A micro-size SLM precision scanning path dynamic optimization distribution method

Abstract

The application relates to the technical field of path optimization, in particular to a micro-size SLM precision scanning path dynamic optimization and distribution method, which comprises the following steps: obtaining a three-dimensional model of a product to be processed; establishing a region prediction model to classify the region to be processed, obtaining a functional partition, and generating a partition mapping table; calculating and generating a processing path scheme according to the functional partition, forming a partition path library; establishing a correlation prediction model to analyze a processing task and material characteristics, obtaining processing parameters; generating an initial processing path distribution scheme based on the processing parameters, monitoring state parameters of the product to be processed in a processing process, calculating optimal processing parameters according to a functional partition and historical processing data when the state parameters deviate from a preset target range, obtaining processing adjustment parameters, and selecting a processing path scheme from the partition path library; and optimizing and adjusting a processing path of a next region according to the processing adjustment parameters in a gap time after current region processing is completed.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of path optimization, in particular to a micro-size SLM precise scanning path dynamic optimization and distribution method. BACKGROUND

[0002] With the performance requirements of micro metal structural parts in the fields of unmanned aerial vehicles, embodied intelligent robots, precision medical devices, 3C electronic components and high-end product accessories increasing, the demand for micro-scale metal technology with high precision, high density and complex structure manufacturing capability is becoming increasingly urgent. Selective laser melting (SLM) technology as the mainstream process of metal additive manufacturing has shown great potential in the field of micro-scale precision manufacturing, but its processing path planning and optimization still faces many challenges.

[0003] The existing scanning path optimization method mainly relies on experience parameter setting or simple geometric algorithm, and cannot realize accurate identification and classification processing of complex microstructure features, and also lacks real-time monitoring and dynamic adjustment mechanism of physical quantities such as heat field distribution and stress change in the processing process. This static path planning method cannot meet the high requirements of micro-scale precision manufacturing on accurate control of processing path.

[0004] Therefore, a micro-size SLM precise scanning path dynamic optimization and distribution method is proposed. SUMMARY

[0005] The purpose of the present application is to provide a micro-size SLM precise scanning path dynamic optimization and distribution method, which generates and dynamically adjusts the processing path and parameters through regional prediction and correlation prediction model, realizes the partition management and real-time optimization of the scanning process, and improves the scanning accuracy and efficiency.

[0006] In order to achieve the above purpose, the present application provides the following technical scheme:

[0007] A micro-size SLM precise scanning path dynamic optimization and distribution method, comprising:

[0008] Obtain the three-dimensional model of the product to be processed, and extract the geometric data, product data and material properties; establish a regional prediction model based on the geometric data and product data to classify the regions to be processed, obtain the functional partition, and generate a partition mapping table; according to the functional partition, pre-calculate and generate a processing path scheme to form a partition path library;

[0009] The correlation prediction model is established to analyze the processing task and material characteristics to obtain processing parameters; an initial processing path allocation scheme is generated based on the processing parameters; the product to be processed is processed through the initial processing path allocation scheme, and the state parameters of the product to be processed are monitored during the processing process; when the state parameters deviate from the preset target range, the optimal processing parameters are calculated according to the functional partition and historical processing data, the processing adjustment parameters are obtained, and the processing path scheme is selected from the partition path library;

[0010] In the gap time after the current area processing is completed, the processing path of the next area is optimized and adjusted according to the processing adjustment parameters.

[0011] Preferably, the geometric data includes the coordinates, surface curvature radius, wall thickness distribution, overhanging angle and local bulk density of the product.

[0012] The product data includes the overall size, surface roughness requirement and structure complexity level of the product.

[0013] The material characteristics include the melting point, thermal conductivity, thermal expansion coefficient, laser absorption rate and crystallization characteristic parameters of the noble metal.

[0014] Preferably, the area prediction model includes a feature extraction layer, a classification decision layer and a mapping generation layer.

[0015] The feature extraction layer extracts feature vectors from the geometric data and product data to obtain a comprehensive feature matrix containing spatial position, geometric complexity and product characteristics.

[0016] The classification decision layer classifies the comprehensive feature matrix to output the area identifier of the precision control area, the transition buffer area and the standard processing area.

[0017] The mapping generation layer associates the area identifier with the coordinates to generate a three-dimensional partition mapping table containing area type, boundary coordinates and priority information.

[0018] Preferably, the process of pre-computing and generating the processing path scheme according to the functional partition includes:

[0019] Based on the crystallization characteristic parameters of the product, a progressive rotation strategy of scanning direction between layers is designed, and the rotation angle of the scanning direction between adjacent layers is set to a first angle to a second angle.

[0020] A short line segment staggered scanning strategy with reduced laser power and improved scanning speed is adopted for the precision control area.

[0021] A transition scanning strategy with a gradual change in scanning direction by a preset step size is adopted for the transition buffer area.

[0022] The standard processing area adopts a unidirectional parallel scanning strategy; the scanning path scheme of each area is stored as a data structure containing a coordinate sequence, power parameters and speed parameters, forming a partition path library.

