged in sequence to generate a numerical control cutting path, and the planning of the numerical control cutting path needs to meet the following requirements: ①For any two adjacent parts, keep , where is the minimum distance between the contour edges of part i and part j, is the minimum gap allowed by the process; ②Based on the thermal expansion effect of steel, the length of the cutting path is corrected and compensated, that is, , where is the compensated path length, is the linear expansion coefficient of steel, is the original length of the cutting path before compensation, is the elevated temperature of the cutting area; According to the numerical control cutting path, the corresponding numerical control code (such as G code or other NC code) is generated, and the code file is uploaded to the numerical control cutting machine controller.

[0016] Preferably, the step 3 uses the 2-Opt algorithm to optimize the cutting sequence, and on the basis of the initial path scheme, the cutting sequence of the local path is repeatedly exchanged to reduce the length of the empty path, until the path cannot be optimized or the optimization increment is lower than the threshold value.

[0017] It should be noted that by introducing the 2-Opt algorithm, the idle time of the cutting head moving between part contours is minimized, thereby improving cutting efficiency and reducing energy consumption.

[0018] Preferably, the step 4 includes: Select at least three reference marker points in the machining area and project them onto the surface of the steel plate to be cut, and record the ideal coordinates of the reference marker points ; Obtain the image of the steel plate to be cut by machine vision, and identify the real coordinates of the projection points on the steel plate ; According to the coordinate difference, calculate the rigid body transformation parameters of each projection point relative to the ideal position, and the expression is as follows: ; , where is the rotation angle, is the horizontal translation, is the vertical translation; The least squares method is used to fit the solution of the rigid body transformation to obtain the estimated values of the translation and rotation angle , calculate the positioning error and compare it with the preset accuracy threshold, and the expression is as follows: ; When all the above conditions are met, the calibration is determined to be passed, and the coordinates of the reference mark points are uploaded to the CNC cutting machine controller, the controller triggers the cutting execution instruction and performs coordinate transformation compensation on the cutting path to be executed. On the contrary, when at least one of the above conditions is not met, the calibration is determined to be failed, and an alarm is prompted to adjust the steel plate placement position or add more reference mark points to recalculate the positioning error until the accuracy requirement is met.

[0019] Preferably, the step 5 comprises: The CNC cutting machine performs cutting operation according to the generated numerical control code, the cutting head cuts out the part profile from the steel plate edge or the pre-punching hole, and the cutting head moves to the next part profile starting point along the shortest path, after the operation is completed, all parts are taken out, and the remaining material is reserved on the cutting machine table; The morphology data of the remaining material is obtained by a laser scanner or machine vision, the coincidence degree index of the real remaining material and the expected theoretical remaining material in the optimal layout scheme is calculated, and the expression is as follows: ; Wherein, IoU is the coincidence degree index, Area is the function of the area of the region, The theoretical remaining material is ; When the IoU is less than the preset coincidence threshold, the difference region between the real remaining material and the theoretical remaining material is discretely processed by rasterization, and the difference raster image is obtained, and the expression is as follows: ; Wherein, is the set difference operation, is the discrete rasterization operation; The difference raster image is input into the neural network of the reinforcement learning intelligent agent for training, so that the algorithm identifies the situation that the difference region is not effectively utilized in the scheme, and adjusts the subsequent layout strategy parameters accordingly.

[0020] Preferably, the function expression for calculating the reward in step 2 is as follows: ; Wherein, is the reward value of the tth action, is the increment of material utilization rate after the tth action compared with the previous step, is a conflict indication function, which takes 1 when the newly placed part overlaps with the placed part, otherwise it takes 0, is the current cumulative cutting time, is the maximum allowed cutting time, ​respectively, are corresponding weight coefficients.

[0021] It should be noted that the present application guides the reinforcement learning agent to optimize the decision towards the goal of zero residual efficient layout by designing a specific reward function, which is an evaluation index considering material utilization, collision conflict and cutting efficiency, the material utilization is the proportion of the total area of the laid-out parts to the steel plate area, represents the proportion of the current accumulated cutting path length or cutting time to the maximum allowed value, reflecting the cutting efficiency, determined by actual experience, to adjust the punishment and reward degree brought by different state changes. Through the above reward function design, the agent can consider improving material utilization, avoiding part conflict and reducing cutting complexity during layout, and gradually learn better layout strategy during the training process.

[0022] Preferably, the deep reinforcement learning-genetic hybrid algorithm comprises reinforcement learning based on policy gradient update and global search of genetic algorithm, and the expression of the policy gradient update is as follows: ; Wherein, is the policy network parameter, is the policy function The gradient of the parameter , is the probability distribution function of the policy network selecting action under the state , is the logarithmic derivative of the probability distribution function with respect to the parameter , is the mathematical expectation of the action-state distribution under the policy , is the advantage function and is used to evaluate the good and bad of a specific action relative to the average level of the state, and the expression is as follows: ; Wherein, is the discount factor, and are the estimated state values of the value network for the current state and the next state .