[0023] Preferably, the correlation prediction model comprises a data preprocessing layer, a feature correlation layer and a parameter prediction layer;

[0024] The data preprocessing layer standardizes the historical processing data, extracts the numerical features of processing power, processing speed, overlap rate and forming quality;

[0025] The feature correlation layer establishes a nonlinear mapping relationship between processing data and product quality;

[0026] The parameter prediction layer obtains product quality according to the geometric data and material properties of the current processing task; based on the product quality, it outputs the laser power range, processing speed range and overlap rate value as the processing parameters.

[0027] Preferably, the process of generating an initial processing path allocation scheme based on the processing parameters includes:

[0028] According to the laser power range and processing speed range determined by the processing parameters, the processing line spacing and processing sequence of each functional partition are calculated; the processing sequence is arranged in the order of precision control area first, transition buffer area second, and standard processing area last;

[0029] A linear transition section of power and speed is set at the boundary between adjacent areas, and the transition length is determined according to the boundary length; an initial processing path allocation scheme containing scanning trajectory coordinates, timestamps and parameter change points is generated.

[0030] Preferably, the process of obtaining processing adjustment parameters includes:

[0031] By collecting the overall temperature distribution data of the processing area and the geometric contour data of the formed layer, when the overall temperature distribution deviates from the target temperature field preset temperature and / or the geometric contour deviation exceeds the preset deviation, the current processing parameters and deviation data are recorded;

[0032] Based on statistical analysis method, the historical processing data is matched to obtain the laser power adjustment value, scanning speed adjustment and path spacing adjustment value suitable for the next layer as the processing adjustment parameters.

[0033] Preferably, the specific process of pre-optimizing the processing path of the next layer according to the processing adjustment parameters includes:

[0034] After each layer of scanning is completed, the temperature distribution data, geometric accuracy data and surface quality data of the whole layer are collected; the processing path of the subsequent same type of functional partition is pre-optimized, and the optimization target is temperature uniformity and geometric accuracy;

[0035] According to the pre-optimization result, the candidate path scheme of the corresponding area in the partition path library is updated; the comparison data of the optimization result and the actual processing result is recorded in the historical database, which is used for offline training of the correlation prediction model and continuous updating of the parameter library.

[0036] Compared with the prior art, the present application has the following advantages:

[0037] 1. The present application constructs a regional prediction model by geometric data and product data, divides the product to be processed into functional partitions such as precision control area, transition buffer area and standard processing area, and generates a three-dimensional partition mapping table. According to the forming characteristics of different functional partitions, a differentiated scanning strategy is designed. The partitioned path scheme is pre-stored as a partition path library, and can be quickly called during processing, realizing on-demand allocation of the path. The partitioning and path library mechanism of the present application significantly improves the adaptability of the scanning path to the thermal field distribution, geometric complexity and quality requirements of different areas, reduces the risk of thermal stress concentration and deformation, and shortens the path generation time.

[0038] 2. The present application introduces a correlation prediction model, standardizes the historical processing data, extracts numerical features, and establishes a nonlinear mapping relationship between input parameters and output quality. During the processing task execution process, the optimal power range, speed range and overlap rate parameters are predicted according to the geometric data and material characteristics, and an initial path allocation scheme is generated. When the processing state parameters (such as temperature distribution, geometric contour accuracy) deviate from the target value, parameter matching is performed based on historical successful cases to obtain adjustment values and apply them to the next processing layer. This mechanism realizes dynamic adaptive adjustment of parameters, can effectively cope with material thermal sensitivity differences, structural complexity changes and working condition disturbances, and avoids defects such as ablation, incomplete fusion or warping caused by fixed parameters.

[0039] 3. In the gap time after each layer of scanning is completed, the temperature distribution, geometric accuracy and surface quality data of the whole layer are collected, the processing path of the subsequent similar functional partition is pre-optimized, and the optimization target is temperature uniformity and geometric accuracy. In the pre-optimization process, the candidate path scheme of the corresponding area in the partition path library is updated, and the comparison data of the optimization result and the actual processing effect is stored in the historical database, which is used for offline training of the correlation prediction model and updating of the parameter library. The scanning path and the processing parameter library are continuously iterated and evolved in actual production, can long-term adapt to the micro-size SLM processing requirements of different batches, different shapes and different materials, and form a dynamic optimization closed loop. BRIEF DESCRIPTION OF DRAWINGS

[0040] Figure 1 A micro-size SLM precision scanning path dynamic optimization allocation method flowchart provided by the present application;

[0041] Figure 2 A schematic diagram of a correlation prediction model structure provided by the present application is shown in FIG. 1.

[0042] Figure 3 A schematic diagram of scanning path dynamic optimization processing logic provided by an embodiment of the present application is shown in FIG. 4. DETAILED DESCRIPTION

[0043] In order to make the purpose, technical solutions and advantages of the present application clearer, the present application is further described in detail below in combination with the drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain the present application and do not limit the present application.