[0023] It should be noted that the reinforcement learning part of the hybrid algorithm in the present application adopts the Actor-Critic architecture, which utilizes the policy network to output the action probability distribution and the value network to estimate the state value, thereby guiding the policy improvement, and through the above policy gradient update formula, the policy network parameter is updated towards the direction of improving the long-term cumulative reward.

[0024] Preferably, the genetic algorithm is based on the current strategy network parameter population, and new strategy individuals are generated through selection, crossover and mutation operations, and the expression is as follows: ; Wherein, is the new strategy network parameter generated after genetic operation, and is the parameter of the i, j-th strategy individual selected from the current strategy population, is the crossover ratio coefficient and , is the Gaussian noise mutation term.

[0025] It should be noted that the disturbance optimization of the strategy network parameter is realized by the above-mentioned genetic update, so that the strategy network parameter is further diversified and explored on the basis of the reinforcement learning gradient update, so as to jump out of the local optimum that the reinforcement learning may fall into, and the probability of finding a globally better solution is improved.

[0026] The application also discloses a ship steel plate digital cutting layout optimization system, comprising: A plate appearance acquisition module acquires the appearance data of the steel plate to be cut and generates a processing area model; An intelligent layout optimization module uses a deep reinforcement learning-genetic hybrid algorithm as a solving strategy based on the processing area model and the specification data of the target part, and outputs an optimal layout scheme; A cutting path generation module generates a cutting path according to the optimal layout scheme and corresponding numerical control code, and performs seam gap constraint and thermal deformation compensation on the numerical control cutting path; A projection positioning calibration module calibrates the steel plate to be cut based on the deviation between the ideal coordinates of the reference marker points in the processing area and the real coordinates of the projection of the reference marker points on the steel plate to be cut; A feedback self-learning module executes the cutting instruction, evaluates the cutting result, and feeds back the deviation data to adjust the optimization strategy parameters.

[0027] In the above system: The plate appearance acquisition module acquires the appearance data of the steel plate to be cut, and outputs a digital representation (such as a polygonal contour or a raster image) of the steel plate area as the input basic data for subsequent layout optimization.

[0028] The intelligent layout optimization module takes the deep reinforcement learning-genetic hybrid algorithm as the core, repeatedly explores through the reinforcement learning agent and globally searches through the genetic algorithm, constantly improves the layout scheme, and finally outputs the optimal or near-optimal layout result that meets the conditions.

[0029] The cutting path generation module generates a cutting path corresponding to the preferred layout scheme, avoids thermal fusion adhesion between adjacent parts due to excessive proximity by lower limit constraint of the slit gap, ensures cutting quality, and eliminates the influence of thermal expansion of the steel plate caused by cutting operation by compensation correction of the cutting path length, and ensures accurate actual cutting size.

[0030] The projection positioning calibration module calibrates the real pose of the steel plate before cutting execution, aligns the real coordinate system of the steel plate with the ideal coordinate system in the preferred layout scheme, ensures accurate matching of the numerical control cutting path and the part contour, avoids cutting deviation, and generates additional excess material.

[0031] The feedback self-learning module forms a closed-loop optimization of adaptive strategy parameters through analysis and feedback of the real cutting effect. The coincidence degree index IoU is essentially the intersection over union of the real excess material area and the theoretical excess material area The value range of IoU is 0 to 1, and the closer to 1, the more consistent the actual excess material and the expected excess material, and the closer to 0, the greater the deviation between the current cutting and the expectation, which needs to be adjusted to improve. Through such a cycle of self-learning process, the system can continuously evolve and optimize the strategy to reduce subsequent layout errors, improve material utilization and cutting precision.

[0032] Preferably, in the system, the data communication between the modules adopts the following standard protocols and data formats: The communication protocol between modules adopts the OPC-UA (Open Platform Communications Unified Architecture) industrial communication standard, and each module (scanning device, computer, numerical control machine tool, etc.) exchanges data as an OPC-UA client / server; The data transmission format adopts a self-defined JSON-NC format, that is, JSON (JavaScript Object Notation) is used to represent numerical control related data (such as plate geometry information, layout result coordinates, cutting sequence and parameters, etc.), so that the data has good readability and universal analysis; The feedback data of the system (such as utilization rate of each cutting, IoU result, algorithm adjustment parameter, etc.) is uploaded to the monitoring terminal or cloud platform in real time through the MQTT (Message Queuing Telemetry Transport) protocol, to realize remote monitoring and big data analysis of the production process.

[0033] It should be noted that the present application can reliably, conveniently and smoothly integrate the system into the existing digital shipbuilding production line by adopting the above standard protocol and data format, realize transparent transmission and sharing of information, and improve the expansibility and maintainability of the system.

[0034] In addition, for small probability anomalies with high risk that occur occasionally in actual production (such as local warping of the plate caused by internal stress release during cutting, temporary invalidation of the projection mark due to slag blocking, path error accumulation caused by instantaneous nozzle deviation, and deviation of the thermal expansion coefficient caused by local abnormal temperature rise), the present application also proposes an online closed-loop bottom-up solution that runs through the execution-detection-correction link, that is, a robust compensation updating method combining deformation field estimation and local path re-planning. Specifically: During the cutting execution process, the online visual / laser displacement sensor is used to intermittently / sliding window sample the uncut area, first, the local attitude drift is estimated by rigid registration (ICP / G-ICP) Then, the thin plate spline (TPS) or bilinear displacement field fitting is used to obtain the non-rigid deformation The unexecuted path curve is mapped and corrected online, and the expression is as follows: ; The corrected path curve is immediately reviewed for process minimum gap and thermal compensation constraints. If it is found that the local minimum gap approaches the threshold due to warping or the size deviation after thermal compensation is out of limits, local dynamic re-planning (only rearranging the order and start and end points between a few to-be-cut contours) is triggered for the affected path segment. The existing 2-Opt / LKH neighborhood can be reused, and only the disturbed subgraph is solved to ensure the shortest downtime. At the same time, online recursive estimation of the temperature compensation coefficient is introduced to improve the robustness of thermal compensation: the current sensing temperature is taken as the independent variable, the executed cutting length / time is taken as the observation, and the RLS / Kalman method is used to update the linear expansion coefficient : ; Wherein, is the online measured profile size deviation, is the adaptive gain; The system sets multiple levels of trigger thresholds: when the positioning error or temperature rise deviation exceeds the first level threshold, only coordinate correction is performed; when the second level threshold is exceeded, local re-planning is executed; when the third level threshold is exceeded, the cooling / pause and re-calibration process (automatic increase in reference marker points or replacement of reference marker points) is inserted, and the abnormal segment is marked and sent back to the feedback self-learning module to improve the prior adaptability of the subsequent layout strategy under similar deformation modes.

[0035] This bottom-up solution ensures the feasibility of maintaining size and gap constraints under abnormal disturbances, significantly reduces the risk of rework and scrap, and absorbs the deviation of compensation parameters into online estimation to avoid error accumulation in long batch processing.

[0036] Advantages In summary, the present application provides a ship steel plate digital cutting layout optimization method and system, which significantly improves the material utilization rate and production efficiency of ship steel plate cutting, and the specific advantages include but are not limited to: 1. Maximize material utilization and reduce cost: intelligent algorithm optimization layout makes the utilization rate of steel plate close to 100% (only cutting gap loss), significantly reduces the generation of waste steel after cutting compared to traditional layout methods, realizes "zero excess" cutting, and greatly reduces material waste and production cost.

[0037] 2. Improve the intelligent level of layout: introduce the optimization strategy of deep reinforcement learning and genetic algorithm fusion, so that the system can autonomously explore the best part layout scheme, automatically generate high-quality layout scheme without human intervention, and further introduce the feedback self-learning mechanism, so that the system can adjust and optimize the strategy from each cutting feedback, become more intelligent, continuously adapt to different steel plate shapes and part combinations, and improve the layout quality.

[0038] 3. Improve cutting precision and quality: accurate calibration of the actual position and attitude of the steel plate before cutting through projection and vision combination, automatic calibration within the precision threshold, no need for manual tedious adjustment, ensure accurate alignment of the cutting path with the steel plate, reduce cutting defects caused by steel plate placement deviation, ensure that even the most compact "zero excess" layout can be safely implemented, cooperate with cutting gap constraints and thermal deformation compensation, ensure accurate part cutting size and good edge quality, and use 2-Opt algorithm to optimize the cutting sequence when planning the cutting path to reduce empty travel, make the motion path of the cutting machine more economical and efficient, reduce machine idling and waiting, and improve cutting operation efficiency.

[0039] 4. System integration and digitization are realized: the functional modules communicate through standard OPC-UA, interact with JSON format data, realize an integrated digital twin system, and seamlessly connect the information flow of scanning, optimization, cutting, and feedback. The delay and error of manual data transmission are avoided, and the real-time data upload of MQTT enables managers to instantly understand the production situation and realize real-time monitoring and optimization of the intelligent manufacturing process.