[0044] Embodiment One

[0045] Referring to FIG. 1, Figure 1 The present application provides a micro-size SLM precision scanning path dynamic optimization distribution method, and the technical solutions are as follows:

[0046] A three-dimensional model of a product to be processed is obtained, and geometric data, product data and material properties are extracted; a regional prediction model is established based on the geometric data and product data to classify the regions to be processed, obtain functional partitions, and generate a partition mapping table; a processing path scheme is pre-calculated based on the functional partitions to form a partition path library.

[0047] In this embodiment, the product to be processed is a precious metal to be processed; the geometric data includes the coordinates, surface curvature radius, wall thickness distribution, overhanging angle and local bulk density of the precious metal to be processed.

[0048] The product data includes the overall size, surface roughness requirement and structure complexity level of the precious metal to be processed.

[0049] The material properties include the melting point, thermal conductivity, thermal expansion coefficient, laser absorption rate and crystallization characteristic parameters of the precious metal.

[0050] In this embodiment, by obtaining the geometric data, product data and material properties of the precious metal to be processed, a regional prediction model is established to perform functional partitioning and generate a partition mapping table, and a partition path library is pre-calculated and formed based on the partition characteristics, thereby realizing accurate partition processing path planning for complex structure precious metals and improving processing precision and efficiency.

[0051] The regional prediction model includes a feature extraction layer, a classification decision layer and a mapping generation layer.

[0052] The feature extraction layer extracts feature vectors from the geometric data and product data to obtain a comprehensive feature matrix containing spatial position, geometric complexity and product characteristics.

[0053] The classification decision layer classifies the comprehensive feature matrix to output region identifiers of the precision control area, the transition buffer area and the standard machining area.

[0054] The mapping generation layer associates the region identifiers with coordinates to generate a three-dimensional partition mapping table containing region types, boundary coordinates and priority information.

[0055] The acquisition process of the comprehensive feature matrix includes: converting the coordinate information in the geometric data into a three-dimensional space position vector, calculating the local geometric complexity indicators of the surface curvature radius and the overhanging angle; normalizing the wall thickness distribution and the local volume density to extract the structural compactness feature; encoding the overall size, roughness requirement and complexity level in the product data into a product feature vector; using a convolutional neural network to extract deep features of the space position vector, and fusing the geometric complexity indicators and the product feature vector through a fully connected layer to generate a comprehensive feature matrix, each row of which corresponds to a multi-dimensional feature description of a spatial sampling point.

[0056] The region prediction model is pre-trained in a supervised learning manner, and the training data set contains three-dimensional model data of the product to be processed. The training samples include: geometric feature data (surface curvature radius, wall thickness distribution, overhanging angle, local volume density), product specification data (overall size, roughness requirement, complexity level) and expert-labeled functional partition labels (precision control area, transition buffer area, standard machining area).

[0057] The pre-training process uses a deep convolutional neural network to convert three-dimensional geometric data into voxel grid input, and expands the training samples through data enhancement techniques (rotation, scaling, mirroring); the training uses the Adam optimizer with a learning rate of 0.001, a batch size of 32, and is trained for 300 rounds; it can accurately identify the functional partition of medical devices with different materials and different structural complexities, and provide a reliable basis for subsequent path planning.

[0058] In this embodiment, the feature extraction layer acquires the comprehensive feature matrix of the geometric and product data, the classification decision layer divides it into the precision control area, the transition buffer area and the standard machining area, and the mapping generation layer generates a three-dimensional partition mapping table containing region types, boundary coordinates and priorities, to realize precise machining area planning.

[0059] The process of pre-computing and generating a processing path scheme according to the functional partition includes:

[0060] Based on the crystallization characteristic parameters of the product, the interlayer scanning direction progressive rotation strategy is designed, and the scanning direction rotation angle between adjacent layers is set to a first angle to a second angle.

[0061] The short line segment staggered scanning strategy with reduced laser power and improved scanning speed is used for the precision control area.

[0062] A transition scanning strategy is adopted for the transition buffer, in which the scanning direction gradually changes according to a preset step size;

[0063] A unidirectional parallel scanning strategy is adopted for the standard processing area; the scanning path scheme of each area is stored as a data structure containing coordinate sequence, power parameters and speed parameters to form a partitioned path library.

[0064] In this embodiment, a progressive rotation strategy for the interlayer scanning direction is designed based on crystallization characteristic parameters, taking into account the functional partition characteristics. This strategy controls the change of the scanning direction of adjacent layers between the first and second angles. Low-power, high-speed short-segment interlaced scanning is used in the precision control area, a step-size gradual transition scanning is used in the transition buffer zone, and a unidirectional parallel scanning is used in the standard processing area. The scanning path schemes for each area are stored in a data structure containing coordinate, power, and speed parameters to form a partitioned path library, thereby achieving high-precision and high-efficiency processing control.

[0065] An association prediction model is established to analyze processing tasks and material properties to obtain processing parameters. An initial processing path allocation scheme is generated based on these parameters. The product to be processed is then processed according to this initial scheme. During processing, the status parameters of the product are monitored. When these parameters deviate from a preset target range, the optimal processing parameters are calculated based on the functional zone and historical processing data to obtain adjustment parameters. A processing path scheme is then selected from the zone path library. A detailed structural diagram of the association prediction model is provided below. Figure 2 .

[0066] The correlation prediction model is pre-trained based on historical processing data of the product to be processed. The training dataset includes: input features (geometric complexity index, material thermophysical parameters, structural thickness distribution, surface area to volume ratio) and output labels (optimal laser power, scanning speed, overlap rate and corresponding forming quality score).

[0067] The pre-training process employs a multilayer perceptron network structure with 5 hidden layers (256-128-64-32-16 neurons), using ReLU activation and Dropout regularization to prevent overfitting. The training data is divided into training, validation, and test sets in an 8:1:1 ratio. The mean squared error loss function is used, along with a learning rate decay strategy (initially 0.01, decreasing by 0.1 every 50 epochs), and the training is conducted for 200 epochs. The model's parameter prediction error on the test set is controlled within a preset range, enabling it to accurately predict the optimal process parameter range based on product characteristics, providing precise parameter guidance for actual production.

[0068] The association prediction model includes a data preprocessing layer, a feature association layer, and a parameter prediction layer;

[0069] The data preprocessing layer standardizes historical processing data, and extracts numerical features of processing power, processing speed, lap joint rate and forming quality.

[0070] The feature correlation layer establishes a nonlinear mapping relationship between processing data and product quality.

[0071] The parameter prediction layer obtains product quality according to the geometric data and material properties of the current processing task, and outputs the laser power range, processing speed range and lap joint rate value as processing parameters based on the product quality.

[0072] In this embodiment, the processing task is analyzed by the correlation prediction model, the power, speed, lap joint rate and forming quality are extracted, the nonlinear mapping between the input and output quality is established, the laser power, speed and lap joint rate range are predicted to generate an initial path scheme; during the processing, the state parameters are monitored in real time, when deviating from the target range, the optimal processing parameters are calculated combined with the functional partition and historical data, and the adaptive scheme is selected from the partition path library, to realize the dynamic optimization and accurate control of the processing process, and improve the product quality and processing stability.

[0073] The process of generating an initial processing path distribution scheme based on processing parameters includes:

[0074] According to the laser power range and processing speed range determined by the processing parameters, the processing line spacing and processing sequence of each functional partition are calculated; the processing sequence is arranged in the order of precision control area first, transition buffer area second, and standard processing area last;

[0075] Linear transition sections of power and speed are set at the boundaries of adjacent regions, and the transition length is determined according to the boundary length; an initial processing path distribution scheme containing scan trajectory coordinates, time stamps and parameter change points is generated.

[0076] In this embodiment, the laser power and speed are determined according to the processing parameters, the line spacing and processing sequence of each functional partition are calculated, the processing is arranged according to the priority of precision control area, transition buffer area and standard processing area, and linear transition sections of power and speed are set at the boundaries of regions, to generate an initial path scheme containing scan trajectory, time stamp and parameter change point, to realize smooth connection and precision guarantee of processing.

[0077] The process of obtaining processing adjustment parameters includes:

[0078] By collecting the overall temperature distribution data of the processing area and the geometric contour data of the formed layer, when the overall temperature distribution deviates from the target temperature field preset temperature and / or the geometric contour deviation exceeds the preset deviation, the current processing parameters and deviation data are recorded;

[0079] Based on the statistical analysis method, the historical processing data are matched with parameters to obtain the laser power adjustment value, scanning speed adjustment value and path spacing adjustment value suitable for the next layer, which are used as the processing adjustment parameters.

[0080] In the embodiment, the temperature distribution and the geometric profile data of the processing area are collected to determine the deviation from the target temperature field and the preset geometric accuracy. When the deviation exceeds a threshold, the current parameters and the deviation information are recorded, and based on the statistical matching of historical successful cases, the laser power, scanning speed and path spacing adjustment values suitable for the next layer are obtained to realize targeted processing parameter optimization and improve the product accuracy and stability.

[0081] The state parameter monitoring adopts a multi-sensor fusion technology, including: arranging an infrared thermal imager array to monitor the temperature field distribution, a high-precision laser displacement sensor to monitor the geometric deformation, and an acoustic emission sensor to monitor the stability of the molten pool; establishing a sensor data fusion algorithm to perform time-space synchronization and weight distribution on multi-source sensor data; setting a sensor fault detection mechanism to automatically switch to a backup sensor or adjust the fusion weight when a single sensor is abnormal.

[0082] Compared with the existing single temperature or deformation monitoring scheme, the multi-sensor fusion monitoring technology can simultaneously obtain three key parameters of temperature field, geometric deformation and molten pool stability, improve the monitoring accuracy, and shorten the abnormal detection response time to milliseconds. The data fusion algorithm has time-space synchronization and weight self-adaptive capabilities, ensuring stable monitoring under different processing speeds, complex curved surfaces and high-reflectivity noble metal scenarios. The built-in sensor fault switching mechanism can maintain a monitoring coverage rate of >95% when the device is abnormal, greatly improving the system robustness and processing process quality consistency.