[0040] 5. Strong versatility and adaptability: the system can be applied to various shapes of whole boards and excess material boards. Whether the size and shape of the board are regular or irregular, automatic processing and optimization can be realized. At the same time, modular design and standard interface facilitate the connection with different cutting equipment and production management systems, and have good versatility. The system can adapt to differences caused by different batch material characteristics and environmental factors through continuous self-learning, and has strong robustness.

[0041] Therefore, the technical scheme of the present application has significant practical value in the field of ship processing, can significantly reduce material waste, reduce labor intensity and human error, improve production efficiency and intelligent level, and conforms to the development direction of green shipbuilding and lean production. BRIEF DESCRIPTION OF DRAWINGS

[0042] Figure 1 FIG. 1 is a flowchart of a ship steel plate digital cutting layout optimization method according to an embodiment of the present application. DETAILED DESCRIPTION

[0043] In order to describe the system structure and working process of the present application in detail below in combination with the drawings and specific embodiments, it should be understood that the following embodiments are intended to help understand the present application, but do not limit the scope of protection of the present application.

[0044] Embodiment 1 The embodiment provides a ship steel plate digital cutting layout optimization system, comprising: A plate appearance acquisition module: acquires the appearance data of the steel plate to be cut and generates a processing area model; An intelligent layout optimization module: based on the processing area model and the specification data of the target part, adopts a deep reinforcement learning-genetic hybrid algorithm as a solving strategy to output an optimal layout scheme; A cutting path generation module: generates a numerical control cutting path and corresponding numerical control code according to the optimal layout scheme, and performs seam gap constraint and thermal deformation compensation on the numerical control cutting path; A projection positioning calibration module: based on the deviation between the ideal coordinates of the reference marker points in the processing area and the real coordinates of their projections on the steel plate to be cut, the steel plate to be cut is positioned and calibrated; A feedback self-learning module: executes the cutting instruction, evaluates the cutting result, and feeds back the deviation data to adjust the optimization strategy parameters.

[0045] The plate appearance acquisition module includes a laser scanner or an industrial camera installed on a numerical control cutting machine, the intelligent layout optimization module and the feedback self-learning module run on an industrial computer or a server, the cutting path generation module is integrated in the control system of the numerical control cutting machine, and the projection positioning calibration module includes a digital projector and an industrial camera. The entire system is connected through an Ethernet or a wireless network, each functional module exchanges data according to an OPC-UA protocol, and a platform for overall collaborative work is formed.

[0046] Embodiment 2 Referring to Figure 1 , the embodiment provides a ship steel plate digital cutting layout optimization method based on the system in embodiment 1, including the following steps. I. Collecting steel plate appearance data Firstly, the steel plate material to be cut is placed on the machine tool of the numerical control cutting machine. If it is an integral plate, it can be directly scanned. If it is a special-shaped scrap plate left after the last cutting, it is also placed on the machine tool. The plate appearance acquisition module is started. The laser scanner moves along the surface of the steel plate to scan and measure the distance of the edge of the steel plate, collects a large number of edge point coordinates, or obtains a steel plate contour image through the camera of the upper camera, extracts the edge through image processing, and the plate appearance acquisition module processes the collected edge data to obtain a digital model (represented by polygon vertex coordinates) of the steel plate processing area, which is expressed as follows: ; Among them, is the boundary coordinate of the processing area.

[0047] In this embodiment, the scanning result shows that the contour of the scrap plate is an irregular polygon, and the approximate size is 2.5 meters long and 1.2 meters wide. At this time, the contour can be divided into several line segments and recorded, and a coordinate system is calibrated for subsequent modules.

[0048] II. Optimizing the part layout scheme The processing area model and the target part list (including the number and respective specifications of the parts to be cut) are input into the intelligent layout optimization module. The module uses a deep reinforcement learning agent and a genetic algorithm to start solving the best layout method, as follows. The intelligent layout optimization module models the layout problem by using a Markov decision process. The state and action can be defined as follows: the state represents the information of the currently laid parts and the remaining space of the steel plate (including the occupied space layout, the shape characteristics of the remaining placeable area, etc.), and the blank processing area without part arrangement is used as the initial state , and the parts are arranged one by one in the processing area to form an action sequence ​, N is the total number of parts to be laid out, action represents selecting a part to be placed and determining its placement on the steel plate, in action , i.e. the part is arranged in the processing area, and the new state is recorded; By designing a reasonable state and action space, the reinforcement learning agent can effectively explore different part placement sequences and pose combinations and correspond to specific plate layouts; In the early stage of action, parts are arranged randomly or based on heuristic rules, such as placing the largest part in a corner of the steel plate first, and then gradually placing smaller parts. The reward is calculated by simulating the result after each placement. If the material utilization rate increases or there is no conflict, a positive reward is given. If the material utilization rate decreases or a conflict occurs, a negative reward is given. The reinforcement learning agent adjusts the strategy accordingly and continuously tries different layout methods. Specifically, the reward function expression is as follows: ; Where, is the reward value of the tth action, is the increment of material utilization rate after the tth action compared to the previous step, is a conflict indicator function, which takes 1 when the newly placed part overlaps with the already placed part, otherwise it takes 0, is the current cumulative cutting time, is the maximum allowed cutting time, are the corresponding weight coefficients, respectively; At the same time, the genetic algorithm maintains a set of possible layouts (as a population of strategy network parameters), and updates the strategy network parameters for a set of feasible layouts given by the reinforcement learning agent. After each iteration, the better ones are crossed and mutated to generate new candidate solutions. Specifically, the deep reinforcement learning-genetic hybrid algorithm includes reinforcement learning based on policy gradient update and global search of genetic algorithm, and the expression of policy gradient update is as follows: ; Where, is the strategy network parameter, is the policy function with respect to the parameter , is the probability distribution function of the policy network selecting action in state , is the logarithmic derivative of the probability distribution function with respect to the parameter , is the policy Mathematical expectation of the next action-state distribution, The advantage function and used to evaluate the goodness of a particular action relative to the average level of that state, expressed as follows: ; Where, is the discount factor, and are the estimated state values of the value network for the current state , next state ; The genetic algorithm is based on the current policy network parameter population, and new policy individuals are generated through selection, crossover, mutation and other operations, expressed as follows: ; Where, is the new policy network parameter generated after genetic operation, and are the parameters of the i, j-th policy individuals selected from the current policy population, is the crossover ratio coefficient and , is the Gaussian noise mutation term (a small random vector with zero mean); After completing the preset number of iterations of training and evaluation, the intelligent layout optimization module outputs a set of preferred part layout parameters as the preferred layout scheme , which includes the rotation angle and the horizontal translation , vertical translation of each part; The preferred layout scheme satisfies the zero tolerance condition, expressed as follows: ; Where, is the area where the t-th part is located after rotation and translation operation, and Area is a function of the area.

[0049] In this embodiment, the target parts list includes 10 parts and the specification parameters are known, and the preferred layout scheme is to place the largest trapezoidal plate in the upper left area of the excess material plate with a rotation angle of 15°, and the second largest arc-shaped plate in the lower right corner with a rotation angle of 33°, and the remaining small parts are cleverly inserted into the gaps. The scheme is compact and reasonable, all parts are placed in the excess material plate area and do not overlap with each other, achieving more than 99% material utilization rate, and the unused plate space is only a very small grid skeleton formed by laser cutting.

[0050] III. Cutting path planning and code generation The cutting path generation module generates a cutting path according to the preferred layout scheme Determine the position of each part in the machining area, divide the contour line of each part into multiple line segments according to the vertex, and arrange all the line segments in a certain order to generate a cutting path; In order to improve the cutting efficiency, this module uses 2-Opt algorithm to optimize the cutting order, that is, on the basis of the initial path scheme, the cutting order of the local path is repeatedly exchanged to reduce the length of the empty path, until the path cannot be optimized or the optimization increment is lower than the threshold value; In this embodiment, the initial cutting order is to cut each part from the top left to the bottom right of the steel plate, and after 2-Opt algorithm optimization, it is adjusted to cut the central part of the steel plate first, and then cut the edge part, reducing the moving distance of the cutting head in the non-working area; Insert the inlet and outlet lead lines, common edge cutting process paths (if needed) on the determined cutting path, and the planning of the cutting path needs to meet the following conditions: ① For any two adjacent parts, keep , where is the minimum distance between the contour edges of part i and part j, is the minimum gap allowed by the process; ② Correct and compensate the cutting path length based on the thermal expansion effect of steel, that is , where is the compensated path length, is the linear expansion coefficient of steel, is the original length of the cutting path before compensation, is the elevated temperature of the cutting area; In this embodiment, the minimum gap allowed by the process is 5mm, and the nearest distance of adjacent parts is controlled to be 5.2mm and above. According to the thermal expansion coefficient of steel and the cutting parameters, the thermal compensation amount is calculated: the linear expansion coefficient of steel , the local temperature rise of plasma cutting ℃, for each straight cutting segment with a length of , the estimated , so when generating numerical control code, each cutting straight line is lengthened in proportion (curve and the like are processed by segmentation approximation); For parts with long contour lines on the steel plate, the cutting path can also use a segmented intermittent cutting method, that is, cut a part first and then pause, and then continue cutting after the material cools down to reduce cumulative deformation; According to the cutting path, the corresponding numerical control code is generated, and the cutting path generation module exports the corresponding code file, which contains the cutting instructions of all part contours, tool position moving path, and speed, perforation, etc. Control command, the code file is immediately uploaded to the numerical control cutting machine controller for execution.