[0083] During the gap time after the current area processing is completed, the processing path of the next area is optimized and adjusted according to the processing adjustment parameters.

[0084] The specific process of pre-optimizing and adjusting the processing path of the next layer according to the processing adjustment parameters includes:

[0085] After each layer scanning is completed, the temperature distribution data, geometric accuracy data and surface quality data of the entire layer are collected; the processing path of the subsequent same type of functional partition is pre-optimized, and the optimization target is temperature uniformity and geometric accuracy;

[0086] The pre-optimization result is used to update the corresponding area of the candidate path scheme in the partition path library; the comparison data of the optimization result and the actual processing result are recorded in the historical database for offline training of the correlation prediction model and continuous updating of the parameter library.

[0087] In this embodiment, multi-sensor fusion is used to achieve comprehensive and accurate state monitoring, improving the real-time and reliability of anomaly detection. In the current regional processing gap, the next regional processing path is optimized according to the state parameters, and temperature, geometric accuracy and surface quality data are collected after each layer scanning. The path of the same type of function is pre-optimized, the partition path library is updated, and the optimization effect and actual results are recorded into the historical database to provide data support for offline training of the correlation prediction model and updating of the parameter library.

[0088] The optimization adjustment in the gap time adopts a multi-task parallel processing mechanism, including: establishing an optimization task queue, simultaneously executing three parallel tasks of current layer quality evaluation, next layer path pre-computation and alternative scheme generation; setting a task priority scheduler to dynamically allocate computing resources according to processing progress and deviation urgency; when a key deviation is detected, interrupt low-priority tasks and prioritize path re-planning tasks.

[0089] Compared with the existing sequential path optimization method, the multi-task parallel processing mechanism can improve the utilization rate of the inter-layer gap time, so that quality evaluation, path pre-computation and alternative scheme generation are completed simultaneously; the task priority scheduler can compress the response time to milliseconds when a key deviation occurs, realizing rapid path re-planning; in the scene of high complexity surface machining, large temperature gradient fluctuation or strict precision requirement, the system can still maintain stable optimization efficiency and processing quality, significantly improving the overall production rhythm and process adaptability. Through the multi-task parallel processing mechanism, the inter-layer gap time is fully utilized, the computing efficiency and response speed are improved, the system can complete more complex optimization calculation in limited gap time, and the overall machining efficiency and quality control ability are improved.

[0090] The dynamic optimization of the processing path adopts a reinforcement learning algorithm, including: establishing an agent model with processing quality as the reward function, current state parameters as input states and path adjustment strategies as action space; setting an experience replay mechanism to store historical optimization decisions and effect feedback; using a deep Q network algorithm to continuously learn the optimal path adjustment strategy, and setting an exploration-exploitation balance mechanism to avoid local optimal solution.

[0091] Compared with the traditional path adjustment technology relying on fixed rules, the reinforcement learning dynamic optimization method can adaptively learn the optimal strategy through the deep Q network, so that the path adjustment accuracy is improved by about 15%-25%, the processing quality fluctuation is reduced to within ±3%, the experience replay mechanism can efficiently utilize historical data, and the convergence speed is accelerated by about 30%, and the exploration-exploitation balance mechanism can effectively avoid falling into local optimization. In the scene of variable temperature field, high complexity geometry and multi-material mixed processing, it can continuously optimize itself, realize the generalization ability and long-term processing stability across tasks. The reinforcement learning algorithm can continuously learn and improve from historical processing experience, adaptively find the optimal path adjustment strategy, gradually improve the optimization effect as the processing task increases, and realize the self-evolution and intelligent level of the system.

[0092] The present application faces the noble metal to be processed, through obtaining its geometric data, product data and material characteristics, establishes a regional prediction model for functional partitioning, generates a three-dimensional partition mapping table containing type, boundary and priority, and pre-generates a partition path library of multiple scanning strategies combined with partition characteristics and crystallization parameters. Using the correlation prediction model to analyze the processing task, the processing parameters such as laser power, speed and overlap rate are predicted, an initial path scheme is generated according to the priority of precision control area, and in the processing process, based on the multi-sensor fusion technology, the temperature field, geometric deformation and molten pool stability are monitored in real time, when the state parameters deviate from the target range, the processing adjustment parameters are calculated combined with the functional partitioning and historical data, and the optimal path is dynamically selected. The system uses a multi-task parallel mechanism to complete quality evaluation, path pre-computation and alternative scheme generation within the interlayer gap time, and combines the reinforcement learning algorithm to continuously optimize the path adjustment strategy, improve the processing precision, efficiency and process stability, and realize the intelligent and high-quality processing of complex structures of noble metals. The overall process schematic diagram is shown in Figure 3 .