[0051] IV. Cutting positioning calibration Before the cutting machine starts actual work, the projection positioning calibration module is started to accurately calibrate the position of the steel plate. First, the digital projector selects three reference marker points in the machining area and projects them onto the surface of the steel plate to be cut (for example, at the lower left corner, the upper right corner and the middle position of the machining area, each projecting a crosshair). The design coordinates of these points in the steel plate coordinate system are known, denoted as Then, the industrial camera fixed on the machine tool captures the steel plate and the projected markers, and the real coordinates of the projected points on the steel plate are calculated using a visual algorithm, denoted as ; According to the coordinate differences of the corresponding points in each group, the rigid body transformation parameters of each projected point relative to the ideal position are calculated, and the expression is as follows: ; Among them, is the rotation angle, is the horizontal translation, is the vertical translation; The least squares method is used to fit the solution of the rigid body transformation to obtain the estimated values of the translation and rotation angle , which are used to describe the offset of the steel plate relative to the ideal position, calculate the positioning error and compare it with the preset accuracy threshold, and the expression is as follows: ; When all the above conditions are met, the calibration is passed, and is uploaded to the numerical control cutting machine controller, which triggers the cutting execution instruction and performs coordinate transformation compensation on the cutting path to be executed; otherwise, when at least one of the above conditions is not met, the calibration is not passed, and an alarm is prompted to adjust the steel plate placement position or add more reference marker points to recalculate the positioning error until the accuracy requirement is met.

[0052] In this embodiment, it is found that the actual steel plate is offset to the right by , downward by , and clockwise by , all of which do not exceed the accuracy threshold, so the calibration is passed, and the projection positioning calibration module sends the above offset parameters to the numerical control cutting machine controller, which performs coordinate transformation compensation on the cutting path to be executed to ensure that the actual cutting trajectory is aligned with the steel plate. ​

[0053] V. Execution of cutting After the positioning calibration is completed, the numerical control cutting machine starts to execute the cutting operation according to the generated numerical control code; In this embodiment, the numerical control cutting machine selects a plasma cutting head, and starts cutting the part contour from the edge of the steel plate or the pre-punching hole. Since the part layout and path sequence in the layout scheme are optimized, the cutting process is smooth, the cutting head sequentially completes the cutting of each part contour, and moves along the shortest path to the starting point of the next part. The cutting of the whole plate is completed in about 15 minutes. During the cutting, the process control module of the cutting machine appropriately adjusts the cutting size according to the thermal compensation instruction, so that the actual cut part size is highly consistent with the drawing requirements. After the cutting is completed, all 10 parts are successfully separated from the plate, and the operator takes them out, while the remaining material is in a grid-shaped fine skeleton structure, which is retained on the table surface for recycling or further processing.

[0054] VI. Result feedback and self-learning After the cutting is completed, the feedback and self-learning module is started to evaluate and collect data of the cutting result. The topographic data of the remaining material is obtained by a laser scanner or machine vision, and the coincidence degree index of the actual remaining material and the expected theoretical remaining material in the optimized layout scheme is calculated. The expression is as follows: ; Wherein, IoU is the coincidence degree index, Area is the function of the area, and is the actual remaining material is the theoretical remaining material ; When the IoU is less than the preset coincidence threshold, the difference between the actual remaining material and the theoretical remaining material is discretely processed by rasterization to obtain a difference raster image, and the expression is as follows: ; Wherein, is the set difference operation, is the discrete rasterization operation; The difference raster image is input into the neural network of the reinforcement learning intelligent agent for training, so that the algorithm identifies the situation that the difference area is not effectively utilized in the scheme, and adjusts the subsequent layout strategy parameters accordingly.