[0093] Embodiment two:

[0094] The present application provides a micro-size SLM precision scanning path dynamic optimization distribution method applied to precision medical instrument manufacturing, taking the manufacturing of a heart pacemaker titanium alloy shell as a specific application scenario, and the technical scheme is as follows:

[0095] For the heart pacemaker shell to be processed, first, the complete three-dimensional CAD model data is obtained. The geometric data includes three-dimensional coordinate point cloud of the shell, surface curvature radius of each part (curvature radius of the shell curved surface part is 8-12 mm, curvature radius of the connecting part is 2-5 mm), wall thickness distribution (main body wall thickness is 1.2 mm, reinforcing rib part wall thickness is 1.8 mm), overhanging angle (overhanging angle of the lead outlet is 45 degrees) and local volume density distribution;

[0096] The product data includes the overall size of the pacemaker shell, surface roughness requirements (the biological contact surface roughness Ra value is not more than 0.8 microns, and the electrode connecting surface roughness Ra value is not more than 0.4 microns) and structure complexity level (defined as high complexity level according to medical device standards); the material properties include the melting point, thermal conductivity, thermal expansion coefficient, laser absorption rate and crystallization characteristic parameters of medical titanium alloy.

[0097] The established regional prediction model includes three core components: feature extraction layer, classification decision layer and mapping generation layer; the feature extraction layer extracts deep feature vectors from the above-mentioned geometric data and product data, generating a comprehensive feature matrix containing spatial position information, geometric complexity index and product features;

[0098] The classification decision layer uses an improved support vector machine algorithm to intelligently classify the comprehensive feature matrix, and divides the pacemaker shell into three functional regions: precision control area (including electrode connecting part, sealing ring groove and biological contact surface), transition buffer area (including the transition area between the main body and the connecting part) and standard processing area (including the main body plane area of the shell);

[0099] The mapping generation layer accurately associates the region identification with the three-dimensional coordinates to generate a three-dimensional partition mapping table containing region type coding, boundary coordinate sequence and processing priority information, which is stored in an octree data structure to support spatial indexing with micron-level processing accuracy.

[0100] According to the characteristics of functional partition, a differentiated processing path scheme is pre-calculated, based on the crystallization characteristic parameters of titanium alloy, a progressive rotation strategy of interlayer scanning direction is designed, the rotation angle of scanning direction between adjacent layers is set to ensure the uniformity of grain structure and the isotropy of mechanical properties.

[0101] For the precision control area, the laser power is reduced to 70% of the rated power, the scanning speed is increased to 1200 mm / s, the line segment length is controlled in the range of 0.1 to 0.3 mm, and the geometric precision of the electrode connecting part is realized; for the transition buffer area, the transition scanning strategy with gradual change of scanning direction according to the preset step length is adopted to ensure smooth transition between different regions and avoid deformation or cracking caused by thermal stress concentration; for the standard processing area, the one-way parallel scanning strategy is adopted, and the scanning line spacing is set to ensure the processing efficiency while ensuring the forming quality; the scanning path scheme of each region is stored in a structured data format, including scanning trajectory coordinate sequence, corresponding laser power parameters and scanning speed parameters, forming a partition path library containing different combinations.

[0102] The established correlation prediction model includes a data preprocessing layer, a feature correlation layer, and a parameter prediction layer; the data preprocessing layer standardizes the historical processing data of the titanium alloy medical device, extracts the numerical features of laser power, scanning speed, line spacing, and overlap rate, and final forming quality;

[0103] The feature correlation layer uses a deep neural network algorithm to establish a nonlinear mapping relationship between the input process parameters and the output quality indicators. The network structure includes 5 hidden layers, each with 256 neurons, and the activation function uses the ReLU function.

[0104] The parameter prediction layer outputs the optimal laser power range, scanning speed range, and overlap rate value based on the geometric characteristics of the current cardiac pacemaker shell and the material properties of titanium alloy, serving as the initial processing parameters.

[0105] Based on the processing parameters output by the correlation prediction model, the specific processing parameters for each functional partition are calculated; the laser power of the precision control area is set to 190 watts, the scanning speed is 1200 mm / s, and the line spacing is 0.06 mm; the laser power of the transition buffer area is set to 205 watts, the scanning speed is 1100 mm / s, and the line spacing is 0.07 mm; the laser power of the standard processing area is set to 215 watts, the scanning speed is 1050 mm / s, and the line spacing is 0.08 mm;

[0106] The processing sequence is arranged in the order of precision control area first, transition buffer area second, and standard processing area last to ensure the processing quality of critical parts; a linear transition section of laser power and scanning speed is set at the boundary between adjacent areas, and the transition length is dynamically adjusted according to the boundary geometry; the generated initial processing path allocation scheme includes complete scanning trajectory coordinate data, timestamp information, and key parameter change node markers, providing basic data support for subsequent dynamic optimization.

[0107] Multi-sensor fusion technology is used to realize real-time monitoring of the entire processing process; an infrared thermal imager is arranged to form a sensor array to monitor the temperature field distribution of the processing area; a high-precision laser displacement sensor is configured to monitor geometric deformation, ensuring accurate control of the geometric accuracy of critical parts; an acoustic emission sensor is installed to monitor the stability of the molten pool, and the uniformity of the melting process is judged by analyzing the acoustic emission signal characteristics;

[0108] A multi-source sensor data fusion algorithm is established, Kalman filter technology is used for spatio-temporal synchronization processing of different sensor data, and fusion weights are allocated according to sensor accuracy and reliability. When the weight of the temperature sensor is 0.4, the weight of the displacement sensor is 0.4, and the weight of the acoustic emission sensor is 0.2, the most accurate comprehensive state evaluation can be provided; an intelligent sensor fault detection mechanism is set, and abnormal sensors are identified by comparing adjacent sensor data and historical data patterns. When a single sensor anomaly is detected, switch to a backup sensor or redistribute the fusion weight to ensure the continuous and stable operation of the monitoring system.

[0109] During processing, when the temperature distribution deviates from the target temperature field preset value by more than 8 degrees, or the geometric profile deviation exceeds 5 microns, the dynamic adjustment mechanism is triggered immediately; record the current specific processing parameters and deviation data, including deviation type, deviation value, occurrence time and spatial position.

[0110] Based on statistical analysis method, the successful processing cases in the historical database are intelligently matched to find historical cases with similar deviation patterns. The laser power adjustment value, scanning speed adjustment value and path spacing adjustment value suitable for the next processing layer are calculated by weighted average algorithm, which are used as processing adjustment parameters for subsequent processing; from the partition path library, select the processing path scheme that best matches the adjustment parameters to ensure that the adjusted processing parameters can effectively correct the deviation and maintain the stability of the processing quality.

[0111] In the interlayer gap time after each layer processing is completed, a multi-task parallel processing mechanism is used to fully utilize this time window, an optimized task queue containing three types of tasks: quality evaluation, path precalculation and alternative scheme generation is established, and the three tasks are executed in parallel to improve calculation efficiency.

[0112] The current layer quality evaluation task is responsible for analyzing the temperature distribution uniformity, geometric accuracy compliance rate and surface quality index of the just completed layer, the next layer path precalculation task predicts the optimal processing path for the next layer according to the current state, and the alternative scheme generation task prepares multiple sets of alternative path schemes for possible deviation conditions.

[0113] An intelligent task priority scheduler is set, which dynamically allocates computing resources according to the current processing progress and the detected deviation urgency; when a critical deviation (such as a geometric deviation exceeding 10 microns or a temperature deviation exceeding 15 degrees) is detected, the system interrupts the low-priority alternative scheme generation task and allocates computing resources to the path re-planning task to ensure timely response and processing of critical abnormalities.

[0114] The deep reinforcement learning algorithm is used to realize continuous optimization of the path adjustment strategy, an intelligent agent model of a reward function is established with the final forming quality as the core, the current temperature distribution, geometric accuracy and surface quality and other state parameters are taken as the input state space, and the laser power adjustment, scanning speed adjustment and path selection are taken as the action space. The experience replay mechanism is set to store the historical optimization decisions and the corresponding effect feedback data in the circular buffer, to provide rich training samples for the agent learning, and the deep Q network algorithm is used to continuously learn and update the optimal path adjustment strategy. The ε-greedy exploration-exploitation balance mechanism is set, the ε value gradually decreases from 0.9 to 0.1, so that the system can fully explore different strategies in the early learning stage and mainly use the learned optimal strategy in the later learning stage, effectively avoiding falling into a local optimal solution. After each layer of scanning is completed, the temperature distribution data, geometric accuracy data and surface quality data of the whole layer are collected; the processing path of the subsequent same type of functional partition is pre-optimized, and the optimization objective function considers the temperature uniformity and geometric accuracy as two core indicators. The genetic algorithm is used to search for the optimal parameter combination, the population size is set to 50, and the evolution number is set to 30 generations.

[0115] According to the pre-optimization calculation result, the candidate path scheme of the corresponding area in the partition path library is updated in real time, so that the path library always contains the latest optimization strategy. The detailed comparison data of the optimization result and the actual processing effect are recorded in the historical database, including the quality index change before and after optimization, parameter adjustment amplitude and optimization effect evaluation and other information, to provide high-quality data support for offline training of the correlation prediction model and continuous updating of the overall parameter library.

[0116] In the embodiment, by applying the complete technical solution to the actual manufacturing process of the titanium alloy shell of the cardiac pacemaker, high-precision and high-efficiency intelligent processing of the complex medical instrument structure is realized. The multi-sensor fusion technology provides comprehensive process monitoring capability, the multi-task parallel mechanism fully utilizes the interlayer time for optimization calculation, and the reinforcement learning algorithm realizes adaptive improvement of the processing strategy, so that the overall technical solution significantly improves the intelligent level and product quality stability of micro-size SLM processing.