[0055] In this embodiment, the actual scanning finds that some areas of the remaining material are slightly larger than the theoretical prediction. For example, a small piece of plate material planned to be cut between two parts is not completely separated due to slight deformation caused by thermal stress, forming a factually remaining material segment. The feedback module analyzes the difference: The intersection area of the remaining material area and the theoretical remaining material area is 200 cm 2 , and the union area is 220 cm 2 , so the calculation Although most areas meet expectations, there is still room for improvement as the IoU is below the preset threshold of 95%. The feedback self-learning module extracts these difference areas and rasterizes them into images, then updates the optimization algorithm strategy according to the difference types: inputting the difference raster image into the neural network of the reinforcement learning agent for training, so that it can identify the situation that this narrow gap area may not have been effectively utilized in the previous scheme, and next time when encountering similar steel plate and part layout requirements, the agent will tend to adjust the part spacing or cutting sequence to avoid this problem; at the same time, the system records the key data of this cutting (utilization rate 99%, IoU about 0.91, cutting time about 15 minutes, etc.) and publishes it to the monitoring server through MQTT for engineers to view and analyze. After the above feedback learning, when the system enters the next plate for layout cutting, the reinforcement learning agent has updated the strategy parameters based on the last experience, and the improved layout scheme factors are included in the initial population of the genetic algorithm, which makes the starting point of this optimization better; after multiple cycles, the self-learning effect of the system gradually emerges: for similar types of scrap plates, the success rate of the layout scheme is improved, and the average value of the IoU after cutting gradually increases (from 91% to 95%, 98%, and even 100%), indicating that the real scrap almost completely meets the expected scheme, the material utilization rate is also maintained at a very high level, and there are no small residual pieces around the parts. Through such continuous evolution, the system of the present application optimizes itself in practice, and achieves the ideal "zero scrap" cutting effect and highly stable cutting quality.

[0056] It should be noted that the above embodiments of the present application are specific application schemes provided for the purpose of facilitating understanding of the present application, but the present application is not limited thereto. For those skilled in the art, various equivalent replacements and improvements can be made to the specific implementation of the system without departing from the principles of the present application, such as using different forms of reinforcement learning algorithms (such as deep Q network DQN, PPO, etc.), more complex evolutionary algorithms, or using other types of scanning and projection devices, etc. These changes should be considered as the protection scope of the present application. The protection scope of the present application is defined by the appended claims.

Claims

1. A method for optimizing digital cutting layout of a ship steel plate, characterized in that, The method comprises the following steps: Step 1, collecting the shape data of the steel plate to be cut, and generating a processing area model; Step 2, based on the processing area model and the specification data of the target part, using a deep reinforcement learning-genetic hybrid algorithm as the solving strategy to output an optimal layout scheme; Step 3, generating a numerical control cutting path and corresponding numerical control code according to the optimal layout scheme, and performing seam gap constraint and thermal deformation compensation on the numerical control cutting path; Step 4, based on the deviation between the ideal coordinates of the reference marker points in the processing area and the real coordinates of their projections on the steel plate to be cut, positioning and calibrating the steel plate to be cut; Step 5, executing the cutting instructions, evaluating the cutting results, and feeding back the deviation data to adjust the optimization strategy parameters.

2. The method of claim 1, wherein, The step 1 comprises: The steel plate to be cut is a whole plate or a remaining plate after the last cutting, and the profile coordinate point cloud of the steel plate is obtained by a laser scanner or machine vision to generate a processing area model, and the expression is as follows: ； wherein, is the coordinate of the boundary of the processing area.

3. The method of claim 1, wherein, The step 2 comprises: The processing area model and the specification data of the target part are input, and a deep reinforcement learning agent and a genetic algorithm are used to explore the best layout method of the target part on the steel plate, that is, all the laid-out parts do not overlap with each other and the unused space in the processing area is the least, and the specific method is as follows: The reinforcement learning agent arranges the parts one by one in the processing area to form a sequence of actions , N is the total number of parts to be arranged, and a blank processing area without part arrangement is taken as the initial state , the action is to arrange the first part in the processing area, and the new state is recorded after the part is arranged In the initial stage of action, the parts are arranged randomly or based on heuristic rules, and the reward is calculated according to the material utilization rate and the part conflict situation after each arrangement, wherein a positive reward is given if the material utilization rate increases or there is no conflict, and a negative reward is given if the material utilization rate decreases or a conflict occurs, and the reinforcement learning agent adjusts the strategy accordingly to try different layout methods; The genetic algorithm is used for strategy network parameter updating of feasible layout modes given by a reinforcement learning agent, and after each iteration, a number of optimal layout modes are subjected to cross and mutation operations to generate new layout modes, and after a preset number of iteration training and evaluation, a set of optimized part layout parameters are output as an optimal layout scheme , the part layout parameters including a rotation angle of each part , a horizontal translation amount , and a vertical translation amount .

4. The method of claim 1, wherein, The step 3 comprises: According to the optimal layout scheme, the positions of the parts in the processing area are determined, the profile line of each part is divided into multiple line segments according to the nodes, and all the line segments are arranged in sequence to generate a numerical control cutting path, and the planning of the numerical control cutting path needs to meet the following requirements: wherein, is the minimum distance between the profile edge of part i and part j, is the minimum gap allowed by the process;​ ② Correcting and compensating the cutting path length based on the thermal expansion effect of steel material, i.e. wherein, is the compensated path length, is the linear expansion coefficient of the steel material, is the original length of the cutting path before compensation, is the elevated temperature of the cutting area; According to the numerical control cutting path, the corresponding numerical control code is generated, and the code file is uploaded to the numerical control cutting machine controller.

5. The method of claim 1, wherein, The step 4 comprises: projecting at least three reference mark points onto the surface of the steel plate to be cut within the machining area, recording ideal coordinates of the reference mark points ; An image of the steel plate to be cut is acquired by machine vision, and the real coordinates of the projection points on the steel plate are identified ; According to the coordinate difference, the rigid transformation parameters of each projection point relative to the ideal position are calculated, and the expression is as follows: ； wherein, is a rotation angle, is a horizontal translation amount, is a vertical translation amount; The least square method is used to fit the results of rigid body transformation to obtain the estimated values of translation and rotation angle , , , the positioning error is calculated and compared with the preset accuracy threshold, and the expression is as follows: ； ； If all the above equations are true, the calibration is considered successful. , , The data is uploaded to the CNC cutting machine controller, which triggers the cutting execution command and performs coordinate transformation compensation on the cutting path to be executed. Conversely, if at least one of the above formulas is not true, the calibration is deemed to have failed, and an alarm is triggered to request adjustment of the steel plate placement position or addition of more reference markers to recalculate the positioning error until the accuracy requirements are met.

6. The method of claim 1, wherein, The step 5 comprises: The numerical control cutting machine performs cutting operation according to the generated numerical control code, the cutting head starts to cut out the part profile from the edge of the steel plate or the pre-punching hole, and the cutting head moves to the starting point of the next part profile along the shortest path, and after the operation is completed, all the parts are taken out and the remaining material is reserved on the cutting machine table; The shape data of the remaining material is obtained by a laser scanner or machine vision, the coincidence degree index of the real remaining material and the expected theoretical remaining material in the optimal layout scheme is calculated, and the expression is as follows: ； Wherein, IoU is the coincidence degree index, Area is the function of area, is is the theoretical excess material ; When the IoU is less than the preset coincidence threshold, the difference area between the real remaining material and the theoretical remaining material is discretely processed by rasterization to obtain a difference raster image, and the expression is as follows: ； wherein, is a set difference operation, is a discrete gridding operation; The difference raster image is input into the neural network of the reinforcement learning agent for training, so that the algorithm can identify the situation that the difference area is not effectively utilized in the scheme, and the subsequent layout strategy parameters are adjusted accordingly.

7. The method of claim 3, wherein the method further comprises: The function expression for calculating the reward in the step 2 is as follows: ； wherein, is the reward value for the t-th action, is the increment of material utilization after the t-th action compared to the previous step, is a conflict indicator function that takes the value 1 when the newly placed part overlaps with an already placed part, and 0 otherwise, is the current accumulated cutting time, is the maximum allowed cutting time, are the corresponding weight coefficients, respectively.

8. The method of claim 3, wherein the method further comprises: The deep reinforcement learning-genetic hybrid algorithm comprises reinforcement learning based on policy gradient update and global search of genetic algorithm, and an expression of the policy gradient update is as follows: ； in, For policy network parameters, For policy function Regarding parameters gradient, For the policy network in state Select action The probability distribution function, The probability distribution function with respect to the parameters The logarithmic derivative, For strategy Mathematical expectation of the next action-state distribution The advantage function, used to evaluate the merits of a particular action relative to the average level of that state, is expressed as follows: ； where, is a discount factor, and are the estimated state values of the value network for the current state , next state , respectively.

9. The method of claim 8, wherein, The genetic algorithm is based on a current policy network parameter population, and generates new policy individuals through selection, crossover and mutation operations, and an expression is as follows: ； wherein, are the new policy network parameters generated after genetic manipulation, and are the parameters of the i, j-th policy individual selected from the current policy population, respectively, is a crossover ratio coefficient and , is a Gaussian noise variation term.

10. A system for optimizing digital cutting layout of a ship steel plate, characterized in that, It comprises: A plate appearance acquisition module: collecting appearance data of a steel plate to be cut and generating a processing area model; An intelligent layout optimization module: based on the processing area model and specification data of a target part, using a deep reinforcement learning-genetic hybrid algorithm as a solving strategy to output an optimal layout scheme; A cutting path generation module: generating a numerical control cutting path and corresponding numerical control code according to the optimal layout scheme, and performing seam gap constraint and thermal deformation compensation on the numerical control cutting path; A projection positioning calibration module: based on the deviation between ideal coordinates of reference marker points in a processing area and real coordinates of projections of the reference marker points on a steel plate to be cut, performing positioning calibration on the steel plate to be cut; A feedback self-learning module: executing a cutting instruction, evaluating a cutting result, and feeding back deviation data to adjust and optimize strategy parameters; To realize the ship plate digital cutting layout optimization method as claimed in any one of claims 1-9.

Images