[0117] Those skilled in the art can easily understand that the above description is only a preferred embodiment of the present application, and is not intended to limit the present application, and any modification, equivalent replacement and improvement made within the spirit and principle of the present application shall be included in the protection scope of the present application.

Claims

1. A method for dynamically optimizing distribution of precision scanning paths for a micro-sized SLM, characterized in that, The method comprises the following steps: acquiring a three-dimensional model of a product to be processed, and extracting geometric data, product data and material characteristics; establishing a region prediction model based on the geometric data and the product data to classify the regions to be processed, obtaining functional partitions, and generating a partition mapping table; pre-computing a processing path scheme according to the functional partitions to form a partition path library; the region prediction model comprises a feature extraction layer, a classification decision layer and a mapping generation layer; the feature extraction layer extracts feature vectors from the geometric data and the product data to obtain a comprehensive feature matrix containing spatial position, geometric complexity and product characteristics; the classification decision layer classifies the comprehensive feature matrix and outputs region identifiers of precision control zones, transition buffer zones and standard machining zones; the mapping generation layer associates the region identifiers with coordinates to generate a three-dimensional partition mapping table containing region types, boundary coordinates and priority information; establishing a correlation prediction model to analyze the processing task and the material characteristics to obtain processing parameters; the correlation prediction model comprises a data preprocessing layer, a feature correlation layer and a parameter prediction layer; the data preprocessing layer standardizes historical processing data to extract numerical features of processing power, processing speed, overlap rate and forming quality; the feature correlation layer establishes a nonlinear mapping relationship between the processing data and the product quality; the parameter prediction layer obtains the product quality according to the geometric data and the material characteristics of the current processing task; outputting a laser power range, a processing speed range and an overlap rate value based on the product quality as the processing parameters; generating an initial processing path allocation scheme based on the processing parameters; processing the product to be processed through the initial processing path allocation scheme, monitoring the state parameters of the product to be processed during the processing, calculating optimal processing parameters according to the functional partition and historical processing data when the state parameters deviate from a preset target range, obtaining processing adjustment parameters, and selecting a processing path scheme from the partition path library; optimizing and adjusting the processing path of the next region according to the processing adjustment parameters during the gap time after the current region is processed.

2. The method according to claim 1, wherein: the geometric data comprises the coordinates, surface curvature radius, wall thickness distribution, overhanging angle and local bulk density of the product; the product data comprises the overall size, surface roughness requirement and structure complexity level of the product; the material characteristics comprise the melting point, thermal conductivity, thermal expansion coefficient, laser absorption rate and crystallization characteristic parameters of the noble metal.

3. The method according to claim 2, wherein: the process of pre-computing a processing path scheme according to the functional partitions comprises: designing a progressive rotation strategy for the interlayer scanning direction based on the crystallization characteristic parameters of the product, and setting the rotation angle of the scanning direction between adjacent layers to be a first angle to a second angle; adopting a short-line segment staggered scanning strategy with reduced laser power and increased scanning speed for the precision control zone; adopting a transition scanning strategy with a gradual change in the scanning direction by a preset step size for the transition buffer zone; The standard processing area adopts a unidirectional parallel scanning strategy; the scanning path scheme of each area is stored as a data structure containing coordinate sequence, power parameter and speed parameter, forming a partition path library.

4. The micro-sized SLM precision scanning path dynamic optimization allocation method according to claim 1, characterized in that: The process of generating an initial processing path allocation scheme based on processing parameters includes: According to the laser power range and processing speed range determined by the processing parameters, the processing line spacing and processing sequence of each functional partition are calculated; the processing sequence is arranged in the order of precision control area first, transition buffer area second, and standard processing area last; Linear transition sections of power and speed are set at the boundaries between adjacent areas, and the transition length is set according to the boundary length; an initial processing path allocation scheme containing scanning trajectory coordinates, time stamps and parameter change points is generated.

5. The micro-sized SLM precision scanning path dynamic optimization allocation method according to claim 1, characterized in that: The process of obtaining processing adjustment parameters includes: By collecting the overall temperature distribution data of the processing area and the geometric contour data of the formed layer, when the overall temperature distribution deviates from the target temperature field preset temperature and / or the geometric contour deviation exceeds the preset deviation, the current processing parameters and deviation data are recorded; Based on statistical analysis method, the historical processing data is matched with parameters to obtain the applicable laser power adjustment value, scanning speed adjustment and path spacing adjustment value for the next layer, as the processing adjustment parameters.

6. The micro-sized SLM precision scanning path dynamic optimization allocation method according to claim 1, characterized in that: The specific process of pre-optimizing the next layer's processing path according to the processing adjustment parameters includes: After each layer is scanned, collect the temperature distribution data, geometric accuracy data and surface quality data of the whole layer; pre-optimize the processing path of the subsequent same type functional partition, and the optimization target is temperature uniformity and geometric accuracy; According to the pre-optimization result, update the alternative path scheme of the corresponding area in the partition path library; record the comparison data of the optimization result and the actual processing result into the historical database, which is used for offline training of the correlation prediction model and continuous updating of the parameter library.

Images